spg_sql/parser.rs
1//! Recursive-descent parser with a Pratt (precedence-climbing) sub-parser for
2//! expressions.
3//!
4//! Precedence (lowest → highest binding):
5//! `OR` (1) `<` `AND` (2) `<` `NOT` unary (3) `<`
6//! comparisons `=` `<>` `<` `<=` `>` `>=` (4) `<`
7//! `+` `-` (5) `<` `*` `/` (6) `<` unary `-` (7) `<` parens / atom.
8//!
9//! This matches PG's behaviour for the operators we support — e.g. `NOT a = b`
10//! parses as `NOT (a = b)` and `-a * b` as `(-a) * b`.
11
12use alloc::boxed::Box;
13use alloc::format;
14use alloc::string::{String, ToString};
15use alloc::vec;
16use alloc::vec::Vec;
17use core::fmt;
18use core::mem;
19
20use crate::ast::{
21 AssignTarget, BinOp, CastTarget, Collation, ColumnDef, ColumnName, ColumnTypeName,
22 CreateFunctionStatement, CreateIndexStatement, CreatePublicationStatement,
23 CreateSubscriptionStatement, CreateTableStatement, CreateTriggerStatement, DiscardTarget, Expr,
24 ExtractField, FkAction, ForeignKeyConstraint, FrameBound, FrameExclusion, FrameKind,
25 FromClause, FromJoin, FunctionArg, FunctionArgMode, FunctionArgType, FunctionAttrs,
26 FunctionBody, FunctionParallel, FunctionReturn, FunctionVolatility, GrantObject, GrantPriv,
27 GrantStatement, IndexMethod, InsertStatement, IsolationLevel, JoinKind, Literal, MysqlIntWidth,
28 NullTreatment, OrderBy, Overriding, PlPgSqlBlock, PlPgSqlDeclare, PlPgSqlStmt,
29 PublicationScope, RaiseLevel, RangeKindAst, ReturnTarget, SelectItem, SelectStatement,
30 Statement, TableRef, TriggerEvent, TriggerForEach, TriggerTiming, UnOp, UnionKind, VecEncoding,
31 WindowFrame,
32};
33use crate::lexer::{self, LexError, Token};
34
35/// v7.38 — a `WINDOW w AS (…)` definition body:
36/// `(PARTITION BY exprs, ORDER BY (expr, desc, nulls_first), frame)`.
37type WindowDef = (
38 Vec<Expr>,
39 Vec<(Expr, bool, Option<bool>)>,
40 Option<WindowFrame>,
41);
42
43/// v7.14.0 — true when the leading keyword of a top-level
44/// statement is one of the dump-emitted DDL forms SPG accepts
45/// as a no-op (no behavioural effect on the single-schema /
46/// single-database model). These statements are consumed up to
47/// the next `;` / EOF and returned as `Statement::Empty`.
48/// v7.39 (read01 round 57) — wrap a parsed GRANT body in the right statement.
49fn finish_grant(grant: bool, g: GrantStatement) -> Statement {
50 if grant {
51 Statement::Grant(g)
52 } else {
53 Statement::Revoke(g)
54 }
55}
56
57fn is_dump_noise_statement(lc: &str) -> bool {
58 matches!(
59 lc,
60 // v7.39 (read01 round 50): "comment" moved OUT — COMMENT ON is now a
61 // real statement with a real store. v7.39 (read01 round 57): "grant" /
62 // "revoke" moved OUT — table privileges are now REAL (stored in
63 // `relacl`, enforced against the session role); a grant on any other
64 // object class still parses and no-ops so dumps restore.
65 // MySQL bulk-load brackets.
66 "unlock"
67 // MySQL OPTIMIZE / ANALYZE TABLE / CHECK TABLE
68 // diagnostics that pg_dump-style tools also emit
69 // post-restore.
70 | "optimize"
71 | "check"
72 // PG psql backslash meta-commands that newer
73 // pg_dump versions emit unescaped (\restrict /
74 // \unrestrict). Real psql intercepts these; SPG's
75 // PG-wire sees them as raw text.
76 | "\\restrict"
77 | "\\unrestrict"
78 // v7.17.0 Phase 4.1 — MySQL `DELIMITER //` and
79 // `DELIMITER ;` directives. Technically client-side
80 // (the `mysql` CLI uses them to set the statement
81 // terminator), not SQL — but mysqldump and stored-
82 // procedure scripts emit them inline. SPG's parser
83 // sees one statement at a time and doesn't care
84 // about the terminator, so consume DELIMITER lines
85 // as Empty.
86 | "delimiter"
87 // v7.37.17 (17.6 siblings) — additional PG maintenance /
88 // session-state statements pg_dump + application startup
89 // scripts emit. SPG has no matching session-state to
90 // discard (no prepared-plan cache surface, no temp
91 // sequences), no matching security-label / storage-
92 // option to apply, no separate CREATE/DROP CAST that
93 // affects execution.
94 // v7.37.17 (17.6 siblings) — PG role-cleanup statements
95 // pg_dump / pg_dumpall emit around DROP ROLE:
96 // REASSIGN OWNED BY <role> [, ...] TO <newrole>
97 // DROP OWNED BY <role> [, ...] [CASCADE | RESTRICT]
98 // Both operate on the role's owned objects; SPG has no
99 // role-owner model, so accept-and-no-op.
100 // v7.37.17 (17.6 sibling) — LOAD '<library>'. pg_dump
101 // + extension scripts use LOAD to preload shared
102 // libraries. SPG doesn't have a shared-library extension
103 // point today (extensions ship as first-class crates
104 // linked at build time); accept as a no-op.
105 | "load"
106 )
107}
108
109/// v7.37.43-T4 — PG-unreserved keywords that are legal identifiers
110/// per `pg_get_keywords()`. SPG tokenizes these as named variants
111/// so the parser can dispatch on them in their owning contexts
112/// (`RELEASE SAVEPOINT`, `SHOW name`, `BEGIN`/`COMMIT`/`ROLLBACK`,
113/// `CREATE INDEX`, etc.), but they MUST stay usable as table /
114/// column / alias names — that's the PG contract for unreserved
115/// keywords (see PG docs Appendix C.1).
116///
117/// Before this generalisation, sentori migration 0001_init.sql
118/// `release TEXT NOT NULL` blew up the parser with "expected
119/// identifier, got Release", and the same gap stalked every
120/// SPG drop-in user whose schema had a column / alias named
121/// `release` / `index` / `tables` / `show` / `savepoint` /
122/// `begin` / `commit` / `rollback` / `drop` / `insert` / `values`
123/// / `limit` / `partition`. PG accepts all of them as identifiers
124/// when unquoted, so SPG must too.
125///
126/// Returns the canonical lowercase identifier text when the token
127/// belongs to PG's unreserved class, `None` otherwise. Used by
128/// `expect_ident_like` (column / table / alias names) so the
129/// generalisation applies everywhere an identifier may appear,
130/// not just in the contexts these tokens were introduced for.
131fn unreserved_keyword_text(tok: &Token) -> Option<String> {
132 let s = match tok {
133 // PG keyword class: unreserved or col_name.
134 //
135 Token::Release => "release",
136 Token::Savepoint => "savepoint",
137 Token::Show => "show",
138 Token::Index => "index",
139 Token::Begin => "begin",
140 Token::Commit => "commit",
141 Token::Rollback => "rollback",
142 Token::Drop => "drop",
143 Token::Insert => "insert",
144 Token::Values => "values",
145 Token::Limit => "limit",
146 Token::Partition => "partition",
147 Token::Tables => "tables",
148 Token::Connection => "connection",
149 Token::Publication => "publication",
150 Token::Subscription => "subscription",
151 Token::Interval => "interval",
152 // `extract` is non-reserved in PG too (it's a function the
153 // parser dispatches via context — outside that context it's
154 // a plain identifier).
155 Token::Extract => "extract",
156 Token::Offset => "offset",
157 // `to` is reserved in PG (used in many "AS … TO …" forms), so
158 // it is NOT relaxed here. Same for `from`, `where`, `as`,
159 // `select`, `not`, `and`, `or`, `null`, `true`, `false`,
160 // `create`, `table`, `into`, `on`, `order`, `by`, `having`,
161 // `group`, `distinct`, `union`, `all`, `join`, `inner`,
162 // `left`, `cross`, `outer`, `default`, `is`, `between`,
163 // `in`, `like`, `for`, `except`, `desc`, `asc`, `partition`
164 // (partial — keep partition as unreserved per modern PG).
165 _ => return None,
166 };
167 Some(s.to_string())
168}
169
170/// v7.9.22 — recognise pgvector / SPG vector-index opclass names
171/// in CREATE INDEX. SPG's HNSW already routes by query operator;
172/// the opclass is accepted for `pg_dump` compatibility (mailrs
173/// migration follow-up G5).
174/// v7.13.0 — extended to recognise PG built-in / pg_trgm opclasses
175/// (mailrs round-5 G5). These are tokens-only acceptance — SPG
176/// doesn't change index behaviour based on them.
177/// v7.37.17 (17.6 siblings) — the four PG `each` SRFs share one
178/// FROM-clause pipeline; the stored name tells the executor whether
179/// the value column keeps JSON rendering (`jsonb_each` / `json_each`)
180/// or unwraps to text (`*_each_text`).
181fn is_json_each_name(s: &str) -> bool {
182 s.eq_ignore_ascii_case("jsonb_each_text")
183 || s.eq_ignore_ascii_case("jsonb_each")
184 || s.eq_ignore_ascii_case("json_each_text")
185 || s.eq_ignore_ascii_case("json_each")
186}
187
188/// v7.38 (read01, T14) — resolve named function arguments (`argname => value`)
189/// to positional order for the `make_*` family (the AST stays positional).
190/// Positional args fill slots left-to-right; a named arg goes to its registered
191/// slot; unfilled slots default to integer 0 (PG's optional make_interval
192/// fields — the make_date/time arity is still checked at eval time).
193fn reorder_named_args(
194 fname: &str,
195 args: Vec<Expr>,
196 names: &[Option<String>],
197) -> Result<Vec<Expr>, String> {
198 let params: &[&str] = match fname.to_ascii_lowercase().as_str() {
199 "make_date" => &["year", "month", "day"],
200 "make_time" => &["hour", "min", "sec"],
201 "make_timestamp" | "make_timestamptz" => &["year", "month", "mday", "hour", "min", "sec"],
202 "make_interval" => &["years", "months", "weeks", "days", "hours", "mins", "secs"],
203 other => {
204 return Err(alloc::format!(
205 "function {other}(...) does not support named arguments"
206 ));
207 }
208 };
209 let mut slots: Vec<Option<Expr>> = (0..params.len()).map(|_| None).collect();
210 let mut next_positional = 0usize;
211 for (arg, name) in args.into_iter().zip(names.iter()) {
212 let idx = match name {
213 Some(n) => params
214 .iter()
215 .position(|p| p.eq_ignore_ascii_case(n))
216 .ok_or_else(|| alloc::format!("{fname}(...) has no argument named \"{n}\""))?,
217 None => {
218 let i = next_positional;
219 next_positional += 1;
220 i
221 }
222 };
223 if idx >= slots.len() {
224 return Err(alloc::format!("too many arguments for {fname}(...)"));
225 }
226 if slots[idx].is_some() {
227 return Err(alloc::format!(
228 "argument \"{}\" specified more than once",
229 params[idx]
230 ));
231 }
232 slots[idx] = Some(arg);
233 }
234 Ok(slots
235 .into_iter()
236 .map(|s| s.unwrap_or(Expr::Literal(Literal::Integer(0))))
237 .collect())
238}
239
240/// v7.38 (read01) — parse a lexer `Token::Numeric` source string (digits with
241/// an optional single `.`, no sign, no exponent) into `(unscaled, scale)` for
242/// `Literal::Numeric`. Returns `None` if the mantissa overflows i128.
243/// v7.39 (read01 numeric.c) — the result of expanding an `1.5e3`-style
244/// scientific literal into PG's plain NUMERIC decimal form.
245#[derive(Debug)]
246pub enum SciExpanded {
247 /// Plain decimal string ("1.5e3" → "1500", "1e-5" → "0.00001").
248 Expanded(String),
249 /// Exponent pushes the value outside PG's numeric format
250 /// (more than 131072 integer digits or 16383 fractional digits).
251 Overflow,
252 /// Not a `[±]digits[.digits]e[±]digits` literal at all.
253 NotScientific,
254}
255
256/// Expand scientific notation into a plain decimal string by moving the
257/// decimal point — no float round-trip, so the value stays exact. PG treats
258/// such literals as NUMERIC; the digit-count caps mirror PG's numeric format
259/// limits ("value overflows numeric format").
260pub fn expand_scientific_literal(s: &str) -> SciExpanded {
261 let s = s.trim();
262 let Some(epos) = s.find(['e', 'E']) else {
263 return SciExpanded::NotScientific;
264 };
265 let (mant, exp_str) = (&s[..epos], &s[epos + 1..]);
266 let Ok(exp) = exp_str.parse::<i64>() else {
267 return SciExpanded::NotScientific;
268 };
269 let (neg, mant) = match mant.strip_prefix('-') {
270 Some(r) => (true, r),
271 None => (false, mant.strip_prefix('+').unwrap_or(mant)),
272 };
273 let (int_part, frac_part) = match mant.split_once('.') {
274 Some((i, f)) => (i, f),
275 None => (mant, ""),
276 };
277 if (int_part.is_empty() && frac_part.is_empty())
278 || !int_part.bytes().all(|b| b.is_ascii_digit())
279 || !frac_part.bytes().all(|b| b.is_ascii_digit())
280 {
281 return SciExpanded::NotScientific;
282 }
283 let mut digits = String::with_capacity(int_part.len() + frac_part.len());
284 digits.push_str(int_part);
285 digits.push_str(frac_part);
286 // Decimal point position within `digits` after applying the exponent.
287 let new_point = int_part.len() as i64 + exp;
288 // PG's numeric format: up to 131072 digits before the point, 16383 after.
289 if new_point > 131_072 {
290 return SciExpanded::Overflow;
291 }
292 if (digits.len() as i64 - new_point) > 16_383 {
293 return SciExpanded::Overflow;
294 }
295 let sign = if neg { "-" } else { "" };
296 let plain = if new_point <= 0 {
297 let mut out = String::with_capacity(digits.len() + 2 + (-new_point) as usize);
298 out.push_str("0.");
299 for _ in 0..(-new_point) {
300 out.push('0');
301 }
302 out.push_str(&digits);
303 out
304 } else if (new_point as usize) >= digits.len() {
305 let mut out = digits;
306 for _ in 0..(new_point as usize - out.len()) {
307 out.push('0');
308 }
309 out
310 } else {
311 let mut out = String::with_capacity(digits.len() + 1);
312 out.push_str(&digits[..new_point as usize]);
313 out.push('.');
314 out.push_str(&digits[new_point as usize..]);
315 out
316 };
317 SciExpanded::Expanded(alloc::format!("{sign}{plain}"))
318}
319
320/// v7.39 (round 367, M20) — lower a MySQL hexadecimal binary-string
321/// literal (`0x…` / `X'…'`) onto the existing bytea cast. The hex digits
322/// are left-padded to an even count (`0x123` → byte string `01 23`, per
323/// MariaDB) and handed to the PG bytea input format (`\x…`).
324#[inline(never)]
325fn hex_literal_to_bytea_expr(hex: &str) -> Expr {
326 let padded = if hex.len() % 2 == 1 {
327 alloc::format!("0{hex}")
328 } else {
329 hex.to_string()
330 };
331 Expr::Cast {
332 expr: alloc::boxed::Box::new(Expr::Literal(Literal::String(alloc::format!(
333 "\\x{padded}"
334 )))),
335 target: CastTarget::Named("bytea".to_string()),
336 }
337}
338
339/// v7.39 (round 367, M20) — lower a MySQL bit-value literal (`b'1010'`)
340/// onto the bytea cast. The bits are read big-endian and left-padded to a
341/// whole number of bytes (`b'1010'` → one byte `0x0A`, per MariaDB).
342#[inline(never)]
343fn bits_literal_to_bytea_expr(bits: &str) -> Expr {
344 let pad = (8 - bits.len() % 8) % 8;
345 let mut hex = String::with_capacity((bits.len() + pad).div_ceil(4));
346 let padded: String = core::iter::repeat_n('0', pad).chain(bits.chars()).collect();
347 for nibble in padded.as_bytes().chunks(4) {
348 let mut v = 0u8;
349 for &b in nibble {
350 v = (v << 1) | (b - b'0');
351 }
352 hex.push(char::from_digit(u32::from(v), 16).unwrap_or('0'));
353 }
354 hex_literal_to_bytea_expr(&hex)
355}
356
357/// Resolve a lexer `Token::Numeric` into its literal. PG semantics: a dotted
358/// or over-i64 literal is exact NUMERIC; scientific notation is NUMERIC too
359/// (expanded to the plain decimal form); only a fractional depth beyond SPG's
360/// scale width (u8) falls back to double precision.
361///
362/// v7.38.19 — that sentence used to end "(recorded delta)". Measured
363/// against PG 18.4: a literal with 300 fractional digits round-trips
364/// identically on both engines, so whatever the note described is gone.
365/// It is RD-7 in `docs/RECORDED_DELTAS.md`, under "corrected by
366/// measurement" rather than under "open".
367/// Kept out of the parse_expr recursion frame — see the call site.
368#[inline(never)]
369fn numeric_token_to_literal(s: String) -> Result<Literal, String> {
370 match parse_decimal_literal(&s) {
371 Some((unscaled, scale)) => Ok(Literal::Numeric { unscaled, scale }),
372 // v7.38 (read01, T3.C3) — a plain decimal too wide for i128 keeps
373 // its exact value as a NumericBig.
374 None if !s.contains(['e', 'E']) => Ok(Literal::NumericBig(s)),
375 // v7.39 (read01 numeric.c) — expand the exponent form.
376 None => match expand_scientific_literal(&s) {
377 SciExpanded::Expanded(plain) => match parse_decimal_literal(&plain) {
378 Some((unscaled, scale)) => Ok(Literal::Numeric { unscaled, scale }),
379 None if plain
380 .split_once('.')
381 .is_none_or(|(_, f)| u8::try_from(f.len()).is_ok()) =>
382 {
383 Ok(Literal::NumericBig(plain))
384 }
385 None => s
386 .parse::<f64>()
387 .map(Literal::Float)
388 .map_err(|_| format!("invalid numeric literal {s:?}")),
389 },
390 SciExpanded::Overflow => Err("value overflows numeric format".to_string()),
391 SciExpanded::NotScientific => s
392 .parse::<f64>()
393 .map(Literal::Float)
394 .map_err(|_| format!("invalid numeric literal {s:?}")),
395 },
396 }
397}
398
399fn parse_decimal_literal(s: &str) -> Option<(i128, u16)> {
400 let (int_part, frac_part) = match s.split_once('.') {
401 Some((i, f)) => (i, f),
402 None => (s, ""),
403 };
404 // v7.39 (round 271) — was u8::MAX. A literal with 256 decimal
405 // places fell out of the numeric path here, which is why
406 // `pg_typeof(1e-256)` answered double precision and a plain
407 // 256-place decimal aborted the query in the big-decimal converter.
408 if frac_part.len() > u16::MAX as usize {
409 return None;
410 }
411 let mut digits = String::with_capacity(int_part.len() + frac_part.len());
412 digits.push_str(int_part);
413 digits.push_str(frac_part);
414 let mantissa: i128 = digits.parse().ok()?;
415 #[allow(clippy::cast_possible_truncation)]
416 Some((mantissa, frac_part.len() as u16))
417}
418
419/// `jsonb_to_record` / `jsonb_to_recordset` (+ `json_` variants) — the
420/// record-returning JSON functions that take a `AS alias(col type, …)`
421/// column-definition list in FROM position.
422fn is_json_to_record_name(s: &str) -> bool {
423 s.eq_ignore_ascii_case("jsonb_to_recordset")
424 || s.eq_ignore_ascii_case("jsonb_to_record")
425 // v7.39 (read01 jsonfuncs.c) — the populate family with an AS
426 // column-definition list desugars identically (the record base
427 // argument only carries the type; a non-NULL base's field
428 // defaults are a recorded delta, RD-6).
429 || s.eq_ignore_ascii_case("json_populate_record")
430 || s.eq_ignore_ascii_case("jsonb_populate_record")
431 || s.eq_ignore_ascii_case("json_populate_recordset")
432 || s.eq_ignore_ascii_case("jsonb_populate_recordset")
433 || s.eq_ignore_ascii_case("json_to_recordset")
434 || s.eq_ignore_ascii_case("json_to_record")
435}
436
437impl Parser {
438 /// Whether what follows an identifier ends an index key, which is how
439 /// an operator class is told from anything else in that position.
440 fn opclass_position_follows(next: Option<&Token>) -> bool {
441 match next {
442 // `ASC` / `DESC` have their own tokens; matching them as
443 // identifiers named "asc" / "desc" — which the first version of
444 // this did — never fires, and `(c text_pattern_ops DESC)` (which
445 // PG18.4 accepts, verified) went on failing to parse.
446 Some(Token::Comma | Token::RParen | Token::Asc | Token::Desc) => true,
447 Some(Token::Ident(w)) => {
448 w.eq_ignore_ascii_case("nulls") || w.eq_ignore_ascii_case("collate")
449 }
450 _ => false,
451 }
452 }
453}
454
455fn is_vector_opclass_name(name: &str) -> bool {
456 let lc = name.to_ascii_lowercase();
457 matches!(
458 lc.as_str(),
459 "vector_cosine_ops"
460 | "vector_l2_ops"
461 | "vector_ip_ops"
462 | "halfvec_cosine_ops"
463 | "halfvec_l2_ops"
464 | "halfvec_ip_ops"
465 | "sq8_cosine_ops"
466 | "sq8_l2_ops"
467 | "sq8_ip_ops"
468 // pg_trgm — trigram operator class. SPG's GIN index
469 // already uses tsvector tokens; trigram-style LIKE
470 // pattern matching still routes through a sequential
471 // scan, but the opclass name is accepted so PG schemas
472 // load.
473 | "gin_trgm_ops"
474 | "gist_trgm_ops"
475 // PG built-in btree opclasses occasionally appear in
476 // pg_dump output for column types with multiple
477 // sort orders (text_pattern_ops, varchar_pattern_ops,
478 // bpchar_pattern_ops).
479 | "text_pattern_ops"
480 | "varchar_pattern_ops"
481 | "bpchar_pattern_ops"
482 | "int4_ops"
483 | "int8_ops"
484 | "text_ops"
485 )
486}
487
488#[derive(Debug, Clone, PartialEq, Eq)]
489pub struct ParseError {
490 pub message: String,
491 /// Index into the token stream where parsing tripped. Not a byte offset.
492 /// v7.39 (read01 round 95) — the byte/char position is NOT stored here: a
493 /// field would grow every `Result<_, ParseError>` slot on the deeply
494 /// recursive parse stack and tip the nesting-budget frame cliff. PG's
495 /// 1-based char position is recovered on the cold error path by
496 /// [`syntax_error_position`], which re-tokenizes to map this token index.
497 pub token_pos: usize,
498}
499
500impl fmt::Display for ParseError {
501 /// v7.39 (round 322/V24) — the message ALONE. It used to be prefixed
502 /// with `parse error at token #N: `, which PG has no equivalent of:
503 /// the message bodies are already PG's verbatim (`LIMIT must not be
504 /// negative`, `invalid input syntax for type bigint: "abc"`), and the
505 /// prefix was SPG's internal token index leaking into every one of
506 /// them. `token_pos` stays a field — the wire recovers PG's 1-based
507 /// character position from it for the ErrorResponse `P`.
508 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
509 f.write_str(&self.message)
510 }
511}
512
513impl From<LexError> for ParseError {
514 fn from(e: LexError) -> Self {
515 Self {
516 message: format!("lex: {e}"),
517 token_pos: 0,
518 }
519 }
520}
521
522/// v7.9.30 — parse a single expression (no trailing junk). Used by
523/// the engine to re-hydrate stored partial-index / unique-index
524/// predicates from their canonical Display form. The same Pratt
525/// parser the statement path uses; this entry point just skips the
526/// statement dispatch.
527pub fn parse_expression(input: &str) -> Result<Expr, ParseError> {
528 let (tokens, offsets) = lexer::tokenize_with_offsets(input, lexer::Dialect::PG)
529 .map_err(|e| shape_lex_error(&e, input))?;
530 let mut p = Parser::new(tokens);
531 let expr = p
532 .parse_expr(0)
533 .and_then(|e| p.expect_eof().map(|()| e))
534 .map_err(|e| shape_syntax_error(e, input, &offsets))?;
535 Ok(expr)
536}
537
538/// Parse exactly one statement, swallow an optional trailing `;`, and require
539/// the token stream to end there. PG string semantics.
540pub fn parse_statement(input: &str) -> Result<Statement, ParseError> {
541 parse_statement_with(input, lexer::Dialect::PG)
542}
543
544/// v7.22 (round-13 T3) — dialect-aware entry: `backslash_escapes`
545/// selects MySQL-style string lexing (see `lexer::tokenize_with`).
546/// The engine threads its session flag through here.
547pub fn parse_statement_with(input: &str, dialect: lexer::Dialect) -> Result<Statement, ParseError> {
548 let (tokens, offsets, merges) =
549 lexer::tokenize_with_merges(input, dialect).map_err(|e| shape_lex_error(&e, input))?;
550 // v7.39.2 — the grammar follows "is this MySQL", the lexer follows
551 // "does backslash escape". They used to be one flag, and a session
552 // that turned escapes off lost the grammar with them.
553 let mut p = Parser::new_with_dialect(tokens, dialect.speaks_mysql)
554 .with_source(input, &offsets)
555 .with_merges(merges);
556 let stmt = (|| {
557 let stmt = p.parse_one_statement()?;
558 if matches!(p.peek(), Token::Semicolon) {
559 p.advance();
560 }
561 p.expect_eof()?;
562 Ok(stmt)
563 })()
564 .map_err(|e: ParseError| shape_syntax_error(e, input, &offsets))?;
565 Ok(stmt)
566}
567
568/// v7.39 (round 340, V56) — PG has exactly two syntax-error wordings:
569/// `syntax error at or near "<token>"` and `syntax error at end of input`
570/// (measured on 18.4 across a dozen shapes). SPG wrote its own per-site
571/// prose — `expected identifier, got Eof`, `unexpected token From in
572/// expression`, `expected end of input, got Ident("with")` — which named
573/// internal token types and, in the Debug forms, leaked the parser's own
574/// enum into a message clients read.
575///
576/// Applied once on the way out, so every construction site is covered and
577/// the token named is the one the error itself points at. Messages whose
578/// bodies are already PG's verbatim (`LIMIT must not be negative`,
579/// `invalid input syntax for type bigint: "abc"`) are left alone — those
580/// are PG's own errors, not its syntax error.
581fn shape_syntax_error(e: ParseError, input: &str, offsets: &[usize]) -> ParseError {
582 if !(e.message.starts_with("expected ") || e.message.starts_with("unexpected token ")) {
583 return e;
584 }
585 let message = match offending_lexeme(input, offsets, e.token_pos) {
586 Some(tok) => alloc::format!("syntax error at or near \"{tok}\""),
587 None => "syntax error at end of input".into(),
588 };
589 ParseError {
590 message,
591 token_pos: e.token_pos,
592 }
593}
594
595/// v7.39 (round 340, V56) — a lexer-level failure the way PG words it.
596/// Measured on 18.4: `unterminated quoted string at or near "'abc"`,
597/// `unterminated quoted identifier at or near ""abc"`, `unterminated /*
598/// comment at or near "/* x"` — the quoted part runs from the opening
599/// delimiter to the end of the input. SPG reported its own internal
600/// shape instead (`unterminated string literal at byte 7`), which named
601/// a byte offset no client can use.
602fn shape_lex_error(e: &lexer::LexError, input: &str) -> ParseError {
603 use lexer::LexErrorKind as K;
604 let from_here = input.get(e.pos..).map(str::trim_end).unwrap_or("");
605 let message = match &e.kind {
606 K::UnterminatedString => {
607 alloc::format!("unterminated quoted string at or near \"{from_here}\"")
608 }
609 K::UnterminatedQuotedIdent => {
610 alloc::format!("unterminated quoted identifier at or near \"{from_here}\"")
611 }
612 K::UnterminatedBlockComment => {
613 alloc::format!("unterminated /* comment at or near \"{from_here}\"")
614 }
615 // PG has no "unknown character" error of its own — the character
616 // is skipped and the parser reports the next token. SPG stops at
617 // the character itself and names it, which is the same shape.
618 K::UnknownChar(c) => alloc::format!("syntax error at or near \"{c}\""),
619 // The number-literal kinds already carry PG's `at or near` form.
620 other => alloc::format!(
621 "{}",
622 lexer::LexError {
623 kind: other.clone(),
624 pos: e.pos,
625 }
626 ),
627 };
628 ParseError {
629 message,
630 token_pos: 0,
631 }
632}
633
634/// The offending token exactly as it appears in the input, or `None` at
635/// end of input. PG echoes the source spelling — a lower-case `frm`
636/// reports as `frm`, not as a canonicalised keyword.
637fn offending_lexeme<'a>(input: &'a str, offsets: &[usize], token_pos: usize) -> Option<&'a str> {
638 let start = *offsets.get(token_pos)?;
639 if start >= input.len() {
640 return None;
641 }
642 let end = offsets
643 .get(token_pos + 1)
644 .copied()
645 .unwrap_or(input.len())
646 .min(input.len());
647 let seg = input.get(start..end)?.trim();
648 if seg.is_empty() {
649 return None;
650 }
651 // A quoted literal / identifier keeps its inner spaces; anything else
652 // ends at the first whitespace (the segment runs to the NEXT token's
653 // start, which may swallow a comment).
654 if seg.starts_with('\'') || seg.starts_with('"') || seg.starts_with('`') {
655 Some(seg)
656 } else {
657 seg.split_whitespace().next()
658 }
659}
660
661/// v7.39 (read01 round 95) — recover PG's 1-based CHARACTER error position for
662/// a [`ParseError::token_pos`]. Kept off the `ParseError` struct (and so off
663/// every recursive `Result` slot) to protect the nesting-budget frame cliff:
664/// this re-tokenizes `input` on the cold error path to map the failing token
665/// index to its start byte, then to a character offset. The dialect
666/// must match the parse that produced `token_pos` (it barely shifts offsets,
667/// but stay consistent). Returns `None` when the index has no offset or the
668/// byte isn't a char boundary. The wire attaches it as the ErrorResponse `P`.
669#[must_use]
670pub fn syntax_error_position(
671 input: &str,
672 dialect: lexer::Dialect,
673 token_pos: usize,
674) -> Option<usize> {
675 let (_, offsets) = lexer::tokenize_with_offsets(input, dialect).ok()?;
676 let byte_off = *offsets.get(token_pos)?;
677 if byte_off > input.len() || !input.is_char_boundary(byte_off) {
678 return None;
679 }
680 Some(input[..byte_off].chars().count() + 1)
681}
682
683struct Parser {
684 tokens: Vec<Token>,
685 pos: usize,
686 /// v7.39 (round 274) — the session's dialect, carried by the same
687 /// signal that drives string-literal escaping: `SET sql_mode` (only
688 /// MySQL clients and mysqldump preambles emit it) turns it on,
689 /// `SET standard_conforming_strings` (every pg_dump preamble) turns
690 /// it off. Needed here because the two dialects disagree about what
691 /// `REAL` means — see the type mapping below.
692 mysql_dialect: bool,
693 /// v7.30.2 (mailrs round-25 ask 2) — live nesting depth of the
694 /// mutually recursive expr/select parsers. Bounded so a deeply
695 /// nested input returns a parse error instead of overflowing
696 /// the stack (embed hosts die on overflow — it is an abort,
697 /// not a catchable error).
698 nest_depth: usize,
699 /// TABLESAMPLE lowering channel: the table-ref parser pushes a
700 /// `random() < p/100` predicate here; the enclosing SELECT
701 /// drains the list after its WHERE parses and ANDs the
702 /// predicates in. parse_bare_select save/restores around its
703 /// FROM+WHERE so nested selects only drain their own.
704 pending_sample_preds: Vec<Expr>,
705 /// v7.38.19 — the target of a `SELECT … INTO <table>`, carried out
706 /// of `parse_bare_select` (which returns a `SelectStatement` and has
707 /// nowhere to put it) to the caller that lowers the pair to the CTAS
708 /// node. `bool` is `TEMP`.
709 pending_select_into: Option<(String, bool)>,
710 /// v7.39 (round 691) — collation lowering channel, the same shape as
711 /// `pending_sample_preds` above. `expr COLLATE "name"` is ORDERING
712 /// information, and `ast::OrderBy` is where this parser keeps ordering
713 /// information (`desc`, `nulls_first`); the alternative — a new `Expr`
714 /// variant — puts a new arm on `eval_expr`, which this repo has
715 /// measured to overflow the debug stack. So while an ORDER BY KEY is
716 /// being parsed the postfix loop drops the name here instead of
717 /// refusing it, and the key's parser takes it.
718 ///
719 /// Only inside an ORDER BY key: everywhere else an unperformable
720 /// collation still errors, because accepting one at a COMPARISON and
721 /// ignoring it is the defect F36 exists to close.
722 in_order_by_key: bool,
723 order_key_collation: Option<String>,
724 /// POSITION(sub IN str) — while parsing the needle, the IN
725 /// keyword is the argument separator, not a membership test.
726 /// The postfix loop leaves IN unconsumed when this is set.
727 suppress_in_tail: bool,
728 /// Index of the token the last `advance()` returned — see
729 /// [`Parser::consumed_pos`].
730 last_consumed: usize,
731 /// v7.39 (round 506) — the statement's own text and the byte each token
732 /// starts at, so a MySQL projection item can report the SOURCE TEXT
733 /// MariaDB reports: `SELECT a + b` names its column `a + b`,
734 /// spacing and all. Only filled for a MySQL session — a PG one names
735 /// columns from the parsed shape and pays nothing for this.
736 src: Option<(String, Vec<usize>)>,
737 /// v7.39.3 — (token index, first-segment byte length) for every
738 /// string literal the lexer built by implicit concatenation.
739 merges: Vec<(usize, usize)>,
740}
741
742/// Max expr/select parser nesting (parens, subqueries, CASE, …).
743/// Real SQL nests a few dozen levels at the extreme. Each nesting level
744/// costs a parse_expr→parse_unary→parse_atom frame chain, so the budget
745/// exists to turn a deep statement into a catchable parse ERROR: a stack
746/// overflow is an abort, and in the server it does not fail one query, it
747/// takes the process down and every other connection with it.
748///
749/// v7.39 (round 507) — measured, because the figure here used to be a
750/// guess ("over 10 KiB in debug … comfortably inside a 2 MiB worker stack
751/// in BOTH debug and release"), and the debug half of that is wrong by
752/// more than an order of magnitude:
753///
754/// * RELEASE, on a 2 MiB worker stack: every recursive shape reaches
755/// this budget and errors. Verified against a live server for nested
756/// derived tables, parens, calls, CASE, IN-subqueries, scalar
757/// subqueries, NOT and unary minus — the server stayed up through all
758/// of them. This is the contract that matters, and it holds.
759/// * DEBUG: nested derived tables cost roughly 235 KiB of stack PER
760/// LEVEL, so parsing alone aborts around 35 levels on an 8 MiB stack
761/// and executing aborts around 8 inside a test thread. The budget is
762/// simply unreachable there, which is why a deep-nesting test has to
763/// ask for a large stack of its own — see `nesting_budget_errors_at`
764/// in the parser tests.
765/// v7.39 (round 541) — the pg_catalog relations SPG synthesises, in
766/// one place.
767///
768/// There were two copies of this fact: a curated list, used for BARE
769/// names, and — in `try_peek_meta_qualified` — no list at all, which
770/// rewrote `pg_catalog.<anything>` to `__spg_pg_<anything>` and left
771/// the engine to complain about a view it could not materialise. So
772/// writing the schema qualifier CHANGED THE ANSWER: `pg_stat_activity`
773/// had rows, `pg_catalog.pg_stat_activity` was an error.
774///
775/// PG puts `pg_catalog` at the implicit front of every search_path, so
776/// the two spellings name the same relation and must resolve the same
777/// way. Names NOT here (`pg_stat_activity`, `pg_locks`,
778/// `pg_stat_statements`, `pg_statio_user_tables`) route through the
779/// meta_view_result path under their own names and must not be
780/// rewritten; a name that is neither reaches the ordinary resolver,
781/// which reports that the relation does not exist — PG's answer.
782const SYNTHESISED_PG_CATALOGS: &[&str] = &[
783 "pg_am",
784 "pg_attrdef",
785 "pg_attribute",
786 "pg_cast",
787 "pg_db_role_setting",
788 "pg_conversion",
789 "pg_default_acl",
790 "pg_shadow",
791 "pg_sequences",
792 "pg_range",
793 "pg_partitioned_table",
794 "pg_language",
795 "pg_group",
796 "pg_authid",
797 "pg_class",
798 "pg_collation",
799 "pg_constraint",
800 "pg_database",
801 "pg_depend",
802 "pg_amop",
803 "pg_amproc",
804 "pg_opclass",
805 "pg_opfamily",
806 // v7.39 (read01 round 50) — COMMENT ON store, PG's pg_description.
807 "pg_description",
808 "pg_enum",
809 "pg_extension",
810 // v7.39 (round 541) — pg_dump reads it for every relation of kind
811 // 'f'. SPG has no foreign tables, so it is empty, which is also
812 // what PG reports on a database that has none.
813 "pg_foreign_table",
814 // v7.39 (round 541) — the empty-by-truth family; see
815 // EMPTY_PG_CATALOGS in spg-engine::system_catalog.
816 "pg_event_trigger",
817 "pg_file_settings",
818 "pg_foreign_data_wrapper",
819 "pg_foreign_server",
820 "pg_hba_file_rules",
821 "pg_ident_file_mappings",
822 "pg_init_privs",
823 "pg_parameter_acl",
824 "pg_prepared_xacts",
825 "pg_publication_namespace",
826 "pg_publication_rel",
827 "pg_publication_tables",
828 "pg_replication_origin",
829 "pg_replication_origin_status",
830 "pg_seclabel",
831 "pg_seclabels",
832 "pg_shdepend",
833 "pg_shdescription",
834 "pg_shmem_allocations",
835 "pg_shmem_allocations_numa",
836 "pg_shseclabel",
837 "pg_statistic_ext_data",
838 "pg_stats_ext",
839 "pg_stats_ext_exprs",
840 "pg_subscription_rel",
841 "pg_transform",
842 "pg_user_mapping",
843 "pg_user_mappings",
844 "pg_index",
845 "pg_indexes",
846 "pg_inherits",
847 // v7.39 (round 650) — the text-search catalogs SPG can fill
848 // honestly. `pg_ts_config_map` is deliberately NOT here: it maps
849 // token types to dictionaries and SPG has no token-type model,
850 // the same gap that leaves `ts_token_type` / `ts_debug` unbuilt.
851 "pg_ts_config",
852 "pg_ts_config_map",
853 "pg_ts_dict",
854 "pg_ts_parser",
855 "pg_ts_template",
856 "pg_matviews",
857 "pg_namespace",
858 // v7.39 (round 621)
859 "pg_operator",
860 "pg_policies",
861 "pg_policy",
862 "pg_proc",
863 "pg_publication",
864 "pg_replication_slots",
865 "pg_roles",
866 // v7.39 (round 143) — the rewrite-rule listing view.
867 // v7.39 (round 312) — and the rule catalogue itself, which
868 // `pg_get_ruledef(oid)` resolves against.
869 "pg_rewrite",
870 "pg_rules",
871 "pg_sequence",
872 "pg_settings",
873 "pg_stat_archiver",
874 "pg_stat_bgwriter",
875 "pg_stat_checkpointer",
876 "pg_stat_database",
877 "pg_stat_io",
878 "pg_stat_progress_analyze",
879 "pg_auth_members",
880 "pg_stat_progress_create_index",
881 "pg_stat_progress_vacuum",
882 "pg_stat_replication",
883 "pg_stat_slru",
884 "pg_stat_subscription_stats",
885 "pg_stat_user_functions",
886 "pg_stat_user_indexes",
887 "pg_stat_user_tables",
888 "pg_stat_wal",
889 "pg_prepared_statements",
890 "pg_largeobject",
891 "pg_largeobject_metadata",
892 "pg_statistic",
893 "pg_statistic_ext",
894 // v7.38.18 — the readable view over pg_statistic.
895 "pg_stats",
896 "pg_subscription",
897 "pg_tables",
898 "pg_tablespace",
899 // v7.39 (round 502) — the timezone catalogues. SPG resolved
900 // named zones correctly but could not list them, so a client
901 // populating a timezone picker got "relation does not exist".
902 "pg_timezone_abbrevs",
903 "pg_timezone_names",
904 "pg_trigger",
905 "pg_type",
906 "pg_user",
907 "pg_views",
908];
909
910const MAX_NEST_DEPTH: usize = 64;
911
912/// Stack accounting for the nesting budget, test-only.
913///
914/// `MAX_NEST_DEPTH` is a fixed count calibrated against a frame size
915/// that MOVES: a compiler upgrade grew the parser's debug frames and
916/// silently ate the margin until `nesting_budget_errors_cleanly` went
917/// from erroring cleanly to aborting on a stack overflow. A count
918/// cannot notice that on its own, so the budget is measured here and
919/// held to a ceiling.
920///
921/// The reading has to come from a helper whose OWN frame is the same at
922/// every call: debug slot placement does not follow source order, so a
923/// local's address inside the function under test is not that
924/// function's frame boundary. Two earlier probes were wrong that way —
925/// one read `&self.nest_depth`, which is the `Parser`'s address and
926/// never moves at all.
927#[cfg(test)]
928mod frame_meter {
929 extern crate std;
930 use std::cell::Cell;
931
932 // Per-THREAD, not global. `cargo test` runs tests in parallel and
933 // plenty of them parse nested expressions, so shared statics get
934 // stack addresses from several threads at once and the subtraction
935 // below turns into noise — it read 229,772 bytes per level that way,
936 // while passing when the test was run on its own.
937 std::thread_local! {
938 static AT_LO: Cell<usize> = const { Cell::new(0) };
939 static AT_HI: Cell<usize> = const { Cell::new(0) };
940 }
941
942 pub(super) const SAMPLE_LO: usize = 4;
943 pub(super) const SAMPLE_HI: usize = 24;
944
945 #[inline(never)]
946 pub(super) fn record(depth: usize) {
947 let anchor = 0u8;
948 let at = core::ptr::from_ref(&anchor) as usize;
949 if depth == SAMPLE_LO {
950 AT_LO.with(|c| c.set(at));
951 } else if depth == SAMPLE_HI {
952 AT_HI.with(|c| c.set(at));
953 }
954 }
955
956 /// Bytes of stack one nesting level costs, averaged over the span.
957 pub(super) fn bytes_per_level() -> usize {
958 let lo = AT_LO.with(Cell::get);
959 let hi = AT_HI.with(Cell::get);
960 assert!(lo > 0 && hi > 0, "meter never sampled: lo={lo} hi={hi}");
961 assert!(lo > hi, "stack grew upwards? lo={lo} hi={hi}");
962 (lo - hi) / (SAMPLE_HI - SAMPLE_LO)
963 }
964
965 pub(super) fn reset() {
966 AT_LO.with(|c| c.set(0));
967 AT_HI.with(|c| c.set(0));
968 }
969}
970
971/// v7.39 (read01 geo_ops.c) — prefix `@@` desugar target, out-of-line so
972/// the constructor's temporaries stay off `parse_unary`'s recursion frame.
973#[inline(never)]
974fn build_center_call(e: Expr) -> Expr {
975 Expr::FunctionCall {
976 name: alloc::string::String::from("center"),
977 args: alloc::vec![e],
978 }
979}
980
981/// Max consecutive binary operators at ONE precedence level
982/// (`a OR b OR c …`, `1+1+1…`). The chain builds iteratively at
983/// parse time but evaluates and drops recursively — depth beyond
984/// this overflows 2 MiB worker stacks (debug eval frames run
985/// multiple KiB). `IN (…)` lists are flat and unaffected.
986const MAX_BINARY_CHAIN: usize = 256;
987
988/// v7.22 (round-13 gap 5) — the kind keyword after `CONSTRAINT
989/// <name>` in a CREATE TABLE column list. FOREIGN KEY is not here:
990/// it keeps its dedicated path (`parse_table_level_fk`).
991enum NamedTableConstraintKind {
992 Check,
993 Unique,
994 PrimaryKey,
995 Exclude,
996}
997
998impl Parser {
999 fn new(tokens: Vec<Token>) -> Self {
1000 Self::new_with_dialect(tokens, false)
1001 }
1002
1003 fn new_with_dialect(tokens: Vec<Token>, mysql_dialect: bool) -> Self {
1004 Self {
1005 tokens,
1006 mysql_dialect,
1007 in_order_by_key: false,
1008 order_key_collation: None,
1009 pos: 0,
1010 nest_depth: 0,
1011 pending_sample_preds: Vec::new(),
1012 pending_select_into: None,
1013 suppress_in_tail: false,
1014 last_consumed: 0,
1015 src: None,
1016 merges: Vec::new(),
1017 }
1018 }
1019
1020 /// Hand the parser the text it is parsing, for [`Parser::source_span`].
1021 fn with_source(mut self, input: &str, offsets: &[usize]) -> Self {
1022 if self.mysql_dialect {
1023 self.src = Some((input.to_string(), offsets.to_vec()));
1024 }
1025 self
1026 }
1027
1028 /// v7.39.3 — the implicit-concatenation log from the lexer, so a
1029 /// merged literal can still be LABELLED by its first segment the way
1030 /// MySQL 9.7.2 labels it.
1031 fn with_merges(mut self, merges: Vec<(usize, usize)>) -> Self {
1032 if self.mysql_dialect {
1033 self.merges = merges;
1034 }
1035 self
1036 }
1037
1038 /// The byte length of the first segment of the literal at `tok`, when
1039 /// that literal was built by implicit concatenation.
1040 fn merged_first_len(&self, tok: usize) -> Option<usize> {
1041 self.merges
1042 .iter()
1043 .find(|(k, _)| *k == tok)
1044 .map(|(_, len)| *len)
1045 }
1046
1047 /// The source text spanning tokens `start ..= end`, trimmed.
1048 ///
1049 /// The offsets are token STARTS, so the span runs to the start of the
1050 /// token after `end` and gives back the whitespace between them —
1051 /// trimming is what makes `a + b FROM t` end at `b`.
1052 fn source_span(&self, start: usize, end: usize) -> Option<&str> {
1053 let (text, offsets) = self.src.as_ref()?;
1054 let from = *offsets.get(start)?;
1055 let to = *offsets.get(end + 1)?;
1056 text.get(from..to).map(str::trim_end)
1057 }
1058
1059 /// v7.30.2 (mailrs round-25 ask 2) — bump the expr/select
1060 /// nesting depth, erroring out cleanly past the budget.
1061 fn enter_nested(&mut self) -> Result<(), ParseError> {
1062 self.nest_depth += 1;
1063 #[cfg(test)]
1064 frame_meter::record(self.nest_depth);
1065 if self.nest_depth > MAX_NEST_DEPTH {
1066 self.nest_depth -= 1;
1067 return Err(self.err(alloc::format!(
1068 "statement nests deeper than {MAX_NEST_DEPTH} levels"
1069 )));
1070 }
1071 Ok(())
1072 }
1073
1074 fn peek(&self) -> &Token {
1075 // tokens always ends with Eof; pos is clamped in advance().
1076 &self.tokens[self.pos]
1077 }
1078
1079 fn advance(&mut self) -> Token {
1080 let t = mem::replace(&mut self.tokens[self.pos], Token::Eof);
1081 self.last_consumed = self.pos;
1082 if self.pos + 1 < self.tokens.len() {
1083 self.pos += 1;
1084 }
1085 t
1086 }
1087
1088 /// v7.39 (round 340, V56) — the index of the token `advance()` just
1089 /// returned. It was computed as `pos - 1`, which is wrong at both
1090 /// ends: `advance()` parks on the final Eof rather than running off
1091 /// the end (so `SELECT * FROM` named `FROM` where PG says `at end of
1092 /// input`), and after backtracking `pos` is no longer one past the
1093 /// token that failed. Recorded by `advance()` itself instead.
1094 const fn consumed_pos(&self) -> usize {
1095 self.last_consumed
1096 }
1097
1098 fn err(&self, message: String) -> ParseError {
1099 ParseError {
1100 message,
1101 token_pos: self.pos,
1102 }
1103 }
1104
1105 /// v7.39.3 — like [`Parser::err`] but pointing at a token the caller
1106 /// names rather than at the current one.
1107 ///
1108 /// The position is not decoration on the MySQL wire: its syntax-error
1109 /// sentence quotes the source from there to the end of the statement,
1110 /// so an error raised after the construct it is about quotes nothing.
1111 fn err_at(&self, token_pos: usize, message: String) -> ParseError {
1112 ParseError { message, token_pos }
1113 }
1114
1115 fn expect_eof(&self) -> Result<(), ParseError> {
1116 if matches!(self.peek(), Token::Eof) {
1117 Ok(())
1118 } else {
1119 Err(self.err(format!("expected end of input, got {:?}", self.peek())))
1120 }
1121 }
1122
1123 /// v7.14.0 — swallow every token up to (but not including) the
1124 /// next semicolon / EOF. Used by the dump-noise dispatcher
1125 /// to consume `COMMENT ON …`, `GRANT …`, `LOCK TABLES …`,
1126 /// etc. without modeling each grammar.
1127 /// v7.39 (read01 round 50) — `COMMENT ON <kind> <name> IS { 'text' | NULL }`.
1128 /// Kinds SPG stores by name: TABLE / COLUMN / INDEX / VIEW / SEQUENCE /
1129 /// SCHEMA / TYPE / DATABASE / FUNCTION. Anything else (and the multi-word
1130 /// `CONSTRAINT c ON t` / `MATERIALIZED VIEW` forms) keeps the old
1131 /// swallow-as-no-op behaviour so a pg_dump tail still loads.
1132 fn parse_comment_on(&mut self) -> Result<Statement, ParseError> {
1133 let start = self.pos;
1134 self.advance(); // COMMENT
1135 if !matches!(self.peek(), Token::On) {
1136 self.pos = start;
1137 self.consume_until_statement_boundary();
1138 return Ok(Statement::Empty);
1139 }
1140 self.advance(); // ON
1141 let kind = match self.peek() {
1142 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
1143 Token::Table => "table".into(),
1144 _ => {
1145 self.consume_until_statement_boundary();
1146 return Ok(Statement::Empty);
1147 }
1148 };
1149 if !matches!(
1150 kind.as_str(),
1151 "table"
1152 | "column"
1153 | "index"
1154 | "view"
1155 | "sequence"
1156 | "schema"
1157 | "type"
1158 | "database"
1159 | "function"
1160 ) {
1161 self.consume_until_statement_boundary();
1162 return Ok(Statement::Empty);
1163 }
1164 self.advance(); // the kind keyword
1165 // The object name. ⚠️ `expect_ident_like` strips a leading
1166 // `<schema>.` qualifier and returns only the trailing ident (SPG is
1167 // single-schema), which would turn `COMMENT ON COLUMN t.c` into just
1168 // `c`. Read the dotted parts from raw tokens instead, then let a
1169 // 3-part `schema.t.c` drop its leading schema like everywhere else.
1170 let mut parts: alloc::vec::Vec<String> = alloc::vec::Vec::new();
1171 loop {
1172 match self.advance() {
1173 Token::Ident(s) | Token::QuotedIdent(s) => parts.push(s),
1174 other if unreserved_keyword_text(&other).is_some() => {
1175 parts.push(unreserved_keyword_text(&other).unwrap());
1176 }
1177 other => {
1178 return Err(ParseError {
1179 message: alloc::format!("expected identifier, got {other:?}"),
1180 token_pos: self.consumed_pos(),
1181 });
1182 }
1183 }
1184 if matches!(self.peek(), Token::Dot) {
1185 self.advance();
1186 } else {
1187 break;
1188 }
1189 }
1190 // COLUMN wants `table.column`; every other kind wants a bare name.
1191 let want = if kind == "column" { 2 } else { 1 };
1192 while parts.len() > want {
1193 parts.remove(0);
1194 }
1195 let name = parts.join(".");
1196 // v7.39 (round 710) — `COMMENT ON FUNCTION f(int, text) IS …`.
1197 // pg_dump writes the SIGNATURE, and the paren list was a syntax
1198 // error here — a dump carrying one function comment failed to
1199 // restore. The list is consumed (the comment store keys by name;
1200 // overload-precise comments are the function-predicate follow-up).
1201 if matches!(self.peek(), Token::LParen)
1202 && matches!(
1203 kind.as_str(),
1204 "function" | "procedure" | "aggregate" | "routine"
1205 )
1206 {
1207 let mut depth = 0usize;
1208 loop {
1209 match self.advance() {
1210 Token::LParen => depth += 1,
1211 Token::RParen => {
1212 depth -= 1;
1213 if depth == 0 {
1214 break;
1215 }
1216 }
1217 Token::Eof => {
1218 return Err(self.err(alloc::string::String::from(
1219 "unterminated argument list in COMMENT ON",
1220 )));
1221 }
1222 _ => {}
1223 }
1224 }
1225 }
1226 // `IS`
1227 if !matches!(self.peek(), Token::Is) {
1228 self.expect_keyword_ident("is")?;
1229 } else {
1230 self.advance();
1231 }
1232 let comment = match self.peek() {
1233 Token::Null => {
1234 self.advance();
1235 None
1236 }
1237 _ => Some(self.expect_string_literal()?),
1238 };
1239 Ok(Statement::CommentOn {
1240 kind,
1241 name,
1242 comment,
1243 })
1244 }
1245
1246 /// v7.39 (read01 round 57) — `GRANT <privs> ON <obj> TO <roles> [WITH GRANT
1247 /// OPTION]` / `REVOKE [GRANT OPTION FOR] <privs> ON <obj> FROM <roles>
1248 /// [CASCADE|RESTRICT]`.
1249 ///
1250 /// TABLE privileges are the real ones (stored, enforced, introspectable).
1251 /// Every other object class — SCHEMA / SEQUENCE / DATABASE / FUNCTION / …,
1252 /// and the no-ON `GRANT role TO role` membership form — parses into
1253 /// `GrantObject::Other` and no-ops in the engine, so a pg_dump that grants
1254 /// on them still restores.
1255 fn parse_grant_or_revoke(&mut self, grant: bool) -> Result<Statement, ParseError> {
1256 self.advance(); // GRANT / REVOKE
1257 // REVOKE's optional `GRANT OPTION FOR` prefix.
1258 let mut grant_option = false;
1259 if !grant
1260 && self.peek_keyword_ident("grant")
1261 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("option"))
1262 {
1263 self.advance(); // GRANT
1264 self.advance(); // OPTION
1265 self.expect_keyword_ident("for")?;
1266 grant_option = true;
1267 }
1268 // The privilege list: `ALL [PRIVILEGES] [(cols)]`, or comma-separated
1269 // words each with an optional COLUMN list.
1270 let mut privileges: Vec<GrantPriv> = Vec::new();
1271 if matches!(self.peek(), Token::All) {
1272 self.advance();
1273 if self.peek_keyword_ident("privileges") {
1274 self.advance();
1275 }
1276 // `GRANT ALL (col) ON t TO r` — every column privilege, on that
1277 // column only.
1278 if matches!(self.peek(), Token::LParen) {
1279 let columns = self.parse_grant_column_list()?;
1280 privileges.push(GrantPriv {
1281 word: "ALL".into(),
1282 columns,
1283 });
1284 }
1285 } else {
1286 loop {
1287 // SELECT and INSERT lex as reserved tokens, so they never
1288 // reach `expect_ident_like` as plain idents; the rest
1289 // (UPDATE / DELETE / TRUNCATE / REFERENCES / TRIGGER /
1290 // MAINTAIN) are ordinary identifiers.
1291 let w = match self.peek() {
1292 Token::Select => {
1293 self.advance();
1294 "SELECT".to_string()
1295 }
1296 Token::Insert => {
1297 self.advance();
1298 "INSERT".to_string()
1299 }
1300 // v7.39 (read01 round 60) — CREATE is a privilege word on a
1301 // schema / database, and it lexes as a reserved token.
1302 Token::Create => {
1303 self.advance();
1304 "CREATE".to_string()
1305 }
1306 // NOT upper-cased: in the no-ON shape (`GRANT devs TO
1307 // alice`) these "privilege words" are ROLE NAMES, and a
1308 // role name is case-sensitive. `priv_from_word` folds case
1309 // itself when they really are privileges.
1310 _ => self.expect_ident_like()?,
1311 };
1312 // v7.39 (read01 round 59) — the optional per-privilege COLUMN
1313 // list: `GRANT SELECT (a, b), INSERT (c) ON t TO dan`.
1314 let columns = if matches!(self.peek(), Token::LParen) {
1315 self.parse_grant_column_list()?
1316 } else {
1317 Vec::new()
1318 };
1319 privileges.push(GrantPriv { word: w, columns });
1320 if matches!(self.peek(), Token::Comma) {
1321 self.advance();
1322 } else {
1323 break;
1324 }
1325 }
1326 }
1327 // v7.39 (read01 round 58) — no ON clause at all = `GRANT devs TO alice`:
1328 // role MEMBERSHIP. The words parsed as "privileges" are the role names.
1329 if !matches!(self.peek(), Token::On) {
1330 let roles: Vec<String> = core::mem::take(&mut privileges)
1331 .into_iter()
1332 .map(|p| p.word)
1333 .collect();
1334 let grantees = self.parse_grantee_list(grant)?;
1335 // `WITH ADMIN OPTION` / `GRANTED BY x` — accepted, ignored (SPG has
1336 // no admin-option layer: a member cannot re-grant).
1337 self.consume_until_statement_boundary();
1338 return Ok(finish_grant(
1339 grant,
1340 GrantStatement {
1341 privileges: Vec::new(),
1342 object: GrantObject::Roles(roles),
1343 grantees,
1344 grant_option,
1345 },
1346 ));
1347 }
1348 self.advance(); // ON
1349 // An optional object-class keyword. `TABLE` (or no keyword at all) is
1350 // the enforced case; anything else parses and no-ops.
1351 let mut class = "TABLE";
1352 match self.peek() {
1353 Token::Table => {
1354 self.advance();
1355 }
1356 Token::All => {
1357 // v7.39 (read01 round 61) — `ALL TABLES IN SCHEMA x` expands to
1358 // every table at GRANT time, like PG. `ALL SEQUENCES/FUNCTIONS
1359 // IN SCHEMA` stay no-ops and keep their own object class.
1360 self.advance(); // ALL
1361 let kind = match self.peek() {
1362 Token::Ident(w) | Token::QuotedIdent(w) => w.to_ascii_lowercase(),
1363 // TABLES has its own token (SHOW TABLES owns it).
1364 Token::Tables | Token::Table => "tables".to_string(),
1365 _ => String::new(),
1366 };
1367 if !kind.is_empty() {
1368 self.advance();
1369 }
1370 // `IN SCHEMA <name>`
1371 if matches!(self.peek(), Token::In) {
1372 self.advance();
1373 if self.peek_keyword_ident("schema") {
1374 self.advance();
1375 let _schema = self.expect_ident_like()?;
1376 }
1377 }
1378 if kind != "tables" {
1379 self.consume_until_statement_boundary();
1380 return Ok(finish_grant(
1381 grant,
1382 GrantStatement {
1383 privileges,
1384 object: GrantObject::Other("ALL … IN SCHEMA".into()),
1385 grantees: Vec::new(),
1386 grant_option,
1387 },
1388 ));
1389 }
1390 let grantees = self.parse_grantee_list(grant)?;
1391 self.consume_until_statement_boundary();
1392 return Ok(finish_grant(
1393 grant,
1394 GrantStatement {
1395 privileges,
1396 object: GrantObject::AllTablesInSchema,
1397 grantees,
1398 grant_option,
1399 },
1400 ));
1401 }
1402 Token::Ident(w) | Token::QuotedIdent(w) => {
1403 let lc = w.to_ascii_lowercase();
1404 // v7.39 (read01 round 60) — SEQUENCE / SCHEMA / DATABASE are
1405 // real objects with real ACLs now.
1406 if matches!(lc.as_str(), "sequence" | "schema" | "database") {
1407 self.advance();
1408 let mut names: Vec<String> = Vec::new();
1409 loop {
1410 let mut parts: Vec<String> = Vec::new();
1411 loop {
1412 parts.push(self.expect_ident_like()?);
1413 if matches!(self.peek(), Token::Dot) {
1414 self.advance();
1415 } else {
1416 break;
1417 }
1418 }
1419 names.push(parts.pop().expect("at least one part"));
1420 if matches!(self.peek(), Token::Comma) {
1421 self.advance();
1422 } else {
1423 break;
1424 }
1425 }
1426 let grantees = self.parse_grantee_list(grant)?;
1427 let mut grant_option = grant_option;
1428 if grant && self.peek_keyword_ident("with") {
1429 self.advance();
1430 self.expect_keyword_ident("grant")?;
1431 self.expect_keyword_ident("option")?;
1432 grant_option = true;
1433 }
1434 self.consume_until_statement_boundary();
1435 let object = match lc.as_str() {
1436 "sequence" => GrantObject::Sequences(names),
1437 "schema" => GrantObject::Schemas(names),
1438 _ => GrantObject::Databases(names),
1439 };
1440 return Ok(finish_grant(
1441 grant,
1442 GrantStatement {
1443 privileges,
1444 object,
1445 grantees,
1446 grant_option,
1447 },
1448 ));
1449 }
1450 // v7.39 (read01 round 61) — `ON FUNCTION f(int)` is real. SPG
1451 // keys functions by NAME, so the argument list parses and is
1452 // dropped (an overload set shares one ACL — recorded residual).
1453 if matches!(lc.as_str(), "function" | "procedure" | "routine") {
1454 self.advance();
1455 let mut names: Vec<(String, Option<Vec<String>>)> = Vec::new();
1456 loop {
1457 let mut parts: Vec<String> = Vec::new();
1458 loop {
1459 parts.push(self.expect_ident_like()?);
1460 if matches!(self.peek(), Token::Dot) {
1461 self.advance();
1462 } else {
1463 break;
1464 }
1465 }
1466 let fname = parts.pop().expect("at least one part");
1467 // v7.39 (read01 round 62) — the signature picks the
1468 // overload, so it is captured.
1469 let sig = if matches!(self.peek(), Token::LParen) {
1470 Some(self.parse_function_signature_types()?)
1471 } else {
1472 None
1473 };
1474 names.push((fname, sig));
1475 if matches!(self.peek(), Token::Comma) {
1476 self.advance();
1477 } else {
1478 break;
1479 }
1480 }
1481 let grantees = self.parse_grantee_list(grant)?;
1482 self.consume_until_statement_boundary();
1483 return Ok(finish_grant(
1484 grant,
1485 GrantStatement {
1486 privileges,
1487 object: GrantObject::Functions(names),
1488 grantees,
1489 grant_option,
1490 },
1491 ));
1492 }
1493 if matches!(
1494 lc.as_str(),
1495 "type"
1496 | "domain"
1497 | "language"
1498 | "tablespace"
1499 | "large"
1500 | "foreign"
1501 | "parameter"
1502 ) {
1503 self.consume_until_statement_boundary();
1504 return Ok(finish_grant(
1505 grant,
1506 GrantStatement {
1507 privileges,
1508 object: GrantObject::Other(lc.to_ascii_uppercase()),
1509 grantees: Vec::new(),
1510 grant_option,
1511 },
1512 ));
1513 }
1514 class = "TABLE";
1515 }
1516 _ => {}
1517 }
1518 let _ = class;
1519 // The table list. Schema-qualified names drop their qualifier (SPG is
1520 // single-schema) — but read the dotted parts from raw tokens, since
1521 // `expect_ident_like` would silently swallow the leading part.
1522 let mut tables: Vec<String> = Vec::new();
1523 loop {
1524 let mut parts: Vec<String> = Vec::new();
1525 loop {
1526 parts.push(self.expect_ident_like()?);
1527 if matches!(self.peek(), Token::Dot) {
1528 self.advance();
1529 } else {
1530 break;
1531 }
1532 }
1533 tables.push(parts.pop().expect("at least one part"));
1534 if matches!(self.peek(), Token::Comma) {
1535 self.advance();
1536 } else {
1537 break;
1538 }
1539 }
1540 let grantees = self.parse_grantee_list(grant)?;
1541 if grant && self.peek_keyword_ident("with") {
1542 self.advance();
1543 self.expect_keyword_ident("grant")?;
1544 self.expect_keyword_ident("option")?;
1545 grant_option = true;
1546 }
1547 // REVOKE's trailing CASCADE / RESTRICT — SPG has no dependent grants
1548 // to cascade to (no re-granting), so both are accepted and ignored.
1549 if !grant && (self.peek_keyword_ident("cascade") || self.peek_keyword_ident("restrict")) {
1550 self.advance();
1551 }
1552 Ok(finish_grant(
1553 grant,
1554 GrantStatement {
1555 privileges,
1556 object: GrantObject::Tables(tables),
1557 grantees,
1558 grant_option,
1559 },
1560 ))
1561 }
1562
1563 /// v7.39 (read01 round 62) — the argument TYPES in a function signature:
1564 /// `(int, text)` or `(x int, y text)` (PG accepts either). Returns the type
1565 /// words; the caller normalises them into a signature key.
1566 fn parse_function_signature_types(&mut self) -> Result<Vec<String>, ParseError> {
1567 self.advance(); // (
1568 let mut types: Vec<String> = Vec::new();
1569 if matches!(self.peek(), Token::RParen) {
1570 self.advance();
1571 return Ok(types);
1572 }
1573 loop {
1574 // Collect the words of one argument up to a comma / close paren.
1575 let mut words: Vec<String> = Vec::new();
1576 loop {
1577 match self.peek() {
1578 Token::Comma | Token::RParen | Token::Eof => break,
1579 _ => {}
1580 }
1581 let tok = self.advance();
1582 match tok {
1583 Token::Ident(w) | Token::QuotedIdent(w) => words.push(w),
1584 other => {
1585 if let Some(w) = unreserved_keyword_text(&other) {
1586 words.push(w);
1587 }
1588 }
1589 }
1590 }
1591 // `name TYPE` or a bare `TYPE`. Several of PG's type names are
1592 // themselves several words (`double precision`, `character
1593 // varying`, `timestamp with time zone`), so "two words means the
1594 // first is a parameter name" reads the type off `f(double
1595 // precision)` as `precision`. v7.39 (round 282): recognise the
1596 // multi-word spellings first — a leading word that STARTS one of
1597 // them is part of the type, not a name.
1598 let joined = words.join(" ");
1599 let ty = if words.len() >= 2 && is_multiword_type_phrase(&joined) {
1600 joined
1601 } else if words.len() >= 2 {
1602 words[1..].join(" ")
1603 } else {
1604 words.first().cloned().unwrap_or_default()
1605 };
1606 types.push(ty);
1607 if matches!(self.peek(), Token::Comma) {
1608 self.advance();
1609 } else {
1610 break;
1611 }
1612 }
1613 if matches!(self.peek(), Token::RParen) {
1614 self.advance();
1615 }
1616 Ok(types)
1617 }
1618
1619 /// v7.39 (read01 round 59) — `( col, col, … )` after a privilege word.
1620 fn parse_grant_column_list(&mut self) -> Result<Vec<String>, ParseError> {
1621 self.advance(); // (
1622 let mut cols = Vec::new();
1623 loop {
1624 cols.push(self.expect_ident_like()?);
1625 if matches!(self.peek(), Token::Comma) {
1626 self.advance();
1627 } else {
1628 break;
1629 }
1630 }
1631 if !matches!(self.peek(), Token::RParen) {
1632 return Err(self.err(alloc::format!(
1633 "expected ')' to close the column list, got {:?}",
1634 self.peek()
1635 )));
1636 }
1637 self.advance(); // )
1638 Ok(cols)
1639 }
1640
1641 /// `TO <roles>` (grant) / `FROM <roles>` (revoke). An empty-string entry is
1642 /// PUBLIC.
1643 fn parse_grantee_list(&mut self, grant: bool) -> Result<Vec<String>, ParseError> {
1644 if grant {
1645 if matches!(self.peek(), Token::To) {
1646 self.advance();
1647 } else {
1648 self.expect_keyword_ident("to")?;
1649 }
1650 } else if matches!(self.peek(), Token::From) {
1651 self.advance();
1652 } else {
1653 self.expect_keyword_ident("from")?;
1654 }
1655 let mut grantees: Vec<String> = Vec::new();
1656 loop {
1657 // `GROUP name` is the legacy spelling of a plain role name.
1658 if self.peek_keyword_ident("group") {
1659 self.advance();
1660 }
1661 if self.peek_keyword_ident("public") {
1662 self.advance();
1663 grantees.push(String::new()); // PUBLIC
1664 } else {
1665 grantees.push(self.expect_ident_like()?);
1666 }
1667 if matches!(self.peek(), Token::Comma) {
1668 self.advance();
1669 } else {
1670 break;
1671 }
1672 }
1673 Ok(grantees)
1674 }
1675
1676 /// v7.39 (round 277) — `PREPARE <name> [(type, …)] AS <statement>`.
1677 /// The body keeps its `$N` placeholders; substitution happens at
1678 /// EXECUTE. The declared types are recorded for
1679 /// `pg_prepared_statements.parameter_types` but are not enforced —
1680 /// PG infers when the list is omitted, and SPG resolves the values
1681 /// at substitution time either way.
1682 fn parse_prepare(&mut self) -> Result<Statement, ParseError> {
1683 let start = self.pos;
1684 self.advance(); // PREPARE
1685 // v7.39 (round 278) — `PREPARE TRANSACTION '<gid>'` is 2PC, a
1686 // different statement that happens to share the keyword. PG
1687 // ships with `max_prepared_transactions = 0` and reports it
1688 // this way; SPG has no prepared-transaction registry, so the
1689 // same wording is the accurate answer rather than a dodge.
1690 // Round 277 turned this from a silent no-op into a confusing
1691 // "expected AS in PREPARE" parse error.
1692 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("transaction")) {
1693 self.advance();
1694 let gid = match self.advance() {
1695 Token::String(g) => g,
1696 other => {
1697 return Err(self.err(alloc::format!(
1698 "expected a transaction identifier after PREPARE TRANSACTION, got {other:?}"
1699 )));
1700 }
1701 };
1702 return Ok(Statement::PrepareTransaction(gid));
1703 }
1704 let name = self.expect_ident_like()?;
1705 let mut param_types = Vec::new();
1706 if matches!(self.peek(), Token::LParen) {
1707 self.advance();
1708 loop {
1709 let mut ty = self.expect_ident_like()?;
1710 // A parameterised type name (`numeric(10,2)`,
1711 // `varchar(20)`) keeps its argument list in the text.
1712 if matches!(self.peek(), Token::LParen) {
1713 let mut depth = 0usize;
1714 let mut buf = String::from("(");
1715 loop {
1716 match self.advance() {
1717 Token::LParen => {
1718 depth += 1;
1719 if depth > 1 {
1720 buf.push('(');
1721 }
1722 }
1723 Token::RParen => {
1724 depth -= 1;
1725 buf.push(')');
1726 if depth == 0 {
1727 break;
1728 }
1729 }
1730 Token::Comma => buf.push(','),
1731 Token::Integer(n) => buf.push_str(&alloc::format!("{n}")),
1732 Token::Eof => break,
1733 _ => {}
1734 }
1735 }
1736 ty.push_str(&buf);
1737 }
1738 // r1049 — `PREPARE p(bigint[]) AS …`: the sixth `[]`
1739 // position, same family as the parameter list above.
1740 let array_suffix = self.consume_array_suffix();
1741 ty.push_str(&array_suffix);
1742 param_types.push(ty);
1743 match self.peek() {
1744 Token::Comma => {
1745 self.advance();
1746 }
1747 Token::RParen => {
1748 self.advance();
1749 break;
1750 }
1751 other => {
1752 return Err(self.err(alloc::format!(
1753 "expected ',' or ')' in PREPARE parameter list, got {other:?}"
1754 )));
1755 }
1756 }
1757 }
1758 }
1759 if !matches!(self.peek(), Token::As) {
1760 return Err(self.err(alloc::format!(
1761 "expected AS in PREPARE, got {:?}",
1762 self.peek()
1763 )));
1764 }
1765 self.advance();
1766 let body = self.parse_one_statement()?;
1767 // The Parser holds tokens, not the source text, so the
1768 // statement PG reports in `pg_prepared_statements.statement`
1769 // is rebuilt from the AST rather than sliced from the input.
1770 let _ = start;
1771 let mut source = alloc::format!("PREPARE {}", crate::ast::quote_ident(&name));
1772 if !param_types.is_empty() {
1773 source.push_str(" (");
1774 source.push_str(¶m_types.join(", "));
1775 source.push(')');
1776 }
1777 source.push_str(" AS ");
1778 source.push_str(&alloc::format!("{body}"));
1779 Ok(Statement::Prepare {
1780 name,
1781 param_types,
1782 body: alloc::boxed::Box::new(body),
1783 source,
1784 })
1785 }
1786
1787 /// v7.39 (round 277) — `EXECUTE <name> [(<expr>, …)]`.
1788 fn parse_execute(&mut self) -> Result<Statement, ParseError> {
1789 self.advance(); // EXECUTE
1790 let name = self.expect_ident_like()?;
1791 let mut args = Vec::new();
1792 if matches!(self.peek(), Token::LParen) {
1793 self.advance();
1794 if matches!(self.peek(), Token::RParen) {
1795 self.advance();
1796 } else {
1797 loop {
1798 args.push(self.parse_expr(0)?);
1799 match self.advance() {
1800 Token::Comma => {}
1801 Token::RParen => break,
1802 other => {
1803 return Err(self.err(alloc::format!(
1804 "expected ',' or ')' in EXECUTE arguments, got {other:?}"
1805 )));
1806 }
1807 }
1808 }
1809 }
1810 }
1811 Ok(Statement::Execute { name, args })
1812 }
1813
1814 /// v7.39 (round 277) — `DEALLOCATE {[PREPARE] <name> | ALL}`.
1815 /// v7.39 (round 278) — `CALL <proc>([args])`. There is no
1816 /// procedure catalog yet, so this reports PG's not-found error
1817 /// (with its HINT) rather than pretending the call ran.
1818 /// v7.39 (round 320, V53) — `DISCARD { ALL | PLANS | SEQUENCES | TEMP }`.
1819 /// Bare `DISCARD` is a syntax error in PG; so it is here.
1820 fn parse_discard(&mut self) -> Result<Statement, ParseError> {
1821 self.advance(); // DISCARD
1822 let target = match self.advance() {
1823 Token::All => DiscardTarget::All,
1824 Token::Ident(w) | Token::QuotedIdent(w) => match w.to_ascii_lowercase().as_str() {
1825 "all" => DiscardTarget::All,
1826 "plans" => DiscardTarget::Plans,
1827 "sequences" => DiscardTarget::Sequences,
1828 "temp" | "temporary" => DiscardTarget::Temp,
1829 other => {
1830 return Err(self.err(format!(
1831 "expected ALL / PLANS / SEQUENCES / TEMP after DISCARD, got {other:?}"
1832 )));
1833 }
1834 },
1835 other => {
1836 return Err(self.err(format!(
1837 "expected ALL / PLANS / SEQUENCES / TEMP after DISCARD, got {other:?}"
1838 )));
1839 }
1840 };
1841 Ok(Statement::Discard(target))
1842 }
1843
1844 /// v7.39 (round 318, V51) — MySQL `KILL [HARD|SOFT] [CONNECTION|QUERY]
1845 /// <expr>`. MariaDB accepts an expression for the id (its own docs use
1846 /// `KILL connection_id()`), and the HARD / SOFT prefixes only pick how
1847 /// aggressively the server interrupts, which SPG does not distinguish.
1848 /// Bare `KILL <id>` means CONNECTION.
1849 fn parse_kill(&mut self) -> Result<Statement, ParseError> {
1850 self.advance(); // KILL
1851 let mut query_only = false;
1852 loop {
1853 // CONNECTION is a reserved keyword token (it also opens
1854 // `CREATE SUBSCRIPTION … CONNECTION '…'`), so it arrives as
1855 // `Token::Connection` rather than a bare ident.
1856 if matches!(self.peek(), Token::Connection) {
1857 self.advance();
1858 break;
1859 }
1860 let (Token::Ident(w) | Token::QuotedIdent(w)) = self.peek() else {
1861 break;
1862 };
1863 match w.to_ascii_lowercase().as_str() {
1864 "hard" | "soft" => {
1865 self.advance();
1866 }
1867 "query" => {
1868 self.advance();
1869 query_only = true;
1870 break;
1871 }
1872 _ => break,
1873 }
1874 }
1875 let id = self.parse_expr(0)?;
1876 Ok(Statement::Kill {
1877 query_only,
1878 id: Box::new(id),
1879 })
1880 }
1881
1882 fn parse_call(&mut self) -> Result<Statement, ParseError> {
1883 self.advance(); // CALL
1884 let name = self.expect_ident_like()?;
1885 self.consume_until_statement_boundary();
1886 Ok(Statement::Call(name))
1887 }
1888
1889 fn parse_deallocate(&mut self) -> Result<Statement, ParseError> {
1890 self.advance(); // DEALLOCATE
1891 // PG accepts an optional noise `PREPARE` keyword here.
1892 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("prepare")) {
1893 self.advance();
1894 }
1895 if matches!(self.peek(), Token::All) {
1896 self.advance();
1897 return Ok(Statement::Deallocate(None));
1898 }
1899 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("all")) {
1900 self.advance();
1901 return Ok(Statement::Deallocate(None));
1902 }
1903 let name = self.expect_ident_like()?;
1904 Ok(Statement::Deallocate(Some(name)))
1905 }
1906
1907 fn consume_until_statement_boundary(&mut self) {
1908 loop {
1909 match self.peek() {
1910 Token::Semicolon | Token::Eof => return,
1911 _ => self.advance(),
1912 };
1913 }
1914 }
1915
1916 /// v7.38.19 — the database name a `CREATE DATABASE` names, skipping
1917 /// an `IF NOT EXISTS`. Consumes only the name; the collation scanner
1918 /// runs after it and eats the rest.
1919 fn scan_database_name(&mut self) -> Option<String> {
1920 // The caller has only PEEKED at `DATABASE`; step past it, or the
1921 // first identifier found is the keyword itself. It was, and
1922 // `pg_database` listed a database called `database`.
1923 if matches!(self.peek(), Token::Ident(w) | Token::QuotedIdent(w) if w.eq_ignore_ascii_case("database"))
1924 {
1925 self.advance();
1926 }
1927 for kw in ["if", "not", "exists"] {
1928 if matches!(self.peek(), Token::Ident(w) | Token::QuotedIdent(w) if w.eq_ignore_ascii_case(kw))
1929 {
1930 self.advance();
1931 }
1932 }
1933 match self.peek().clone() {
1934 Token::Ident(w) | Token::QuotedIdent(w) => {
1935 self.advance();
1936 Some(w)
1937 }
1938 _ => None,
1939 }
1940 }
1941
1942 /// v7.38.18 — consume to the statement boundary like
1943 /// `consume_until_statement_boundary`, but pick out the collation a
1944 /// `CREATE DATABASE` asked for on the way.
1945 ///
1946 /// `LC_COLLATE 'de_DE.utf8'` and `LOCALE 'de_DE.utf8'` both count;
1947 /// `LC_CTYPE` does not, because SPG has no separate ctype and
1948 /// pretending to honour it would be the more misleading answer. An
1949 /// `=` between the keyword and the value is optional, as in PG.
1950 ///
1951 /// The whole statement used to be thrown away. Being single-database
1952 /// makes the NAME a no-op; it does not make the collation one.
1953 fn scan_database_collation_until_boundary(&mut self) -> Option<String> {
1954 let mut want_value = false;
1955 let mut found: Option<String> = None;
1956 loop {
1957 let tok = self.peek().clone();
1958 match &tok {
1959 Token::Semicolon | Token::Eof => break,
1960 Token::Ident(w) | Token::QuotedIdent(w)
1961 if w.eq_ignore_ascii_case("lc_collate") || w.eq_ignore_ascii_case("locale") =>
1962 {
1963 want_value = true;
1964 }
1965 Token::Eq if want_value => {}
1966 Token::String(v) if want_value => {
1967 found = Some(v.clone());
1968 want_value = false;
1969 }
1970 Token::Ident(v) | Token::QuotedIdent(v) if want_value => {
1971 found = Some(v.clone());
1972 want_value = false;
1973 }
1974 _ => want_value = false,
1975 }
1976 self.advance();
1977 }
1978 found
1979 }
1980
1981 /// v7.22 (round-13 T2) — consume to the statement boundary like
1982 /// `consume_until_statement_boundary`, but pick out the sequence
1983 /// name on the way: either `SEQUENCE NAME <ident>` (identity
1984 /// columns) or the first string literal (`nextval('<seq>')`).
1985 /// Schema qualifiers and `::regclass` casts are stripped.
1986 fn scan_sequence_name_until_boundary(&mut self) -> Option<String> {
1987 let mut seq: Option<String> = None;
1988 let mut after_sequence_kw = false;
1989 let mut after_name_kw = false;
1990 loop {
1991 match self.peek().clone() {
1992 Token::Semicolon | Token::Eof => break,
1993 Token::Ident(s) | Token::QuotedIdent(s) => {
1994 if after_name_kw && seq.is_none() {
1995 self.advance();
1996 let mut name = s;
1997 // `SEQUENCE NAME public.groups_id_seq` — keep
1998 // the bare name, drop qualifiers.
1999 while matches!(self.peek(), Token::Dot) {
2000 self.advance();
2001 if let Token::Ident(n) | Token::QuotedIdent(n) = self.advance() {
2002 name = n;
2003 }
2004 }
2005 seq = Some(name);
2006 after_name_kw = false;
2007 continue;
2008 }
2009 if after_sequence_kw && s.eq_ignore_ascii_case("name") {
2010 after_name_kw = true;
2011 after_sequence_kw = false;
2012 } else {
2013 after_sequence_kw = s.eq_ignore_ascii_case("sequence");
2014 }
2015 self.advance();
2016 }
2017 Token::String(s) => {
2018 if seq.is_none() {
2019 // `nextval('public.groups_id_seq'::regclass)`
2020 let bare = s
2021 .rsplit_once('.')
2022 .map_or_else(|| s.clone(), |(_, b)| b.to_string());
2023 seq = Some(bare);
2024 }
2025 self.advance();
2026 }
2027 _ => {
2028 after_sequence_kw = false;
2029 after_name_kw = false;
2030 self.advance();
2031 }
2032 }
2033 }
2034 seq
2035 }
2036
2037 /// v7.39 (round 621) — is the next token the keyword `BY`?
2038 ///
2039 /// `pg_get_keywords()` classes `by` as `U` (unreserved), so it is a legal
2040 /// column, table and alias name — and SPG lexed it into a dedicated
2041 /// `Token::By`, which made it unusable as a name ANYWHERE. Of the seven
2042 /// two-letter keywords the lexer knew, this was the only one PG leaves
2043 /// unreserved (`as`, `in`, `on`, `or`, `to` are reserved and `is` is `T`).
2044 ///
2045 /// The token is gone; the three clauses that own the word — GROUP BY,
2046 /// ORDER BY, PARTITION BY — and the handful of other places that expect it
2047 /// ask this instead. Adding it to the unreserved-identifier table was not
2048 /// enough on its own: identifier positions that match the token shape
2049 /// directly (an index's column list, a table alias) never consult that
2050 /// table, so `CREATE INDEX … ON t(by)` and `FROM t AS by` still failed.
2051 /// Not lexing it as a keyword closes the whole class rather than the two
2052 /// positions that happened to be noticed.
2053 fn peek_is_by(&self) -> bool {
2054 matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("by"))
2055 }
2056
2057 /// v7.39 (round 621) — the optional `CASCADE` / `RESTRICT` trailer a DROP
2058 /// takes. Accepted and dropped: SPG tracks no dependents to cascade to,
2059 /// which is what `DROP TABLE` and `DROP INDEX` have done since v7.14.
2060 fn consume_drop_behaviour(&mut self) {
2061 if matches!(
2062 self.peek(),
2063 Token::Ident(s) if s.eq_ignore_ascii_case("cascade") || s.eq_ignore_ascii_case("restrict")
2064 ) {
2065 self.advance();
2066 }
2067 }
2068
2069 fn expect_ident_like(&mut self) -> Result<String, ParseError> {
2070 let first = match self.advance() {
2071 Token::Ident(s) | Token::QuotedIdent(s) => s,
2072 // v7.37.43-T4 — PG-unreserved keywords are legal identifiers
2073 // per PG's `pg_get_keywords()` classification. SPG tokenizes
2074 // these as named variants for parsing leverage in the
2075 // contexts that own them (`RELEASE SAVEPOINT`, `SHOW name`,
2076 // `BEGIN`, etc.), but they MUST still be usable as table /
2077 // column / alias names in DDL+DML. Sentori migrations like
2078 // 0001_init.sql ship `release TEXT NOT NULL` in the events
2079 // table — the `events.release` column carries the release
2080 // identifier string. Pre-T4 this triggered "expected
2081 // identifier, got Release" and blocked every drop-in user
2082 // whose schema had a column / alias with one of these names.
2083 other if unreserved_keyword_text(&other).is_some() => {
2084 unreserved_keyword_text(&other).unwrap()
2085 }
2086 other => {
2087 return Err(ParseError {
2088 message: format!("expected identifier, got {other:?}"),
2089 token_pos: self.consumed_pos(),
2090 });
2091 }
2092 };
2093 // v7.14.0 — strip optional `<schema>.` prefix. PG dumps
2094 // qualify every name with `public.` (and pg_catalog.* for
2095 // functions); SPG is single-schema so we discard the
2096 // prefix and return only the trailing ident. Same shape
2097 // also handles MySQL `db.tbl` cross-database refs (SPG
2098 // ignores the db part).
2099 if matches!(self.peek(), Token::Dot) {
2100 self.advance();
2101 match self.advance() {
2102 Token::Ident(s) | Token::QuotedIdent(s) => return Ok(s),
2103 other if unreserved_keyword_text(&other).is_some() => {
2104 return Ok(unreserved_keyword_text(&other).unwrap());
2105 }
2106 other => {
2107 return Err(ParseError {
2108 message: format!("expected identifier after '{first}.', got {other:?}"),
2109 token_pos: self.consumed_pos(),
2110 });
2111 }
2112 }
2113 }
2114 Ok(first)
2115 }
2116
2117 #[allow(clippy::too_many_lines)]
2118 fn parse_one_statement(&mut self) -> Result<Statement, ParseError> {
2119 // v7.14.0 — empty / comment-only / semicolon-only input
2120 // (after the lexer strips line + block + MySQL
2121 // conditional comments) lands as Statement::Empty.
2122 // pg_dump and mysqldump emit several wrappers that
2123 // collapse to nothing after stripping (`/*!40101 SET …
2124 // */;`, blank lines between statements); the engine
2125 // returns CommandOk no-op so the dump loads cleanly.
2126 if matches!(self.peek(), Token::Eof | Token::Semicolon) {
2127 return Ok(Statement::Empty);
2128 }
2129 // v7.14.0 — pg_dump / mysqldump "noise" statements:
2130 // catalog / metadata DDL that has no behavioural effect
2131 // on SPG's single-schema, single-database, single-user
2132 // model. Consume the whole statement up to the next
2133 // semicolon / EOF and return Empty. This is broader than
2134 // the per-keyword DROP / SET / COMMENT arms but lets the
2135 // long tail of `LOCK TABLES`, `UNLOCK TABLES`, `GRANT`,
2136 // `REVOKE`, `ALTER OWNER TO`, `\restrict`, `\unrestrict`,
2137 // `BEGIN; COMMIT;` wrappers, etc. all pass through.
2138 if let Token::Ident(s) | Token::QuotedIdent(s) = self.peek() {
2139 let lc = s.to_ascii_lowercase();
2140 // v7.39 (read01 round 50) — COMMENT ON is a real statement now.
2141 if lc == "comment" {
2142 return self.parse_comment_on();
2143 }
2144 // v7.39 (read01 round 57) — so is GRANT / REVOKE.
2145 if lc == "grant" || lc == "revoke" {
2146 return self.parse_grant_or_revoke(lc == "grant");
2147 }
2148 // v7.39 (round 277) — the SQL-level prepared-statement
2149 // surface is REAL now. It used to be accepted and dropped
2150 // on the theory that "real execution still happens via the
2151 // extended-query flow" — true only for a driver that uses
2152 // that flow; a plain SQL PREPARE / EXECUTE returned no
2153 // rows at all.
2154 if lc == "prepare" {
2155 return self.parse_prepare();
2156 }
2157 if lc == "execute" {
2158 return self.parse_execute();
2159 }
2160 if lc == "deallocate" {
2161 return self.parse_deallocate();
2162 }
2163 // v7.39 (round 278) — `CALL <proc>(<args>)` used to be
2164 // accepted and dropped, so an application's stored-procedure
2165 // invocation reported success and did nothing. SPG has no
2166 // procedure catalog, so every CALL names a procedure that
2167 // does not exist — which is exactly what PG says.
2168 if lc == "call" {
2169 return self.parse_call();
2170 }
2171 // v7.39 (round 318, V51) — MySQL `KILL`. Not dump noise: it
2172 // names one connection and acts on it.
2173 if lc == "kill" {
2174 return self.parse_kill();
2175 }
2176 if lc == "discard" {
2177 return self.parse_discard();
2178 }
2179 // v7.39 (round 696) — REASSIGN OWNED BY <role> [, …] TO <role>.
2180 // Still performs nothing; the roles are carried out so a name
2181 // that does not exist is refused, as PG18 refuses it.
2182 if lc == "reassign" {
2183 self.advance();
2184 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("owned")) {
2185 self.advance();
2186 }
2187 if self.peek_is_by() {
2188 self.advance();
2189 }
2190 // Only the roles BEFORE the TO are the ones that must
2191 // exist — `TO` names the new owner, which PG checks as
2192 // well, so both lists are collected.
2193 let mut names = self.take_comma_separated_names();
2194 if matches!(self.peek(), Token::To) {
2195 self.advance();
2196 names.extend(self.take_comma_separated_names());
2197 }
2198 self.consume_until_statement_boundary();
2199 return Ok(Statement::ValidateOnly {
2200 kind: crate::ast::ValidateOnlyKind::RoleName,
2201 names,
2202 });
2203 }
2204 // v7.39 (round 696) — SECURITY LABEL. PG18 refuses it
2205 // unconditionally with `no security label providers have been
2206 // loaded`, whatever object it names, because none is loaded.
2207 // SPG has none either; accepting it told the caller a label had
2208 // been applied when nothing anywhere records one.
2209 if lc == "security" {
2210 self.consume_until_statement_boundary();
2211 return Ok(Statement::ValidateOnly {
2212 kind: crate::ast::ValidateOnlyKind::SecurityLabel,
2213 names: Vec::new(),
2214 });
2215 }
2216 // v7.39.2 — `USE <db>` is a real statement now, and only in
2217 // the MySQL dialect. It used to be swallowed here with the
2218 // dump noise, so `USE myapp; SELECT DATABASE()` answered the
2219 // same constant it answered before — MySQL 9.7.2 answers
2220 // `myapp`. PostgreSQL has no USE at all, and pg_dump does not
2221 // emit one, but the swallow stays on that side: it was put
2222 // there for restores and taking it away is not this defect.
2223 if lc == "use" {
2224 if self.mysql_dialect {
2225 self.advance();
2226 let name = self.expect_ident_like()?;
2227 return Ok(Statement::UseDatabase(name));
2228 }
2229 self.consume_until_statement_boundary();
2230 return Ok(Statement::Empty);
2231 }
2232 if is_dump_noise_statement(&lc) {
2233 self.consume_until_statement_boundary();
2234 return Ok(Statement::Empty);
2235 }
2236 }
2237 match self.peek() {
2238 Token::Select => self.parse_select_stmt(),
2239 // v7.37.17 (17.6 siblings) — a statement opening with a
2240 // parenthesized query group: `(SELECT … UNION …)
2241 // INTERSECT …`. parse_bare_select's group arm consumes
2242 // the parens; the select parser handles the outer chain
2243 // and tail.
2244 Token::LParen
2245 if matches!(
2246 self.tokens.get(self.pos + 1),
2247 Some(Token::Select | Token::LParen | Token::Values)
2248 ) =>
2249 {
2250 self.parse_select_stmt()
2251 }
2252 // v7.37.17 (17.6 siblings) — top-level bare VALUES
2253 // statement (`VALUES (1), (2) [ORDER BY …] [LIMIT …]`).
2254 // Lowers to the same UNION ALL chain the FROM-position
2255 // form uses, then reuses the shared SELECT tail.
2256 Token::Values => {
2257 self.advance(); // VALUES
2258 let mut head = self.parse_values_rows_body()?;
2259 self.parse_select_tail_into(&mut head)?;
2260 Ok(Statement::Select(head))
2261 }
2262 // SQL-standard `TABLE name` shorthand for
2263 // `SELECT * FROM name` — pg_dump never emits it, but
2264 // psql users and PG docs use it constantly. Set-op
2265 // chains and the ORDER BY/LIMIT tail compose like any
2266 // SELECT head.
2267 Token::Table
2268 if matches!(
2269 self.tokens.get(self.pos + 1),
2270 Some(Token::Ident(_) | Token::QuotedIdent(_))
2271 ) =>
2272 {
2273 let mut head = self.parse_table_shorthand()?;
2274 self.parse_setop_chain_into(&mut head)?;
2275 self.parse_select_tail_into(&mut head)?;
2276 Ok(Statement::Select(head))
2277 }
2278 // v7.9.27 — `DO $$ … $$ [LANGUAGE plpgsql]`. The
2279 // body is a dollar-quoted plpgsql block (lexer already
2280 // collapsed `$$…$$` into a single Token::String).
2281 // v7.16.2 — mailrs round-10 A.2: parse the body as a
2282 // real PlPgSqlBlock so the engine can EXECUTE it at
2283 // top level instead of silently swallowing. Pre-
2284 // v7.16.2 the parser threw the body away and the
2285 // engine returned CommandOk for the entire DO; that
2286 // turned `DO BEGIN … IF EXISTS ... THEN ALTER …; END
2287 // $$` into a SEV-1 silent no-op (the IF + the rename
2288 // were both invisible — mailrs's migrate-042 didn't
2289 // actually run). Now the body parses + executes;
2290 // EmbeddedSql inside the block runs immediately
2291 // against the engine (not deferred — we're at top
2292 // level, not inside a trigger row-write loop).
2293 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("do") => {
2294 self.advance();
2295 let body_text = match self.advance() {
2296 Token::String(s) => s,
2297 other => {
2298 return Err(self.err(alloc::format!(
2299 "expected dollar-quoted body after DO, got {other:?}"
2300 )));
2301 }
2302 };
2303 // Optional `LANGUAGE <name>` trailer (idents only).
2304 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("language")) {
2305 self.advance();
2306 let _ = self.expect_ident_like()?;
2307 }
2308 // Parse the body — same shape CREATE FUNCTION
2309 // uses for trigger function bodies. If the body
2310 // doesn't parse cleanly we surface the error
2311 // (better than silent no-op).
2312 let block = parse_plpgsql_body(&body_text)?;
2313 Ok(Statement::DoBlock(block))
2314 }
2315 // v4.11: `WITH name AS (SELECT ...) [, ...] SELECT ...`.
2316 // WITH isn't a reserved token in our lexer — comes through
2317 // as `Token::Ident("with")` (case-insensitive).
2318 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("with") => {
2319 self.advance();
2320 self.parse_with_cte_then_select()
2321 }
2322 // v4.26: `EXPLAIN [ANALYZE] <select>`. Comes through as
2323 // an identifier — not a reserved keyword.
2324 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("explain") => {
2325 self.advance();
2326 let mut analyze = false;
2327 let mut suggest = false;
2328 let mut costs_off = false;
2329 let mut buffers = false;
2330 let mut timing_off = false;
2331 let mut settings = false;
2332 let mut wal = false;
2333 let mut summary_off = false;
2334 let mut format = crate::ast::ExplainFormat::Text;
2335 // v6.8.3 + v7.37.7 — `EXPLAIN (option [, option…])`
2336 // syntax accepts SUGGEST + COSTS ON|OFF. Multiple
2337 // options are comma-separated. Booleans default to ON
2338 // when the value token is omitted (matches PG).
2339 if matches!(self.peek(), Token::LParen) {
2340 self.advance();
2341 loop {
2342 let opt = match self.peek().clone() {
2343 Token::Ident(s) | Token::QuotedIdent(s) => s,
2344 other => {
2345 return Err(self.err(format!(
2346 "expected option keyword inside EXPLAIN (…), got {other:?}"
2347 )));
2348 }
2349 };
2350 self.advance();
2351 if opt.eq_ignore_ascii_case("suggest") {
2352 suggest = true;
2353 // SUGGEST takes no explicit value today.
2354 } else if opt.eq_ignore_ascii_case("costs") {
2355 // PG syntax: `COSTS [ON | OFF]`. Default
2356 // when value omitted is ON, so plain
2357 // `COSTS` is a no-op. `COSTS OFF` flips.
2358 // `ON` lexes to `Token::On` (reserved
2359 // keyword in JOIN ... ON contexts); accept
2360 // it alongside the bare Ident form so the
2361 // grammar matches PG verbatim.
2362 let value = match self.peek().clone() {
2363 Token::On => {
2364 self.advance();
2365 true
2366 }
2367 Token::Ident(v) | Token::QuotedIdent(v)
2368 if v.eq_ignore_ascii_case("off") =>
2369 {
2370 self.advance();
2371 false
2372 }
2373 Token::Ident(v) | Token::QuotedIdent(v)
2374 if v.eq_ignore_ascii_case("true") =>
2375 {
2376 self.advance();
2377 true
2378 }
2379 _ => true,
2380 };
2381 costs_off = !value;
2382 } else if opt.eq_ignore_ascii_case("analyze")
2383 || opt.eq_ignore_ascii_case("analyse")
2384 {
2385 // v7.37.22 — `EXPLAIN (ANALYZE [ON|OFF]) <S>`.
2386 // Same default-ON rule as ANALYZE keyword form.
2387 let value = match self.peek().clone() {
2388 Token::On => {
2389 self.advance();
2390 true
2391 }
2392 Token::Ident(v) | Token::QuotedIdent(v)
2393 if v.eq_ignore_ascii_case("off") =>
2394 {
2395 self.advance();
2396 false
2397 }
2398 Token::Ident(v) | Token::QuotedIdent(v)
2399 if v.eq_ignore_ascii_case("true") =>
2400 {
2401 self.advance();
2402 true
2403 }
2404 _ => true,
2405 };
2406 analyze = value;
2407 } else if opt.eq_ignore_ascii_case("buffers") {
2408 // v7.37.22 — `BUFFERS [ON|OFF]`.
2409 let value = match self.peek().clone() {
2410 Token::On => {
2411 self.advance();
2412 true
2413 }
2414 Token::Ident(v) | Token::QuotedIdent(v)
2415 if v.eq_ignore_ascii_case("off") =>
2416 {
2417 self.advance();
2418 false
2419 }
2420 Token::Ident(v) | Token::QuotedIdent(v)
2421 if v.eq_ignore_ascii_case("true") =>
2422 {
2423 self.advance();
2424 true
2425 }
2426 _ => true,
2427 };
2428 buffers = value;
2429 } else if opt.eq_ignore_ascii_case("timing") {
2430 // v7.37.22 — `TIMING [ON|OFF]`. OFF strips
2431 // the measured wall-clock annotation.
2432 let value = match self.peek().clone() {
2433 Token::On => {
2434 self.advance();
2435 true
2436 }
2437 Token::Ident(v) | Token::QuotedIdent(v)
2438 if v.eq_ignore_ascii_case("off") =>
2439 {
2440 self.advance();
2441 false
2442 }
2443 Token::Ident(v) | Token::QuotedIdent(v)
2444 if v.eq_ignore_ascii_case("true") =>
2445 {
2446 self.advance();
2447 true
2448 }
2449 _ => true,
2450 };
2451 timing_off = !value;
2452 } else if opt.eq_ignore_ascii_case("settings") {
2453 settings = true;
2454 } else if opt.eq_ignore_ascii_case("wal") {
2455 wal = true;
2456 } else if opt.eq_ignore_ascii_case("summary") {
2457 // v7.39 (round 227) — `SUMMARY [ON|OFF]` really
2458 // gates the trailing Planning/Execution Time
2459 // lines now (was accept-and-no-op).
2460 let value = match self.peek().clone() {
2461 Token::On => {
2462 self.advance();
2463 true
2464 }
2465 Token::Ident(v) | Token::QuotedIdent(v)
2466 if v.eq_ignore_ascii_case("off") =>
2467 {
2468 self.advance();
2469 false
2470 }
2471 Token::Ident(v) | Token::QuotedIdent(v)
2472 if v.eq_ignore_ascii_case("true") =>
2473 {
2474 self.advance();
2475 true
2476 }
2477 _ => true,
2478 };
2479 summary_off = !value;
2480 } else if opt.eq_ignore_ascii_case("verbose")
2481 || opt.eq_ignore_ascii_case("format")
2482 {
2483 // v7.37.22 — accept-but-no-op the remaining
2484 // PG options so EXPLAIN-using clients
2485 // (pgAdmin / DataGrip) don't see syntax
2486 // errors. FORMAT takes a value (text /
2487 // json / yaml / xml); skip the next token
2488 // if it's an ident.
2489 if opt.eq_ignore_ascii_case("format") {
2490 if let Token::Ident(v) | Token::QuotedIdent(v) = self.peek().clone()
2491 {
2492 self.advance();
2493 format = match v.to_ascii_lowercase().as_str() {
2494 "text" => crate::ast::ExplainFormat::Text,
2495 "json" => crate::ast::ExplainFormat::Json,
2496 "xml" => crate::ast::ExplainFormat::Xml,
2497 "yaml" => crate::ast::ExplainFormat::Yaml,
2498 other => {
2499 return Err(self.err(format!(
2500 "EXPLAIN (FORMAT …): unknown format {other:?}; \
2501 supports text, json, xml, yaml"
2502 )));
2503 }
2504 };
2505 }
2506 } else {
2507 // VERBOSE / SUMMARY take optional ON/OFF;
2508 // consume if present.
2509 if matches!(self.peek(), Token::On) {
2510 self.advance();
2511 } else if let Token::Ident(v) | Token::QuotedIdent(v) =
2512 self.peek().clone()
2513 && (v.eq_ignore_ascii_case("off")
2514 || v.eq_ignore_ascii_case("true"))
2515 {
2516 self.advance();
2517 let _ = v;
2518 }
2519 }
2520 } else {
2521 return Err(self.err(format!(
2522 "unknown EXPLAIN option {opt:?}; supports ANALYZE, COSTS, BUFFERS, TIMING, SETTINGS, WAL, SUGGEST, VERBOSE, FORMAT, SUMMARY"
2523 )));
2524 }
2525 if matches!(self.peek(), Token::Comma) {
2526 self.advance();
2527 continue;
2528 }
2529 break;
2530 }
2531 if !matches!(self.peek(), Token::RParen) {
2532 return Err(self.err(format!(
2533 "expected ')' after EXPLAIN options, got {:?}",
2534 self.peek()
2535 )));
2536 }
2537 self.advance();
2538 } else if let Token::Ident(s) | Token::QuotedIdent(s) = self.peek()
2539 && (s.eq_ignore_ascii_case("analyze") || s.eq_ignore_ascii_case("analyse"))
2540 {
2541 self.advance();
2542 analyze = true;
2543 }
2544 // v7.39 (round 224) — the body may open with WITH (CTEs);
2545 // route through the same CTE-then-SELECT path the top-level
2546 // WITH statement uses. v7.39 (round 225) — DML bodies parse
2547 // too (PG explains INSERT / UPDATE / DELETE).
2548 let inner = match self.peek().clone() {
2549 Token::Ident(s) if s.eq_ignore_ascii_case("with") => {
2550 self.advance();
2551 self.parse_with_cte_then_select()?
2552 }
2553 Token::Insert => self.parse_insert_stmt(false)?,
2554 Token::Ident(s) if s.eq_ignore_ascii_case("update") => {
2555 self.advance();
2556 self.parse_update_after_keyword()?
2557 }
2558 Token::Ident(s) if s.eq_ignore_ascii_case("delete") => {
2559 self.advance();
2560 self.parse_delete_after_keyword()?
2561 }
2562 _ => self.parse_select_stmt()?,
2563 };
2564 if !matches!(
2565 inner,
2566 Statement::Select(_)
2567 | Statement::Insert(_)
2568 | Statement::Update(_)
2569 | Statement::Delete(_)
2570 ) {
2571 return Err(self.err(format!(
2572 "EXPLAIN body must be SELECT / INSERT / UPDATE / DELETE, got {inner:?}"
2573 )));
2574 }
2575 Ok(Statement::Explain(crate::ast::ExplainStatement {
2576 analyze,
2577 inner: Box::new(inner),
2578 suggest,
2579 costs_off,
2580 buffers,
2581 timing_off,
2582 settings,
2583 wal,
2584 summary_off,
2585 format,
2586 }))
2587 }
2588 Token::Create => self.parse_create_stmt(),
2589 Token::Insert => self.parse_insert_stmt(false),
2590 // MySQL `DESCRIBE t` / `DESC t` — the SHOW COLUMNS
2591 // spelling; route to the same handler. DESC is the
2592 // reserved ORDER BY token, so it gets its own arm.
2593 Token::Ident(s)
2594 if s.eq_ignore_ascii_case("describe")
2595 && matches!(
2596 self.tokens.get(self.pos + 1),
2597 Some(Token::Ident(_) | Token::QuotedIdent(_))
2598 ) =>
2599 {
2600 self.advance();
2601 let table = self.expect_ident_like()?;
2602 Ok(Statement::ShowColumns(table))
2603 }
2604 Token::Desc
2605 if matches!(
2606 self.tokens.get(self.pos + 1),
2607 Some(Token::Ident(_) | Token::QuotedIdent(_))
2608 ) =>
2609 {
2610 self.advance();
2611 let table = self.expect_ident_like()?;
2612 Ok(Statement::ShowColumns(table))
2613 }
2614 // `COPY table [(cols)] TO STDOUT` — the export half of
2615 // pg_dump's COPY pair (the FROM stdin half rides the
2616 // embed import path). Options need a format design and
2617 // error honestly.
2618 Token::Ident(s)
2619 if s.eq_ignore_ascii_case("copy")
2620 && matches!(
2621 self.tokens.get(self.pos + 1),
2622 Some(Token::Ident(_) | Token::QuotedIdent(_))
2623 ) =>
2624 {
2625 self.advance(); // COPY
2626 let table = self.expect_ident_like()?;
2627 let columns = if matches!(self.peek(), Token::LParen) {
2628 self.advance();
2629 let mut cols = alloc::vec![self.expect_ident_like()?];
2630 while matches!(self.peek(), Token::Comma) {
2631 self.advance();
2632 cols.push(self.expect_ident_like()?);
2633 }
2634 if !matches!(self.peek(), Token::RParen) {
2635 return Err(self.err(format!(
2636 "expected ')' after COPY column list, got {:?}",
2637 self.peek()
2638 )));
2639 }
2640 self.advance();
2641 Some(cols)
2642 } else {
2643 None
2644 };
2645 // v7.39 (round 249) — `COPY t FROM '<path>'`: the file
2646 // endpoint. (FROM STDIN still rides the wire/import path —
2647 // its data arrives out of band.)
2648 if matches!(self.peek(), Token::From)
2649 && matches!(self.tokens.get(self.pos + 1), Some(Token::String(_)))
2650 {
2651 self.advance(); // FROM
2652 let Token::String(path) = self.advance() else {
2653 unreachable!()
2654 };
2655 let options = self.parse_copy_to_options()?;
2656 return Ok(Statement::CopyFromFile {
2657 table,
2658 columns,
2659 path,
2660 options,
2661 });
2662 }
2663 if !matches!(self.peek(), Token::To) {
2664 return Err(self.err(format!(
2665 "COPY: only TO STDOUT is supported here (FROM stdin \
2666 rides the import path); got {:?}",
2667 self.peek()
2668 )));
2669 }
2670 self.advance();
2671 if matches!(self.peek(), Token::String(_)) {
2672 let Token::String(path) = self.advance() else { unreachable!() };
2673 let options = self.parse_copy_to_options()?;
2674 return Ok(Statement::CopyToFile {
2675 table,
2676 columns,
2677 query: None,
2678 path,
2679 options,
2680 });
2681 }
2682 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("stdout")) {
2683 return Err(self.err(format!(
2684 "COPY TO supports STDOUT only (no file endpoints), got {:?}",
2685 self.peek()
2686 )));
2687 }
2688 self.advance();
2689 let options = self.parse_copy_to_options()?;
2690 Ok(Statement::CopyTo {
2691 table,
2692 columns,
2693 query: None,
2694 options,
2695 })
2696 }
2697 // v7.39 (read01 round 94) — `COPY (<query>) TO STDOUT [WITH (…)]`.
2698 // The parens directly after COPY wrap a SELECT/VALUES/CTE whose
2699 // result set is streamed in COPY format (PG's query form).
2700 Token::Ident(s)
2701 if s.eq_ignore_ascii_case("copy")
2702 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen)) =>
2703 {
2704 self.advance(); // COPY
2705 self.advance(); // (
2706 let query = self.parse_select_stmt()?;
2707 if !matches!(self.peek(), Token::RParen) {
2708 return Err(self.err(format!(
2709 "expected ')' after COPY query, got {:?}",
2710 self.peek()
2711 )));
2712 }
2713 self.advance(); // )
2714 if !matches!(self.peek(), Token::To) {
2715 return Err(self.err(format!(
2716 "COPY (query): only TO STDOUT is supported, got {:?}",
2717 self.peek()
2718 )));
2719 }
2720 self.advance();
2721 if matches!(self.peek(), Token::String(_)) {
2722 let Token::String(path) = self.advance() else { unreachable!() };
2723 let options = self.parse_copy_to_options()?;
2724 return Ok(Statement::CopyToFile {
2725 table: String::new(),
2726 columns: None,
2727 query: Some(alloc::boxed::Box::new(query)),
2728 path,
2729 options,
2730 });
2731 }
2732 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("stdout")) {
2733 return Err(self.err(format!(
2734 "COPY (query): TO supports STDOUT only, got {:?}",
2735 self.peek()
2736 )));
2737 }
2738 self.advance();
2739 let options = self.parse_copy_to_options()?;
2740 Ok(Statement::CopyTo {
2741 table: String::new(),
2742 columns: None,
2743 query: Some(alloc::boxed::Box::new(query)),
2744 options,
2745 })
2746 }
2747 // MySQL `REPLACE INTO t …` — delete-then-insert upsert.
2748 // Shares the INSERT body; the replace flag lowers it
2749 // onto ON CONFLICT DO UPDATE with an empty assignment
2750 // list (engine: replace the whole row).
2751 Token::Ident(s)
2752 if s.eq_ignore_ascii_case("replace")
2753 && matches!(self.tokens.get(self.pos + 1), Some(Token::Into)) =>
2754 {
2755 self.parse_insert_stmt(true)
2756 }
2757 Token::Begin => {
2758 self.advance();
2759 // v7.38 轴 4 / v7.39 (read01 round 118, B3) — PG-standard
2760 // `BEGIN [WORK|TRANSACTION] [ISOLATION LEVEL …] [READ ONLY|WRITE]
2761 // [[NOT] DEFERRABLE]`. The optional WORK/TRANSACTION noise word
2762 // is consumed first, then the trailing modes — including the
2763 // case where `ISOLATION LEVEL …` follows BEGIN directly (no
2764 // WORK/TRANSACTION). The explicit level, when present, rides the
2765 // statement so `exec_begin` applies it for this transaction.
2766 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("work") || s.eq_ignore_ascii_case("transaction"))
2767 {
2768 self.advance();
2769 }
2770 let iso = self.parse_isolation_level_clauses()?;
2771 Ok(Statement::Begin(iso))
2772 }
2773 // v7.38 轴 4 — PG-standard `START TRANSACTION …` synonym
2774 // for BEGIN. START is contextual in PG too; pattern-match
2775 // on the ident here. Iso clauses are parse-and-ignored,
2776 // same as BEGIN above.
2777 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("start") => {
2778 self.advance();
2779 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("transaction"))
2780 {
2781 return Err(self.err(alloc::format!(
2782 "expected TRANSACTION after START, got {:?}",
2783 self.peek()
2784 )));
2785 }
2786 self.advance();
2787 let iso = self.parse_isolation_level_clauses()?;
2788 Ok(Statement::Begin(iso))
2789 }
2790 Token::Commit => {
2791 self.advance();
2792 // PG: `COMMIT [WORK | TRANSACTION]`.
2793 if let Token::Ident(w) = self.peek()
2794 && (w.eq_ignore_ascii_case("work") || w.eq_ignore_ascii_case("transaction"))
2795 {
2796 self.advance();
2797 }
2798 Ok(Statement::Commit)
2799 }
2800 // r1066 (7.38 S5.1) — `END [WORK | TRANSACTION]` is PG's
2801 // COMMIT synonym; pgbench's builtin tpcb-like script closes
2802 // every transaction with `END;` and the drop-in aborted on
2803 // it. Only reachable at statement start (CASE … END lives
2804 // inside expressions), so no ambiguity.
2805 Token::Ident(s) if s.eq_ignore_ascii_case("end") => {
2806 self.advance();
2807 if let Token::Ident(w) = self.peek()
2808 && (w.eq_ignore_ascii_case("work") || w.eq_ignore_ascii_case("transaction"))
2809 {
2810 self.advance();
2811 }
2812 Ok(Statement::Commit)
2813 }
2814 Token::Rollback => {
2815 self.advance();
2816 // `ROLLBACK TO [SAVEPOINT] <name>` returns to that
2817 // savepoint without ending the transaction. Bare
2818 // `ROLLBACK` drops the whole TX.
2819 if matches!(self.peek(), Token::To) {
2820 self.advance();
2821 if matches!(self.peek(), Token::Savepoint) {
2822 self.advance();
2823 }
2824 let name = self.expect_ident_like()?;
2825 Ok(Statement::RollbackToSavepoint(name))
2826 } else {
2827 Ok(Statement::Rollback)
2828 }
2829 }
2830 Token::Savepoint => {
2831 self.advance();
2832 let name = self.expect_ident_like()?;
2833 Ok(Statement::Savepoint(name))
2834 }
2835 Token::Release => {
2836 self.advance();
2837 // `RELEASE [SAVEPOINT] <name>` — the `SAVEPOINT` keyword
2838 // is optional in standard SQL.
2839 if matches!(self.peek(), Token::Savepoint) {
2840 self.advance();
2841 }
2842 let name = self.expect_ident_like()?;
2843 Ok(Statement::ReleaseSavepoint(name))
2844 }
2845 Token::Show => {
2846 self.advance();
2847 // `SHOW TABLES` / `SHOW USERS` / `SHOW COLUMNS FROM <table>`.
2848 // v6.1.2 promoted TABLES to a reserved keyword (for
2849 // `CREATE PUBLICATION … FOR ALL TABLES`), so it now
2850 // arrives as `Token::Tables` rather than a bare ident.
2851 // USERS / COLUMNS remain bare idents.
2852 let target = match self.advance() {
2853 Token::Tables => "tables".to_string(),
2854 // v7.17.0 Phase 3.P0-59 — CREATE is a reserved
2855 // keyword token; recognise it as the SHOW CREATE
2856 // dispatch keyword too.
2857 Token::Create => "create".to_string(),
2858 // v7.17.0 Phase 3.P0-60 — INDEX is a reserved
2859 // keyword too; let SHOW INDEX FROM parse.
2860 Token::Index => "index".to_string(),
2861 // v7.37.17 (17.6 sibling) — SHOW ALL. ALL is
2862 // reserved (used in aggregate function calls);
2863 // recognise it here so the parser dispatches
2864 // to ShowParameter("all") — the engine returns
2865 // the curated parameter inventory.
2866 Token::All => "all".to_string(),
2867 // v7.38.18 (C12) — `SHOW COUNT(*) WARNINGS`, MySQL's
2868 // spelling for the size of the diagnostics area.
2869 // MySQL-dialect only: PostgreSQL 18.4 answers this
2870 // phrase with `syntax error at or near "("`, and a
2871 // PG session must keep getting exactly that rather
2872 // than a message about an unknown parameter.
2873 // `COUNT` arrives as a bare ident; the `(*)` and the
2874 // trailing keyword are consumed here so the whole
2875 // form reaches the engine as one parameter name.
2876 Token::Ident(ref c)
2877 if self.mysql_dialect
2878 && c.eq_ignore_ascii_case("count")
2879 && matches!(self.peek(), Token::LParen) =>
2880 {
2881 self.advance();
2882 if matches!(self.peek(), Token::Star) {
2883 self.advance();
2884 }
2885 if matches!(self.peek(), Token::RParen) {
2886 self.advance();
2887 }
2888 match self.advance() {
2889 Token::Ident(w) if w.eq_ignore_ascii_case("warnings") => {
2890 return Ok(Statement::ShowParameter(
2891 "count(*) warnings".to_string(),
2892 ));
2893 }
2894 other => {
2895 return Err(self.err(format!(
2896 "expected WARNINGS after SHOW COUNT(*), got {other:?}"
2897 )));
2898 }
2899 }
2900 }
2901 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
2902 other => {
2903 return Err(self.err(format!(
2904 "expected SHOW target, got {other:?}"
2905 )));
2906 }
2907 };
2908 match target.as_str() {
2909 "tables" => Ok(Statement::ShowTables),
2910 "users" => Ok(Statement::ShowUsers),
2911 // v7.38 轴 4 — `SHOW transaction_isolation`
2912 // returns the currently-selected isolation level.
2913 "transaction_isolation" => Ok(Statement::ShowParameter(
2914 "transaction_isolation".to_string(),
2915 )),
2916 // v7.17.0 Phase 3.P0-59 — MySQL `SHOW CREATE
2917 // TABLE <t>` returns a 2-column row: (Table,
2918 // Create Table). mysqldump emits this for every
2919 // table at scrape time; without it the dump
2920 // round-trip stalls.
2921 // v7.17.0 Phase 3.P0-60 — MySQL `SHOW INDEXES
2922 // FROM <t>` (also spelled `SHOW INDEX` and
2923 // `SHOW KEYS`). admin / mysqldump probes use
2924 // it to list per-table indexes.
2925 "indexes" | "index" | "keys" => {
2926 if !matches!(self.peek(), Token::From) {
2927 return Err(self.err(format!(
2928 "expected FROM after SHOW INDEXES, got {:?}",
2929 self.peek()
2930 )));
2931 }
2932 self.advance();
2933 let table = self.expect_ident_like()?;
2934 Ok(Statement::ShowIndexes(table))
2935 }
2936 // v7.17.0 Phase 3.P0-61 — MySQL `SHOW STATUS` /
2937 // `SHOW VARIABLES`. Both return a 2-column row
2938 // set listing server-side state; clients probe
2939 // them at connect time.
2940 "status" => Ok(Statement::ShowStatus),
2941 "variables" => {
2942 // r1067 — `SHOW VARIABLES LIKE 'pat'`.
2943 if matches!(self.peek(), Token::Like) {
2944 self.advance();
2945 let pat = match self.advance() {
2946 Token::String(p) => p,
2947 other => {
2948 return Err(self.err(format!(
2949 "SHOW VARIABLES LIKE expects a quoted pattern, got {other:?}"
2950 )));
2951 }
2952 };
2953 return Ok(Statement::ShowVariablesLike(pat));
2954 }
2955 Ok(Statement::ShowVariables)
2956 }
2957 // v7.17.0 Phase 3.P0-62 — MySQL `SHOW PROCESSLIST`.
2958 "processlist" => Ok(Statement::ShowProcesslist),
2959 "create" => {
2960 // SHOW CREATE TABLE / VIEW / DATABASE — only
2961 // TABLE is supported in v7.17.
2962 let kind = match self.advance() {
2963 Token::Ident(s) | Token::QuotedIdent(s) => s,
2964 Token::Table => "table".to_string(),
2965 other => {
2966 return Err(self.err(format!(
2967 "expected TABLE after SHOW CREATE, got {other:?}"
2968 )));
2969 }
2970 };
2971 if !kind.eq_ignore_ascii_case("table") {
2972 return Err(self.err(format!(
2973 "unsupported SHOW CREATE {kind:?}; v7.17 supports TABLE only"
2974 )));
2975 }
2976 let name = self.expect_ident_like()?;
2977 Ok(Statement::ShowCreateTable(name))
2978 }
2979 // v7.17.0 Phase 3.P0-58 — MySQL `SHOW DATABASES`
2980 // (and `SHOW SCHEMAS` alias). The mysql client uses
2981 // it to populate the database selector at connect
2982 // time; without it `mysql -p` errors before the
2983 // first user query.
2984 "databases" | "schemas" => Ok(Statement::ShowDatabases),
2985 // v6.1.3 — PUBLICATIONS plural is NOT a reserved
2986 // keyword on its own; it lands here as a bare
2987 // ident. Returning all publications + their
2988 // scope summary.
2989 "publications" => Ok(Statement::ShowPublications),
2990 // v6.1.4 — same shape for SUBSCRIPTIONS plural.
2991 "subscriptions" => Ok(Statement::ShowSubscriptions),
2992 "columns" => {
2993 if !matches!(self.peek(), Token::From) {
2994 return Err(self.err(format!(
2995 "expected FROM after SHOW COLUMNS, got {:?}",
2996 self.peek()
2997 )));
2998 }
2999 self.advance();
3000 let table = self.expect_ident_like()?;
3001 Ok(Statement::ShowColumns(table))
3002 }
3003 // v7.38 轴 4 surface — `SHOW <param>` for any
3004 // remaining session / preset parameter name
3005 // (server_version, search_path, client_encoding,
3006 // …). The engine's ShowParameter handler does the
3007 // dispatch; unrecognised names error there with
3008 // a pointer to pg_settings, not at parse time —
3009 // so a driver that issues `SHOW spam_setting`
3010 // gets a clear runtime error instead of a
3011 // confusing "unknown SHOW target".
3012 other => {
3013 // v7.38 (read01 P3.20) — a custom namespaced GUC
3014 // (`SHOW app.foo`) arrives as `app` + `.` + `foo`;
3015 // consume the dotted tail so it round-trips with
3016 // `SET app.foo` / `current_setting('app.foo')`.
3017 let mut full = other.to_string();
3018 while matches!(self.peek(), Token::Dot) {
3019 self.advance();
3020 let seg = self.expect_ident_like()?;
3021 full.push('.');
3022 full.push_str(&seg.to_ascii_lowercase());
3023 }
3024 Ok(Statement::ShowParameter(full))
3025 }
3026 }
3027 }
3028 // v6.1.2: `DROP` is now a reserved keyword (it dispatches
3029 // to DROP USER and DROP PUBLICATION today; DROP TABLE /
3030 // DROP INDEX are still SHOW-shaped admin ops). Pre-6.1.2
3031 // arrived as a bare ident; tokenising it dedicatedly
3032 // keeps the dispatch tree small.
3033 Token::Drop => {
3034 self.advance();
3035 match self.peek() {
3036 // v7.37.17 (17.6 sibling) — DROP OWNED BY <role>
3037 // [, ...] [CASCADE | RESTRICT]. pg_dumpall emits
3038 // around DROP ROLE cleanup. SPG has no role-owner
3039 // model, so consume to boundary as a no-op.
3040 Token::Ident(s) | Token::QuotedIdent(s)
3041 if s.eq_ignore_ascii_case("owned") =>
3042 {
3043 // v7.39 (round 696) — still a no-op (SPG has no
3044 // role-owner model), but the ROLE is carried out so
3045 // the engine can refuse one that does not exist,
3046 // which is what PG18 does.
3047 self.advance();
3048 if self.peek_is_by() {
3049 self.advance();
3050 }
3051 let names = self.take_comma_separated_names();
3052 self.consume_until_statement_boundary();
3053 Ok(Statement::ValidateOnly {
3054 kind: crate::ast::ValidateOnlyKind::RoleName,
3055 names,
3056 })
3057 }
3058 // v7.39 (round 436) — MySQL's `DROP TEMPORARY TABLE t`.
3059 // It drops only a TEMPORARY table, and name resolution
3060 // already prefers the session's own, so the keyword is
3061 // consumed and the ordinary DROP TABLE path runs.
3062 Token::Ident(s) | Token::QuotedIdent(s)
3063 if s.eq_ignore_ascii_case("temporary") || s.eq_ignore_ascii_case("temp") =>
3064 {
3065 self.advance();
3066 if !matches!(self.peek(), Token::Table) {
3067 return Err(self.err(alloc::format!(
3068 "expected TABLE after DROP TEMPORARY, got {:?}",
3069 self.peek()
3070 )));
3071 }
3072 self.parse_drop_table_after_keyword()
3073 }
3074 Token::Publication => {
3075 self.advance();
3076 // v7.39 (round 754, F31-B4) — the round-753
3077 // audit probe tripped over the missing
3078 // `IF EXISTS` here (syntax error).
3079 let if_exists = self.consume_if_exists();
3080 let name = self.expect_ident_or_string()?;
3081 Ok(Statement::DropPublication { name, if_exists })
3082 }
3083 Token::Subscription => {
3084 self.advance();
3085 let if_exists = self.consume_if_exists();
3086 let name = self.expect_ident_or_string()?;
3087 Ok(Statement::DropSubscription { name, if_exists })
3088 }
3089 Token::Ident(s) | Token::QuotedIdent(s)
3090 if s.eq_ignore_ascii_case("user") || s.eq_ignore_ascii_case("role") =>
3091 {
3092 self.advance();
3093 // v7.39 (read01 round 58) — DROP ROLE is DROP USER: a
3094 // login user IS a role in PG, and SPG's store holds
3095 // both. `IF EXISTS` is accepted on either spelling.
3096 let if_exists = self.consume_if_exists();
3097 let name = self.expect_ident_or_string()?;
3098 Ok(Statement::DropUser { name, if_exists })
3099 }
3100 // v7.39 (round 806) — DROP DATABASE [IF EXISTS] <name>.
3101 // CREATE DATABASE has parsed since v7.14 and this did
3102 // not, so `DROP DATABASE IF EXISTS x` — what every
3103 // teardown script and pg_dumpall preamble opens with —
3104 // came back as a syntax error, which IF EXISTS cannot
3105 // soften. The name is carried so the engine can answer
3106 // the way PG does; PG never lets this succeed on a
3107 // single-database server, since the name is either
3108 // unknown ("database … does not exist", or a notice
3109 // under IF EXISTS) or the one you are connected to
3110 // ("cannot drop the currently open database").
3111 Token::Ident(s) | Token::QuotedIdent(s)
3112 if s.eq_ignore_ascii_case("database") =>
3113 {
3114 self.advance();
3115 let if_exists = self.consume_if_exists();
3116 let name = self.expect_ident_or_string()?;
3117 self.consume_until_statement_boundary();
3118 Ok(Statement::DropDatabase { name, if_exists })
3119 }
3120 // v7.12.4 — DROP TRIGGER [IF EXISTS] name ON table.
3121 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("trigger") => {
3122 self.advance();
3123 let if_exists = self.consume_if_exists();
3124 let name = self.expect_ident_like()?;
3125 // ON <table>
3126 if !matches!(self.peek(), Token::On) {
3127 return Err(self.err(alloc::format!(
3128 "expected ON <table> after DROP TRIGGER {name:?}, got {:?}",
3129 self.peek()
3130 )));
3131 }
3132 self.advance();
3133 let table = self.expect_ident_like()?;
3134 Ok(Statement::DropTrigger {
3135 name,
3136 table,
3137 if_exists,
3138 })
3139 }
3140 // v7.39 (round 139) — DROP RULE [IF EXISTS] name ON table
3141 // [CASCADE|RESTRICT]. Mirrors DROP TRIGGER's shape.
3142 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("rule") => {
3143 self.advance();
3144 let if_exists = self.consume_if_exists();
3145 let name = self.expect_ident_like()?;
3146 if !matches!(self.peek(), Token::On) {
3147 return Err(self.err(alloc::format!(
3148 "expected ON <table> after DROP RULE {name:?}, got {:?}",
3149 self.peek()
3150 )));
3151 }
3152 self.advance();
3153 let table = self.expect_ident_like()?;
3154 // Optional CASCADE / RESTRICT — accepted, no effect.
3155 self.consume_until_statement_boundary();
3156 Ok(Statement::DropRule {
3157 name,
3158 table,
3159 if_exists,
3160 })
3161 }
3162 // v7.12.4 — DROP FUNCTION [IF EXISTS] name [(args)].
3163 // v7.12.4 ignores any optional arg-list (signature-
3164 // based overload disambiguation lands in v7.12.5+).
3165 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("function") => {
3166 self.advance();
3167 let if_exists = self.consume_if_exists();
3168 let name = self.expect_ident_like()?;
3169 // v7.39 (read01 round 62) — the argument list identifies
3170 // WHICH overload to drop, so it is captured, not
3171 // discarded. `DROP FUNCTION f` (no list) is legal when
3172 // the name is unambiguous; the engine enforces that.
3173 let args = if matches!(self.peek(), Token::LParen) {
3174 Some(self.parse_function_signature_types()?)
3175 } else {
3176 None
3177 };
3178 // v7.39 (round 621) — the `CASCADE` / `RESTRICT`
3179 // trailer, which `DROP TABLE` and `DROP INDEX` have
3180 // accepted since v7.14 and this one refused outright.
3181 // pg_dump writes it, so refusing was a parse error in
3182 // the middle of a restore. SPG drops the function
3183 // either way — it tracks no dependents to cascade to —
3184 // which is the same reading the other two give it.
3185 self.consume_drop_behaviour();
3186 Ok(Statement::DropFunction {
3187 name,
3188 args,
3189 if_exists,
3190 })
3191 }
3192 // v7.14.0 — DROP TABLE [IF EXISTS] name [, name…]
3193 // [CASCADE|RESTRICT]. pg_dump and mysqldump both
3194 // emit DROP TABLE IF EXISTS at the head of every
3195 // CREATE TABLE block so re-importing a dump
3196 // overwrites prior state. SPG accepts and removes
3197 // matching tables; CASCADE/RESTRICT trailers
3198 // accepted silently.
3199 Token::Table => self.parse_drop_table_after_keyword(),
3200 // v7.14.0 — DROP INDEX [IF EXISTS] name
3201 // [CASCADE|RESTRICT]. PG / mysqldump emit this
3202 // for partial-index renames and pgvector
3203 // migrations. SPG removes the matching index;
3204 // IF EXISTS makes the drop idempotent.
3205 Token::Index => {
3206 self.advance();
3207 let if_exists_at = self.pos;
3208 let if_exists = self.consume_if_exists();
3209 let name = self.expect_ident_like()?;
3210 // v7.39.7 — MySQL's own spelling, which SPG
3211 // refused.
3212 //
3213 // `DROP INDEX i ON t` is how MySQL drops an
3214 // index; its names live inside a table, so the
3215 // statement names the table. Measured against
3216 // MySQL 9.7.2: the form above works, and the
3217 // bare `DROP INDEX i` PostgreSQL uses is a 1064
3218 // there. SPG had it exactly backwards on the
3219 // MySQL wire — the bare form accepted, MySQL's
3220 // own a syntax error — so a migration that drops
3221 // an index failed against the drop-in and not
3222 // against the thing it replaces.
3223 let table = if matches!(self.peek(), Token::On) {
3224 self.advance();
3225 Some(self.expect_ident_like()?)
3226 } else {
3227 None
3228 };
3229 if self.mysql_dialect {
3230 // MySQL has no `IF EXISTS` here either:
3231 // `DROP INDEX IF EXISTS i ON t` is a 1064.
3232 if if_exists {
3233 return Err(self.err_at(
3234 if_exists_at,
3235 "MySQL has no IF EXISTS on DROP INDEX".into(),
3236 ));
3237 }
3238 if table.is_none() {
3239 return Err(self.err("expected ON after the index name".into()));
3240 }
3241 }
3242 if matches!(
3243 self.peek(),
3244 Token::Ident(s) if s.eq_ignore_ascii_case("cascade")
3245 || s.eq_ignore_ascii_case("restrict")
3246 ) {
3247 self.advance();
3248 }
3249 Ok(Statement::DropIndex {
3250 name,
3251 if_exists,
3252 table,
3253 })
3254 }
3255 // v7.14.0 — DROP SCHEMA [IF EXISTS] name
3256 // [CASCADE|RESTRICT]. SPG is single-database;
3257 // v7.17.0 Phase 1.6 — DROP SCHEMA [IF EXISTS]
3258 // name [, name…] [CASCADE | RESTRICT]. Real
3259 // unregister (was silent no-op pre-v7.17).
3260 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("schema") => {
3261 self.advance();
3262 let if_exists = self.consume_if_exists();
3263 let mut names = vec![self.expect_ident_like()?];
3264 while matches!(self.peek(), Token::Comma) {
3265 self.advance();
3266 names.push(self.expect_ident_like()?);
3267 }
3268 if matches!(
3269 self.peek(),
3270 Token::Ident(s) if s.eq_ignore_ascii_case("cascade")
3271 || s.eq_ignore_ascii_case("restrict")
3272 ) {
3273 self.advance();
3274 }
3275 Ok(Statement::DropSchema { names, if_exists })
3276 }
3277 // v7.17.0 Phase 1.4 — DROP TYPE [IF EXISTS]
3278 // name [, name…] [CASCADE|RESTRICT].
3279 Token::Ident(s) | Token::QuotedIdent(s)
3280 if s.eq_ignore_ascii_case("type") =>
3281 {
3282 self.advance();
3283 let if_exists = self.consume_if_exists();
3284 let mut names = vec![self.expect_ident_like()?];
3285 while matches!(self.peek(), Token::Comma) {
3286 self.advance();
3287 names.push(self.expect_ident_like()?);
3288 }
3289 if matches!(
3290 self.peek(),
3291 Token::Ident(s) if s.eq_ignore_ascii_case("cascade")
3292 || s.eq_ignore_ascii_case("restrict")
3293 ) {
3294 self.advance();
3295 }
3296 Ok(Statement::DropType { names, if_exists })
3297 }
3298 // v7.17.0 Phase 1.5 — DROP DOMAIN [IF EXISTS]
3299 // name [, name…] [CASCADE|RESTRICT].
3300 Token::Ident(s) | Token::QuotedIdent(s)
3301 if s.eq_ignore_ascii_case("domain") =>
3302 {
3303 self.advance();
3304 let if_exists = self.consume_if_exists();
3305 let mut names = vec![self.expect_ident_like()?];
3306 while matches!(self.peek(), Token::Comma) {
3307 self.advance();
3308 names.push(self.expect_ident_like()?);
3309 }
3310 if matches!(
3311 self.peek(),
3312 Token::Ident(s) if s.eq_ignore_ascii_case("cascade")
3313 || s.eq_ignore_ascii_case("restrict")
3314 ) {
3315 self.advance();
3316 }
3317 Ok(Statement::DropDomain { names, if_exists })
3318 }
3319 // v7.17.0 Phase 1.3 — DROP MATERIALIZED VIEW
3320 // [IF EXISTS] name [, name…] [CASCADE|RESTRICT].
3321 Token::Ident(s) | Token::QuotedIdent(s)
3322 if s.eq_ignore_ascii_case("materialized") =>
3323 {
3324 self.advance();
3325 let nxt = self.peek().clone();
3326 if !matches!(&nxt, Token::Ident(s2) | Token::QuotedIdent(s2) if s2.eq_ignore_ascii_case("view"))
3327 {
3328 return Err(self.err(alloc::format!(
3329 "expected VIEW after DROP MATERIALIZED, got {nxt:?}"
3330 )));
3331 }
3332 self.advance();
3333 let if_exists = self.consume_if_exists();
3334 let mut names = vec![self.expect_ident_like()?];
3335 while matches!(self.peek(), Token::Comma) {
3336 self.advance();
3337 names.push(self.expect_ident_like()?);
3338 }
3339 if matches!(
3340 self.peek(),
3341 Token::Ident(s) if s.eq_ignore_ascii_case("cascade")
3342 || s.eq_ignore_ascii_case("restrict")
3343 ) {
3344 self.advance();
3345 }
3346 Ok(Statement::DropMaterializedView { names, if_exists })
3347 }
3348 // v7.17.0 Phase 1.2 — DROP VIEW [IF EXISTS]
3349 // name [, name…] [CASCADE|RESTRICT].
3350 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("view") => {
3351 self.advance();
3352 let if_exists = self.consume_if_exists();
3353 let mut names = vec![self.expect_ident_like()?];
3354 while matches!(self.peek(), Token::Comma) {
3355 self.advance();
3356 names.push(self.expect_ident_like()?);
3357 }
3358 if matches!(
3359 self.peek(),
3360 Token::Ident(s) if s.eq_ignore_ascii_case("cascade")
3361 || s.eq_ignore_ascii_case("restrict")
3362 ) {
3363 self.advance();
3364 }
3365 Ok(Statement::DropView { names, if_exists })
3366 }
3367 // v7.17.0 — DROP SEQUENCE [IF EXISTS] name [,name…]
3368 // [CASCADE|RESTRICT]. Real removal from catalog
3369 // (was a silent no-op pre-v7.17).
3370 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("sequence") => {
3371 self.advance();
3372 let if_exists = self.consume_if_exists();
3373 let mut names = vec![self.expect_ident_like()?];
3374 while matches!(self.peek(), Token::Comma) {
3375 self.advance();
3376 names.push(self.expect_ident_like()?);
3377 }
3378 if matches!(
3379 self.peek(),
3380 Token::Ident(s) if s.eq_ignore_ascii_case("cascade")
3381 || s.eq_ignore_ascii_case("restrict")
3382 ) {
3383 self.advance();
3384 }
3385 Ok(Statement::DropSequence { names, if_exists })
3386 }
3387 // v7.39 (RLS) — DROP POLICY [IF EXISTS] name ON table.
3388 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("policy") => {
3389 self.advance();
3390 self.parse_drop_policy_after_keyword()
3391 }
3392 // v7.37.17 (17.6 siblings) — DROP <target> for
3393 // targets SPG doesn't natively track. pg_dump
3394 // emits DROP EXTENSION / DROP TYPE / DROP DOMAIN
3395 // / DROP AGGREGATE / DROP OPERATOR / DROP CAST /
3396 // DROP COLLATION / DROP LANGUAGE / DROP CONVERSION
3397 // / DROP TEXT SEARCH / DROP FOREIGN * / DROP
3398 // SERVER / DROP MATERIALIZED VIEW / DROP EVENT
3399 // TRIGGER / DROP TABLESPACE / DROP RULE / DROP
3400 // POLICY / DROP LARGE OBJECT / DROP ROLE / DROP
3401 // ACCESS METHOD / DROP OPERATOR CLASS/FAMILY /
3402 // etc. — accept + Empty-return so pg_dump tails
3403 // load through. Materialized-view drop dispatches
3404 // to the existing DropTable path when the token
3405 // is Materialized-View-shaped (elsewhere in
3406 // this parser).
3407 Token::Ident(s) | Token::QuotedIdent(s)
3408 if s.eq_ignore_ascii_case("text")
3409 // The DROP dispatch matches on PEEK — `text` is
3410 // not yet consumed, so SEARCH/CONFIGURATION sit
3411 // at pos+1/pos+2 (the round-695 trap's mirror).
3412 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(k)) if k.eq_ignore_ascii_case("search"))
3413 && matches!(self.tokens.get(self.pos + 2), Some(Token::Ident(k)) if k.eq_ignore_ascii_case("configuration")) =>
3414 {
3415 // v7.39 (round 709) — DROP TEXT SEARCH CONFIGURATION
3416 // validates the name; DICTIONARY / PARSER / TEMPLATE
3417 // stay in the noise arm below.
3418 self.advance(); // TEXT
3419 self.advance(); // SEARCH
3420 self.advance(); // CONFIGURATION
3421 let if_exists = self.consume_if_exists();
3422 let names = self.take_comma_separated_names();
3423 self.consume_until_statement_boundary();
3424 if if_exists {
3425 return Ok(Statement::Empty);
3426 }
3427 Ok(Statement::ValidateOnly {
3428 kind: crate::ast::ValidateOnlyKind::TsConfigName,
3429 names,
3430 })
3431 }
3432 Token::Ident(s) | Token::QuotedIdent(s)
3433 if matches!(
3434 s.to_ascii_lowercase().as_str(),
3435 "type"
3436 | "domain"
3437 | "operator"
3438 | "cast"
3439 // `text` = TEXT SEARCH DICTIONARY / PARSER /
3440 // TEMPLATE (CONFIGURATION intercepted above).
3441 | "text"
3442 | "materialized"
3443 | "large"
3444 | "role"
3445 | "access"
3446 | "procedure"
3447 | "routine"
3448 ) =>
3449 {
3450 self.consume_until_statement_boundary();
3451 Ok(Statement::Empty)
3452 }
3453 // v7.39 (round 709) — DROP COLLATION / EVENT TRIGGER /
3454 // TABLESPACE / TEXT SEARCH CONFIGURATION validate their
3455 // NAME; DROP SERVER / DROP FOREIGN TABLE join the
3456 // foreign-data warning family (round 706) so a
3457 // CREATE→DROP sequence in a dump stays consistent.
3458 Token::Ident(s) | Token::QuotedIdent(s)
3459 if s.eq_ignore_ascii_case("server")
3460 || s.eq_ignore_ascii_case("foreign") =>
3461 {
3462 self.advance();
3463 self.consume_until_statement_boundary();
3464 Ok(Statement::ValidateOnly {
3465 kind: crate::ast::ValidateOnlyKind::ForeignInfra,
3466 names: Vec::new(),
3467 })
3468 }
3469 Token::Ident(s) | Token::QuotedIdent(s)
3470 if s.eq_ignore_ascii_case("collation")
3471 || s.eq_ignore_ascii_case("tablespace") =>
3472 {
3473 let kind = if s.eq_ignore_ascii_case("collation") {
3474 crate::ast::ValidateOnlyKind::CollationName
3475 } else {
3476 crate::ast::ValidateOnlyKind::TablespaceName
3477 };
3478 self.advance();
3479 let if_exists = self.consume_if_exists();
3480 let names = self.take_comma_separated_names();
3481 self.consume_until_statement_boundary();
3482 if if_exists {
3483 return Ok(Statement::Empty);
3484 }
3485 Ok(Statement::ValidateOnly { kind, names })
3486 }
3487 Token::Ident(s) | Token::QuotedIdent(s)
3488 if s.eq_ignore_ascii_case("event") =>
3489 {
3490 self.advance();
3491 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("trigger"))
3492 {
3493 self.advance();
3494 }
3495 let if_exists = self.consume_if_exists();
3496 let names = self.take_comma_separated_names();
3497 self.consume_until_statement_boundary();
3498 if if_exists {
3499 return Ok(Statement::Empty);
3500 }
3501 Ok(Statement::ValidateOnly {
3502 kind: crate::ast::ValidateOnlyKind::EventTriggerName,
3503 names,
3504 })
3505 }
3506 // v7.39 (round 708) — DROP CONVERSION / DROP LANGUAGE
3507 // leave the noise list; see the ValidateOnly kinds.
3508 Token::Ident(s) | Token::QuotedIdent(s)
3509 if s.eq_ignore_ascii_case("conversion")
3510 || s.eq_ignore_ascii_case("language")
3511 // `DROP PROCEDURAL LANGUAGE` puts the modifier
3512 // FIRST — the first draft looked for it after.
3513 || s.eq_ignore_ascii_case("procedural") =>
3514 {
3515 let kind = if s.eq_ignore_ascii_case("conversion") {
3516 crate::ast::ValidateOnlyKind::ConversionName
3517 } else {
3518 crate::ast::ValidateOnlyKind::LanguageName
3519 };
3520 self.advance();
3521 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("language"))
3522 {
3523 self.advance();
3524 }
3525 let if_exists = self.consume_if_exists();
3526 let names = self.take_comma_separated_names();
3527 self.consume_until_statement_boundary();
3528 if if_exists {
3529 return Ok(Statement::Empty);
3530 }
3531 Ok(Statement::ValidateOnly { kind, names })
3532 }
3533 // v7.39 (round 707) — `DROP AGGREGATE [IF EXISTS]
3534 // name(argtypes)[, …]`. Parsed for real so the engine
3535 // can answer as PG does; see Statement::DropAggregate.
3536 Token::Ident(s) | Token::QuotedIdent(s)
3537 if s.eq_ignore_ascii_case("aggregate") =>
3538 {
3539 self.advance();
3540 let if_exists = self.consume_if_exists();
3541 let mut items: Vec<(String, Option<Vec<String>>)> = Vec::new();
3542 loop {
3543 let name = self.expect_ident_like()?;
3544 if !matches!(self.peek(), Token::LParen) {
3545 return Err(self.err(alloc::format!(
3546 "expected argument list after DROP AGGREGATE {name}"
3547 )));
3548 }
3549 self.advance();
3550 let mut args: Vec<String> = Vec::new();
3551 let mut star = false;
3552 loop {
3553 match self.peek().clone() {
3554 Token::RParen => {
3555 self.advance();
3556 break;
3557 }
3558 Token::Star => {
3559 self.advance();
3560 star = true;
3561 }
3562 Token::Comma => {
3563 self.advance();
3564 }
3565 _ => {
3566 // A type name may be multi-token
3567 // (`double precision`); glue idents
3568 // until , or ).
3569 let mut t = self.expect_ident_like()?;
3570 while let Token::Ident(nx) = self.peek() {
3571 let nx = nx.clone();
3572 self.advance();
3573 t.push(' ');
3574 t.push_str(&nx);
3575 }
3576 args.push(t);
3577 }
3578 }
3579 }
3580 items.push((name, if star { None } else { Some(args) }));
3581 if matches!(self.peek(), Token::Comma) {
3582 self.advance();
3583 } else {
3584 break;
3585 }
3586 }
3587 self.consume_until_statement_boundary();
3588 Ok(Statement::DropAggregate { if_exists, items })
3589 }
3590 // v7.39 (round 697) — `DROP EXTENSION [IF EXISTS] <e>
3591 // [, …] [CASCADE|RESTRICT]`. PG refuses one that is not
3592 // installed; `IF EXISTS` is the spelling that says do
3593 // not, and it keeps the no-op.
3594 Token::Ident(s) | Token::QuotedIdent(s)
3595 if s.eq_ignore_ascii_case("extension") =>
3596 {
3597 self.advance();
3598 let if_exists = self.consume_if_exists();
3599 let names = self.take_comma_separated_names();
3600 self.consume_until_statement_boundary();
3601 if if_exists {
3602 return Ok(Statement::Empty);
3603 }
3604 Ok(Statement::ValidateOnly {
3605 kind: crate::ast::ValidateOnlyKind::ExtensionInstalled,
3606 names,
3607 })
3608 }
3609 Token::Ident(s) | Token::QuotedIdent(s)
3610 if s.eq_ignore_ascii_case("statistics") =>
3611 {
3612 self.parse_drop_statistics_after_drop()
3613 }
3614 other => Err(self.err(format!(
3615 "expected TABLE / INDEX / SCHEMA / SEQUENCE / USER / PUBLICATION / \
3616 SUBSCRIPTION / TRIGGER / FUNCTION / STATISTICS after DROP, got {other:?}"
3617 ))),
3618 }
3619 }
3620 // v7.17.0 Phase 1.3 — REFRESH MATERIALIZED VIEW name [WITH [NO] DATA].
3621 // v7.37.19 (19.8) — `CONCURRENTLY` modifier (PG 9.4+) parsed
3622 // and accepted before the view name. SPG materialised
3623 // views re-evaluate on read (always-fresh semantics), so
3624 // the CONCURRENTLY-vs-serial distinction has no runtime
3625 // effect — the refresh body does not block readers either
3626 // way. Same accept-and-no-op pattern as DETACH PARTITION
3627 // CONCURRENTLY (16.5).
3628 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("refresh") => {
3629 self.advance();
3630 let nxt = self.peek().clone();
3631 if !matches!(&nxt, Token::Ident(s2) | Token::QuotedIdent(s2) if s2.eq_ignore_ascii_case("materialized"))
3632 {
3633 return Err(self.err(alloc::format!(
3634 "expected MATERIALIZED after REFRESH, got {nxt:?}"
3635 )));
3636 }
3637 self.advance();
3638 let nxt2 = self.peek().clone();
3639 if !matches!(&nxt2, Token::Ident(s2) | Token::QuotedIdent(s2) if s2.eq_ignore_ascii_case("view"))
3640 {
3641 return Err(self.err(alloc::format!(
3642 "expected VIEW after REFRESH MATERIALIZED, got {nxt2:?}"
3643 )));
3644 }
3645 self.advance();
3646 // Optional CONCURRENTLY noise word — consumed without
3647 // changing semantics.
3648 if matches!(self.peek(), Token::Ident(s2) | Token::QuotedIdent(s2) if s2.eq_ignore_ascii_case("concurrently"))
3649 {
3650 self.advance();
3651 }
3652 let name = self.expect_ident_like()?;
3653 let with_data = self.parse_optional_with_data(true)?;
3654 Ok(Statement::RefreshMaterializedView { name, with_data })
3655 }
3656 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("update") => {
3657 self.advance();
3658 self.parse_update_after_keyword()
3659 }
3660 // v7.37.17 (17.6 sibling) — TRUNCATE [TABLE] [ONLY]
3661 // <name> [, ...] [RESTART IDENTITY | CONTINUE IDENTITY]
3662 // [CASCADE | RESTRICT]. Clears every row from each named
3663 // table. Parses at the top level; the engine dispatcher
3664 // walks Statement::Truncate.
3665 // v7.39.9 — MySQL's top-level `RENAME TABLE a TO b [, c TO d]`.
3666 //
3667 // PostgreSQL renames a table through `ALTER TABLE … RENAME
3668 // TO`, which SPG already had, so this spelling answered 1064
3669 // — and it is what a MySQL migration writes. Measured on
3670 // 9.7.2: several pairs in one statement are accepted, and
3671 // renaming onto a name that exists is 1050.
3672 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("rename") => {
3673 self.advance();
3674 if matches!(self.peek(), Token::Table) {
3675 self.advance();
3676 }
3677 let mut pairs: Vec<(String, String)> = Vec::new();
3678 loop {
3679 let from = self.expect_ident_like()?;
3680 if matches!(self.peek(), Token::To) {
3681 self.advance();
3682 } else {
3683 self.expect_keyword_ident("to")?;
3684 }
3685 let to = self.expect_ident_like()?;
3686 pairs.push((from, to));
3687 if matches!(self.peek(), Token::Comma) {
3688 self.advance();
3689 } else {
3690 break;
3691 }
3692 }
3693 Ok(Statement::RenameTables(pairs))
3694 }
3695 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("truncate") => {
3696 self.advance();
3697 // Optional TABLE noise word — PG accepts both the reserved
3698 // token and the bare identifier spelling.
3699 if matches!(self.peek(), Token::Table)
3700 || matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("table"))
3701 {
3702 self.advance();
3703 }
3704 // v7.39 (round 647) — `TRUNCATE ONLY t` is carried now,
3705 // not absorbed. The lookahead keeps a table genuinely
3706 // called `only` working: the keyword is a keyword only
3707 // when a name follows it.
3708 let only = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
3709 if s.eq_ignore_ascii_case("only"))
3710 && matches!(
3711 self.tokens.get(self.pos + 1),
3712 Some(Token::Ident(_) | Token::QuotedIdent(_))
3713 );
3714 if only {
3715 self.advance();
3716 }
3717 // Table names (comma-separated).
3718 let mut tables = Vec::new();
3719 loop {
3720 tables.push(self.expect_ident_like()?);
3721 if matches!(self.peek(), Token::Comma) {
3722 self.advance();
3723 continue;
3724 }
3725 break;
3726 }
3727 // Optional RESTART IDENTITY / CONTINUE IDENTITY.
3728 let mut restart_identity = false;
3729 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("restart"))
3730 {
3731 self.advance();
3732 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("identity"))
3733 {
3734 self.advance();
3735 restart_identity = true;
3736 }
3737 } else if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("continue"))
3738 {
3739 self.advance();
3740 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("identity"))
3741 {
3742 self.advance();
3743 }
3744 }
3745 // Optional CASCADE / RESTRICT.
3746 let mut cascade = false;
3747 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("cascade"))
3748 {
3749 self.advance();
3750 cascade = true;
3751 } else if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("restrict"))
3752 {
3753 self.advance();
3754 }
3755 Ok(Statement::Truncate {
3756 tables,
3757 restart_identity,
3758 cascade,
3759 only,
3760 })
3761 }
3762 // v7.37.17 (17.6 sibling) — REINDEX [(OPTION [, ...])]
3763 // [CONCURRENTLY] { INDEX | TABLE | SCHEMA | DATABASE |
3764 // SYSTEM } [IF EXISTS] <name>. SPG rebuilds indexes as
3765 // rows change so the index tree is always up-to-date;
3766 // REINDEX is a strict no-op. Accept the whole statement
3767 // shape to boundary for pg_dump round-trip compatibility.
3768 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("reindex") => {
3769 // v7.39 (round 535) — the target is CARRIED now. SPG has no
3770 // index bloat to rebuild, so the work stays a no-op, but PG
3771 // validates what it was pointed at and this swallowed the
3772 // name at parse time — `REINDEX TABLE typo` reported
3773 // success. Measured on PG18: INDEX / TABLE name a relation,
3774 // SCHEMA a schema, SYSTEM nothing.
3775 self.advance();
3776 self.parse_reindex_tail()
3777 }
3778 // v7.37.17 (17.6 sibling) — VACUUM [(OPTION [, ...])]
3779 // [FULL] [FREEZE] [VERBOSE] [ANALYZE] [<table> [(cols)]].
3780 // SPG has no MVCC bloat today (Phase D visibility map
3781 // queues with v7.38); the freezer collapses hot-tier
3782 // rows into cold segments automatically. VACUUM is a
3783 // no-op — pg_dump maintenance scripts and Discourse's
3784 // periodic-maintenance path both emit it.
3785 // v7.39 (round 169) — VACUUM is REAL now: with the in-place
3786 // MVCC gate ON (v7.37.15 flip), tombstoned versions are
3787 // actual bloat, so the pre-MVCC accept-and-ignore posture
3788 // became a silent no-op on a customer's manual reclaim.
3789 // Grammar: VACUUM [(opts)] [FULL] [FREEZE] [VERBOSE]
3790 // [ANALYZE] [<table> [(cols)]] — option words are absorbed,
3791 // ANALYZE is captured, the optional table name is captured.
3792 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("vacuum") => {
3793 self.advance();
3794 // Parenthesised option list: absorb it.
3795 if matches!(self.peek(), Token::LParen) {
3796 let mut depth = 0usize;
3797 loop {
3798 match self.advance() {
3799 Token::LParen => depth += 1,
3800 Token::RParen => {
3801 depth -= 1;
3802 if depth == 0 {
3803 break;
3804 }
3805 }
3806 Token::Eof => break,
3807 _ => {}
3808 }
3809 }
3810 }
3811 let mut analyze = false;
3812 let mut table: Option<String> = None;
3813 loop {
3814 match self.peek() {
3815 // v7.39 (round 535) — `FULL` lexes as a keyword, not
3816 // an identifier, so the loop below broke out on it and
3817 // dropped the table name: `VACUUM FULL nosuch` was
3818 // accepted where `VACUUM nosuch` was refused.
3819 Token::Full => {
3820 self.advance();
3821 }
3822 Token::Ident(w) | Token::QuotedIdent(w) => {
3823 let wl = w.to_ascii_lowercase();
3824 match wl.as_str() {
3825 "full" | "freeze" | "verbose" => {
3826 self.advance();
3827 }
3828 "analyze" | "analyse" => {
3829 analyze = true;
3830 self.advance();
3831 }
3832 _ => {
3833 table = Some(self.expect_ident_like()?);
3834 break;
3835 }
3836 }
3837 }
3838 _ => break,
3839 }
3840 }
3841 // Optional trailing column list / anything else to the
3842 // statement boundary (PG accepts per-column ANALYZE).
3843 self.consume_until_statement_boundary();
3844 Ok(Statement::Vacuum { table, analyze })
3845 }
3846 // v7.37.17 (17.6 sibling) — CLUSTER [VERBOSE] <table>
3847 // [USING <index>] / CLUSTER (VERBOSE) <table> USING
3848 // <index>. PG stores rows in physical order matching
3849 // an index; SPG's hot-tier is append-only + cold-tier
3850 // is segment-frozen, so clustering has no persistent
3851 // effect. Accept-and-no-op for pg_dump compat.
3852 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("cluster") => {
3853 // v7.39 (round 535) — same as REINDEX above: the relation is
3854 // carried so the engine can refuse one that does not exist.
3855 // A bare `CLUSTER [VERBOSE]` names nothing and is accepted.
3856 self.advance();
3857 self.parse_cluster_tail()
3858 }
3859 // v7.39 (round 222) — LISTEN / NOTIFY / UNLISTEN with real
3860 // delivery (was accept-and-drop since v7.37.17). NOTIFY takes an
3861 // optional string payload; UNLISTEN takes a channel or `*`.
3862 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("listen") => {
3863 self.advance();
3864 let ch = match self.advance() {
3865 Token::Ident(c) | Token::QuotedIdent(c) => c,
3866 other => {
3867 return Err(self.err(format!(
3868 "expected channel name after LISTEN, got {other:?}"
3869 )));
3870 }
3871 };
3872 Ok(Statement::Listen(ch))
3873 }
3874 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("notify") => {
3875 self.advance();
3876 let channel = match self.advance() {
3877 Token::Ident(c) | Token::QuotedIdent(c) => c,
3878 other => {
3879 return Err(self.err(format!(
3880 "expected channel name after NOTIFY, got {other:?}"
3881 )));
3882 }
3883 };
3884 let payload = if matches!(self.peek(), Token::Comma) {
3885 self.advance();
3886 match self.advance() {
3887 Token::String(p) => Some(p),
3888 other => {
3889 return Err(self.err(format!(
3890 "expected string payload after NOTIFY <channel>, got {other:?}"
3891 )));
3892 }
3893 }
3894 } else {
3895 None
3896 };
3897 Ok(Statement::Notify { channel, payload })
3898 }
3899 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("unlisten") => {
3900 self.advance();
3901 match self.advance() {
3902 Token::Star => Ok(Statement::Unlisten(None)),
3903 Token::Ident(c) | Token::QuotedIdent(c) => Ok(Statement::Unlisten(Some(c))),
3904 other => Err(self.err(format!(
3905 "expected channel name or * after UNLISTEN, got {other:?}"
3906 ))),
3907 }
3908 }
3909 // v7.37.17 (17.6 sibling) — LOCK [TABLE] [ONLY] <table>
3910 // [IN <mode> MODE] [NOWAIT]. SPG's engine holds a
3911 // process-wide write lock today; explicit LOCK has no
3912 // effect. Accept-and-no-op for pg_dump / migration
3913 // compat.
3914 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("lock") => {
3915 self.advance();
3916 // v7.39 (round 696) — the LOCK still has no effect (SPG's
3917 // engine holds a process-wide write lock), but the TABLE
3918 // NAME is now carried out so the engine can refuse one that
3919 // does not exist, as PG18 does. MySQL's `LOCK TABLES …
3920 // READ|WRITE` is a different statement with the same first
3921 // word; it keeps the old no-op, because a MySQL dump's
3922 // bracket names tables it is about to create.
3923 let mysql_tables = matches!(self.peek(), Token::Ident(k)
3924 if k.eq_ignore_ascii_case("tables"));
3925 if mysql_tables {
3926 self.consume_until_statement_boundary();
3927 return Ok(Statement::Empty);
3928 }
3929 if matches!(self.peek(), Token::Table) {
3930 self.advance();
3931 }
3932 let names = self.take_comma_separated_names();
3933 self.consume_until_statement_boundary();
3934 Ok(Statement::ValidateOnly {
3935 kind: crate::ast::ValidateOnlyKind::LockTable,
3936 names,
3937 })
3938 }
3939 // v7.37.17 (17.6 sibling) — CHECKPOINT. Forces a WAL
3940 // durability marker + snapshot in PG. SPG has WAL
3941 // checkpointing on a byte / time schedule (v7.37.10
3942 // 60s / 4 MiB defaults). The bare statement parses to
3943 // `Statement::Empty` here (the no_std engine owns no
3944 // WAL / snapshot); v7.37 Epic Du wires the HOST
3945 // (embedded `Database::execute_buffered`, via
3946 // `sql_is_checkpoint`) to force an immediate synchronous
3947 // checkpoint through `Database::checkpoint` — a real
3948 // durability barrier, matching PG.
3949 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("checkpoint") => {
3950 self.advance();
3951 self.consume_until_statement_boundary();
3952 Ok(Statement::Empty)
3953 }
3954 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("delete") => {
3955 self.advance();
3956 self.parse_delete_after_keyword()
3957 }
3958 // v6.0.4: ALTER INDEX <name> REBUILD [WITH (encoding = ...)].
3959 // ALTER is not a reserved keyword in the lexer — handled
3960 // as a bare ident here.
3961 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("alter") => {
3962 self.advance();
3963 self.parse_alter_after_keyword()
3964 }
3965 // v6.1.7: WAIT FOR WAL POSITION <pos> [WITH TIMEOUT <ms>].
3966 // WAIT / POSITION / TIMEOUT are bare idents — no lexer
3967 // additions needed.
3968 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("wait") => {
3969 self.advance();
3970 self.parse_wait_after_keyword()
3971 }
3972 // v6.2.0: ANALYZE [<table>]. ANALYZE is a bare ident.
3973 // Bare ANALYZE → analyse every user table; ANALYZE
3974 // <name> → re-stats one. The argument is an optional
3975 // ident (or quoted ident); anything else is a parse
3976 // error.
3977 // v6.7.3 — `COMPACT COLD SEGMENTS`. No arguments, no
3978 // `WHERE` filter (carved out per V6_7_DESIGN.md
3979 // STABILITY). Lex order: identifier "compact" → "cold"
3980 // → "segments". Anything else after `COMPACT` is a
3981 // parse error.
3982 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("compact") => {
3983 self.advance();
3984 let next = self.peek().clone();
3985 let cold = match next {
3986 Token::Ident(s) | Token::QuotedIdent(s) => s,
3987 _ => {
3988 return Err(
3989 self.err(format!("expected COLD after COMPACT, got {:?}", self.peek()))
3990 );
3991 }
3992 };
3993 if !cold.eq_ignore_ascii_case("cold") {
3994 return Err(self.err(format!("expected COLD after COMPACT, got {cold:?}")));
3995 }
3996 self.advance();
3997 let next = self.peek().clone();
3998 let segments = match next {
3999 Token::Ident(s) | Token::QuotedIdent(s) => s,
4000 _ => {
4001 return Err(self.err(format!(
4002 "expected SEGMENTS after COMPACT COLD, got {:?}",
4003 self.peek()
4004 )));
4005 }
4006 };
4007 if !segments.eq_ignore_ascii_case("segments") {
4008 return Err(self.err(format!(
4009 "expected SEGMENTS after COMPACT COLD, got {segments:?}"
4010 )));
4011 }
4012 self.advance();
4013 Ok(Statement::CompactColdSegments)
4014 }
4015 // v7.17.0 Phase 3.P0-42 — SQL:2003 / PG 15+ MERGE.
4016 // Parsed as a case-insensitive identifier since MERGE
4017 // isn't a reserved lexer keyword (collides with the
4018 // mysqldump `ALGORITHM = MERGE` view clause if it
4019 // were); the inner parser drives the rest of the
4020 // surface (USING / ON / WHEN [NOT] MATCHED / THEN).
4021 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("merge") => {
4022 self.advance();
4023 self.parse_merge_after_keyword()
4024 }
4025 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("analyze") => {
4026 self.advance();
4027 // v7.39.9 — MySQL spells it `ANALYZE TABLE t`. The
4028 // keyword is noise to the parse; what differs is the
4029 // ANSWER, which MySQL returns as a result set — see the
4030 // executor.
4031 let mysql_table_kw = matches!(self.peek(), Token::Table);
4032 if mysql_table_kw {
4033 self.advance();
4034 }
4035 let target = match self.peek() {
4036 Token::Eof | Token::Semicolon => None,
4037 Token::Ident(_) | Token::QuotedIdent(_) => {
4038 Some(self.expect_ident_like()?)
4039 }
4040 other => {
4041 return Err(self.err(format!(
4042 "expected table name or end of statement after ANALYZE, got {other:?}"
4043 )));
4044 }
4045 };
4046 // v7.39 (round 776, F31 J7) — the per-column form
4047 // (`ANALYZE t (x, y)`, PG-accepted) was a syntax error
4048 // here while the VACUUM arm already consumed it; SPG
4049 // analyzes whole tables, so the list parses and is
4050 // accepted like the VACUUM path's.
4051 if target.is_some() && matches!(self.peek(), Token::LParen) {
4052 self.advance();
4053 loop {
4054 let _ = self.expect_ident_like()?;
4055 match self.peek() {
4056 Token::Comma => {
4057 self.advance();
4058 }
4059 Token::RParen => {
4060 self.advance();
4061 break;
4062 }
4063 other => {
4064 return Err(self.err(format!(
4065 "expected ',' or ')' in ANALYZE column list, got {other:?}"
4066 )));
4067 }
4068 }
4069 }
4070 }
4071 Ok(Statement::Analyze(target))
4072 }
4073 // v7.12.1 — `SET <name> [TO|=] <value>`. The
4074 // `default_text_search_config` parameter is consumed
4075 // by the FTS function dispatcher; other parameter
4076 // names are recorded but treated as a no-op so PG
4077 // dump output loads.
4078 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("set") => {
4079 self.advance();
4080 // PG allows `SET LOCAL` / `SET SESSION` qualifiers; MySQL
4081 // adds `SET GLOBAL` too (and the alias `SET @@global.name =
4082 // …` which the SessionVar path handles). `LOCAL` is the only
4083 // one that changes semantics — it scopes the change to the
4084 // current transaction — so capture it; SESSION / GLOBAL are
4085 // accepted and treated as the default session scope.
4086 let mut set_local = false;
4087 if let Token::Ident(s) | Token::QuotedIdent(s) = self.peek() {
4088 let q = s.to_ascii_lowercase();
4089 if q == "local" || q == "session" || q == "global" {
4090 set_local = q == "local";
4091 self.advance();
4092 }
4093 }
4094 // 7.38.1 S5.2 — PG `SET [SESSION] AUTHORIZATION
4095 // { DEFAULT | <role> }`. pg_dump's ACL section switches
4096 // to the object owner with it. SPG maps it onto the
4097 // session-role machinery (recorded delta RD-10: PG moves
4098 // session_user too; SPG moves the effective role).
4099 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
4100 if s.eq_ignore_ascii_case("authorization"))
4101 {
4102 self.advance(); // AUTHORIZATION
4103 let role = match self.peek().clone() {
4104 Token::Default => {
4105 self.advance();
4106 None
4107 }
4108 Token::String(s) | Token::Ident(s) | Token::QuotedIdent(s) => {
4109 self.advance();
4110 Some(s)
4111 }
4112 _ => None,
4113 };
4114 return Ok(Statement::SetRole(role));
4115 }
4116 // v7.14.0 — MySQL `SET NAMES <charset> [COLLATE
4117 // <collation>]` — change the connection client
4118 // charset. SPG stores UTF-8 always and orders
4119 // bytewise; accept as a no-op.
4120 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("names"))
4121 {
4122 self.advance();
4123 // v7.39 — this used to parse the clause and throw it
4124 // away ("SPG stores UTF-8 always and orders
4125 // bytewise; accept as a no-op"). That sentence
4126 // stopped being true when collations arrived, and
4127 // once `collation_connection` began driving literal
4128 // comparison, dropping the COLLATE clause became a
4129 // silently wrong answer: `SET NAMES utf8mb4 COLLATE
4130 // utf8mb4_general_ci` reported back
4131 // `utf8mb4_0900_ai_ci` and compared as NO PAD.
4132 //
4133 // The charset name is emitted as `names` and the
4134 // ENGINE expands it, because which collation a
4135 // charset defaults to is MySQL semantics and belongs
4136 // beside the rest of them, not in the parser.
4137 let mut pairs = alloc::vec::Vec::new();
4138 if matches!(
4139 self.peek(),
4140 Token::Ident(_) | Token::QuotedIdent(_) | Token::String(_)
4141 ) {
4142 let charset = match self.advance() {
4143 Token::Ident(s) | Token::QuotedIdent(s) | Token::String(s) => s,
4144 _ => unreachable!("peeked an ident-or-string"),
4145 };
4146 pairs.push((String::from("names"), crate::ast::SetValue::Ident(charset)));
4147 }
4148 // Optional `COLLATE <name>` — emitted AFTER `names`
4149 // so it overrides the charset's default, which is
4150 // what MySQL does.
4151 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("collate"))
4152 {
4153 self.advance();
4154 if matches!(
4155 self.peek(),
4156 Token::Ident(_) | Token::QuotedIdent(_) | Token::String(_)
4157 ) {
4158 let coll = match self.advance() {
4159 Token::Ident(s) | Token::QuotedIdent(s) | Token::String(s) => s,
4160 _ => unreachable!("peeked an ident-or-string"),
4161 };
4162 pairs.push((
4163 String::from("collation_connection"),
4164 crate::ast::SetValue::Ident(coll),
4165 ));
4166 }
4167 }
4168 if pairs.is_empty() {
4169 return Ok(Statement::Empty);
4170 }
4171 return Ok(Statement::SetParameterList(pairs));
4172 }
4173 // v7.37.17 (17.6 sibling) — PG `SET ROLE
4174 // { NONE | DEFAULT | <role_name> }`. pg_dump preamble
4175 // uses this to switch to the object owner before
4176 // recreating tables. SPG has no role system so this
4177 // is a no-op.
4178 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("role"))
4179 {
4180 self.advance(); // ROLE
4181 // v7.39 (RLS) — real session-role switch. NONE / DEFAULT
4182 // reset to the login identity; a name / string sets the
4183 // effective role that drives current_user + RLS.
4184 let role = match self.peek().clone() {
4185 Token::Default => {
4186 self.advance();
4187 None
4188 }
4189 Token::Ident(s) | Token::QuotedIdent(s)
4190 if s.eq_ignore_ascii_case("none") =>
4191 {
4192 self.advance();
4193 None
4194 }
4195 Token::String(s) | Token::Ident(s) | Token::QuotedIdent(s) => {
4196 self.advance();
4197 Some(s)
4198 }
4199 _ => None,
4200 };
4201 return Ok(Statement::SetRole(role));
4202 }
4203 // v7.37.17 (17.6 sibling) — PG `SET SESSION
4204 // CHARACTERISTICS AS TRANSACTION <mode>` (per PG
4205 // ISO SQL surface). pg_dump prepends this to fix
4206 // the isolation level for the restore session. SPG
4207 // defaults to READ COMMITTED and doesn't yet honor
4208 // session-set isolation across statements — accept
4209 // and no-op. SET (LOCAL/SESSION) TRANSACTION AS ...
4210 // per-tx form is handled elsewhere.
4211 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("characteristics"))
4212 {
4213 self.advance(); // CHARACTERISTICS
4214 if matches!(self.peek(), Token::As) {
4215 self.advance();
4216 }
4217 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("transaction")) {
4218 self.advance();
4219 }
4220 // v7.39 — no longer a no-op. The note above said SPG
4221 // "doesn't yet honor session-set isolation across
4222 // statements"; it does now, through
4223 // `default_transaction_isolation`, and measured on
4224 // PG 18.6 this statement is exactly a way to set it:
4225 //
4226 // SET SESSION CHARACTERISTICS AS TRANSACTION
4227 // ISOLATION LEVEL REPEATABLE READ;
4228 // current_setting('default_transaction_isolation')
4229 // -> repeatable read
4230 //
4231 // pg_dump prepends this to fix the level for a
4232 // restore session, so accepting it and doing nothing
4233 // meant the restore ran at a level nobody chose.
4234 //
4235 // The trailing READ ONLY / [NOT] DEFERRABLE modes are
4236 // still consumed and dropped. `default_transaction_read_only`
4237 // exists in the GUC inventory but nothing enforces it,
4238 // and setting a value no code honours is the very
4239 // defect this version is about — a session told it
4240 // holds a guarantee it does not.
4241 let modes = self.parse_isolation_level_clauses()?;
4242 self.consume_until_statement_boundary();
4243 let mut pairs: alloc::vec::Vec<(
4244 alloc::string::String,
4245 crate::ast::SetValue,
4246 )> = alloc::vec::Vec::new();
4247 if let Some(level) = modes.isolation {
4248 pairs.push((
4249 alloc::string::String::from("default_transaction_isolation"),
4250 crate::ast::SetValue::String(alloc::string::String::from(
4251 level.as_pg_str(),
4252 )),
4253 ));
4254 }
4255 if let Some(ro) = modes.read_only {
4256 pairs.push((
4257 alloc::string::String::from("default_transaction_read_only"),
4258 crate::ast::SetValue::Ident(alloc::string::String::from(if ro {
4259 "on"
4260 } else {
4261 "off"
4262 })),
4263 ));
4264 }
4265 return Ok(if pairs.is_empty() {
4266 Statement::Empty
4267 } else {
4268 Statement::SetParameterList(pairs)
4269 });
4270 }
4271 // v7.37.17 (17.6 sibling) — PG `SET CONSTRAINTS
4272 // { ALL | <name>[, ...] } { DEFERRED | IMMEDIATE }`.
4273 // pg_dump emits this to control the deferrability of
4274 // FK / UNIQUE constraints across a bulk restore. SPG
4275 // has no deferrable-constraint machinery today; the
4276 // FK checker is strict-immediate. Accept-and-no-op
4277 // for pg_dump round-trip compatibility.
4278 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("constraints"))
4279 {
4280 self.advance(); // CONSTRAINTS
4281 // v7.39 (round 288) — no longer a no-op: the trailing
4282 // DEFERRED / IMMEDIATE sets the transaction's timing.
4283 // v7.39 (round 308, V29) — and the names are kept.
4284 // They used to be skipped over on the way to the
4285 // DEFERRED keyword, so a named form silently behaved
4286 // as ALL: `SET CONSTRAINTS fk_a DEFERRED` deferred
4287 // every deferrable constraint in the transaction.
4288 let mut names: alloc::vec::Vec<alloc::string::String> =
4289 alloc::vec::Vec::new();
4290 if matches!(self.peek(), Token::All) {
4291 self.advance();
4292 } else {
4293 loop {
4294 let mut n = self.expect_ident_like()?;
4295 // A schema-qualified name (`public.fk_a`)
4296 // identifies the same constraint; PG resolves
4297 // it by the trailing segment.
4298 while matches!(self.peek(), Token::Dot) {
4299 self.advance();
4300 n = self.expect_ident_like()?;
4301 }
4302 names.push(n);
4303 if matches!(self.peek(), Token::Comma) {
4304 self.advance();
4305 } else {
4306 break;
4307 }
4308 }
4309 }
4310 let deferred = match self.peek() {
4311 Token::Ident(s) | Token::QuotedIdent(s)
4312 if s.eq_ignore_ascii_case("deferred") =>
4313 {
4314 true
4315 }
4316 Token::Ident(s) | Token::QuotedIdent(s)
4317 if s.eq_ignore_ascii_case("immediate") =>
4318 {
4319 false
4320 }
4321 other => {
4322 return Err(self.err(alloc::format!(
4323 "expected DEFERRED or IMMEDIATE after SET CONSTRAINTS, got {other:?}"
4324 )));
4325 }
4326 };
4327 self.advance();
4328 return Ok(Statement::SetConstraints { names, deferred });
4329 }
4330 // v7.16.2 — PG `SET [SESSION] AUTHORIZATION
4331 // { DEFAULT | '<role>' | <ident> }` (mailrs
4332 // round-10 A.1). pg_dump preamble emits the
4333 // `DEFAULT` form to reset session authorization.
4334 //
4335 // v7.39 (round 697) — this said "SPG has no role system so
4336 // this is a strict no-op". SPG has had one since round 58;
4337 // the comment outlived it, and with it the reason a name
4338 // that is not a role was accepted here. It still switches
4339 // no authorization — what it does now is refuse a role
4340 // that does not exist, as PG18 does. PG also accepts `RESET SESSION
4341 // AUTHORIZATION` (handled by the RESET parser
4342 // elsewhere). Reference:
4343 // <https://www.postgresql.org/docs/current/sql-set-session-authorization.html>
4344 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("authorization"))
4345 {
4346 self.advance(); // AUTHORIZATION
4347 match self.peek().clone() {
4348 Token::Default => {
4349 self.advance();
4350 }
4351 Token::String(r) | Token::Ident(r) | Token::QuotedIdent(r) => {
4352 self.advance();
4353 return Ok(Statement::ValidateOnly {
4354 kind: crate::ast::ValidateOnlyKind::RoleName,
4355 names: alloc::vec![r],
4356 });
4357 }
4358 other => {
4359 return Err(self.err(alloc::format!(
4360 "expected DEFAULT / '<role>' / <ident> after SET SESSION AUTHORIZATION, got {other:?}"
4361 )));
4362 }
4363 }
4364 return Ok(Statement::Empty);
4365 }
4366 // v7.38 轴 4 — `SET [SESSION] TRANSACTION
4367 // ISOLATION LEVEL { READ COMMITTED | READ
4368 // UNCOMMITTED | REPEATABLE READ | SERIALIZABLE }
4369 // [, READ {ONLY|WRITE}] [, [NOT] DEFERRABLE]`.
4370 // PG-standard surface. v7.37.8 accepts the syntax
4371 // and tracks the selected level on
4372 // `Engine::current_isolation_level()`; the actual
4373 // MVCC / SSI semantics implementation lands in
4374 // the 轴 4 isolation framework (separate train).
4375 // PG itself maps READ UNCOMMITTED to READ COMMITTED
4376 // internally; SPG behaves the same (effectively
4377 // READ COMMITTED at every level today).
4378 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("transaction"))
4379 {
4380 self.advance(); // TRANSACTION
4381 let modes = self.parse_isolation_level_clauses()?;
4382 return Ok(Statement::SetTransaction { modes });
4383 }
4384 // v7.14.0 — MySQL `SET CHARACTER SET <charset>`
4385 // alias — same accept-as-no-op as SET NAMES.
4386 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("character"))
4387 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("set"))
4388 {
4389 self.advance(); // CHARACTER
4390 self.advance(); // SET
4391 if matches!(
4392 self.peek(),
4393 Token::Ident(_) | Token::QuotedIdent(_) | Token::String(_)
4394 ) {
4395 self.advance();
4396 }
4397 return Ok(Statement::Empty);
4398 }
4399 // v7.39 (GUC) — PG spells the timezone GUC as two
4400 // keywords: `SET [LOCAL|SESSION] TIME ZONE <value>`,
4401 // where <value> is a string/ident or the LOCAL /
4402 // DEFAULT keyword (both mean "back to the default").
4403 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("time"))
4404 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("zone"))
4405 {
4406 self.advance(); // TIME
4407 self.advance(); // ZONE
4408 let value = match self.peek().clone() {
4409 Token::Ident(s)
4410 if s.eq_ignore_ascii_case("local")
4411 || s.eq_ignore_ascii_case("default") =>
4412 {
4413 self.advance();
4414 crate::ast::SetValue::Default
4415 }
4416 Token::Default => {
4417 self.advance();
4418 crate::ast::SetValue::Default
4419 }
4420 _ => self.parse_set_value()?,
4421 };
4422 return Ok(Statement::SetParameter {
4423 name: "timezone".into(),
4424 value,
4425 local: set_local,
4426 });
4427 }
4428 // v7.39 (round 430) — `SET @x = <expr> [, @y := <expr>]` is a
4429 // MySQL USER-variable assignment: its own per-session
4430 // namespace, an arbitrary expression on the right, and `:=`
4431 // as a second spelling of `=`. It used to fall into the
4432 // session-PARAMETER list below, whose values are literals and
4433 // whose store nothing reads back under a `@` name — so the
4434 // assignment reported success and vanished.
4435 //
4436 // A `@@`-prefixed LHS is a real engine setting and keeps the
4437 // old path.
4438 if matches!(self.peek(), Token::SessionVar(s) if !s.starts_with("@@")) {
4439 return self.parse_set_user_vars();
4440 }
4441 // v7.14.0 — multi-assignment form
4442 // `SET a = 1, b = 2, …`. Single-assignment is the
4443 // 1-element case. Each LHS may be a regular ident
4444 // or a SessionVar (`@VAR` / `@@VAR`).
4445 let mut pairs: Vec<(String, crate::ast::SetValue)> = Vec::new();
4446 loop {
4447 let lhs = match self.peek().clone() {
4448 Token::SessionVar(s) => {
4449 self.advance();
4450 s
4451 }
4452 Token::Ident(_) | Token::QuotedIdent(_) => self.parse_set_param_name()?,
4453 other => {
4454 return Err(self.err(format!(
4455 "expected parameter name after SET, got {other:?}"
4456 )));
4457 }
4458 };
4459 // Accept either `=` or the bare `TO` keyword.
4460 match self.peek() {
4461 Token::Eq => {
4462 self.advance();
4463 }
4464 Token::To => {
4465 self.advance();
4466 }
4467 other => {
4468 return Err(self.err(format!(
4469 "expected `=` or TO after SET {lhs}, got {other:?}"
4470 )));
4471 }
4472 }
4473 let mut value = self.parse_set_value()?;
4474 // v7.39 (GUC) — disambiguate the comma: `, name =` /
4475 // `, name TO` continues a MySQL-style multi-assign,
4476 // anything else is a PG list VALUE
4477 // (`SET search_path = myschema, public`) folded into
4478 // one comma-joined string.
4479 while matches!(self.peek(), Token::Comma) {
4480 let is_assign = matches!(
4481 self.tokens.get(self.pos + 1),
4482 Some(Token::Ident(_) | Token::QuotedIdent(_) | Token::SessionVar(_))
4483 ) && matches!(
4484 self.tokens.get(self.pos + 2),
4485 Some(Token::Eq | Token::To)
4486 );
4487 if is_assign {
4488 break;
4489 }
4490 self.advance(); // comma
4491 let next = self.parse_set_value()?;
4492 let joined = alloc::format!(
4493 "{}, {}",
4494 set_value_text(&value),
4495 set_value_text(&next)
4496 );
4497 value = crate::ast::SetValue::String(joined);
4498 }
4499 pairs.push((lhs, value));
4500 if matches!(self.peek(), Token::Comma) {
4501 self.advance();
4502 continue;
4503 }
4504 break;
4505 }
4506 if pairs.len() == 1 {
4507 let (name, value) = pairs.into_iter().next().unwrap();
4508 Ok(Statement::SetParameter {
4509 name,
4510 value,
4511 local: set_local,
4512 })
4513 } else {
4514 Ok(Statement::SetParameterList(pairs))
4515 }
4516 }
4517 // v7.12.1 — `RESET <name>` / `RESET ALL`.
4518 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("reset") => {
4519 self.advance();
4520 match self.peek().clone() {
4521 Token::All => {
4522 self.advance();
4523 Ok(Statement::ResetParameter(None))
4524 }
4525 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("all") => {
4526 self.advance();
4527 Ok(Statement::ResetParameter(None))
4528 }
4529 // v7.39 (RLS) — `RESET ROLE` clears the session role.
4530 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("role") => {
4531 self.advance();
4532 Ok(Statement::SetRole(None))
4533 }
4534 // 7.38.1 S5.2 — `RESET SESSION AUTHORIZATION`
4535 // (pg_dump's return from the owner switch).
4536 Token::Ident(s) | Token::QuotedIdent(s)
4537 if s.eq_ignore_ascii_case("session")
4538 && matches!(
4539 self.tokens.get(self.pos + 1),
4540 Some(Token::Ident(a) | Token::QuotedIdent(a))
4541 if a.eq_ignore_ascii_case("authorization")
4542 ) =>
4543 {
4544 self.advance(); // SESSION
4545 self.advance(); // AUTHORIZATION
4546 Ok(Statement::SetRole(None))
4547 }
4548 _ => {
4549 let name = self.parse_set_param_name()?;
4550 Ok(Statement::ResetParameter(Some(name)))
4551 }
4552 }
4553 }
4554 // v7.39 (round 218) — server-side cursors.
4555 Token::Ident(s) if s.eq_ignore_ascii_case("declare") => self.parse_declare_cursor(),
4556 Token::Ident(s) if s.eq_ignore_ascii_case("fetch") => self.parse_fetch_or_move(false),
4557 Token::Ident(s) if s.eq_ignore_ascii_case("move") => self.parse_fetch_or_move(true),
4558 Token::Ident(s) if s.eq_ignore_ascii_case("close") => {
4559 self.advance();
4560 match self.peek().clone() {
4561 Token::All => {
4562 self.advance();
4563 Ok(Statement::CloseCursor { name: None })
4564 }
4565 Token::Ident(s) if s.eq_ignore_ascii_case("all") => {
4566 self.advance();
4567 Ok(Statement::CloseCursor { name: None })
4568 }
4569 Token::Ident(n) | Token::QuotedIdent(n) => {
4570 self.advance();
4571 Ok(Statement::CloseCursor { name: Some(n) })
4572 }
4573 other => Err(self.err(format!(
4574 "expected cursor name or ALL after CLOSE, got {other:?}"
4575 ))),
4576 }
4577 }
4578 other => Err(self.err(format!(
4579 "expected SELECT / CREATE / DROP / INSERT / UPDATE / DELETE / ALTER / BEGIN / COMMIT / \
4580 ROLLBACK / SAVEPOINT / RELEASE / SHOW at start of statement, got {other:?}"
4581 ))),
4582 }
4583 }
4584
4585 /// v7.39 (round 218) — `DECLARE <name> [BINARY] [INSENSITIVE] [ASENSITIVE]
4586 /// [[NO] SCROLL] CURSOR [{WITH|WITHOUT} HOLD] FOR <select>`. BINARY /
4587 /// (IN|A)SENSITIVE are accepted and ignored (SPG cursors materialize at
4588 /// DECLARE, which is INSENSITIVE — PG's only actual behaviour too).
4589 fn parse_declare_cursor(&mut self) -> Result<Statement, ParseError> {
4590 self.advance(); // DECLARE
4591 let name = match self.advance() {
4592 Token::Ident(n) | Token::QuotedIdent(n) => n,
4593 other => {
4594 return Err(self.err(format!("expected cursor name after DECLARE, got {other:?}")));
4595 }
4596 };
4597 let mut scroll: Option<bool> = None;
4598 loop {
4599 match self.peek() {
4600 Token::Ident(s)
4601 if s.eq_ignore_ascii_case("binary")
4602 || s.eq_ignore_ascii_case("insensitive")
4603 || s.eq_ignore_ascii_case("asensitive") =>
4604 {
4605 self.advance();
4606 }
4607 Token::Ident(s) if s.eq_ignore_ascii_case("scroll") => {
4608 self.advance();
4609 scroll = Some(true);
4610 }
4611 Token::Not | Token::Ident(_) if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("no")) =>
4612 {
4613 self.advance(); // NO
4614 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("scroll")) {
4615 return Err(self.err(format!(
4616 "expected SCROLL after NO in DECLARE, got {:?}",
4617 self.peek()
4618 )));
4619 }
4620 self.advance();
4621 scroll = Some(false);
4622 }
4623 _ => break,
4624 }
4625 }
4626 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("cursor")) {
4627 return Err(self.err(format!("expected CURSOR in DECLARE, got {:?}", self.peek())));
4628 }
4629 self.advance();
4630 let mut hold = false;
4631 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with")) {
4632 self.advance();
4633 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("hold")) {
4634 return Err(self.err(format!(
4635 "expected HOLD after WITH in DECLARE, got {:?}",
4636 self.peek()
4637 )));
4638 }
4639 self.advance();
4640 hold = true;
4641 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("without")) {
4642 self.advance();
4643 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("hold")) {
4644 return Err(self.err(format!(
4645 "expected HOLD after WITHOUT in DECLARE, got {:?}",
4646 self.peek()
4647 )));
4648 }
4649 self.advance();
4650 }
4651 if !matches!(self.peek(), Token::For) {
4652 return Err(self.err(format!(
4653 "expected FOR before the cursor query, got {:?}",
4654 self.peek()
4655 )));
4656 }
4657 self.advance();
4658 let query = self.parse_one_statement()?;
4659 Ok(Statement::DeclareCursor {
4660 name,
4661 scroll,
4662 hold,
4663 query: alloc::boxed::Box::new(query),
4664 })
4665 }
4666
4667 /// v7.39 (round 218) — `FETCH`/`MOVE` `[<direction>] [FROM|IN] <name>`.
4668 /// Direction: NEXT | PRIOR | FIRST | LAST | ABSOLUTE n | RELATIVE n | n |
4669 /// ALL | FORWARD [n|ALL] | BACKWARD [n|ALL]; bare `FETCH <name>` = NEXT.
4670 fn parse_fetch_or_move(&mut self, is_move: bool) -> Result<Statement, ParseError> {
4671 use crate::ast::CursorDirection as D;
4672 self.advance(); // FETCH / MOVE
4673 let mut signed_count = |this: &mut Self| -> Result<i64, ParseError> {
4674 let neg = if matches!(this.peek(), Token::Minus) {
4675 this.advance();
4676 true
4677 } else {
4678 false
4679 };
4680 match this.advance() {
4681 Token::Integer(v) => Ok(if neg { -v } else { v }),
4682 other => Err(this.err(format!("expected count, got {other:?}"))),
4683 }
4684 };
4685 let direction = match self.peek().clone() {
4686 Token::Ident(s) if s.eq_ignore_ascii_case("next") => {
4687 self.advance();
4688 D::Next
4689 }
4690 Token::Ident(s) if s.eq_ignore_ascii_case("prior") => {
4691 self.advance();
4692 D::Prior
4693 }
4694 Token::Ident(s) if s.eq_ignore_ascii_case("first") => {
4695 self.advance();
4696 D::First
4697 }
4698 Token::Ident(s) if s.eq_ignore_ascii_case("last") => {
4699 self.advance();
4700 D::Last
4701 }
4702 Token::Ident(s) if s.eq_ignore_ascii_case("absolute") => {
4703 self.advance();
4704 D::Absolute(signed_count(self)?)
4705 }
4706 Token::Ident(s) if s.eq_ignore_ascii_case("relative") => {
4707 self.advance();
4708 D::Relative(signed_count(self)?)
4709 }
4710 Token::Ident(s) if s.eq_ignore_ascii_case("forward") => {
4711 self.advance();
4712 match self.peek().clone() {
4713 Token::All => {
4714 self.advance();
4715 D::All
4716 }
4717 Token::Ident(s) if s.eq_ignore_ascii_case("all") => {
4718 self.advance();
4719 D::All
4720 }
4721 Token::Integer(_) | Token::Minus => D::Count(signed_count(self)?),
4722 _ => D::Next, // bare FORWARD = FORWARD 1
4723 }
4724 }
4725 Token::Ident(s) if s.eq_ignore_ascii_case("backward") => {
4726 self.advance();
4727 match self.peek().clone() {
4728 Token::All => {
4729 self.advance();
4730 D::BackwardAll
4731 }
4732 Token::Ident(s) if s.eq_ignore_ascii_case("all") => {
4733 self.advance();
4734 D::BackwardAll
4735 }
4736 Token::Integer(_) | Token::Minus => D::Backward(signed_count(self)?),
4737 _ => D::Backward(1), // bare BACKWARD = BACKWARD 1
4738 }
4739 }
4740 Token::All => {
4741 self.advance();
4742 D::All
4743 }
4744 Token::Ident(s) if s.eq_ignore_ascii_case("all") => {
4745 self.advance();
4746 D::All
4747 }
4748 Token::Integer(_) | Token::Minus => D::Count(signed_count(self)?),
4749 // Bare `FETCH <name>` — direction defaults to NEXT.
4750 _ => D::Next,
4751 };
4752 // Optional FROM / IN.
4753 if matches!(self.peek(), Token::From)
4754 || matches!(self.peek(), Token::In)
4755 || matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("in"))
4756 {
4757 self.advance();
4758 }
4759 let name = match self.advance() {
4760 Token::Ident(n) | Token::QuotedIdent(n) => n,
4761 other => {
4762 return Err(self.err(format!("expected cursor name, got {other:?}")));
4763 }
4764 };
4765 Ok(if is_move {
4766 Statement::MoveCursor { name, direction }
4767 } else {
4768 Statement::FetchCursor { name, direction }
4769 })
4770 }
4771
4772 /// v7.39 (round 280) — `CREATE STATISTICS [IF NOT EXISTS] <name>
4773 /// [(kind, …)] ON <col>, … FROM <table>`.
4774 fn parse_create_statistics_after_create(&mut self) -> Result<Statement, ParseError> {
4775 self.advance(); // STATISTICS
4776 // `IF` / `EXISTS` lex as plain identifiers; only NOT is a keyword.
4777 let mut if_not_exists = false;
4778 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if"))
4779 && matches!(self.tokens.get(self.pos + 1), Some(Token::Not))
4780 {
4781 self.advance();
4782 self.advance();
4783 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("exists")) {
4784 self.advance();
4785 if_not_exists = true;
4786 }
4787 }
4788 let name = self.expect_ident_like()?;
4789 let mut kinds = Vec::new();
4790 if matches!(self.peek(), Token::LParen) {
4791 self.advance();
4792 loop {
4793 let k = self.expect_ident_like()?;
4794 // PG stores the single letters; accept the spelled-out
4795 // names the SQL uses and record what PG records.
4796 kinds.push(match k.to_ascii_lowercase().as_str() {
4797 "ndistinct" => String::from("d"),
4798 "dependencies" => String::from("f"),
4799 "mcv" => String::from("m"),
4800 other => {
4801 return Err(
4802 self.err(alloc::format!("unrecognized statistics kind \"{other}\""))
4803 );
4804 }
4805 });
4806 match self.advance() {
4807 Token::Comma => {}
4808 Token::RParen => break,
4809 other => {
4810 return Err(self.err(alloc::format!(
4811 "expected ',' or ')' in statistics kind list, got {other:?}"
4812 )));
4813 }
4814 }
4815 }
4816 }
4817 if !matches!(self.peek(), Token::On) {
4818 return Err(self.err(alloc::format!(
4819 "expected ON in CREATE STATISTICS, got {:?}",
4820 self.peek()
4821 )));
4822 }
4823 self.advance();
4824 let mut columns = Vec::new();
4825 loop {
4826 columns.push(self.expect_ident_like()?);
4827 if matches!(self.peek(), Token::Comma) {
4828 self.advance();
4829 } else {
4830 break;
4831 }
4832 }
4833 if !matches!(self.peek(), Token::From) {
4834 return Err(self.err(alloc::format!(
4835 "expected FROM in CREATE STATISTICS, got {:?}",
4836 self.peek()
4837 )));
4838 }
4839 self.advance();
4840 let table = self.expect_ident_like()?;
4841 Ok(Statement::CreateStatistics {
4842 name,
4843 if_not_exists,
4844 kinds,
4845 columns,
4846 table,
4847 })
4848 }
4849
4850 /// v7.39 (round 280) — `DROP STATISTICS [IF EXISTS] <name>`.
4851 /// v7.39 (round 436) — the body of `DROP TABLE [IF EXISTS] a[, b] …`,
4852 /// entered with the `TABLE` keyword still unconsumed. Extracted so
4853 /// `DROP TEMPORARY TABLE` (MySQL) runs the identical grammar instead of
4854 /// a second copy — the parser cannot rewind, so re-dispatch has to be a
4855 /// forward call.
4856 fn parse_drop_table_after_keyword(&mut self) -> Result<Statement, ParseError> {
4857 self.advance(); // TABLE
4858 let if_exists = self.consume_if_exists();
4859 let mut names: Vec<String> = Vec::new();
4860 loop {
4861 names.push(self.expect_ident_like()?);
4862 if matches!(self.peek(), Token::Comma) {
4863 self.advance();
4864 continue;
4865 }
4866 break;
4867 }
4868 if matches!(
4869 self.peek(),
4870 Token::Ident(s) if s.eq_ignore_ascii_case("cascade")
4871 || s.eq_ignore_ascii_case("restrict")
4872 ) {
4873 self.advance();
4874 }
4875 Ok(Statement::DropTable { names, if_exists })
4876 }
4877
4878 fn parse_drop_statistics_after_drop(&mut self) -> Result<Statement, ParseError> {
4879 self.advance(); // STATISTICS
4880 let mut if_exists = false;
4881 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if"))
4882 && matches!(self.tokens.get(self.pos + 1),
4883 Some(Token::Ident(e)) if e.eq_ignore_ascii_case("exists"))
4884 {
4885 self.advance();
4886 self.advance();
4887 if_exists = true;
4888 }
4889 let name = self.expect_ident_like()?;
4890 Ok(Statement::DropStatistics { name, if_exists })
4891 }
4892
4893 fn parse_create_stmt(&mut self) -> Result<Statement, ParseError> {
4894 debug_assert!(matches!(self.peek(), Token::Create));
4895 self.advance();
4896 match self.peek() {
4897 Token::Table => self.parse_create_table_stmt_after_create(),
4898 Token::Index => self.parse_create_index_stmt_after_create(false),
4899 // v7.39 (round 280) — CREATE STATISTICS is a real catalog
4900 // object now. It used to be consumed by the CREATE-noise
4901 // arm, so a pg_dump that declares extended statistics
4902 // restored silently without them and reflection showed
4903 // nothing.
4904 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("statistics") => {
4905 self.parse_create_statistics_after_create()
4906 }
4907 // v7.9.29 — `CREATE UNIQUE INDEX … [WHERE pred]`.
4908 // The `UNIQUE` modifier turns a partial index into a
4909 // partial-uniqueness invariant (only rows matching the
4910 // WHERE predicate are checked for duplicates). mailrs
4911 // K1 (3 hits: email_templates default, calendar_events
4912 // master, calendar_events instance).
4913 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("unique") => {
4914 self.advance();
4915 if !matches!(self.peek(), Token::Index) {
4916 return Err(self.err(alloc::format!(
4917 "expected INDEX after CREATE UNIQUE, got {:?}",
4918 self.peek()
4919 )));
4920 }
4921 self.parse_create_index_stmt_after_create(true)
4922 }
4923 Token::Publication => {
4924 self.advance();
4925 self.parse_create_publication_after_keyword()
4926 }
4927 Token::Subscription => {
4928 self.advance();
4929 self.parse_create_subscription_after_keyword()
4930 }
4931 // v4.1: CREATE USER 'name' WITH PASSWORD 'pw' [ROLE 'role'].
4932 // USER isn't a reserved keyword — we look for the bare
4933 // identifier so the lexer doesn't have to grow a token.
4934 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("user") => {
4935 self.advance();
4936 self.parse_create_user_after_keyword(true)
4937 }
4938 // v7.39 (read01 round 58) — `CREATE ROLE name [WITH] [options]`.
4939 // PG's CREATE USER *is* CREATE ROLE … LOGIN; the only difference is
4940 // the default of the LOGIN attribute.
4941 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("role") => {
4942 self.advance();
4943 self.parse_create_user_after_keyword(false)
4944 }
4945 // v7.39 (RLS) — `CREATE POLICY name ON table …`.
4946 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("policy") => {
4947 self.advance();
4948 self.parse_create_policy_after_keyword()
4949 }
4950 // v7.9.15 — `CREATE EXTENSION [IF NOT EXISTS] <name>
4951 // [WITH SCHEMA …] [VERSION '…'] [CASCADE]` as a
4952 // no-op. mailrs follow-up F3.
4953 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("extension") => {
4954 self.advance();
4955 self.parse_create_extension_after_keyword()
4956 }
4957 // v7.12.4 — `CREATE [OR REPLACE] FUNCTION …` and
4958 // `CREATE [OR REPLACE] TRIGGER …`. `OR REPLACE` is
4959 // optional; absorb it here and forward to the
4960 // per-kind parsers with the flag. OR is a reserved
4961 // keyword token.
4962 Token::Or => {
4963 self.advance();
4964 let next = self.peek();
4965 let (Token::Ident(s2) | Token::QuotedIdent(s2)) = next else {
4966 return Err(self.err(alloc::format!(
4967 "expected REPLACE after CREATE OR, got {next:?}"
4968 )));
4969 };
4970 if !s2.eq_ignore_ascii_case("replace") {
4971 return Err(self.err(alloc::format!(
4972 "expected REPLACE after CREATE OR, got {s2:?}"
4973 )));
4974 }
4975 self.advance();
4976 self.parse_create_function_or_trigger_after_or_replace(true)
4977 }
4978 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("function") => {
4979 self.advance();
4980 self.parse_create_function_after_keyword(false)
4981 }
4982 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("trigger") => {
4983 self.advance();
4984 self.parse_create_trigger_after_keyword(false)
4985 }
4986 // v7.39 (round 139) — CREATE RULE …
4987 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("rule") => {
4988 self.advance();
4989 self.parse_create_rule_after_keyword(false)
4990 }
4991 // v7.39 (read01 round 82) — CREATE CONSTRAINT TRIGGER. A constraint
4992 // trigger is a row-level AFTER trigger that additionally carries
4993 // DEFERRABLE / INITIALLY DEFERRED timing; the `parse_create_trigger`
4994 // path already tolerates and skips those clauses, so consuming the
4995 // CONSTRAINT keyword and reusing it makes the statement parse and the
4996 // trigger fire. (The deferral timing itself is not yet honoured —
4997 // SPG fires it as a plain AFTER trigger, which is correct behaviour
4998 // for every non-deferred use.)
4999 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("constraint") => {
5000 self.advance();
5001 if !matches!(self.peek(), Token::Ident(t) | Token::QuotedIdent(t)
5002 if t.eq_ignore_ascii_case("trigger"))
5003 {
5004 return Err(self.err(alloc::format!(
5005 "expected TRIGGER after CREATE CONSTRAINT, got {:?}",
5006 self.peek()
5007 )));
5008 }
5009 self.advance();
5010 self.parse_create_trigger_after_keyword(false)
5011 }
5012 // v7.17.0 — CREATE [TEMPORARY] SEQUENCE …
5013 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("sequence") => {
5014 self.advance();
5015 self.parse_create_sequence_after_keyword(false)
5016 }
5017 // v7.17.0 Phase 1.2 — CREATE [TEMPORARY] VIEW …
5018 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("view") => {
5019 self.advance();
5020 self.parse_create_view_after_keyword(false, false, false)
5021 }
5022 // v7.17.0 Phase 2.6 — MySQL view prefix clauses
5023 // `ALGORITHM = {UNDEFINED|MERGE|TEMPTABLE}` /
5024 // `DEFINER = <user>` / `SQL SECURITY {DEFINER|INVOKER}`
5025 // appear (in any order) between `CREATE` and `VIEW` in
5026 // every mysqldump-emitted view. Pre-2.6 the parser
5027 // rejected the prefix and the customer's whole view
5028 // backup failed on the first view. The hints are pure
5029 // planner / permission metadata; SPG's view-rewrite
5030 // path is semantically equivalent for all three
5031 // algorithms in v7.17 (TEMPTABLE differs only in
5032 // perf for huge views — out of v7.17 scope), and
5033 // DEFINER / SQL SECURITY are pure single-user
5034 // permissioning that SPG ignores by design.
5035 Token::Ident(s) | Token::QuotedIdent(s)
5036 if s.eq_ignore_ascii_case("algorithm")
5037 || s.eq_ignore_ascii_case("definer")
5038 || s.eq_ignore_ascii_case("sql") =>
5039 {
5040 self.consume_mysql_view_prefix()?;
5041 // After absorbing ALGORITHM / DEFINER / SQL SECURITY
5042 // (in any order, in any combination), the next
5043 // keyword must be VIEW. mysqldump never emits these
5044 // prefixes on non-view statements.
5045 let next = self.peek().clone();
5046 if matches!(&next, Token::Ident(s2) | Token::QuotedIdent(s2)
5047 if s2.eq_ignore_ascii_case("view"))
5048 {
5049 self.advance();
5050 self.parse_create_view_after_keyword(false, false, false)
5051 } else {
5052 Err(self.err(alloc::format!(
5053 "expected VIEW after MySQL view prefix (ALGORITHM/DEFINER/SQL SECURITY), got {next:?}"
5054 )))
5055 }
5056 }
5057 // v7.17.0 Phase 1.4 — CREATE TYPE name AS ENUM (…).
5058 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("type") => {
5059 self.advance();
5060 self.parse_create_type_after_keyword()
5061 }
5062 // v7.17.0 Phase 1.5 — CREATE DOMAIN name AS base
5063 // [DEFAULT expr] [NOT NULL] [CHECK (expr)]*.
5064 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("domain") => {
5065 self.advance();
5066 self.parse_create_domain_after_keyword()
5067 }
5068 // v7.17.0 Phase 1.6 — CREATE SCHEMA [IF NOT EXISTS]
5069 // name [AUTHORIZATION user]. Real catalog registry
5070 // (was silent-no-op'd pre-v7.17).
5071 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("schema") => {
5072 self.advance();
5073 let if_not_exists = self.parse_if_not_exists();
5074 let name = self.expect_ident_like()?;
5075 // Optional `AUTHORIZATION <user>` trailer — accepted,
5076 // ignored (single-user catalog).
5077 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
5078 if s.eq_ignore_ascii_case("authorization"))
5079 {
5080 self.advance();
5081 let _ = self.expect_ident_like()?;
5082 }
5083 Ok(Statement::CreateSchema { name, if_not_exists })
5084 }
5085 // v7.17.0 Phase 1.3 — CREATE MATERIALIZED VIEW …
5086 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("materialized") => {
5087 self.advance();
5088 let next = self.peek().clone();
5089 if matches!(&next, Token::Ident(s2) | Token::QuotedIdent(s2) if s2.eq_ignore_ascii_case("view"))
5090 {
5091 self.advance();
5092 self.parse_create_materialized_view_after_keyword()
5093 } else {
5094 Err(self.err(alloc::format!(
5095 "expected VIEW after CREATE MATERIALIZED, got {next:?}"
5096 )))
5097 }
5098 }
5099 // v7.38 (read01 P6.57) — CREATE UNLOGGED TABLE. Unlike TEMP (a
5100 // no-op below), an UNLOGGED table is a real, fully-usable table in
5101 // PG — it only skips WAL. SPG creates a normal table (the WAL-skip
5102 // durability optimisation is a follow-up), so a dump / app that
5103 // declares UNLOGGED tables works instead of failing to parse.
5104 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("unlogged") => {
5105 self.advance(); // UNLOGGED
5106 if matches!(self.peek(), Token::Table) {
5107 self.parse_create_table_stmt_after_create()
5108 } else {
5109 Err(self.err(format!(
5110 "expected TABLE after CREATE UNLOGGED, got {:?}",
5111 self.peek()
5112 )))
5113 }
5114 }
5115 Token::Ident(s) | Token::QuotedIdent(s)
5116 if s.eq_ignore_ascii_case("temporary") || s.eq_ignore_ascii_case("temp") =>
5117 {
5118 self.advance();
5119 // TEMPORARY/TEMP followed by SEQUENCE / VIEW.
5120 let next = self.peek().clone();
5121 if matches!(&next, Token::Ident(s2) | Token::QuotedIdent(s2) if s2.eq_ignore_ascii_case("sequence"))
5122 {
5123 self.advance();
5124 self.parse_create_sequence_after_keyword(true)
5125 } else if matches!(&next, Token::Ident(s2) | Token::QuotedIdent(s2) if s2.eq_ignore_ascii_case("view"))
5126 {
5127 self.advance();
5128 self.parse_create_view_after_keyword(false, false, true)
5129 } else {
5130 // v7.39 (round 436) — `CREATE TEMPORARY TABLE` used to be
5131 // consumed and answered OK while creating nothing, so
5132 // every statement that touched the table afterwards failed
5133 // with "table not found" — the DDL itself lied. It is a
5134 // real CREATE TABLE now, marked temporary so the executor
5135 // puts it in the session's own namespace. An optional
5136 // TABLE keyword may or may not be present (`CREATE TEMP t`
5137 // is not legal, but the keyword is consumed by the
5138 // CREATE TABLE parser itself).
5139 let stmt = self.parse_create_table_stmt_after_create()?;
5140 match stmt {
5141 Statement::CreateTable(mut c) => {
5142 c.temporary = true;
5143 Ok(Statement::CreateTable(c))
5144 }
5145 // `CREATE TEMPORARY TABLE x AS <select>` lowers to the
5146 // CTAS node, which needs the same session namespace.
5147 Statement::CreateMaterializedView(mut m) if m.as_plain_table => {
5148 m.temporary = true;
5149 Ok(Statement::CreateMaterializedView(m))
5150 }
5151 other => Ok(other),
5152 }
5153 }
5154 }
5155 // v7.17.0 Phase 4.2 — MySQL `CREATE PROCEDURE name (…)
5156 // BEGIN <body> END`. The body may reference `@var`
5157 // session variables, SET statements, internal `;`
5158 // terminators, etc. SPG has no procedure runtime, so
5159 // consume the whole `CREATE PROCEDURE … END` block as
5160 // a no-op so mysqldump scripts that include stored
5161 // routines load through. The matching-END consumer
5162 // tracks BEGIN/END nesting depth to handle nested
5163 // BEGIN blocks correctly.
5164 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("procedure") => {
5165 self.consume_mysql_routine_body();
5166 Ok(Statement::Empty)
5167 }
5168 // v7.14.0 — pg_dump / mysqldump emit
5169 // `CREATE SCHEMA / VIEW / MATERIALIZED VIEW /
5170 // TYPE / DOMAIN / DATABASE / ROLE / POLICY / OPERATOR`.
5171 // SPG is single-schema / single-database; these have
5172 // no behavioural effect, so consume + return Empty.
5173 // v7.17.0 NOTE: SEQUENCE / VIEW / MATERIALIZED VIEW /
5174 // TYPE / DOMAIN / SCHEMA were here pre-v7.17; all
5175 // moved up to real parser branches. DATABASE / ROLE /
5176 // POLICY / OPERATOR stay no-op forever
5177 // (single-database, hardcoded roles).
5178 Token::Ident(s) | Token::QuotedIdent(s)
5179 if matches!(
5180 s.to_ascii_lowercase().as_str(),
5181 "database"
5182 | "role"
5183 | "operator"
5184 | "cast"
5185 | "aggregate"
5186 | "language"
5187 | "collation"
5188 | "conversion"
5189 // v7.17.0 Phase 8 (audit N6) — rarely-
5190 // emitted pg_dump shapes that should
5191 // load through without a parser error.
5192 // SPG has no planner statistics catalog,
5193 // no event-trigger hooks, no foreign-
5194 // data-wrapper infrastructure; consume
5195 // + return Empty.
5196 | "statistics"
5197 | "event"
5198 // v7.37.17 (17.6 siblings) — additional CREATE
5199 // targets pg_dump / operator install scripts
5200 // may emit that SPG has no matching machinery
5201 // for. Consume + Empty-return.
5202 | "text"
5203 | "tablespace"
5204 | "access"
5205 | "large"
5206 ) =>
5207 {
5208 // DATABASE is the one member of this list PG refuses
5209 // inside a transaction block; the rest (ROLE, CAST,
5210 // TABLESPACE, …) it runs there quite happily, so only
5211 // this one is named. Still a no-op otherwise — SPG is
5212 // single-database.
5213 let is_database = s.eq_ignore_ascii_case("database");
5214 // The name is the first token after DATABASE, past an
5215 // `IF NOT EXISTS`.
5216 let name = if is_database {
5217 self.scan_database_name()
5218 } else {
5219 None
5220 };
5221 let collation = if is_database {
5222 self.scan_database_collation_until_boundary()
5223 } else {
5224 self.consume_until_statement_boundary();
5225 None
5226 };
5227 if is_database {
5228 return Ok(Statement::NoOpPreventedInTransaction {
5229 what: String::from("CREATE DATABASE"),
5230 collation,
5231 name,
5232 });
5233 }
5234 Ok(Statement::Empty)
5235 }
5236 // v7.39 (round 706) — the foreign-data family leaves the silent
5237 // list: `CREATE SERVER …`, `CREATE FOREIGN TABLE …`, `CREATE
5238 // FOREIGN DATA WRAPPER …` are still consumed whole (SPG has no
5239 // FDW machinery), but the ENGINE now warns, so a restore log
5240 // says what will not function instead of reporting success.
5241 Token::Ident(s) | Token::QuotedIdent(s)
5242 if s.eq_ignore_ascii_case("server") || s.eq_ignore_ascii_case("foreign") =>
5243 {
5244 self.consume_until_statement_boundary();
5245 Ok(Statement::ValidateOnly {
5246 kind: crate::ast::ValidateOnlyKind::ForeignInfra,
5247 names: Vec::new(),
5248 })
5249 }
5250 other => Err(self.err(format!(
5251 "expected TABLE / INDEX / USER / EXTENSION / PUBLICATION / SUBSCRIPTION / FUNCTION / TRIGGER / SEQUENCE / SCHEMA / VIEW / TYPE / DOMAIN [OR REPLACE …] after CREATE, got {other:?}"
5252 ))),
5253 }
5254 }
5255
5256 /// v7.12.4 — `CREATE OR REPLACE` already consumed; the next
5257 /// keyword decides whether we parse a function or trigger
5258 /// body. PG accepts other `OR REPLACE`-able objects (VIEW,
5259 /// PROCEDURE) — those land in later releases.
5260 fn parse_create_function_or_trigger_after_or_replace(
5261 &mut self,
5262 or_replace: bool,
5263 ) -> Result<Statement, ParseError> {
5264 let tok = self.peek();
5265 let (Token::Ident(s) | Token::QuotedIdent(s)) = tok else {
5266 return Err(self.err(alloc::format!(
5267 "expected FUNCTION / TRIGGER / RULE / VIEW after CREATE OR REPLACE, got {tok:?}"
5268 )));
5269 };
5270 if s.eq_ignore_ascii_case("function") {
5271 self.advance();
5272 self.parse_create_function_after_keyword(or_replace)
5273 } else if s.eq_ignore_ascii_case("trigger") {
5274 self.advance();
5275 self.parse_create_trigger_after_keyword(or_replace)
5276 } else if s.eq_ignore_ascii_case("rule") {
5277 // v7.39 (round 143) — CREATE OR REPLACE RULE name AS ON …
5278 self.advance();
5279 self.parse_create_rule_after_keyword(or_replace)
5280 } else if s.eq_ignore_ascii_case("view") {
5281 // v7.17.0 Phase 1.2 — CREATE OR REPLACE VIEW name AS SELECT …
5282 self.advance();
5283 self.parse_create_view_after_keyword(or_replace, false, false)
5284 } else if s.eq_ignore_ascii_case("temporary") || s.eq_ignore_ascii_case("temp") {
5285 // CREATE OR REPLACE TEMPORARY VIEW … (rare but legal).
5286 self.advance();
5287 let nxt = self.peek().clone();
5288 if matches!(&nxt, Token::Ident(n) | Token::QuotedIdent(n) if n.eq_ignore_ascii_case("view"))
5289 {
5290 self.advance();
5291 self.parse_create_view_after_keyword(or_replace, false, true)
5292 } else {
5293 Err(self.err(alloc::format!(
5294 "expected VIEW after CREATE OR REPLACE TEMPORARY, got {nxt:?}"
5295 )))
5296 }
5297 } else {
5298 Err(self.err(alloc::format!(
5299 "expected FUNCTION / TRIGGER / RULE / VIEW after CREATE OR REPLACE, got {s:?}"
5300 )))
5301 }
5302 }
5303
5304 /// v7.9.15 — accept and discard `CREATE EXTENSION` DDL.
5305 /// SPG doesn't have a registry; pgvector / similar are
5306 /// either builtin (VECTOR(N) ↔ pgvector) or n/a. Parsing
5307 /// the syntax lets dual-target schemas keep the line.
5308 fn parse_create_extension_after_keyword(&mut self) -> Result<Statement, ParseError> {
5309 // Optional `IF NOT EXISTS`.
5310 self.consume_if_not_exists();
5311 let name = self.expect_ident_like()?;
5312 // Drain optional WITH SCHEMA <ident> / VERSION '<v>' /
5313 // CASCADE / FROM '<v>' clauses; we don't model them.
5314 loop {
5315 match self.peek() {
5316 Token::Ident(s) if s.eq_ignore_ascii_case("with") => {
5317 self.advance();
5318 continue;
5319 }
5320 Token::Ident(s) if s.eq_ignore_ascii_case("schema") => {
5321 self.advance();
5322 let _ = self.expect_ident_like()?;
5323 continue;
5324 }
5325 Token::Ident(s) if s.eq_ignore_ascii_case("version") => {
5326 self.advance();
5327 // String or ident literal.
5328 let _ = self.advance();
5329 continue;
5330 }
5331 Token::Ident(s) if s.eq_ignore_ascii_case("from") => {
5332 self.advance();
5333 let _ = self.advance();
5334 continue;
5335 }
5336 Token::Ident(s) if s.eq_ignore_ascii_case("cascade") => {
5337 self.advance();
5338 continue;
5339 }
5340 _ => break,
5341 }
5342 }
5343 // v7.39 (round 697) — the NAME is checked now. `CREATE EXTENSION
5344 // nosuch` reported success and `pg_extension` then did not list it,
5345 // which is the accept-and-do-nothing shape F31 exists to find.
5346 Ok(Statement::ValidateOnly {
5347 kind: crate::ast::ValidateOnlyKind::ExtensionAvailable,
5348 names: alloc::vec![name],
5349 })
5350 }
5351
5352 /// v7.12.4 — body of `CREATE [OR REPLACE] FUNCTION`. The
5353 /// `[OR REPLACE]` flag (and the `FUNCTION` keyword) have
5354 /// already been consumed by the caller. Grammar accepted:
5355 ///
5356 /// name `(` arg-list `)`
5357 /// `RETURNS` return-type
5358 /// [ `LANGUAGE` ident ]
5359 /// `AS` $$ body $$
5360 /// [ `LANGUAGE` ident ]
5361 ///
5362 /// Either `LANGUAGE` position is allowed; PG accepts both.
5363 fn parse_create_function_after_keyword(
5364 &mut self,
5365 or_replace: bool,
5366 ) -> Result<Statement, ParseError> {
5367 let name = self.expect_ident_like()?;
5368 // Argument list. v7.12.4 commonly sees the empty `()`
5369 // (trigger functions); typed args parse and round-trip
5370 // but the executor only invokes nullary functions.
5371 if !matches!(self.peek(), Token::LParen) {
5372 return Err(self.err(alloc::format!(
5373 "expected '(' after function name {name:?}, got {:?}",
5374 self.peek()
5375 )));
5376 }
5377 self.advance();
5378 let args = self.parse_function_arg_list()?;
5379 // RETURNS clause.
5380 let tok = self.peek();
5381 let (Token::Ident(s) | Token::QuotedIdent(s)) = tok else {
5382 return Err(self.err(alloc::format!(
5383 "expected RETURNS after function arg list, got {tok:?}"
5384 )));
5385 };
5386 if !s.eq_ignore_ascii_case("returns") {
5387 return Err(self.err(alloc::format!(
5388 "expected RETURNS after function arg list, got {s:?}"
5389 )));
5390 }
5391 self.advance();
5392 let returns = self.parse_function_return()?;
5393 // Optional LANGUAGE clause (PG also accepts after AS — we'll
5394 // re-check after the body too).
5395 let mut language: Option<String> = self.parse_optional_language()?;
5396 // v7.39 (round 322, V46) — attribute clauses. PG allows them on
5397 // either side of the body and in any order, interleaved with
5398 // LANGUAGE; `CREATE FUNCTION f() RETURNS int LANGUAGE sql
5399 // IMMUTABLE STRICT AS $$…$$` used to be a parse error, which meant
5400 // PG's own pg_dump output did not restore.
5401 let mut attrs = FunctionAttrs::default();
5402 loop {
5403 let before = self.pos;
5404 self.parse_function_attrs_into(&mut attrs)?;
5405 if language.is_none() {
5406 language = self.parse_optional_language()?;
5407 }
5408 if self.pos == before {
5409 break;
5410 }
5411 }
5412 // `AS` followed by a $$-quoted body (lexer already
5413 // collapses both `$$…$$` and `$tag$…$tag$` to a single
5414 // Token::String). AS is a reserved keyword (Token::As).
5415 if !matches!(self.peek(), Token::As) {
5416 return Err(self.err(alloc::format!(
5417 "expected AS before function body, got {:?}",
5418 self.peek()
5419 )));
5420 }
5421 self.advance();
5422 let body_text = match self.peek() {
5423 Token::String(s) => {
5424 let body = s.clone();
5425 self.advance();
5426 body
5427 }
5428 other => {
5429 return Err(self.err(alloc::format!(
5430 "expected $$-quoted function body after AS, got {other:?}"
5431 )));
5432 }
5433 };
5434 // Trailing clauses — PG's other accepted position for both the
5435 // LANGUAGE and the attributes.
5436 loop {
5437 let before = self.pos;
5438 self.parse_function_attrs_into(&mut attrs)?;
5439 if language.is_none() {
5440 language = self.parse_optional_language()?;
5441 }
5442 if self.pos == before {
5443 break;
5444 }
5445 }
5446 let language = language.unwrap_or_else(|| String::from("sql"));
5447 // PL/pgSQL bodies get structure-parsed. Other languages
5448 // (or PL/pgSQL bodies the v7.12.4 parser doesn't yet
5449 // recognise) round-trip as Raw text — the executor errors
5450 // when invoked with a clear unsupported message.
5451 let body = if language.eq_ignore_ascii_case("plpgsql") {
5452 match parse_plpgsql_body(&body_text) {
5453 Ok(block) => FunctionBody::PlPgSql(block),
5454 // Best-effort: if the body parser doesn't yet
5455 // support a construct used inside, fall back to
5456 // raw — keeps `CREATE FUNCTION` itself working
5457 // (catalogue accepts), executor errors on
5458 // invocation only.
5459 Err(_) => FunctionBody::Raw(body_text),
5460 }
5461 } else {
5462 FunctionBody::Raw(body_text)
5463 };
5464 Ok(Statement::CreateFunction(CreateFunctionStatement {
5465 name,
5466 or_replace,
5467 args,
5468 returns,
5469 language,
5470 body,
5471 attrs,
5472 }))
5473 }
5474
5475 /// v7.39 (round 322, V46) — consume any run of `CREATE FUNCTION`
5476 /// attribute clauses into `attrs`, stopping at the first token that
5477 /// is not one. Measured against PG 18.4, which accepts them in any
5478 /// order and on either side of the body.
5479 fn parse_function_attrs_into(&mut self, attrs: &mut FunctionAttrs) -> Result<(), ParseError> {
5480 loop {
5481 let word = match self.peek() {
5482 Token::Ident(w) | Token::QuotedIdent(w) => w.to_ascii_lowercase(),
5483 // NOT LEAKPROOF — NOT is a reserved keyword token.
5484 Token::Not
5485 if matches!(
5486 self.tokens.get(self.pos + 1),
5487 Some(Token::Ident(w)) if w.eq_ignore_ascii_case("leakproof")
5488 ) =>
5489 {
5490 self.advance();
5491 self.advance();
5492 attrs.leakproof = false;
5493 continue;
5494 }
5495 _ => return Ok(()),
5496 };
5497 match word.as_str() {
5498 "immutable" => {
5499 self.advance();
5500 attrs.volatility = FunctionVolatility::Immutable;
5501 }
5502 "stable" => {
5503 self.advance();
5504 attrs.volatility = FunctionVolatility::Stable;
5505 }
5506 "volatile" => {
5507 self.advance();
5508 attrs.volatility = FunctionVolatility::Volatile;
5509 }
5510 "strict" => {
5511 self.advance();
5512 attrs.strict = true;
5513 }
5514 "leakproof" => {
5515 self.advance();
5516 attrs.leakproof = true;
5517 }
5518 // RETURNS NULL ON NULL INPUT / CALLED ON NULL INPUT — the
5519 // spelled-out forms of STRICT and its opposite.
5520 "returns" | "called" => {
5521 let strict = word == "returns";
5522 let mut probe = self.pos + 1;
5523 if strict {
5524 // RETURNS *NULL* ON NULL INPUT; a bare RETURNS here
5525 // is not ours.
5526 match self.tokens.get(probe) {
5527 Some(Token::Null) => probe += 1,
5528 Some(Token::Ident(w)) if w.eq_ignore_ascii_case("null") => probe += 1,
5529 _ => return Ok(()),
5530 }
5531 }
5532 let ok = matches!(self.tokens.get(probe), Some(Token::On))
5533 || matches!(self.tokens.get(probe), Some(Token::Ident(w)) if w.eq_ignore_ascii_case("on"));
5534 if !ok {
5535 return Ok(());
5536 }
5537 probe += 1;
5538 match self.tokens.get(probe) {
5539 Some(Token::Null) => probe += 1,
5540 Some(Token::Ident(w)) if w.eq_ignore_ascii_case("null") => probe += 1,
5541 _ => return Ok(()),
5542 }
5543 match self.tokens.get(probe) {
5544 Some(Token::Ident(w)) if w.eq_ignore_ascii_case("input") => probe += 1,
5545 _ => return Ok(()),
5546 }
5547 self.pos = probe;
5548 attrs.strict = strict;
5549 }
5550 "security" | "external" => {
5551 // [EXTERNAL] SECURITY { INVOKER | DEFINER }
5552 let mut probe = self.pos + 1;
5553 if word == "external" {
5554 match self.tokens.get(probe) {
5555 Some(Token::Ident(w)) if w.eq_ignore_ascii_case("security") => {
5556 probe += 1;
5557 }
5558 _ => return Ok(()),
5559 }
5560 }
5561 let definer = match self.tokens.get(probe) {
5562 Some(Token::Ident(w)) if w.eq_ignore_ascii_case("definer") => true,
5563 Some(Token::Ident(w)) if w.eq_ignore_ascii_case("invoker") => false,
5564 _ => return Ok(()),
5565 };
5566 self.pos = probe + 1;
5567 attrs.security_definer = definer;
5568 }
5569 "parallel" => {
5570 let level = match self.tokens.get(self.pos + 1) {
5571 Some(Token::Ident(w)) if w.eq_ignore_ascii_case("safe") => {
5572 FunctionParallel::Safe
5573 }
5574 Some(Token::Ident(w)) if w.eq_ignore_ascii_case("restricted") => {
5575 FunctionParallel::Restricted
5576 }
5577 Some(Token::Ident(w)) if w.eq_ignore_ascii_case("unsafe") => {
5578 FunctionParallel::Unsafe
5579 }
5580 _ => return Ok(()),
5581 };
5582 self.pos += 2;
5583 attrs.parallel = level;
5584 }
5585 "cost" | "rows" => {
5586 let Some(n) = self.peek_number_at(self.pos + 1) else {
5587 return Ok(());
5588 };
5589 self.pos += 2;
5590 if word == "cost" {
5591 attrs.cost = Some(n);
5592 } else {
5593 attrs.rows = Some(n);
5594 }
5595 }
5596 _ => return Ok(()),
5597 }
5598 }
5599 }
5600
5601 /// The numeric literal at `idx`, if there is one.
5602 fn peek_number_at(&self, idx: usize) -> Option<f64> {
5603 match self.tokens.get(idx)? {
5604 Token::Integer(n) => Some(*n as f64),
5605 Token::Float(f) => Some(*f),
5606 Token::Numeric(t) => t.parse::<f64>().ok(),
5607 _ => None,
5608 }
5609 }
5610
5611 /// Closing `)`-terminated argument list. v7.12.4 commonly
5612 /// sees the empty `()`; typed args round-trip but the
5613 /// executor (yet) doesn't invoke them.
5614 /// v7.39 (round 344) — consume a `( n [, m] )` type modifier and throw
5615 /// it away, which is what PG does with one on a function parameter.
5616 fn skip_type_modifier(&mut self) {
5617 if !matches!(self.peek(), Token::LParen) {
5618 return;
5619 }
5620 // Only a numeric modifier — anything else is not one, and eating
5621 // it would swallow real grammar.
5622 let mut i = self.pos + 1;
5623 let mut seen_number = false;
5624 loop {
5625 match self.tokens.get(i) {
5626 Some(Token::Integer(_)) => seen_number = true,
5627 Some(Token::Comma) => {}
5628 Some(Token::RParen) => break,
5629 _ => return,
5630 }
5631 i += 1;
5632 }
5633 if !seen_number {
5634 return;
5635 }
5636 while self.pos <= i {
5637 self.advance();
5638 }
5639 }
5640
5641 fn parse_function_arg_list(&mut self) -> Result<Vec<FunctionArg>, ParseError> {
5642 let mut args: Vec<FunctionArg> = Vec::new();
5643 if matches!(self.peek(), Token::RParen) {
5644 self.advance();
5645 return Ok(args);
5646 }
5647 loop {
5648 // Optional `IN` / `OUT` / `INOUT` mode keyword. IN is
5649 // a reserved token; OUT / INOUT are bare idents.
5650 let mode = if matches!(self.peek(), Token::In) {
5651 self.advance();
5652 FunctionArgMode::In
5653 } else if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("out"))
5654 {
5655 self.advance();
5656 FunctionArgMode::Out
5657 } else if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("inout"))
5658 {
5659 self.advance();
5660 FunctionArgMode::InOut
5661 } else {
5662 FunctionArgMode::In
5663 };
5664 // Optional name. The next token is either a name
5665 // (followed by a type ident) or the type itself.
5666 // Disambiguate by peeking ahead: if the token after
5667 // the next ident is also an ident, we treat the
5668 // first as the name.
5669 // v7.39 (round 315, V19) — take EVERY ident-like word up to
5670 // the comma or paren, then decide. Reading at most two of
5671 // them could not spell `x double precision` at all, and
5672 // silently mis-read the bare `double precision` as a
5673 // parameter named "double" — which is what made the same
5674 // signature key two different ways.
5675 let (name, ty_token) = {
5676 let mut words: Vec<String> = alloc::vec![self.expect_ident_like()?];
5677 while matches!(self.peek(), Token::Ident(_) | Token::QuotedIdent(_)) {
5678 words.push(self.expect_ident_like()?);
5679 }
5680 // v7.39 (round 344) — a length / precision modifier on the
5681 // type: `f(character varying(9))`, `f(numeric(10,2))`. PG
5682 // accepts it and DROPS it — `pg_get_function_arguments`
5683 // reports plain `character varying` / `numeric`, measured on
5684 // 18.4 — but SPG raised `syntax error at or near "("`,
5685 // because the modifier's parens were never consumed.
5686 self.skip_type_modifier();
5687 // r1049 — `f(v bigint[])`. The array suffix parsed in
5688 // the column position, the cast position and (r1038)
5689 // the RETURNS position, but not here: the fifth
5690 // member of the same family, reported by sentori as
5691 // presumably the same code. It is now.
5692 let array_suffix = self.consume_array_suffix();
5693 let whole = words.join(" ");
5694 let (name, mut ty_token) = if words.len() >= 2 && !is_multiword_type_phrase(&whole)
5695 {
5696 (Some(words[0].clone()), words[1..].join(" "))
5697 } else {
5698 (None, whole)
5699 };
5700 ty_token.push_str(&array_suffix);
5701 (name, ty_token)
5702 };
5703 // Type — try to map to ColumnTypeName, else Raw.
5704 let ty = match map_type_ident_to_column_type_name(&ty_token) {
5705 Some(t) => FunctionArgType::Typed(t),
5706 None => FunctionArgType::Raw(ty_token),
5707 };
5708 args.push(FunctionArg { mode, name, ty });
5709 match self.peek() {
5710 Token::Comma => {
5711 self.advance();
5712 continue;
5713 }
5714 Token::RParen => {
5715 self.advance();
5716 return Ok(args);
5717 }
5718 other => {
5719 return Err(self.err(alloc::format!(
5720 "expected , or ) in function arg list, got {other:?}"
5721 )));
5722 }
5723 }
5724 }
5725 }
5726
5727 fn parse_function_return(&mut self) -> Result<FunctionReturn, ParseError> {
5728 // v7.39 (read01 round 65) — `RETURNS TABLE(col type, …)`: a set-returning
5729 // function whose row shape is named inline.
5730 if matches!(self.peek(), Token::Table)
5731 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
5732 {
5733 self.advance(); // TABLE
5734 self.advance(); // (
5735 let mut cols: Vec<String> = Vec::new();
5736 loop {
5737 let cname = self.expect_ident_like()?;
5738 let mut ty: Vec<String> = Vec::new();
5739 loop {
5740 match self.peek() {
5741 Token::Comma | Token::RParen | Token::Eof => break,
5742 _ => {}
5743 }
5744 match self.advance() {
5745 Token::Ident(w) | Token::QuotedIdent(w) => ty.push(w),
5746 other => {
5747 if let Some(w) = unreserved_keyword_text(&other) {
5748 ty.push(w);
5749 }
5750 }
5751 }
5752 }
5753 cols.push(alloc::format!("{cname} {}", ty.join(" ")));
5754 if matches!(self.peek(), Token::Comma) {
5755 self.advance();
5756 } else {
5757 break;
5758 }
5759 }
5760 if matches!(self.peek(), Token::RParen) {
5761 self.advance();
5762 }
5763 return Ok(FunctionReturn::Other(alloc::format!(
5764 "TABLE({})",
5765 cols.join(", ")
5766 )));
5767 }
5768 let ident = self.expect_ident_like()?;
5769 // v7.39 (read01 round 65) — `RETURNS SETOF <type>`.
5770 if ident.eq_ignore_ascii_case("setof") {
5771 let inner = self.expect_ident_like()?;
5772 let inner = alloc::format!("{inner}{}", self.consume_array_suffix());
5773 return Ok(FunctionReturn::Other(alloc::format!("SETOF {inner}")));
5774 }
5775 if ident.eq_ignore_ascii_case("trigger") {
5776 return Ok(FunctionReturn::Trigger);
5777 }
5778 if ident.eq_ignore_ascii_case("void") {
5779 return Ok(FunctionReturn::Void);
5780 }
5781 // r1038 — `RETURNS bigint[]`. An array COLUMN type parsed; the
5782 // RETURN position did not, so the `[` was a syntax error and the
5783 // whole migration stopped. sentori worked around it by returning
5784 // zero-padded text.
5785 let suffix = self.consume_array_suffix();
5786 if !suffix.is_empty() {
5787 return Ok(FunctionReturn::Other(alloc::format!("{ident}{suffix}")));
5788 }
5789 match map_type_ident_to_column_type_name(&ident) {
5790 Some(t) => Ok(FunctionReturn::Type(t)),
5791 None => Ok(FunctionReturn::Other(ident)),
5792 }
5793 }
5794
5795 /// Consume any `[]` / `[N]` array markers after a type name and give
5796 /// back their text. Empty when there are none.
5797 fn consume_array_suffix(&mut self) -> String {
5798 let mut out = String::new();
5799 while matches!(self.peek(), Token::LBracket) {
5800 self.advance();
5801 // `[N]` is accepted and, as in PG, the length is not enforced.
5802 if let Token::Integer(n) = self.peek().clone() {
5803 self.advance();
5804 out.push_str(&alloc::format!("[{n}]"));
5805 } else {
5806 out.push_str("[]");
5807 }
5808 if matches!(self.peek(), Token::RBracket) {
5809 self.advance();
5810 }
5811 }
5812 out
5813 }
5814
5815 fn parse_optional_language(&mut self) -> Result<Option<String>, ParseError> {
5816 match self.peek() {
5817 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("language") => {
5818 self.advance();
5819 let lang = self.expect_ident_like()?;
5820 Ok(Some(lang.to_ascii_lowercase()))
5821 }
5822 _ => Ok(None),
5823 }
5824 }
5825
5826 /// v7.17.0 Phase 1.5 — body of `CREATE DOMAIN name AS
5827 /// base_type [DEFAULT expr] [NOT NULL | NULL] [CHECK
5828 /// (expr)]*`. The `DOMAIN` keyword has already been
5829 /// consumed. PG allows the trailing constraints in any
5830 /// order; we approximate with a small loop.
5831 fn parse_create_domain_after_keyword(&mut self) -> Result<Statement, ParseError> {
5832 let name = self.expect_ident_like()?;
5833 // Optional `AS`.
5834 if matches!(self.peek(), Token::As) {
5835 self.advance();
5836 }
5837 // v7.39 (round 259) — keep the raw type NAME when the base is not
5838 // a builtin: it is how `CREATE DOMAIN child AS parent` records its
5839 // parent domain.
5840 let (base_type, _, _, base_user_ref, _, _, _, _, _, _, _, _, _, _) =
5841 self.parse_type_with_implied_flags()?;
5842 let mut default: Option<Expr> = None;
5843 let mut not_null = false;
5844 let mut checks: Vec<Expr> = Vec::new();
5845 loop {
5846 match self.peek() {
5847 Token::Default => {
5848 if default.is_some() {
5849 return Err(self.err("DOMAIN DEFAULT specified twice".into()));
5850 }
5851 self.advance();
5852 default = Some(self.parse_expr(0)?);
5853 }
5854 Token::Not => {
5855 self.advance();
5856 if !matches!(self.peek(), Token::Null) {
5857 return Err(self.err(alloc::format!(
5858 "expected NULL after NOT in DOMAIN, got {:?}",
5859 self.peek()
5860 )));
5861 }
5862 self.advance();
5863 not_null = true;
5864 }
5865 Token::Null => {
5866 self.advance();
5867 // v7.39 (round 761, F31 tranche 2 #31) — bare NULL
5868 // is the default-nullable marker (PG accepts it),
5869 // but AFTER a NOT NULL it is a conflict PG refuses
5870 // (`conflicting NULL/NOT NULL constraints`,
5871 // PG18-measured); the old arm no-opped both ways.
5872 if not_null {
5873 return Err(self.err(alloc::string::String::from(
5874 "conflicting NULL/NOT NULL constraints",
5875 )));
5876 }
5877 }
5878 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("check") => {
5879 self.advance();
5880 if !matches!(self.peek(), Token::LParen) {
5881 return Err(self.err(alloc::format!(
5882 "expected '(' after CHECK in DOMAIN, got {:?}",
5883 self.peek()
5884 )));
5885 }
5886 self.advance();
5887 let expr = self.parse_expr(0)?;
5888 if !matches!(self.peek(), Token::RParen) {
5889 return Err(self.err(alloc::format!(
5890 "expected ')' after CHECK expr, got {:?}",
5891 self.peek()
5892 )));
5893 }
5894 self.advance();
5895 checks.push(expr);
5896 }
5897 // CONSTRAINT <name> CHECK (…) — PG accepts a name
5898 // prefix on the constraint; we drop the name and
5899 // recurse into the constraint parsing.
5900 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("constraint") => {
5901 self.advance();
5902 let _ = self.expect_ident_like()?;
5903 }
5904 _ => break,
5905 }
5906 }
5907 Ok(Statement::CreateDomain(crate::ast::CreateDomainStatement {
5908 name,
5909 base_type,
5910 base_domain: base_user_ref,
5911 default,
5912 not_null,
5913 checks,
5914 }))
5915 }
5916
5917 /// v7.17.0 Phase 1.4 — body of `CREATE TYPE name AS ENUM
5918 /// ('a', 'b', …)`. The `TYPE` keyword has already been
5919 /// consumed.
5920 fn parse_create_type_after_keyword(&mut self) -> Result<Statement, ParseError> {
5921 let name = self.expect_ident_like()?;
5922 // Required `AS`.
5923 if !matches!(self.peek(), Token::As) {
5924 return Err(self.err(alloc::format!(
5925 "expected AS after CREATE TYPE {name:?}, got {:?}",
5926 self.peek()
5927 )));
5928 }
5929 self.advance();
5930 // v7.37.x (ζ-B composite Phase 1) — `AS (` is the composite-
5931 // type shape: `CREATE TYPE foo AS (a INT, b TEXT)`. Branch
5932 // on the next token: `(` = composite, ident `ENUM` = enum.
5933 if matches!(self.peek(), Token::LParen) {
5934 self.advance();
5935 let mut fields: Vec<(String, ColumnTypeName)> = Vec::new();
5936 let mut field_user_types: Vec<Option<String>> = Vec::new();
5937 // v7.39 (round 769, F31 tranche 5 #140) — `CREATE TYPE x AS ()`
5938 // is legal PG (an attribute-less composite; measured — the old
5939 // e2e note claimed PG requires at least one attribute).
5940 if matches!(self.peek(), Token::RParen) {
5941 self.advance();
5942 return Ok(Statement::CreateType(crate::ast::CreateTypeStatement {
5943 name,
5944 kind: crate::ast::TypeKind::Composite {
5945 fields,
5946 field_user_types,
5947 },
5948 }));
5949 }
5950 loop {
5951 let field_name = self.expect_ident_like()?;
5952 // v7.39 (round 264) — keep the raw type name when it is not
5953 // a builtin: that is how a NESTED composite field records
5954 // which composite it holds.
5955 let (field_type, _, _, field_user_ref, _, _, _, _, _, _, _, _, _, _) =
5956 self.parse_type_with_implied_flags()?;
5957 fields.push((field_name, field_type));
5958 field_user_types.push(field_user_ref);
5959 if matches!(self.peek(), Token::Comma) {
5960 self.advance();
5961 continue;
5962 }
5963 if matches!(self.peek(), Token::RParen) {
5964 self.advance();
5965 break;
5966 }
5967 return Err(self.err(alloc::format!(
5968 "expected , or ) in composite field list, got {:?}",
5969 self.peek()
5970 )));
5971 }
5972 if fields.is_empty() {
5973 return Err(self.err("CREATE TYPE … AS (…) must declare at least one field".into()));
5974 }
5975 return Ok(Statement::CreateType(crate::ast::CreateTypeStatement {
5976 name,
5977 kind: crate::ast::TypeKind::Composite {
5978 fields,
5979 field_user_types,
5980 },
5981 }));
5982 }
5983 // Required `ENUM` ident.
5984 let kind_ident = match self.peek().clone() {
5985 Token::Ident(s) | Token::QuotedIdent(s) => s,
5986 other => {
5987 return Err(self.err(alloc::format!(
5988 "expected ENUM or '(' after CREATE TYPE {name:?} AS, got {other:?}"
5989 )));
5990 }
5991 };
5992 if !kind_ident.eq_ignore_ascii_case("enum") {
5993 return Err(self.err(alloc::format!(
5994 "Phase 1.4 only supports ENUM or composite '(…)'; got {kind_ident:?}"
5995 )));
5996 }
5997 self.advance();
5998 if !matches!(self.peek(), Token::LParen) {
5999 return Err(self.err(alloc::format!(
6000 "expected '(' after ENUM, got {:?}",
6001 self.peek()
6002 )));
6003 }
6004 self.advance();
6005 let mut labels: Vec<String> = Vec::new();
6006 loop {
6007 match self.peek().clone() {
6008 Token::String(s) => {
6009 self.advance();
6010 labels.push(s);
6011 }
6012 other => {
6013 return Err(
6014 self.err(alloc::format!("expected enum label string, got {other:?}"))
6015 );
6016 }
6017 }
6018 if matches!(self.peek(), Token::Comma) {
6019 self.advance();
6020 continue;
6021 }
6022 if matches!(self.peek(), Token::RParen) {
6023 self.advance();
6024 break;
6025 }
6026 return Err(self.err(alloc::format!(
6027 "expected , or ) in ENUM label list, got {:?}",
6028 self.peek()
6029 )));
6030 }
6031 if labels.is_empty() {
6032 return Err(self.err("CREATE TYPE … AS ENUM must declare at least one label".into()));
6033 }
6034 Ok(Statement::CreateType(crate::ast::CreateTypeStatement {
6035 name,
6036 kind: crate::ast::TypeKind::Enum { labels },
6037 }))
6038 }
6039
6040 /// v7.17.0 Phase 1.3 — body of `CREATE MATERIALIZED VIEW
6041 /// [IF NOT EXISTS] name [(col, …)] AS <SELECT …> [WITH [NO] DATA]`.
6042 /// The `CREATE MATERIALIZED VIEW` keywords have already been
6043 /// consumed.
6044 fn parse_create_materialized_view_after_keyword(&mut self) -> Result<Statement, ParseError> {
6045 let if_not_exists = self.parse_if_not_exists();
6046 let name = self.expect_ident_like()?;
6047 let mut columns: Vec<String> = Vec::new();
6048 if matches!(self.peek(), Token::LParen) {
6049 self.advance();
6050 loop {
6051 let c = self.expect_ident_like()?;
6052 columns.push(c);
6053 if matches!(self.peek(), Token::Comma) {
6054 self.advance();
6055 continue;
6056 }
6057 if matches!(self.peek(), Token::RParen) {
6058 self.advance();
6059 break;
6060 }
6061 return Err(self.err(alloc::format!(
6062 "expected , or ) in MATERIALIZED VIEW column list, got {:?}",
6063 self.peek()
6064 )));
6065 }
6066 }
6067 if !matches!(self.peek(), Token::As) {
6068 return Err(self.err(alloc::format!(
6069 "expected AS <SELECT …> after CREATE MATERIALIZED VIEW {name:?}, got {:?}",
6070 self.peek()
6071 )));
6072 }
6073 self.advance();
6074 // v7.39 (round 151) — a WITH-headed body is legal (read-only
6075 // CTEs only; the engine rejects data-modifying ones with PG's
6076 // message). A trailing `WITH [NO] DATA` can't START the body,
6077 // so WITH here heads the query.
6078 let body = if self.peek_is_with_kw() {
6079 self.advance();
6080 self.parse_nested_with_select()?
6081 } else {
6082 let body_stmt = self.parse_select_stmt()?;
6083 let Statement::Select(body) = body_stmt else {
6084 return Err(self.err(alloc::format!(
6085 "CREATE MATERIALIZED VIEW body must be a SELECT, got {body_stmt:?}"
6086 )));
6087 };
6088 body
6089 };
6090 // Optional trailing `WITH [NO] DATA`.
6091 let with_data = self.parse_optional_with_data(true)?;
6092 Ok(Statement::CreateMaterializedView(
6093 crate::ast::CreateMaterializedViewStatement {
6094 temporary: false,
6095 name,
6096 if_not_exists,
6097 columns,
6098 body,
6099 with_data,
6100 as_plain_table: false,
6101 },
6102 ))
6103 }
6104
6105 /// v7.17.0 Phase 1.3 — `WITH [NO] DATA` trailer.
6106 /// `default_when_absent` is what to return if the tail is
6107 /// missing (CREATE defaults to WITH DATA, REFRESH defaults to
6108 /// WITH DATA).
6109 fn parse_optional_with_data(&mut self, default_when_absent: bool) -> Result<bool, ParseError> {
6110 let save = self.pos;
6111 // `WITH` is an Ident (not reserved in the lexer).
6112 let is_with = match self.peek() {
6113 Token::Ident(s) | Token::QuotedIdent(s) => s.eq_ignore_ascii_case("with"),
6114 _ => false,
6115 };
6116 if !is_with {
6117 return Ok(default_when_absent);
6118 }
6119 self.advance();
6120 // Optional `NO`.
6121 let mut with_data = true;
6122 let is_no = match self.peek() {
6123 Token::Ident(s) | Token::QuotedIdent(s) => s.eq_ignore_ascii_case("no"),
6124 _ => false,
6125 };
6126 if is_no {
6127 self.advance();
6128 with_data = false;
6129 }
6130 // Required `DATA` ident.
6131 let is_data = match self.peek() {
6132 Token::Ident(s) | Token::QuotedIdent(s) => s.eq_ignore_ascii_case("data"),
6133 _ => false,
6134 };
6135 if is_data {
6136 self.advance();
6137 Ok(with_data)
6138 } else {
6139 // Caller's WITH wasn't WITH-DATA — rewind so the outer
6140 // parser can interpret it.
6141 self.pos = save;
6142 Ok(default_when_absent)
6143 }
6144 }
6145
6146 /// v7.17.0 Phase 1.2 — body of `CREATE [OR REPLACE]
6147 /// [TEMPORARY] VIEW [IF NOT EXISTS] name [(col, …)] AS <SELECT>`.
6148 /// All keyword prefixes have already been consumed; the flags
6149 /// say which were present.
6150 fn parse_create_view_after_keyword(
6151 &mut self,
6152 or_replace: bool,
6153 _materialized_unused: bool,
6154 temporary: bool,
6155 ) -> Result<Statement, ParseError> {
6156 let if_not_exists = self.parse_if_not_exists();
6157 let name = self.expect_ident_like()?;
6158 // Optional `(col, col, …)` rename list.
6159 let mut columns: Vec<String> = Vec::new();
6160 if matches!(self.peek(), Token::LParen) {
6161 self.advance();
6162 loop {
6163 let c = self.expect_ident_like()?;
6164 columns.push(c);
6165 if matches!(self.peek(), Token::Comma) {
6166 self.advance();
6167 continue;
6168 }
6169 if matches!(self.peek(), Token::RParen) {
6170 self.advance();
6171 break;
6172 }
6173 return Err(self.err(alloc::format!(
6174 "expected , or ) in VIEW column list, got {:?}",
6175 self.peek()
6176 )));
6177 }
6178 }
6179 // Required `AS`.
6180 if !matches!(self.peek(), Token::As) {
6181 return Err(self.err(alloc::format!(
6182 "expected AS <SELECT …> after CREATE VIEW {name:?}, got {:?}",
6183 self.peek()
6184 )));
6185 }
6186 self.advance();
6187 // Body: a regular SELECT statement. v7.39 (round 151) — a
6188 // WITH-headed body is legal too (read-only CTEs only; the
6189 // engine rejects data-modifying ones with PG's message).
6190 // Disambiguation vs `WITH CHECK OPTION`: a body can't START
6191 // with the check-option clause, so WITH here heads the query.
6192 let body = if self.peek_is_with_kw() {
6193 self.advance();
6194 self.parse_nested_with_select()?
6195 } else {
6196 let body_stmt = self.parse_select_stmt()?;
6197 let Statement::Select(body) = body_stmt else {
6198 return Err(self.err(alloc::format!(
6199 "CREATE VIEW body must be a SELECT statement, got {body_stmt:?}"
6200 )));
6201 };
6202 body
6203 };
6204 // v7.39 (round 132) — optional `WITH [ LOCAL | CASCADED ] CHECK OPTION`.
6205 // The SELECT parser stops before a trailing WITH, so it lands here.
6206 let check_option = if matches!(self.peek(),
6207 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("with"))
6208 {
6209 self.advance(); // WITH
6210 let opt = match self.peek() {
6211 Token::Ident(s) if s.eq_ignore_ascii_case("local") => {
6212 self.advance();
6213 crate::ast::ViewCheckOption::Local
6214 }
6215 Token::Ident(s) if s.eq_ignore_ascii_case("cascaded") => {
6216 self.advance();
6217 crate::ast::ViewCheckOption::Cascaded
6218 }
6219 // Bare `WITH CHECK OPTION` defaults to CASCADED (PG).
6220 _ => crate::ast::ViewCheckOption::Cascaded,
6221 };
6222 if !matches!(self.peek(),
6223 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("check"))
6224 {
6225 return Err(self.err(alloc::format!(
6226 "expected CHECK in CREATE VIEW … WITH [LOCAL|CASCADED] CHECK OPTION, got {:?}",
6227 self.peek()
6228 )));
6229 }
6230 self.advance(); // CHECK
6231 if !matches!(self.peek(),
6232 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("option"))
6233 {
6234 return Err(self.err(alloc::format!(
6235 "expected OPTION after WITH CHECK in CREATE VIEW, got {:?}",
6236 self.peek()
6237 )));
6238 }
6239 self.advance(); // OPTION
6240 Some(opt)
6241 } else {
6242 None
6243 };
6244 Ok(Statement::CreateView(crate::ast::CreateViewStatement {
6245 name,
6246 or_replace,
6247 if_not_exists,
6248 temporary,
6249 columns,
6250 body,
6251 check_option,
6252 }))
6253 }
6254
6255 /// v7.17.0 — body of `CREATE [TEMPORARY] SEQUENCE`. The
6256 /// `[TEMPORARY]` and `SEQUENCE` tokens have already been
6257 /// consumed; `temporary` carries whether TEMPORARY was seen.
6258 fn parse_create_sequence_after_keyword(
6259 &mut self,
6260 temporary: bool,
6261 ) -> Result<Statement, ParseError> {
6262 let if_not_exists = self.parse_if_not_exists();
6263 let name = self.expect_ident_like()?;
6264 // Optional `AS data_type`.
6265 let data_type = if matches!(self.peek(), Token::As) {
6266 self.advance();
6267 Some(self.parse_sequence_data_type()?)
6268 } else {
6269 None
6270 };
6271 let options = self.parse_sequence_options(/* allow_restart = */ false)?;
6272 Ok(Statement::CreateSequence(
6273 crate::ast::CreateSequenceStatement {
6274 name,
6275 if_not_exists,
6276 temporary,
6277 data_type,
6278 options,
6279 },
6280 ))
6281 }
6282
6283 /// v7.17.0 — body of `ALTER SEQUENCE`. The `ALTER` keyword has
6284 /// already been consumed; this is reached after `SEQUENCE`.
6285 /// v7.39 (round 260) — `ALTER DOMAIN name <action>`.
6286 fn parse_alter_domain_after_keyword(&mut self) -> Result<Statement, ParseError> {
6287 use crate::ast::AlterDomainAction as A;
6288 let name = self.expect_ident_like()?;
6289 // DROP / SET / ADD lex as reserved keyword tokens, not idents.
6290 let kw = match self.peek() {
6291 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
6292 Token::Drop => alloc::string::String::from("drop"),
6293 Token::Default => alloc::string::String::from("default"),
6294 other => {
6295 return Err(self.err(alloc::format!(
6296 "expected an ALTER DOMAIN action, got {other:?}"
6297 )));
6298 }
6299 };
6300 let action = match kw.as_str() {
6301 "add" => {
6302 self.advance();
6303 let cname = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint"))
6304 {
6305 self.advance();
6306 Some(self.expect_ident_like()?)
6307 } else {
6308 None
6309 };
6310 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("check")) {
6311 return Err(self.err(alloc::format!(
6312 "ALTER DOMAIN ADD supports CHECK only, got {:?}",
6313 self.peek()
6314 )));
6315 }
6316 self.advance();
6317 if !matches!(self.peek(), Token::LParen) {
6318 return Err(self.err("expected '(' after CHECK".into()));
6319 }
6320 self.advance();
6321 let check = self.parse_expr(0)?;
6322 if !matches!(self.peek(), Token::RParen) {
6323 return Err(self.err("expected ')' after CHECK expression".into()));
6324 }
6325 self.advance();
6326 A::AddConstraint { name: cname, check }
6327 }
6328 "drop" => {
6329 self.advance();
6330 match self.peek() {
6331 Token::Ident(s) if s.eq_ignore_ascii_case("constraint") => {
6332 self.advance();
6333 let if_exists = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if"))
6334 {
6335 self.advance();
6336 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("exists"))
6337 {
6338 return Err(self.err("expected EXISTS after IF".into()));
6339 }
6340 self.advance();
6341 true
6342 } else {
6343 false
6344 };
6345 let cn = self.expect_ident_like()?;
6346 A::DropConstraint {
6347 name: cn,
6348 if_exists,
6349 }
6350 }
6351 Token::Default => {
6352 self.advance();
6353 A::DropDefault
6354 }
6355 Token::Not => {
6356 self.advance();
6357 if !matches!(self.peek(), Token::Null) {
6358 return Err(self.err("expected NULL after NOT".into()));
6359 }
6360 self.advance();
6361 A::DropNotNull
6362 }
6363 other => {
6364 return Err(self.err(alloc::format!(
6365 "ALTER DOMAIN DROP expects CONSTRAINT / DEFAULT / NOT NULL, got {other:?}"
6366 )));
6367 }
6368 }
6369 }
6370 "set" => {
6371 self.advance();
6372 match self.peek() {
6373 Token::Default => {
6374 self.advance();
6375 A::SetDefault(self.parse_expr(0)?)
6376 }
6377 Token::Not => {
6378 self.advance();
6379 if !matches!(self.peek(), Token::Null) {
6380 return Err(self.err("expected NULL after NOT".into()));
6381 }
6382 self.advance();
6383 A::SetNotNull
6384 }
6385 other => {
6386 return Err(self.err(alloc::format!(
6387 "ALTER DOMAIN SET expects DEFAULT / NOT NULL, got {other:?}"
6388 )));
6389 }
6390 }
6391 }
6392 "rename" => {
6393 self.advance();
6394 if !matches!(self.peek(), Token::To) {
6395 return Err(self.err("expected TO after RENAME".into()));
6396 }
6397 self.advance();
6398 A::RenameTo(self.expect_ident_like()?)
6399 }
6400 other => {
6401 return Err(self.err(alloc::format!("unsupported ALTER DOMAIN action {other:?}")));
6402 }
6403 };
6404 Ok(Statement::AlterDomain { name, action })
6405 }
6406
6407 fn parse_alter_sequence_after_keyword(&mut self) -> Result<Statement, ParseError> {
6408 let if_exists = self.parse_if_exists();
6409 let name = self.expect_ident_like()?;
6410 // v7.39 (read01 round 49) — `RENAME TO new`; mutually exclusive with
6411 // the option list (PG allows only one or the other).
6412 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("rename")) {
6413 self.advance();
6414 if matches!(self.peek(), Token::To) {
6415 self.advance();
6416 } else {
6417 self.expect_keyword_ident("to")?;
6418 }
6419 let new = self.expect_ident_like()?;
6420 return Ok(Statement::AlterSequence(
6421 crate::ast::AlterSequenceStatement {
6422 name,
6423 if_exists,
6424 options: crate::ast::SequenceOptions::default(),
6425 rename_to: Some(new),
6426 },
6427 ));
6428 }
6429 let options = self.parse_sequence_options(/* allow_restart = */ true)?;
6430 Ok(Statement::AlterSequence(
6431 crate::ast::AlterSequenceStatement {
6432 name,
6433 if_exists,
6434 options,
6435 rename_to: None,
6436 },
6437 ))
6438 }
6439
6440 fn parse_sequence_data_type(&mut self) -> Result<crate::ast::SequenceDataType, ParseError> {
6441 let kw = self.expect_ident_like()?;
6442 match kw.to_ascii_lowercase().as_str() {
6443 "smallint" | "int2" => Ok(crate::ast::SequenceDataType::SmallInt),
6444 "integer" | "int" | "int4" => Ok(crate::ast::SequenceDataType::Int),
6445 "bigint" | "int8" => Ok(crate::ast::SequenceDataType::BigInt),
6446 other => Err(self.err(alloc::format!(
6447 "expected SMALLINT / INTEGER / BIGINT after SEQUENCE AS, got {other:?}"
6448 ))),
6449 }
6450 }
6451
6452 fn parse_sequence_options(
6453 &mut self,
6454 allow_restart: bool,
6455 ) -> Result<crate::ast::SequenceOptions, ParseError> {
6456 use crate::ast::{SeqBound, SequenceOptions, SequenceOwnedBy};
6457 let mut opts = SequenceOptions::default();
6458 #[allow(clippy::while_let_loop)]
6459 loop {
6460 // Match an ident; stop at any non-ident token (sentinel,
6461 // semicolon, end of statement).
6462 let kw_lc = match self.peek() {
6463 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
6464 _ => break,
6465 };
6466 match kw_lc.as_str() {
6467 "increment" => {
6468 self.advance();
6469 // Optional BY.
6470 if self.peek_is_by() {
6471 self.advance();
6472 }
6473 opts.increment = Some(self.expect_signed_int()?);
6474 }
6475 "minvalue" => {
6476 self.advance();
6477 opts.min_value = Some(SeqBound::Value(self.expect_signed_int()?));
6478 }
6479 "maxvalue" => {
6480 self.advance();
6481 opts.max_value = Some(SeqBound::Value(self.expect_signed_int()?));
6482 }
6483 "no" => {
6484 self.advance();
6485 let what = self.expect_ident_like()?;
6486 match what.to_ascii_lowercase().as_str() {
6487 "minvalue" => opts.min_value = Some(SeqBound::NoBound),
6488 "maxvalue" => opts.max_value = Some(SeqBound::NoBound),
6489 "cycle" => opts.cycle = Some(false),
6490 other => {
6491 return Err(self.err(alloc::format!(
6492 "expected MINVALUE / MAXVALUE / CYCLE after NO, got {other:?}"
6493 )));
6494 }
6495 }
6496 }
6497 "start" => {
6498 self.advance();
6499 // Optional WITH.
6500 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
6501 if s.eq_ignore_ascii_case("with"))
6502 {
6503 self.advance();
6504 }
6505 opts.start = Some(self.expect_signed_int()?);
6506 }
6507 "restart" if allow_restart => {
6508 self.advance();
6509 // Optional WITH n; bare RESTART means restart at START.
6510 let mut with_val: Option<i64> = None;
6511 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
6512 if s.eq_ignore_ascii_case("with"))
6513 {
6514 self.advance();
6515 with_val = Some(self.expect_signed_int()?);
6516 } else if matches!(self.peek(), Token::Integer(_) | Token::Minus) {
6517 with_val = Some(self.expect_signed_int()?);
6518 }
6519 opts.restart = Some(with_val);
6520 }
6521 "cache" => {
6522 self.advance();
6523 opts.cache = Some(self.expect_signed_int()?);
6524 }
6525 "cycle" => {
6526 self.advance();
6527 opts.cycle = Some(true);
6528 }
6529 "owned" => {
6530 self.advance();
6531 match self.peek() {
6532 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("by") => {
6533 self.advance();
6534 }
6535 other => {
6536 return Err(
6537 self.err(alloc::format!("expected BY after OWNED, got {other:?}"))
6538 );
6539 }
6540 }
6541 // OWNED BY {NONE | tab.col}. Read just one ident
6542 // (NOT expect_ident_like which would auto-strip
6543 // a schema prefix and consume the `.col` we need).
6544 let first = match self.advance() {
6545 Token::Ident(s) | Token::QuotedIdent(s) => s,
6546 other => {
6547 return Err(self.err(alloc::format!(
6548 "expected identifier or NONE after OWNED BY, got {other:?}"
6549 )));
6550 }
6551 };
6552 if first.eq_ignore_ascii_case("none") {
6553 opts.owned_by = Some(SequenceOwnedBy::None);
6554 } else if matches!(self.peek(), Token::Dot) {
6555 self.advance();
6556 let second = match self.advance() {
6557 Token::Ident(s) | Token::QuotedIdent(s) => s,
6558 other => {
6559 return Err(self.err(alloc::format!(
6560 "expected column name after OWNED BY {first}., got {other:?}"
6561 )));
6562 }
6563 };
6564 // v7.17 dump-compat fix — pg_dump emits
6565 // OWNED BY clauses as
6566 // `schema.table.column` (three segments).
6567 // If a third `.<ident>` follows, treat the
6568 // first ident as schema (drop it; SPG is
6569 // single-schema) and the middle / last
6570 // pair as table.column. Otherwise it's
6571 // the two-segment form table.column.
6572 if matches!(self.peek(), Token::Dot) {
6573 self.advance();
6574 let third = match self.advance() {
6575 Token::Ident(s) | Token::QuotedIdent(s) => s,
6576 other => {
6577 return Err(self.err(alloc::format!(
6578 "expected column name after OWNED BY {first}.{second}., got {other:?}"
6579 )));
6580 }
6581 };
6582 let _ = first; // schema prefix discarded
6583 opts.owned_by = Some(SequenceOwnedBy::Column {
6584 table: second,
6585 column: third,
6586 });
6587 } else {
6588 opts.owned_by = Some(SequenceOwnedBy::Column {
6589 table: first,
6590 column: second,
6591 });
6592 }
6593 } else {
6594 return Err(self.err(alloc::format!(
6595 "expected table.column or NONE after OWNED BY, got {first:?}"
6596 )));
6597 }
6598 }
6599 _ => break,
6600 }
6601 }
6602 Ok(opts)
6603 }
6604
6605 fn expect_signed_int(&mut self) -> Result<i64, ParseError> {
6606 let neg = if matches!(self.peek(), Token::Minus) {
6607 self.advance();
6608 true
6609 } else {
6610 false
6611 };
6612 match self.peek() {
6613 Token::Integer(n) => {
6614 let v = *n;
6615 self.advance();
6616 Ok(if neg { -v } else { v })
6617 }
6618 other => Err(self.err(alloc::format!("expected signed integer, got {other:?}"))),
6619 }
6620 }
6621
6622 /// v7.17.0 Phase 3.1 — absorb `[NOT] DEFERRABLE [INITIALLY
6623 /// {DEFERRED | IMMEDIATE}]` constraint-timing clauses. Each
6624 /// clause is fully accepted and discarded — SPG always runs
6625 /// constraint checks immediately (single-writer model). The
6626 /// loop allows DEFERRABLE and the INITIALLY suffix to appear
6627 /// in either order (per the SQL spec they're independent),
6628 /// though pg_dump always emits them in the canonical
6629 /// `[NOT] DEFERRABLE INITIALLY {DEFERRED|IMMEDIATE}` shape.
6630 /// Stops at the first token that isn't part of the clause.
6631 fn consume_optional_deferrable_clauses(&mut self) -> Result<(), ParseError> {
6632 self.consume_deferrable_clauses_timed().map(|_| ())
6633 }
6634
6635 /// v7.39 (round 288) — the same scan, but reporting what it saw:
6636 /// `(deferrable, initially_deferred)`. The clauses were parsed and
6637 /// dropped, so `DEFERRABLE INITIALLY DEFERRED` on an FK behaved as
6638 /// NOT DEFERRABLE and a circular-FK migration could not load.
6639 fn consume_deferrable_clauses_timed(&mut self) -> Result<(bool, bool), ParseError> {
6640 let mut deferrable = false;
6641 let mut initially_deferred = false;
6642 loop {
6643 // Bare `DEFERRABLE` (Phase 3.1 — was hard-error pre-3.1).
6644 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("deferrable")) {
6645 self.advance();
6646 deferrable = true;
6647 if self.consume_optional_initially_clause()? {
6648 initially_deferred = true;
6649 }
6650 continue;
6651 }
6652 // `NOT DEFERRABLE` — already worked pre-3.1.
6653 if matches!(self.peek(), Token::Not) {
6654 let look = self.tokens.get(self.pos + 1);
6655 if matches!(look, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("deferrable")) {
6656 self.advance(); // NOT
6657 self.advance(); // DEFERRABLE
6658 deferrable = false;
6659 initially_deferred = false;
6660 let _ = self.consume_optional_initially_clause()?;
6661 continue;
6662 }
6663 break;
6664 }
6665 // Standalone `INITIALLY {DEFERRED|IMMEDIATE}` — PG
6666 // accepts this without a leading [NOT] DEFERRABLE
6667 // (the timing keyword alone). pg_dump occasionally
6668 // emits it on FK constraints that inherit timing.
6669 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("initially")) {
6670 if self.consume_optional_initially_clause()? {
6671 initially_deferred = true;
6672 // PG: a bare `INITIALLY DEFERRED` implies DEFERRABLE.
6673 deferrable = true;
6674 }
6675 continue;
6676 }
6677 break;
6678 }
6679 Ok((deferrable, initially_deferred))
6680 }
6681
6682 /// Helper for [`consume_optional_deferrable_clauses`]. When the
6683 /// next token is `INITIALLY`, consume it plus the required
6684 /// `DEFERRED` | `IMMEDIATE` trailer. No-op otherwise.
6685 /// Returns true when the timing seen was `DEFERRED`.
6686 fn consume_optional_initially_clause(&mut self) -> Result<bool, ParseError> {
6687 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("initially")) {
6688 return Ok(false);
6689 }
6690 self.advance(); // INITIALLY
6691 match self.advance() {
6692 Token::Ident(s)
6693 if s.eq_ignore_ascii_case("deferred") || s.eq_ignore_ascii_case("immediate") =>
6694 {
6695 Ok(s.eq_ignore_ascii_case("deferred"))
6696 }
6697 other => Err(self.err(alloc::format!(
6698 "expected DEFERRED or IMMEDIATE after INITIALLY, got {other:?}"
6699 ))),
6700 }
6701 }
6702
6703 /// v7.17.0 Phase 4.2 — consume a MySQL `CREATE PROCEDURE` body
6704 /// in its entirety so the parser returns Empty without
6705 /// touching the runtime. The CREATE+PROCEDURE keywords are
6706 /// already consumed; this swallows everything from the
6707 /// procedure name through the matching `END`, including
6708 /// nested `BEGIN`/`END` blocks, internal `;` terminators
6709 /// (DELIMITER `//` makes the script splitter forward the
6710 /// whole block as one statement), `@var` session-variable
6711 /// references, and the trailing terminator.
6712 ///
6713 /// Tracks nesting depth so:
6714 /// BEGIN
6715 /// IF cond THEN
6716 /// BEGIN ... END;
6717 /// END IF;
6718 /// END
6719 /// terminates at the outer END.
6720 fn consume_mysql_routine_body(&mut self) {
6721 // Outer skeleton: name, (...), optional clauses, BEGIN
6722 // <body> END [;]. Scan for the first BEGIN — anything
6723 // before it is signature decoration we don't care about.
6724 // Once inside BEGIN, count up on BEGIN, down on END.
6725 let mut depth: i32 = 0;
6726 let mut started = false;
6727 loop {
6728 match self.peek().clone() {
6729 Token::Begin => {
6730 self.advance();
6731 depth += 1;
6732 started = true;
6733 }
6734 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("end") => {
6735 self.advance();
6736 if started {
6737 depth -= 1;
6738 if depth <= 0 {
6739 // Optional trailing ident (`END IF`,
6740 // `END LOOP`, `END WHILE`, `END CASE`,
6741 // `END label_name`) — eat the next
6742 // ident if present so we don't
6743 // mistake `END IF;` for the outer
6744 // close.
6745 if matches!(self.peek(), Token::Ident(_) | Token::QuotedIdent(_)) {
6746 // If the next token is one of the
6747 // PL/SQL block-closer keywords,
6748 // the END belongs to an inner
6749 // block; bump depth back up.
6750 let is_inner_close = matches!(
6751 self.peek(),
6752 Token::Ident(s) | Token::QuotedIdent(s)
6753 if matches!(
6754 s.to_ascii_lowercase().as_str(),
6755 "if" | "loop" | "while" | "case" | "repeat"
6756 )
6757 );
6758 if is_inner_close {
6759 self.advance();
6760 depth += 1;
6761 continue;
6762 }
6763 }
6764 // Eat optional trailing `;`.
6765 if matches!(self.peek(), Token::Semicolon) {
6766 self.advance();
6767 }
6768 return;
6769 }
6770 }
6771 }
6772 Token::Eof => return,
6773 _ => {
6774 self.advance();
6775 }
6776 }
6777 }
6778 }
6779
6780 /// v7.17.0 Phase 2.6 — absorb the MySQL view-prefix clauses
6781 /// that appear between `CREATE` and `VIEW` in mysqldump output:
6782 ///
6783 /// * `ALGORITHM = {UNDEFINED|MERGE|TEMPTABLE}`
6784 /// * `DEFINER = <user>` (user may be a quoted string, a bare
6785 /// ident, or `ident @ ident-or-quoted-string` host form)
6786 /// * `SQL SECURITY {DEFINER|INVOKER}`
6787 ///
6788 /// Each clause may appear at most once but in any order.
6789 /// The hints are pure planner / permission metadata that
6790 /// SPG's view-rewrite engine handles uniformly; we accept
6791 /// and discard. Returns `Ok(())` once a non-clause token is
6792 /// peeked (the caller then checks for the `VIEW` keyword).
6793 fn consume_mysql_view_prefix(&mut self) -> Result<(), ParseError> {
6794 loop {
6795 match self.peek().clone() {
6796 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("algorithm") => {
6797 self.advance(); // ALGORITHM
6798 // Optional `=`. MySQL spec requires it but be
6799 // generous.
6800 if matches!(self.peek(), Token::Eq) {
6801 self.advance();
6802 }
6803 // UNDEFINED / MERGE / TEMPTABLE — accept any
6804 // bare ident; unknown values still parse so
6805 // future MySQL versions don't break.
6806 if matches!(
6807 self.peek(),
6808 Token::Ident(_) | Token::QuotedIdent(_) | Token::String(_)
6809 ) {
6810 self.advance();
6811 }
6812 }
6813 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("definer") => {
6814 self.advance(); // DEFINER
6815 if matches!(self.peek(), Token::Eq) {
6816 self.advance();
6817 }
6818 // User: quoted string, ident, OR ident @ host
6819 // (host may itself be quoted or bare).
6820 match self.peek().clone() {
6821 Token::String(_) | Token::Ident(_) | Token::QuotedIdent(_) => {
6822 self.advance();
6823 // Optional `@host`.
6824 if matches!(self.peek(), Token::At) {
6825 self.advance();
6826 if matches!(
6827 self.peek(),
6828 Token::Ident(_) | Token::QuotedIdent(_) | Token::String(_)
6829 ) {
6830 self.advance();
6831 }
6832 }
6833 }
6834 _ => {}
6835 }
6836 }
6837 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("sql") => {
6838 // `SQL SECURITY {DEFINER|INVOKER}`. Only honoured
6839 // when followed by SECURITY — the dispatcher must
6840 // not consume a bare `SQL` token (it's not a
6841 // legal CREATE prefix on its own).
6842 let save = self.pos;
6843 self.advance(); // SQL
6844 if matches!(self.peek(), Token::Ident(s2) | Token::QuotedIdent(s2)
6845 if s2.eq_ignore_ascii_case("security"))
6846 {
6847 self.advance(); // SECURITY
6848 // DEFINER / INVOKER trailing ident.
6849 if matches!(self.peek(), Token::Ident(_) | Token::QuotedIdent(_)) {
6850 self.advance();
6851 }
6852 } else {
6853 // Not a SQL SECURITY clause — roll back and
6854 // bail; the caller will error out cleanly.
6855 self.pos = save;
6856 return Ok(());
6857 }
6858 }
6859 _ => return Ok(()),
6860 }
6861 }
6862 }
6863
6864 fn parse_if_not_exists(&mut self) -> bool {
6865 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("if"))
6866 {
6867 let save = self.pos;
6868 self.advance();
6869 if matches!(self.peek(), Token::Not) {
6870 self.advance();
6871 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("exists"))
6872 {
6873 self.advance();
6874 return true;
6875 }
6876 }
6877 self.pos = save;
6878 }
6879 false
6880 }
6881
6882 fn parse_if_exists(&mut self) -> bool {
6883 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("if"))
6884 {
6885 let save = self.pos;
6886 self.advance();
6887 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("exists"))
6888 {
6889 self.advance();
6890 return true;
6891 }
6892 self.pos = save;
6893 }
6894 false
6895 }
6896
6897 /// v7.12.4 — body of `CREATE [OR REPLACE] TRIGGER`. The
6898 /// `[OR REPLACE]` flag and the `TRIGGER` keyword have already
6899 /// been consumed.
6900 fn parse_create_trigger_after_keyword(
6901 &mut self,
6902 or_replace: bool,
6903 ) -> Result<Statement, ParseError> {
6904 let name = self.expect_ident_like()?;
6905 let timing = {
6906 let ident = self.expect_ident_like()?;
6907 if ident.eq_ignore_ascii_case("before") {
6908 TriggerTiming::Before
6909 } else if ident.eq_ignore_ascii_case("after") {
6910 TriggerTiming::After
6911 } else if ident.eq_ignore_ascii_case("instead") {
6912 let next = self.expect_ident_like()?;
6913 if !next.eq_ignore_ascii_case("of") {
6914 return Err(self.err(alloc::format!(
6915 "expected OF after INSTEAD in trigger timing, got {next:?}"
6916 )));
6917 }
6918 TriggerTiming::InsteadOf
6919 } else {
6920 return Err(self.err(alloc::format!(
6921 "expected BEFORE / AFTER / INSTEAD OF in trigger timing, got {ident:?}"
6922 )));
6923 }
6924 };
6925 // Events: INSERT [ OR UPDATE [ OR DELETE [ OR TRUNCATE ] ] ].
6926 // OR is a reserved keyword token (Token::Or), not an Ident.
6927 // v7.13.0 — after an UPDATE event we may optionally see
6928 // `OF col, col, …` (mailrs round-5 G7). Columns are
6929 // captured into `update_columns` once across the whole
6930 // events list; multiple `UPDATE OF` clauses are rejected.
6931 let mut events: Vec<TriggerEvent> = Vec::new();
6932 let mut update_columns: Vec<String> = Vec::new();
6933 let (first_ev, first_cols) = self.parse_trigger_event_with_optional_of()?;
6934 events.push(first_ev);
6935 if !first_cols.is_empty() {
6936 update_columns = first_cols;
6937 }
6938 while matches!(self.peek(), Token::Or) {
6939 self.advance();
6940 let (ev, cols) = self.parse_trigger_event_with_optional_of()?;
6941 events.push(ev);
6942 if !cols.is_empty() {
6943 if !update_columns.is_empty() {
6944 return Err(
6945 self.err("CREATE TRIGGER: `UPDATE OF cols` may appear at most once".into())
6946 );
6947 }
6948 update_columns = cols;
6949 }
6950 }
6951 // ON <table>
6952 let tok = self.peek();
6953 let Token::On = tok else {
6954 return Err(self.err(alloc::format!(
6955 "expected ON after trigger events, got {tok:?}"
6956 )));
6957 };
6958 self.advance();
6959 let table = self.expect_ident_like()?;
6960 // v7.39 (read01 round 82) — a CONSTRAINT TRIGGER may carry `FROM
6961 // reftable` and `[NOT] DEFERRABLE [INITIALLY {DEFERRED|IMMEDIATE}]`
6962 // between the table and FOR EACH ROW. Accept and skip them: SPG fires
6963 // the trigger as a plain AFTER trigger (correct for every non-deferred
6964 // use; deferral timing is not yet honoured).
6965 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
6966 if s.eq_ignore_ascii_case("from"))
6967 {
6968 self.advance();
6969 let _reftable = self.expect_ident_like()?;
6970 }
6971 self.consume_optional_deferrable_clauses()?;
6972 // FOR EACH ROW / FOR EACH STATEMENT. FOR is a reserved
6973 // keyword (Token::For); EACH / ROW / STATEMENT are bare
6974 // idents.
6975 if !matches!(self.peek(), Token::For) {
6976 return Err(self.err(alloc::format!(
6977 "expected FOR EACH ROW / STATEMENT, got {:?}",
6978 self.peek()
6979 )));
6980 }
6981 self.advance();
6982 let for_each = {
6983 let e = self.expect_ident_like()?;
6984 if !e.eq_ignore_ascii_case("each") {
6985 return Err(self.err(alloc::format!("expected EACH after FOR, got {e:?}")));
6986 }
6987 let unit = self.expect_ident_like()?;
6988 if unit.eq_ignore_ascii_case("row") {
6989 TriggerForEach::Row
6990 } else if unit.eq_ignore_ascii_case("statement") {
6991 TriggerForEach::Statement
6992 } else {
6993 return Err(self.err(alloc::format!(
6994 "expected ROW / STATEMENT after FOR EACH, got {unit:?}"
6995 )));
6996 }
6997 };
6998 // v7.39 (round 138) — optional `WHEN ( condition )` before EXECUTE.
6999 let when_condition = if matches!(self.peek(),
7000 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("when"))
7001 {
7002 self.advance();
7003 Some(self.parse_paren_expr("WHEN")?)
7004 } else {
7005 None
7006 };
7007 // EXECUTE FUNCTION/PROCEDURE name(...)
7008 let exec = self.expect_ident_like()?;
7009 if !exec.eq_ignore_ascii_case("execute") {
7010 return Err(self.err(alloc::format!(
7011 "expected EXECUTE FUNCTION/PROCEDURE in CREATE TRIGGER, got {exec:?}"
7012 )));
7013 }
7014 let fn_or_proc = self.expect_ident_like()?;
7015 if !(fn_or_proc.eq_ignore_ascii_case("function")
7016 || fn_or_proc.eq_ignore_ascii_case("procedure"))
7017 {
7018 return Err(self.err(alloc::format!(
7019 "expected FUNCTION / PROCEDURE after EXECUTE, got {fn_or_proc:?}"
7020 )));
7021 }
7022 let function = self.expect_ident_like()?;
7023 // Optional empty arg list `()`.
7024 if matches!(self.peek(), Token::LParen) {
7025 self.advance();
7026 if !matches!(self.peek(), Token::RParen) {
7027 return Err(self.err(alloc::format!(
7028 "v7.12.4 trigger function calls take no args; got {:?}",
7029 self.peek()
7030 )));
7031 }
7032 self.advance();
7033 }
7034 Ok(Statement::CreateTrigger(CreateTriggerStatement {
7035 name,
7036 or_replace,
7037 timing,
7038 events,
7039 table,
7040 for_each,
7041 function,
7042 update_columns,
7043 when_condition,
7044 }))
7045 }
7046
7047 /// v7.39 (round 139) — `CREATE RULE <name> AS ON <event> TO <table>
7048 /// [WHERE <cond>] DO [ALSO|INSTEAD] { NOTHING | cmd | ( cmd; … ) }`.
7049 fn parse_create_rule_after_keyword(
7050 &mut self,
7051 or_replace: bool,
7052 ) -> Result<Statement, ParseError> {
7053 let name = self.expect_ident_like()?;
7054 if !matches!(self.peek(), Token::As) {
7055 return Err(self.err(alloc::format!(
7056 "expected AS in CREATE RULE, got {:?}",
7057 self.peek()
7058 )));
7059 }
7060 self.advance();
7061 if !matches!(self.peek(), Token::On) {
7062 return Err(self.err(alloc::format!(
7063 "expected ON in CREATE RULE, got {:?}",
7064 self.peek()
7065 )));
7066 }
7067 self.advance();
7068 let event = self.parse_rule_event()?;
7069 if !matches!(self.peek(), Token::To)
7070 && !matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("to"))
7071 {
7072 return Err(self.err(alloc::format!(
7073 "expected TO after rule event, got {:?}",
7074 self.peek()
7075 )));
7076 }
7077 self.advance();
7078 let table = self.expect_ident_like()?;
7079 // Optional `WHERE <cond>` (no parentheses, unlike a trigger WHEN).
7080 let when_condition = if matches!(self.peek(), Token::Where) {
7081 self.advance();
7082 Some(self.parse_expr(0)?)
7083 } else {
7084 None
7085 };
7086 if !matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("do"))
7087 {
7088 return Err(self.err(alloc::format!(
7089 "expected DO in CREATE RULE, got {:?}",
7090 self.peek()
7091 )));
7092 }
7093 self.advance();
7094 // `DO [ ALSO | INSTEAD ]` — ALSO is the default when neither is written.
7095 let instead = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("instead"))
7096 {
7097 self.advance();
7098 true
7099 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("also")) {
7100 self.advance();
7101 false
7102 } else {
7103 false
7104 };
7105 // `NOTHING` | `( cmd; … )` | `cmd`.
7106 let commands = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("nothing"))
7107 {
7108 self.advance();
7109 Vec::new()
7110 } else if matches!(self.peek(), Token::LParen) {
7111 self.advance();
7112 let mut cmds = Vec::new();
7113 loop {
7114 cmds.push(self.parse_one_statement()?);
7115 if matches!(self.peek(), Token::Semicolon) {
7116 self.advance();
7117 if matches!(self.peek(), Token::RParen) {
7118 break;
7119 }
7120 continue;
7121 }
7122 break;
7123 }
7124 if !matches!(self.peek(), Token::RParen) {
7125 return Err(self.err(alloc::format!(
7126 "expected ) closing the CREATE RULE command list, got {:?}",
7127 self.peek()
7128 )));
7129 }
7130 self.advance();
7131 cmds
7132 } else {
7133 alloc::vec![self.parse_one_statement()?]
7134 };
7135 Ok(Statement::CreateRule(crate::ast::CreateRuleStatement {
7136 name,
7137 or_replace,
7138 event,
7139 table,
7140 instead,
7141 when_condition,
7142 commands,
7143 }))
7144 }
7145
7146 /// v7.39 (round 139) — a rule event keyword → uppercase string.
7147 fn parse_rule_event(&mut self) -> Result<alloc::string::String, ParseError> {
7148 if matches!(self.peek(), Token::Insert) {
7149 self.advance();
7150 return Ok(alloc::string::String::from("INSERT"));
7151 }
7152 if matches!(self.peek(), Token::Select) {
7153 self.advance();
7154 return Ok(alloc::string::String::from("SELECT"));
7155 }
7156 match self.peek() {
7157 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("update") => {
7158 self.advance();
7159 Ok(alloc::string::String::from("UPDATE"))
7160 }
7161 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("delete") => {
7162 self.advance();
7163 Ok(alloc::string::String::from("DELETE"))
7164 }
7165 other => Err(self.err(alloc::format!(
7166 "expected INSERT / UPDATE / DELETE / SELECT in CREATE RULE, got {other:?}"
7167 ))),
7168 }
7169 }
7170
7171 /// v7.13.0 — parse one trigger event, then optionally consume
7172 /// `OF col, col, …` after `UPDATE` (mailrs round-5 G7). Other
7173 /// events (INSERT/DELETE/TRUNCATE) don't accept the OF tail.
7174 fn parse_trigger_event_with_optional_of(
7175 &mut self,
7176 ) -> Result<(TriggerEvent, Vec<String>), ParseError> {
7177 let ev = self.parse_trigger_event()?;
7178 if !matches!(ev, TriggerEvent::Update) {
7179 return Ok((ev, Vec::new()));
7180 }
7181 // `OF` is a bare ident.
7182 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("of")) {
7183 return Ok((ev, Vec::new()));
7184 }
7185 self.advance(); // OF
7186 let mut cols: Vec<String> = Vec::new();
7187 loop {
7188 cols.push(self.expect_ident_like()?);
7189 if matches!(self.peek(), Token::Comma) {
7190 self.advance();
7191 continue;
7192 }
7193 break;
7194 }
7195 if cols.is_empty() {
7196 return Err(
7197 self.err("CREATE TRIGGER: `UPDATE OF` requires at least one column name".into())
7198 );
7199 }
7200 Ok((ev, cols))
7201 }
7202
7203 /// v7.12.4 — `BEGIN stmt; stmt; … END[;]` PL/pgSQL block.
7204 /// v7.12.6 — optional `DECLARE var TYPE [:= init];` prelude
7205 /// before `BEGIN`, and IF / RAISE / embedded SQL statements
7206 /// inside the body.
7207 /// Called by [`parse_plpgsql_body`] after the body's tokens
7208 /// have been lexed into this temporary parser.
7209 pub(crate) fn parse_plpgsql_block(&mut self) -> Result<PlPgSqlBlock, ParseError> {
7210 // v7.12.6 — optional DECLARE prelude.
7211 let declarations = if matches!(
7212 self.peek(),
7213 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("declare")
7214 ) {
7215 self.advance();
7216 self.parse_plpgsql_declare_block()?
7217 } else {
7218 Vec::new()
7219 };
7220 // BEGIN keyword (PL/pgSQL — distinct from the SQL
7221 // `BEGIN` transaction-start, but we can reuse the
7222 // reserved Token::Begin since the body is a separate
7223 // lex/parse context).
7224 if !matches!(self.peek(), Token::Begin) {
7225 return Err(self.err(alloc::format!(
7226 "expected BEGIN at start of plpgsql block, got {:?}",
7227 self.peek()
7228 )));
7229 }
7230 self.advance();
7231 let statements = self.parse_plpgsql_stmt_list_until_end()?;
7232 // v7.37.20 (20.10) — optional EXCEPTION clause between the
7233 // body's last statement and the trailing END. When present
7234 // it's a series of `WHEN <cond> [OR <cond>]* THEN <body>`
7235 // arms terminated by END.
7236 let exception_handlers = if matches!(
7237 self.peek(),
7238 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("exception")
7239 ) {
7240 self.advance();
7241 self.parse_plpgsql_exception_handlers()?
7242 } else {
7243 Vec::new()
7244 };
7245 Ok(PlPgSqlBlock {
7246 declarations,
7247 statements,
7248 exception_handlers,
7249 })
7250 }
7251
7252 /// v7.37.20 (20.10) — parse EXCEPTION handlers `WHEN <cond>
7253 /// [OR <cond>]* THEN <body>` sequence up to the trailing END.
7254 fn parse_plpgsql_exception_handlers(
7255 &mut self,
7256 ) -> Result<Vec<crate::ast::ExceptionHandler>, ParseError> {
7257 let mut out: Vec<crate::ast::ExceptionHandler> = Vec::new();
7258 loop {
7259 // Stop at END — the block-level trailing END LOOP / END;
7260 // is handled by the caller.
7261 if matches!(
7262 self.peek(),
7263 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("end")
7264 ) {
7265 return Ok(out);
7266 }
7267 // WHEN <cond> [OR <cond>]* THEN <body>
7268 if !matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("when"))
7269 {
7270 return Err(self.err(alloc::format!(
7271 "expected WHEN or END inside EXCEPTION clause, got {:?}",
7272 self.peek()
7273 )));
7274 }
7275 self.advance();
7276 let mut conditions: Vec<String> = Vec::new();
7277 conditions.push(self.expect_ident_like()?);
7278 while matches!(self.peek(), Token::Or) {
7279 self.advance();
7280 conditions.push(self.expect_ident_like()?);
7281 }
7282 let then_kw = self.expect_ident_like()?;
7283 if !then_kw.eq_ignore_ascii_case("then") {
7284 return Err(self.err(alloc::format!(
7285 "expected THEN after WHEN condition list, got {then_kw:?}"
7286 )));
7287 }
7288 let body = self.parse_plpgsql_stmt_list_until_end()?;
7289 out.push(crate::ast::ExceptionHandler { conditions, body });
7290 }
7291 }
7292
7293 /// v7.12.6 — parse the `DECLARE ... [var TYPE [:= init];]+`
7294 /// prelude. Caller has already consumed `DECLARE`. We stop
7295 /// reading entries when we hit `BEGIN`.
7296 fn parse_plpgsql_declare_block(&mut self) -> Result<Vec<PlPgSqlDeclare>, ParseError> {
7297 let mut out: Vec<PlPgSqlDeclare> = Vec::new();
7298 loop {
7299 if matches!(self.peek(), Token::Begin) {
7300 return Ok(out);
7301 }
7302 let name = self.expect_ident_like()?;
7303 // v7.37.20 (20.7) — type inference: if the next token is
7304 // `:=` or `=` (no explicit type), infer from the default
7305 // expression. Otherwise the ident that follows is the
7306 // declared type.
7307 //
7308 // v7.37.20 (20.8) — `<table>.<col>%TYPE` / `<table>%ROWTYPE`
7309 // (PG-standard). SPG parse-accepts and treats identically
7310 // to inference — the eventual runtime value determines
7311 // the local's type, which is faithful to how SPG handles
7312 // untyped locals today (see 20.7). Full compile-time
7313 // catalog lookup queues with v7.40 PL/pgSQL epic.
7314 let ty = if matches!(self.peek(), Token::ColonEq | Token::Eq) {
7315 // Sentinel: `FunctionArgType::Raw("_infer_")` tells the
7316 // downstream declaration walker to type the local by
7317 // the runtime type of the default expression.
7318 FunctionArgType::Raw("_infer_".into())
7319 } else {
7320 let ty_token = self.expect_ident_like()?;
7321 // Detect `<ident>[.<ident>][%TYPE | %ROWTYPE]`:
7322 // consume optional `.<ident>` qualifier + `%<KW>`
7323 // suffix. Both qualifier and suffix map to _infer_.
7324 if matches!(self.peek(), Token::Dot) {
7325 self.advance();
7326 let _ = self.expect_ident_like()?;
7327 }
7328 if matches!(self.peek(), Token::Percent) {
7329 self.advance();
7330 // Consume the trailing TYPE / ROWTYPE ident.
7331 let _ = self.expect_ident_like()?;
7332 FunctionArgType::Raw("_infer_".into())
7333 } else {
7334 match map_type_ident_to_column_type_name(&ty_token) {
7335 Some(t) => FunctionArgType::Typed(t),
7336 None => FunctionArgType::Raw(ty_token),
7337 }
7338 }
7339 };
7340 let default = match self.peek() {
7341 Token::ColonEq => {
7342 self.advance();
7343 Some(self.parse_expr(0)?)
7344 }
7345 Token::Eq => {
7346 // PL/pgSQL also accepts `=` for the
7347 // DECLARE default (PG treats them the same
7348 // in this position).
7349 self.advance();
7350 Some(self.parse_expr(0)?)
7351 }
7352 _ => None,
7353 };
7354 // Mandatory `;` between declarations.
7355 if !matches!(self.peek(), Token::Semicolon) {
7356 return Err(self.err(alloc::format!(
7357 "expected ; after DECLARE entry for {name:?}, got {:?}",
7358 self.peek()
7359 )));
7360 }
7361 self.advance();
7362 out.push(PlPgSqlDeclare { name, ty, default });
7363 }
7364 }
7365
7366 /// v7.12.6 — parse PL/pgSQL statements up to (and consuming)
7367 /// the terminating `END;` (or `END IF;` etc — handled by the
7368 /// per-construct sub-parsers). Used by both the outer block
7369 /// and the IF/ELSE branch bodies.
7370 fn parse_plpgsql_stmt_list_until_end(&mut self) -> Result<Vec<PlPgSqlStmt>, ParseError> {
7371 let mut statements: Vec<PlPgSqlStmt> = Vec::new();
7372 loop {
7373 // Allow trailing semicolons + END.
7374 while matches!(self.peek(), Token::Semicolon) {
7375 self.advance();
7376 }
7377 // END / ELSE / ELSIF / EXCEPTION — handled by the caller.
7378 if matches!(
7379 self.peek(),
7380 Token::Ident(s) | Token::QuotedIdent(s)
7381 if s.eq_ignore_ascii_case("end")
7382 || s.eq_ignore_ascii_case("else")
7383 || s.eq_ignore_ascii_case("elsif")
7384 || s.eq_ignore_ascii_case("elseif")
7385 || s.eq_ignore_ascii_case("exception")
7386 || s.eq_ignore_ascii_case("when")
7387 ) {
7388 return Ok(statements);
7389 }
7390 // Otherwise: one statement, then expect `;` or
7391 // a block-terminator keyword.
7392 let stmt = self.parse_plpgsql_stmt()?;
7393 statements.push(stmt);
7394 match self.peek() {
7395 Token::Semicolon => {
7396 self.advance();
7397 }
7398 Token::Ident(s) | Token::QuotedIdent(s)
7399 if s.eq_ignore_ascii_case("end")
7400 || s.eq_ignore_ascii_case("else")
7401 || s.eq_ignore_ascii_case("elsif")
7402 || s.eq_ignore_ascii_case("elseif")
7403 || s.eq_ignore_ascii_case("exception")
7404 || s.eq_ignore_ascii_case("when") =>
7405 {
7406 // Final statement of the block without `;`.
7407 }
7408 other => {
7409 return Err(self.err(alloc::format!(
7410 "expected ; or END/ELSE/ELSIF after plpgsql statement, got {other:?}"
7411 )));
7412 }
7413 }
7414 }
7415 }
7416
7417 fn parse_plpgsql_stmt(&mut self) -> Result<PlPgSqlStmt, ParseError> {
7418 // RETURN keyword?
7419 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("return"))
7420 {
7421 self.advance();
7422 return self.parse_plpgsql_return();
7423 }
7424 // v7.12.6 — IF block.
7425 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("if"))
7426 {
7427 self.advance();
7428 return self.parse_plpgsql_if();
7429 }
7430 // v7.37.20 (20.6) — FOR <var> IN EXECUTE <string_expr> LOOP.
7431 // Detected by peeking that token pos+3 is Ident("execute").
7432 if matches!(self.peek(), Token::For)
7433 && matches!(
7434 self.tokens.get(self.pos + 1),
7435 Some(Token::Ident(_) | Token::QuotedIdent(_))
7436 )
7437 && matches!(self.tokens.get(self.pos + 2), Some(Token::In))
7438 && matches!(
7439 self.tokens.get(self.pos + 3),
7440 Some(Token::Ident(s) | Token::QuotedIdent(s)) if s.eq_ignore_ascii_case("execute")
7441 )
7442 {
7443 self.advance(); // FOR
7444 let var = self.expect_ident_like()?;
7445 self.advance(); // IN
7446 self.advance(); // EXECUTE
7447 // Prescan for LOOP at paren depth 0 so parse_expr stops
7448 // before the LOOP keyword (same trick as the bare-SELECT
7449 // ForQuery arm).
7450 let mut depth: i32 = 0;
7451 let mut loop_pos: Option<usize> = None;
7452 let mut scan = self.pos;
7453 while scan < self.tokens.len() {
7454 match self.tokens.get(scan) {
7455 Some(Token::LParen) => depth += 1,
7456 Some(Token::RParen) => depth -= 1,
7457 Some(Token::Ident(s) | Token::QuotedIdent(s))
7458 if depth == 0 && s.eq_ignore_ascii_case("loop") =>
7459 {
7460 loop_pos = Some(scan);
7461 break;
7462 }
7463 _ => {}
7464 }
7465 scan += 1;
7466 }
7467 let loop_pos = loop_pos.ok_or_else(|| {
7468 self.err(alloc::format!(
7469 "FOR <var> IN EXECUTE <expr> ... LOOP: no LOOP keyword found"
7470 ))
7471 })?;
7472 let saved_loop = self.tokens[loop_pos].clone();
7473 self.tokens[loop_pos] = Token::Semicolon;
7474 let expr_result = self.parse_expr(0);
7475 self.tokens[loop_pos] = saved_loop;
7476 let sql_expr = expr_result?;
7477 let loop_kw = self.expect_ident_like()?;
7478 if !loop_kw.eq_ignore_ascii_case("loop") {
7479 return Err(self.err(alloc::format!(
7480 "expected LOOP after FOR <var> IN EXECUTE <expr>, got {loop_kw:?}"
7481 )));
7482 }
7483 let body = self.parse_plpgsql_stmt_list_until_end()?;
7484 let end_kw = self.expect_ident_like()?;
7485 if !end_kw.eq_ignore_ascii_case("end") {
7486 return Err(self.err(alloc::format!(
7487 "expected END LOOP after FOR IN EXECUTE body, got {end_kw:?}"
7488 )));
7489 }
7490 let loop_kw2 = self.expect_ident_like()?;
7491 if !loop_kw2.eq_ignore_ascii_case("loop") {
7492 return Err(self.err(alloc::format!(
7493 "expected END LOOP after FOR IN EXECUTE body, got END {loop_kw2:?}"
7494 )));
7495 }
7496 return Ok(PlPgSqlStmt::ForExecute {
7497 var,
7498 sql_expr,
7499 body,
7500 });
7501 }
7502 // v7.37.20 (20.5) — FOR <var> IN <SELECT> LOOP.
7503 //
7504 // Two syntactic forms:
7505 // FOR var IN SELECT ... ORDER BY ... LOOP ...
7506 // FOR var IN (SELECT ...) LOOP ...
7507 //
7508 // Bare-SELECT form: to keep parse_select_stmt from swallowing
7509 // the trailing `LOOP` keyword as a table alias, we prescan
7510 // forward to find LOOP at paren depth 0, splice a fake
7511 // Semicolon at that position (so SELECT parses cleanly),
7512 // then re-splice LOOP back in.
7513 //
7514 // Paren-wrapped form: parse `(` `SELECT ...` `)` then expect
7515 // LOOP directly — no scan required.
7516 if matches!(self.peek(), Token::For)
7517 && matches!(
7518 self.tokens.get(self.pos + 1),
7519 Some(Token::Ident(_) | Token::QuotedIdent(_))
7520 )
7521 && matches!(self.tokens.get(self.pos + 2), Some(Token::In))
7522 && (matches!(self.tokens.get(self.pos + 3), Some(Token::Select))
7523 || matches!(self.tokens.get(self.pos + 3), Some(Token::LParen)))
7524 {
7525 self.advance(); // FOR
7526 let var = self.expect_ident_like()?;
7527 // IN
7528 self.advance();
7529 let query = if matches!(self.peek(), Token::LParen) {
7530 // Paren-wrapped SELECT.
7531 self.advance();
7532 let inner = self.parse_select_stmt()?;
7533 let Statement::Select(q) = inner else {
7534 return Err(self.err(alloc::format!(
7535 "expected SELECT inside (…), got {:?}",
7536 self.peek()
7537 )));
7538 };
7539 if !matches!(self.peek(), Token::RParen) {
7540 return Err(self.err(alloc::format!(
7541 "expected ')' after FOR-IN-SELECT body, got {:?}",
7542 self.peek()
7543 )));
7544 }
7545 self.advance();
7546 q
7547 } else {
7548 // Bare SELECT: prescan to find the LOOP boundary.
7549 let mut depth: i32 = 0;
7550 let mut loop_pos: Option<usize> = None;
7551 let mut scan = self.pos;
7552 while scan < self.tokens.len() {
7553 match self.tokens.get(scan) {
7554 Some(Token::LParen) => depth += 1,
7555 Some(Token::RParen) => depth -= 1,
7556 Some(Token::Ident(s) | Token::QuotedIdent(s))
7557 if depth == 0 && s.eq_ignore_ascii_case("loop") =>
7558 {
7559 loop_pos = Some(scan);
7560 break;
7561 }
7562 _ => {}
7563 }
7564 scan += 1;
7565 }
7566 let loop_pos = loop_pos.ok_or_else(|| {
7567 self.err(alloc::format!(
7568 "FOR <var> IN <SELECT> ... LOOP: no LOOP keyword found"
7569 ))
7570 })?;
7571 // Swap the LOOP token with a synthetic Semicolon so
7572 // parse_select_stmt stops there, then restore afterward.
7573 let saved_loop = self.tokens[loop_pos].clone();
7574 self.tokens[loop_pos] = Token::Semicolon;
7575 let parse_result = self.parse_select_stmt();
7576 self.tokens[loop_pos] = saved_loop;
7577 let inner = parse_result?;
7578 let Statement::Select(q) = inner else {
7579 return Err(self.err(alloc::format!(
7580 "expected SELECT after FOR <var> IN, got {:?}",
7581 self.peek()
7582 )));
7583 };
7584 q
7585 };
7586 let loop_kw = self.expect_ident_like()?;
7587 if !loop_kw.eq_ignore_ascii_case("loop") {
7588 return Err(self.err(alloc::format!(
7589 "expected LOOP after FOR <var> IN <SELECT>, got {loop_kw:?}"
7590 )));
7591 }
7592 let body = self.parse_plpgsql_stmt_list_until_end()?;
7593 let end_kw = self.expect_ident_like()?;
7594 if !end_kw.eq_ignore_ascii_case("end") {
7595 return Err(self.err(alloc::format!(
7596 "expected END LOOP after FOR IN SELECT body, got {end_kw:?}"
7597 )));
7598 }
7599 let loop_kw2 = self.expect_ident_like()?;
7600 if !loop_kw2.eq_ignore_ascii_case("loop") {
7601 return Err(self.err(alloc::format!(
7602 "expected END LOOP after FOR IN SELECT body, got END {loop_kw2:?}"
7603 )));
7604 }
7605 return Ok(PlPgSqlStmt::ForQuery {
7606 var,
7607 query: Box::new(query),
7608 body,
7609 });
7610 }
7611 // v7.37.20 (20.4) — FOR <var> IN [REVERSE] <start>..<end> LOOP.
7612 // FOR is a reserved keyword token (Token::For).
7613 if matches!(self.peek(), Token::For)
7614 && matches!(
7615 self.tokens.get(self.pos + 1),
7616 Some(Token::Ident(_) | Token::QuotedIdent(_))
7617 )
7618 && matches!(self.tokens.get(self.pos + 2), Some(Token::In))
7619 {
7620 self.advance(); // FOR
7621 let var = self.expect_ident_like()?;
7622 if !matches!(self.peek(), Token::In) {
7623 return Err(self.err(alloc::format!(
7624 "expected IN after FOR <var>, got {:?}",
7625 self.peek()
7626 )));
7627 }
7628 self.advance();
7629 let reverse = matches!(
7630 self.peek(),
7631 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("reverse")
7632 );
7633 if reverse {
7634 self.advance();
7635 }
7636 let start = self.parse_expr(0)?;
7637 if !matches!(self.peek(), Token::DotDot) {
7638 return Err(self.err(alloc::format!(
7639 "expected '..' between FOR loop bounds, got {:?}",
7640 self.peek()
7641 )));
7642 }
7643 self.advance();
7644 let end = self.parse_expr(0)?;
7645 let loop_kw = self.expect_ident_like()?;
7646 if !loop_kw.eq_ignore_ascii_case("loop") {
7647 return Err(self.err(alloc::format!(
7648 "expected LOOP after FOR <var> IN start..end, got {loop_kw:?}"
7649 )));
7650 }
7651 let body = self.parse_plpgsql_stmt_list_until_end()?;
7652 let end_kw = self.expect_ident_like()?;
7653 if !end_kw.eq_ignore_ascii_case("end") {
7654 return Err(self.err(alloc::format!(
7655 "expected END LOOP after FOR body, got {end_kw:?}"
7656 )));
7657 }
7658 let loop_kw2 = self.expect_ident_like()?;
7659 if !loop_kw2.eq_ignore_ascii_case("loop") {
7660 return Err(self.err(alloc::format!(
7661 "expected END LOOP after FOR body, got END {loop_kw2:?}"
7662 )));
7663 }
7664 return Ok(PlPgSqlStmt::ForRange {
7665 var,
7666 start,
7667 end,
7668 reverse,
7669 body,
7670 });
7671 }
7672 // v7.37.20 (20.2) — bare `LOOP <body> END LOOP;`.
7673 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("loop"))
7674 {
7675 self.advance();
7676 let body = self.parse_plpgsql_stmt_list_until_end()?;
7677 let end_kw = self.expect_ident_like()?;
7678 if !end_kw.eq_ignore_ascii_case("end") {
7679 return Err(self.err(alloc::format!(
7680 "expected END LOOP after LOOP body, got {end_kw:?}"
7681 )));
7682 }
7683 let loop_kw = self.expect_ident_like()?;
7684 if !loop_kw.eq_ignore_ascii_case("loop") {
7685 return Err(self.err(alloc::format!(
7686 "expected END LOOP after LOOP body, got END {loop_kw:?}"
7687 )));
7688 }
7689 return Ok(PlPgSqlStmt::Loop { body });
7690 }
7691 // v7.37.20 (20.2) — `EXIT [WHEN <cond>]` inside a loop.
7692 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("exit"))
7693 {
7694 self.advance();
7695 let when = if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("when"))
7696 {
7697 self.advance();
7698 Some(self.parse_expr(0)?)
7699 } else {
7700 None
7701 };
7702 return Ok(PlPgSqlStmt::Exit { when });
7703 }
7704 // v7.37.20 (20.13) — `EXECUTE <string_expr>`. Dispatches an
7705 // already-parsed Statement or a runtime-computed SQL string.
7706 // The disambiguator vs the extended-query-protocol `EXECUTE
7707 // <stmt_name>` (which is a top-level Statement, not a
7708 // plpgsql line) is that inside a DO block / trigger body the
7709 // EXECUTE keyword ALWAYS refers to dynamic SQL.
7710 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("execute"))
7711 {
7712 self.advance();
7713 let sql = self.parse_expr(0)?;
7714 return Ok(PlPgSqlStmt::ExecuteDynamic { sql });
7715 }
7716 // v7.37.20 (20.2) — `CONTINUE [WHEN <cond>]` inside a loop.
7717 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("continue"))
7718 {
7719 self.advance();
7720 let when = if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("when"))
7721 {
7722 self.advance();
7723 Some(self.parse_expr(0)?)
7724 } else {
7725 None
7726 };
7727 return Ok(PlPgSqlStmt::Continue { when });
7728 }
7729 // v7.37.20 (20.3) — WHILE <cond> LOOP <body> END LOOP.
7730 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("while"))
7731 {
7732 self.advance();
7733 let condition = self.parse_expr(0)?;
7734 let loop_kw = self.expect_ident_like()?;
7735 if !loop_kw.eq_ignore_ascii_case("loop") {
7736 return Err(self.err(alloc::format!(
7737 "expected LOOP after WHILE <condition>, got {loop_kw:?}"
7738 )));
7739 }
7740 let body = self.parse_plpgsql_stmt_list_until_end()?;
7741 // Expect END LOOP.
7742 let end_kw = self.expect_ident_like()?;
7743 if !end_kw.eq_ignore_ascii_case("end") {
7744 return Err(self.err(alloc::format!(
7745 "expected END LOOP after WHILE body, got {end_kw:?}"
7746 )));
7747 }
7748 let loop_kw2 = self.expect_ident_like()?;
7749 if !loop_kw2.eq_ignore_ascii_case("loop") {
7750 return Err(self.err(alloc::format!(
7751 "expected END LOOP after WHILE body, got END {loop_kw2:?}"
7752 )));
7753 }
7754 return Ok(PlPgSqlStmt::While { condition, body });
7755 }
7756 // v7.12.6 — RAISE.
7757 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("raise"))
7758 {
7759 self.advance();
7760 return self.parse_plpgsql_raise();
7761 }
7762 // v7.37.20 (20.14) — ASSERT <cond> [, <msg>].
7763 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("assert"))
7764 {
7765 self.advance();
7766 let condition = self.parse_expr(0)?;
7767 let message = if matches!(self.peek(), Token::Comma) {
7768 self.advance();
7769 Some(self.parse_expr(0)?)
7770 } else {
7771 None
7772 };
7773 return Ok(PlPgSqlStmt::Assert { condition, message });
7774 }
7775 // v7.37.20 (20.12) — PERFORM <select>. Per PG docs:
7776 // "PERFORM is equivalent to SELECT but discards the
7777 // result." Side effects (function calls, RAISE inside
7778 // SQL functions, etc.) still execute. We desugar to
7779 // `SELECT <body>` and wrap in EmbeddedSql so the engine's
7780 // existing embedded-statement path handles execution +
7781 // result-discard cleanly. The result is naturally
7782 // discarded because EmbeddedSql doesn't propagate row
7783 // sets back to the plpgsql interpreter.
7784 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("perform"))
7785 {
7786 self.advance();
7787 // Splice a synthetic Token::Select into the stream at
7788 // the current position so parse_select_stmt parses the
7789 // remainder as a normal SELECT body. Token-stream
7790 // surgery mirrors the try_parse_plpgsql_select_into
7791 // pattern used for SELECT … INTO desugaring.
7792 self.tokens.insert(self.pos, Token::Select);
7793 let select = self.parse_select_stmt()?;
7794 let Statement::Select(s) = select else {
7795 return Err(self.err(alloc::format!(
7796 "expected SELECT body after PERFORM, got {:?}",
7797 self.peek()
7798 )));
7799 };
7800 return Ok(PlPgSqlStmt::EmbeddedSql(Box::new(Statement::Select(s))));
7801 }
7802 // v7.16.2 — `SELECT <projection> INTO <var> [FROM …]`
7803 // plpgsql-specific shape (mailrs round-10 migrate-042).
7804 // PG's SELECT INTO at top-level SQL would CREATE a new
7805 // table; inside plpgsql it ASSIGNS the query result to
7806 // a local variable. We detect the INTO at paren-depth
7807 // 0 between SELECT and the statement boundary; if
7808 // found, split the token stream into "pre-INTO
7809 // projection" + "var" + "post-INTO FROM/WHERE…" and
7810 // rebuild as a SelectInto with a regular SELECT body
7811 // (no INTO clause).
7812 if matches!(self.peek(), Token::Select)
7813 && let Some((select_body, var_name)) = self.try_parse_plpgsql_select_into()?
7814 {
7815 return Ok(PlPgSqlStmt::SelectInto {
7816 var: var_name,
7817 body: Box::new(select_body),
7818 });
7819 }
7820 // v7.12.6 — embedded SQL statements. INSERT/UPDATE/DELETE/
7821 // SELECT can appear directly inside a trigger body; we
7822 // recurse into the regular Statement parser, which will
7823 // stop at the trailing `;` (which our caller then
7824 // consumes).
7825 // v7.16.2 — top-level DO blocks (mailrs round-10 A.2)
7826 // also embed ALTER / CREATE / DROP statements; route
7827 // those through the same parser so the DO body parses
7828 // cleanly.
7829 if matches!(self.peek(), Token::Insert)
7830 || matches!(self.peek(), Token::Select)
7831 || matches!(self.peek(), Token::Create)
7832 || matches!(self.peek(), Token::Drop)
7833 || matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
7834 if s.eq_ignore_ascii_case("update")
7835 || s.eq_ignore_ascii_case("delete")
7836 || s.eq_ignore_ascii_case("alter"))
7837 {
7838 let stmt = self.parse_one_statement()?;
7839 return Ok(PlPgSqlStmt::EmbeddedSql(Box::new(stmt)));
7840 }
7841 // Otherwise: assignment. `NEW.col` / `OLD.col` / `var`
7842 // followed by `:=` and an expression.
7843 let target = self.parse_plpgsql_assign_target()?;
7844 // PL/pgSQL assignment uses `:=`. The lexer represents
7845 // this as a colon followed by `=`; check both shapes.
7846 match self.peek() {
7847 Token::ColonEq => {
7848 self.advance();
7849 }
7850 Token::Colon => {
7851 self.advance();
7852 if !matches!(self.peek(), Token::Eq) {
7853 return Err(self.err(alloc::format!(
7854 "expected := after plpgsql assign target, got `:` then {:?}",
7855 self.peek()
7856 )));
7857 }
7858 self.advance();
7859 }
7860 other => {
7861 return Err(self.err(alloc::format!(
7862 "expected := after plpgsql assign target, got {other:?}"
7863 )));
7864 }
7865 }
7866 let value = self.parse_expr(0)?;
7867 Ok(PlPgSqlStmt::Assign { target, value })
7868 }
7869
7870 /// v7.12.6 — `IF cond THEN body [ELSIF cond THEN body]*
7871 /// [ELSE body] END IF`. `IF` keyword already consumed.
7872 fn parse_plpgsql_if(&mut self) -> Result<PlPgSqlStmt, ParseError> {
7873 let mut branches: Vec<(Expr, Vec<PlPgSqlStmt>)> = Vec::new();
7874 let mut else_branch: Vec<PlPgSqlStmt> = Vec::new();
7875 loop {
7876 // <expr> THEN
7877 let cond = self.parse_expr(0)?;
7878 let then_kw = self.expect_ident_like()?;
7879 if !then_kw.eq_ignore_ascii_case("then") {
7880 return Err(self.err(alloc::format!(
7881 "expected THEN after IF/ELSIF condition, got {then_kw:?}"
7882 )));
7883 }
7884 let body = self.parse_plpgsql_stmt_list_until_end()?;
7885 branches.push((cond, body));
7886 // Look at terminator: ELSIF/ELSEIF, ELSE, or END IF.
7887 match self.peek() {
7888 Token::Ident(s) | Token::QuotedIdent(s)
7889 if s.eq_ignore_ascii_case("elsif") || s.eq_ignore_ascii_case("elseif") =>
7890 {
7891 self.advance();
7892 continue;
7893 }
7894 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("else") => {
7895 self.advance();
7896 else_branch = self.parse_plpgsql_stmt_list_until_end()?;
7897 break;
7898 }
7899 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("end") => {
7900 break;
7901 }
7902 other => {
7903 return Err(self.err(alloc::format!(
7904 "expected ELSIF / ELSE / END after IF branch body, got {other:?}"
7905 )));
7906 }
7907 }
7908 }
7909 // Expect `END IF` (the END keyword is the one we're
7910 // looking at right now).
7911 let end_kw = self.expect_ident_like()?;
7912 if !end_kw.eq_ignore_ascii_case("end") {
7913 return Err(self.err(alloc::format!("expected END IF, got {end_kw:?}")));
7914 }
7915 let if_kw = self.expect_ident_like()?;
7916 if !if_kw.eq_ignore_ascii_case("if") {
7917 return Err(self.err(alloc::format!("expected END IF, got END {if_kw:?}")));
7918 }
7919 Ok(PlPgSqlStmt::If {
7920 branches,
7921 else_branch,
7922 })
7923 }
7924
7925 /// v7.12.6 — `RAISE { NOTICE | WARNING | INFO | LOG | DEBUG
7926 /// | EXCEPTION } '<message>' [, args]*`. The `RAISE` keyword
7927 /// is already consumed.
7928 fn parse_plpgsql_raise(&mut self) -> Result<PlPgSqlStmt, ParseError> {
7929 let lvl_ident = self.expect_ident_like()?;
7930 let level = match lvl_ident.to_ascii_lowercase().as_str() {
7931 "notice" => RaiseLevel::Notice,
7932 "warning" => RaiseLevel::Warning,
7933 "info" => RaiseLevel::Info,
7934 "log" => RaiseLevel::Log,
7935 "debug" => RaiseLevel::Debug,
7936 "exception" => RaiseLevel::Exception,
7937 other => {
7938 return Err(self.err(alloc::format!(
7939 "expected RAISE level (NOTICE/WARNING/INFO/LOG/DEBUG/EXCEPTION), got {other:?}"
7940 )));
7941 }
7942 };
7943 // Message: required for v7.12.6. PG accepts a bare
7944 // RAISE-rethrow form (no message), reserved for future
7945 // RAISE-no-args support.
7946 let Token::String(msg) = self.peek() else {
7947 return Err(self.err(alloc::format!(
7948 "expected RAISE message string, got {:?}",
7949 self.peek()
7950 )));
7951 };
7952 let message = msg.clone();
7953 self.advance();
7954 // Optional comma-separated args (PG `%` format substitution).
7955 let mut args: Vec<Expr> = Vec::new();
7956 while matches!(self.peek(), Token::Comma) {
7957 self.advance();
7958 args.push(self.parse_expr(0)?);
7959 }
7960 Ok(PlPgSqlStmt::Raise {
7961 level,
7962 message,
7963 args,
7964 })
7965 }
7966
7967 /// v7.16.2 — scan ahead for a plpgsql-flavoured `SELECT
7968 /// <projection> INTO <var> [FROM …]` (mailrs round-10
7969 /// migrate-042). Returns `(rebuilt_select_without_into,
7970 /// var_name)` when the pattern matches; `None` for
7971 /// regular SELECTs (those go through the embedded-SQL
7972 /// path). Token-stream surgery so the rebuilt SELECT
7973 /// parses through the regular `parse_select_stmt`.
7974 #[allow(clippy::too_many_lines)]
7975 fn try_parse_plpgsql_select_into(
7976 &mut self,
7977 ) -> Result<Option<(SelectStatement, String)>, ParseError> {
7978 // Scan forward from `self.pos + 1` (past Token::Select)
7979 // for Token::Into at paren-depth 0, stopping at the
7980 // first `;`, `END`, `ELSE`, `ELSIF` keyword that would
7981 // end the plpgsql statement.
7982 let start = self.pos;
7983 let mut into_pos: Option<usize> = None;
7984 let mut depth: i32 = 0;
7985 let mut i = start + 1;
7986 while i < self.tokens.len() {
7987 match &self.tokens[i] {
7988 Token::LParen => depth += 1,
7989 Token::RParen => depth -= 1,
7990 Token::Semicolon if depth == 0 => break,
7991 Token::Ident(s)
7992 if depth == 0
7993 && (s.eq_ignore_ascii_case("end")
7994 || s.eq_ignore_ascii_case("else")
7995 || s.eq_ignore_ascii_case("elsif")) =>
7996 {
7997 break;
7998 }
7999 Token::Into if depth == 0 => {
8000 into_pos = Some(i);
8001 break;
8002 }
8003 _ => {}
8004 }
8005 i += 1;
8006 }
8007 let Some(into_at) = into_pos else {
8008 return Ok(None);
8009 };
8010 // The token immediately after INTO must be the target
8011 // var ident; anything else (e.g. INSERT INTO table)
8012 // ruled out by the depth-0 check above. Capture it.
8013 let var = match self.tokens.get(into_at + 1) {
8014 Some(Token::Ident(s) | Token::QuotedIdent(s)) => s.clone(),
8015 other => {
8016 return Err(self.err(alloc::format!(
8017 "expected variable name after SELECT … INTO, got {other:?}"
8018 )));
8019 }
8020 };
8021 // Find the end of the plpgsql SELECT INTO statement —
8022 // same boundary rules as the depth-0 scan above.
8023 let mut end = into_at + 2;
8024 let mut depth2: i32 = 0;
8025 while end < self.tokens.len() {
8026 match &self.tokens[end] {
8027 Token::LParen => depth2 += 1,
8028 Token::RParen => depth2 -= 1,
8029 Token::Semicolon if depth2 == 0 => break,
8030 Token::Ident(s)
8031 if depth2 == 0
8032 && (s.eq_ignore_ascii_case("end")
8033 || s.eq_ignore_ascii_case("else")
8034 || s.eq_ignore_ascii_case("elsif")) =>
8035 {
8036 break;
8037 }
8038 _ => {}
8039 }
8040 end += 1;
8041 }
8042 // Rebuild a token stream that represents the SELECT
8043 // WITHOUT the INTO clause: [SELECT .. up-to-INTO] + [
8044 // post-var tokens up to statement end]. Run the
8045 // regular `parse_select_stmt` against it.
8046 let mut rebuilt: Vec<Token> = Vec::with_capacity(end - start);
8047 for j in start..into_at {
8048 rebuilt.push(self.tokens[j].clone());
8049 }
8050 for j in (into_at + 2)..end {
8051 rebuilt.push(self.tokens[j].clone());
8052 }
8053 rebuilt.push(Token::Eof);
8054 let saved_pos = self.pos;
8055 let saved_tokens = core::mem::replace(&mut self.tokens, rebuilt);
8056 self.pos = 0;
8057 // parse_select_stmt → parse_bare_select consumes Token::Select itself.
8058 if !matches!(self.peek(), Token::Select) {
8059 self.tokens = saved_tokens;
8060 self.pos = saved_pos;
8061 return Err(self.err("plpgsql SELECT … INTO: rebuilt stream missing SELECT".into()));
8062 }
8063 let sel = self.parse_select_stmt();
8064 self.tokens = saved_tokens;
8065 self.pos = end;
8066 let sel = sel?;
8067 let Statement::Select(body) = sel else {
8068 return Err(self.err(alloc::format!(
8069 "plpgsql SELECT … INTO: rebuilt SELECT did not produce a Select node, got {sel:?}"
8070 )));
8071 };
8072 Ok(Some((body, var)))
8073 }
8074
8075 fn parse_plpgsql_assign_target(&mut self) -> Result<AssignTarget, ParseError> {
8076 // v7.16.1 — read the head token DIRECTLY rather than
8077 // via `expect_ident_like`. The v7.14.0 schema-qualifier
8078 // strip (`public.t` → `t`) inside `expect_ident_like`
8079 // greedily consumes any `ident . ident` pair, which
8080 // silently turned every `NEW.col := …` /
8081 // `OLD.col := …` plpgsql assignment into a Local("col")
8082 // assignment — the head "new"/"old" was eaten as if it
8083 // were a schema name and the Dot was consumed too, so
8084 // this function's own `peek() == Token::Dot` check
8085 // below never fired. Every BEFORE trigger that rewrote
8086 // a NEW cell was a silent no-op for two major releases
8087 // (v7.14.0 + v7.15.0) until the e2e_trigger workspace-
8088 // gate failures were investigated as v7.16.1 backlog.
8089 let head = match self.advance() {
8090 Token::Ident(s) | Token::QuotedIdent(s) => s,
8091 other => {
8092 return Err(self.err(alloc::format!(
8093 "expected NEW / OLD / <local_var> as plpgsql assign target, got {other:?}"
8094 )));
8095 }
8096 };
8097 if matches!(self.peek(), Token::Dot) {
8098 self.advance();
8099 let col = self.expect_ident_like()?;
8100 if head.eq_ignore_ascii_case("new") {
8101 return Ok(AssignTarget::NewColumn(col));
8102 }
8103 if head.eq_ignore_ascii_case("old") {
8104 return Ok(AssignTarget::OldColumn(col));
8105 }
8106 return Err(self.err(alloc::format!(
8107 "plpgsql assign target must be NEW.<col> / OLD.<col> / <local_var>; \
8108 got {head:?}.<col>"
8109 )));
8110 }
8111 Ok(AssignTarget::Local(head))
8112 }
8113
8114 fn parse_plpgsql_return(&mut self) -> Result<PlPgSqlStmt, ParseError> {
8115 // RETURN NEW / OLD / NULL — bare-ident forms.
8116 match self.peek() {
8117 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("new") => {
8118 self.advance();
8119 return Ok(PlPgSqlStmt::Return(ReturnTarget::New));
8120 }
8121 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("old") => {
8122 self.advance();
8123 return Ok(PlPgSqlStmt::Return(ReturnTarget::Old));
8124 }
8125 Token::Null => {
8126 self.advance();
8127 return Ok(PlPgSqlStmt::Return(ReturnTarget::Null));
8128 }
8129 // Bare `RETURN;` (no value) — treated as `RETURN NULL`
8130 // per PL/pgSQL convention.
8131 Token::Semicolon => {
8132 return Ok(PlPgSqlStmt::Return(ReturnTarget::Null));
8133 }
8134 _ => {}
8135 }
8136 // v7.37.20 (20.11) — RETURN QUERY <select> / RETURN QUERY
8137 // EXECUTE <expr>. In a DO block context RETURN QUERY has no
8138 // caller-visible effect (blocks don't return sets), so we
8139 // desugar it identically to PERFORM: parse the SELECT (or
8140 // EXECUTE dynamic) as embedded SQL that runs for side
8141 // effects and discards the result. RETURN NEXT <expr>
8142 // (single-row accumulator) queues with v7.40 SETOF function
8143 // infrastructure.
8144 // v7.39 (read01 round 66) — `RETURN NEXT <expr>`: append a row to the set
8145 // and keep going.
8146 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("next"))
8147 {
8148 self.advance();
8149 let e = self.parse_expr(0)?;
8150 return Ok(PlPgSqlStmt::ReturnNext(e));
8151 }
8152 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("query"))
8153 {
8154 self.advance();
8155 // v7.39 (read01 round 68) — `RETURN QUERY EXECUTE <sql expr>`: the
8156 // rows go to the set, like the static form. It used to desugar to a
8157 // bare ExecuteDynamic, whose result was DISCARDED.
8158 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("execute"))
8159 {
8160 self.advance();
8161 let sql = self.parse_expr(0)?;
8162 return Ok(PlPgSqlStmt::ReturnQueryExecute { sql });
8163 }
8164 // Bare RETURN QUERY <select>. If the current token is
8165 // not already SELECT (e.g., the user wrote `RETURN QUERY
8166 // <projection> FROM ...` in a shorthand — rare but PG
8167 // accepts a bare projection here), splice one in. Same
8168 // trick as PERFORM.
8169 if !matches!(self.peek(), Token::Select) {
8170 self.tokens.insert(self.pos, Token::Select);
8171 }
8172 let select = self.parse_select_stmt()?;
8173 let Statement::Select(s) = select else {
8174 return Err(self.err(alloc::format!(
8175 "expected SELECT body after RETURN QUERY, got {:?}",
8176 self.peek()
8177 )));
8178 };
8179 // v7.39 (read01 round 66) — a REAL statement now. It used to desugar
8180 // to an embedded side-effect SELECT whose rows were DISCARDED, which
8181 // in a SETOF function is the entire answer thrown away.
8182 return Ok(PlPgSqlStmt::ReturnQuery(Box::new(s)));
8183 }
8184 // Fall through: parse a full expression.
8185 let e = self.parse_expr(0)?;
8186 Ok(PlPgSqlStmt::Return(ReturnTarget::Expr(e)))
8187 }
8188
8189 fn parse_trigger_event(&mut self) -> Result<TriggerEvent, ParseError> {
8190 // INSERT is a reserved Token; UPDATE / DELETE / TRUNCATE
8191 // are ident-shaped (the parser keys off case-insensitive
8192 // match — same shape used by the top-level Update / Delete
8193 // dispatchers at parse_one_statement).
8194 if matches!(self.peek(), Token::Insert) {
8195 self.advance();
8196 return Ok(TriggerEvent::Insert);
8197 }
8198 match self.peek() {
8199 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("update") => {
8200 self.advance();
8201 Ok(TriggerEvent::Update)
8202 }
8203 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("delete") => {
8204 self.advance();
8205 Ok(TriggerEvent::Delete)
8206 }
8207 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("truncate") => {
8208 self.advance();
8209 Ok(TriggerEvent::Truncate)
8210 }
8211 other => Err(self.err(alloc::format!(
8212 "expected INSERT / UPDATE / DELETE / TRUNCATE in trigger event list, got {other:?}"
8213 ))),
8214 }
8215 }
8216
8217 /// v6.1.2 → v6.1.3 — `CREATE PUBLICATION <name>` body. Accepts:
8218 /// - (no clause) → implicit `FOR ALL TABLES`
8219 /// - `FOR ALL TABLES`
8220 /// - `FOR ALL TABLES EXCEPT t1, t2, …` (v6.1.3)
8221 /// - `FOR TABLE t1, t2, …` (v6.1.3) — `FOR TABLES …` also
8222 /// accepted as an SPG lenience. PG18-measured (round 753): PG
8223 /// REJECTS the bare plural (`invalid publication object list`,
8224 /// TABLES only pairs with IN SCHEMA); the old note claimed an
8225 /// unverifiable "PG 19 accepts both". Ledgered, not load-bearing.
8226 fn parse_create_publication_after_keyword(&mut self) -> Result<Statement, ParseError> {
8227 let name = self.expect_ident_or_string()?;
8228 // Bare DDL maps to FOR ALL TABLES — matches the v6.1.2
8229 // shape so existing publications keep parsing identically.
8230 let scope = if matches!(self.peek(), Token::For) {
8231 self.advance();
8232 if matches!(self.peek(), Token::All) {
8233 self.advance();
8234 if !matches!(self.peek(), Token::Tables) {
8235 return Err(self.err(format!(
8236 "expected TABLES after FOR ALL, got {:?}",
8237 self.peek()
8238 )));
8239 }
8240 self.advance();
8241 if matches!(self.peek(), Token::Except) {
8242 self.advance();
8243 let tables = self.parse_publication_table_list()?;
8244 PublicationScope::AllTablesExcept(tables)
8245 } else {
8246 PublicationScope::AllTables
8247 }
8248 } else if matches!(self.peek(), Token::Table) {
8249 self.advance();
8250 let tables = self.parse_publication_table_list()?;
8251 PublicationScope::ForTables(tables)
8252 } else if matches!(self.peek(), Token::Tables) {
8253 // v7.39 (round 754, F31-B5) — PG18-measured: the bare
8254 // plural (`FOR TABLES t`) is REJECTED (`invalid
8255 // publication object list`); TABLES only pairs with
8256 // `IN SCHEMA`. The old arm accepted it on an
8257 // unverifiable "PG 19 accepts both" claim.
8258 self.advance();
8259 if !matches!(self.peek(), Token::In) {
8260 return Err(self.err(alloc::string::String::from(
8261 "invalid publication object list",
8262 )));
8263 }
8264 self.advance();
8265 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("schema")) {
8266 return Err(self.err(format!(
8267 "expected SCHEMA after FOR TABLES IN, got {:?}",
8268 self.peek()
8269 )));
8270 }
8271 self.advance();
8272 let schema = self.expect_ident_or_string()?;
8273 PublicationScope::TablesInSchema(schema)
8274 } else {
8275 return Err(self.err(format!(
8276 "expected ALL TABLES or TABLE <list> after FOR, got {:?}",
8277 self.peek()
8278 )));
8279 }
8280 } else {
8281 PublicationScope::AllTables
8282 };
8283 Ok(Statement::CreatePublication(CreatePublicationStatement {
8284 name,
8285 scope,
8286 }))
8287 }
8288
8289 /// v6.1.3 — Comma-separated identifier list for the publication
8290 /// FOR-clause. Requires at least one entry; empty list is a
8291 /// parse error (PG behaviour). Quoted idents are accepted; the
8292 /// names round-trip through `Display` as `quote_ident(name)`.
8293 ///
8294 /// v7.37.21 (21.2 + 21.3) — accept-and-discard the per-table
8295 /// `(col_list) WHERE (predicate)` modifiers PG 15+ emits in
8296 /// pg_dump output. SPG's publication state today is per-table
8297 /// only (matching the pre-PG-15 surface); the col list + WHERE
8298 /// are parsed so dumps load through and the table name reaches
8299 /// `PublicationScope::ForTables`, but the filter is not enforced
8300 /// at publish time. Re-open when a customer dogfood gate
8301 /// requires per-row-filter or column-subset publish semantics
8302 /// (which gates on persistent slot state landing first, 21.12).
8303 fn parse_publication_table_list(&mut self) -> Result<Vec<String>, ParseError> {
8304 let first = self.parse_publication_table_entry()?;
8305 let mut out = alloc::vec![first];
8306 while matches!(self.peek(), Token::Comma) {
8307 self.advance();
8308 out.push(self.parse_publication_table_entry()?);
8309 }
8310 Ok(out)
8311 }
8312
8313 /// One table entry inside a FOR TABLE clause:
8314 /// tab_name [ (col, col, …) ] [ WHERE (predicate) ]
8315 /// Returns just the table name; the column list + WHERE predicate
8316 /// are consumed and discarded per the parse-accept-discard
8317 /// commitment above.
8318 fn parse_publication_table_entry(&mut self) -> Result<String, ParseError> {
8319 let name = self.expect_ident_like()?;
8320 // Optional column list — `(col, col, …)`.
8321 if matches!(self.peek(), Token::LParen) {
8322 self.advance();
8323 // Empty parens are a PG error too; require ≥ 1 column.
8324 let _ = self.expect_ident_like()?;
8325 while matches!(self.peek(), Token::Comma) {
8326 self.advance();
8327 let _ = self.expect_ident_like()?;
8328 }
8329 if !matches!(self.peek(), Token::RParen) {
8330 return Err(self.err(alloc::format!(
8331 "expected ')' to close publication column list, got {:?}",
8332 self.peek()
8333 )));
8334 }
8335 self.advance();
8336 }
8337 // Optional row filter — `WHERE (predicate)`.
8338 if matches!(self.peek(), Token::Where) {
8339 self.advance();
8340 if !matches!(self.peek(), Token::LParen) {
8341 return Err(self.err(alloc::format!(
8342 "expected '(' after WHERE in publication row filter, got {:?}",
8343 self.peek()
8344 )));
8345 }
8346 self.advance();
8347 let _ = self.parse_expr(0)?;
8348 if !matches!(self.peek(), Token::RParen) {
8349 return Err(self.err(alloc::format!(
8350 "expected ')' to close publication WHERE filter, got {:?}",
8351 self.peek()
8352 )));
8353 }
8354 self.advance();
8355 }
8356 Ok(name)
8357 }
8358
8359 /// v6.1.4 — `CREATE SUBSCRIPTION <name>
8360 /// CONNECTION '<conn>'
8361 /// PUBLICATION <pub> [, <pub> ...]`.
8362 ///
8363 /// The clause order is fixed (CONNECTION first, then
8364 /// PUBLICATION) to match PG. No WITH-options accepted in
8365 /// v6.1.4 — `enabled` defaults to true, no other knobs ship.
8366 fn parse_create_subscription_after_keyword(&mut self) -> Result<Statement, ParseError> {
8367 let name = self.expect_ident_or_string()?;
8368 if !matches!(self.peek(), Token::Connection) {
8369 return Err(self.err(format!(
8370 "expected CONNECTION after CREATE SUBSCRIPTION <name>, got {:?}",
8371 self.peek()
8372 )));
8373 }
8374 self.advance();
8375 let conn_str = self.expect_string_literal()?;
8376 if !matches!(self.peek(), Token::Publication) {
8377 return Err(self.err(format!(
8378 "expected PUBLICATION after CONNECTION '<conn>', got {:?}",
8379 self.peek()
8380 )));
8381 }
8382 self.advance();
8383 // Reuse the publication FOR-list parser shape: at least one
8384 // identifier, comma-separated.
8385 let first = self.expect_ident_like()?;
8386 let mut publications = alloc::vec![first];
8387 while matches!(self.peek(), Token::Comma) {
8388 self.advance();
8389 publications.push(self.expect_ident_like()?);
8390 }
8391 Ok(Statement::CreateSubscription(CreateSubscriptionStatement {
8392 name,
8393 conn_str,
8394 publications,
8395 }))
8396 }
8397
8398 /// v6.1.7 — `WAIT FOR WAL POSITION <pos> [WITH TIMEOUT <ms>]`.
8399 /// All keywords after `WAIT` are bare idents in v6.1.x; no
8400 /// lexer churn. Both `<pos>` and `<ms>` are positive integers
8401 /// that fit `u64`.
8402 /// Parameter name in `SET <name>`. A GUC name may be dotted, but the
8403 /// qualifier is a *namespace* the app owns (`app.user_id`,
8404 /// `myapp.tenant` — the request-context / RLS pattern), NOT a schema
8405 /// to discard. So parse the raw segments here instead of
8406 /// `expect_ident_like`, which strips a leading `schema.` qualifier
8407 /// and would collapse `SET app.foo` to just `foo`. Standard GUCs are
8408 /// a single segment and round-trip unchanged.
8409 fn parse_set_param_name(&mut self) -> Result<String, ParseError> {
8410 let mut parts: alloc::vec::Vec<String> = alloc::vec::Vec::new();
8411 loop {
8412 let seg = match self.advance() {
8413 Token::Ident(s) | Token::QuotedIdent(s) => s,
8414 other if unreserved_keyword_text(&other).is_some() => {
8415 unreserved_keyword_text(&other).unwrap()
8416 }
8417 other => {
8418 return Err(ParseError {
8419 message: format!("expected parameter name, got {other:?}"),
8420 token_pos: self.consumed_pos(),
8421 });
8422 }
8423 };
8424 parts.push(seg);
8425 if matches!(self.peek(), Token::Dot) {
8426 self.advance();
8427 continue;
8428 }
8429 break;
8430 }
8431 Ok(parts.join(".").to_ascii_lowercase())
8432 }
8433
8434 fn parse_set_value(&mut self) -> Result<crate::ast::SetValue, ParseError> {
8435 Self::parse_set_value_inner(self)
8436 }
8437
8438 fn parse_set_value_inner(&mut self) -> Result<crate::ast::SetValue, ParseError> {
8439 match self.advance() {
8440 Token::String(s) => Ok(crate::ast::SetValue::String(s)),
8441 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("default") => {
8442 Ok(crate::ast::SetValue::Default)
8443 }
8444 Token::Ident(s) | Token::QuotedIdent(s) => {
8445 let mut accum = s;
8446 while matches!(self.peek(), Token::Dot) {
8447 self.advance();
8448 let next = self.expect_ident_like()?;
8449 accum.push('.');
8450 accum.push_str(&next);
8451 }
8452 Ok(crate::ast::SetValue::Ident(accum))
8453 }
8454 Token::Integer(n) => Ok(crate::ast::SetValue::Number(n.to_string())),
8455 Token::Float(f) => Ok(crate::ast::SetValue::Number(f.to_string())),
8456 // v7.22 (mailrs round-13 gap 2) — PG boolean parameter
8457 // spellings that lex as keyword tokens, not idents:
8458 // `SET standard_conforming_strings = on` is in every
8459 // pg_dump preamble (`off` already lexes as an ident).
8460 // v7.39 (round 769, F31 tranche 5 #150) — `SET x TO DEFAULT`:
8461 // DEFAULT lexes as its keyword token, so the ident arm above
8462 // never saw it and the everyday reset form was a syntax error.
8463 Token::Default => Ok(crate::ast::SetValue::Default),
8464 Token::On => Ok(crate::ast::SetValue::Ident("on".to_string())),
8465 Token::True => Ok(crate::ast::SetValue::Ident("true".to_string())),
8466 Token::False => Ok(crate::ast::SetValue::Ident("false".to_string())),
8467 // v7.14.0 — MySQL session/user variable RHS
8468 // (e.g. `SET OLD_FOREIGN_KEY_CHECKS = @@FOREIGN_KEY_CHECKS`).
8469 // Wrap as Ident so the SET handler can record it; the
8470 // engine treats `@VAR` / `@@VAR` values as opaque
8471 // strings.
8472 Token::SessionVar(s) => Ok(crate::ast::SetValue::Ident(s)),
8473 // v7.14.0 — `SET sql_mode = 'NO_AUTO_VALUE_ON_ZERO,STRICT_TRANS_TABLES'`
8474 // is the common MySQL preamble shape. Allow a `+` or
8475 // `-` prefix on negative numerics for parity with PG
8476 // (some param defaults are negative).
8477 Token::Minus => match self.advance() {
8478 Token::Integer(n) => Ok(crate::ast::SetValue::Number(alloc::format!("-{n}"))),
8479 Token::Float(f) => Ok(crate::ast::SetValue::Number(alloc::format!("-{f}"))),
8480 other => Err(self.err(format!(
8481 "expected numeric after `-` in SET value, got {other:?}"
8482 ))),
8483 },
8484 other => Err(self.err(format!(
8485 "expected literal, identifier, or DEFAULT after `=` in SET, got {other:?}"
8486 ))),
8487 }
8488 }
8489
8490 /// v7.38 轴 4 — `[ISOLATION LEVEL …] [READ ONLY|WRITE]
8491 /// [[NOT] DEFERRABLE]` modes after `SET TRANSACTION` or
8492 /// `START TRANSACTION` / `BEGIN`. Returns the isolation level
8493 /// (default `ReadCommitted` if no `ISOLATION LEVEL` clause was
8494 /// present). Modes are comma-separated per PG; SPG also
8495 /// accepts space-separated for tolerance. READ ONLY / WRITE
8496 /// / DEFERRABLE are parsed-and-ignored (recorded for future
8497 /// surface but not behaviorally honoured today).
8498 /// Parse the trailing `[ISOLATION LEVEL …] [READ ONLY|WRITE]
8499 /// [[NOT] DEFERRABLE]` modes of BEGIN / START TRANSACTION / SET
8500 /// TRANSACTION. Returns `Some(level)` only when an explicit `ISOLATION
8501 /// LEVEL` clause was given, so a bare `BEGIN` / `BEGIN READ ONLY` keeps the
8502 /// session default rather than forcing READ COMMITTED.
8503 fn parse_isolation_level_clauses(
8504 &mut self,
8505 ) -> Result<crate::ast::TransactionModes, ParseError> {
8506 let mut level = IsolationLevel::default();
8507 let mut have_level = false;
8508 // v7.39 — READ ONLY / READ WRITE used to be consumed and dropped,
8509 // so `BEGIN READ ONLY` opened an ordinary read-write transaction.
8510 let mut read_only: Option<bool> = None;
8511 loop {
8512 // ISOLATION LEVEL …
8513 let saw_isolation =
8514 matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("isolation"));
8515 if saw_isolation {
8516 self.advance(); // ISOLATION
8517 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("level")) {
8518 return Err(self.err(alloc::format!(
8519 "expected LEVEL after ISOLATION, got {:?}",
8520 self.peek()
8521 )));
8522 }
8523 self.advance(); // LEVEL
8524 // SERIALIZABLE | REPEATABLE READ | READ COMMITTED | READ UNCOMMITTED
8525 let w1 = self
8526 .expect_ident_like()
8527 .map_err(|e| self.err(alloc::format!("isolation level: {e:?}")))?;
8528 let lc = w1.to_ascii_lowercase();
8529 level = match lc.as_str() {
8530 "serializable" => IsolationLevel::Serializable,
8531 "repeatable" => {
8532 // Expect READ
8533 let w2 = self
8534 .expect_ident_like()
8535 .map_err(|e| self.err(alloc::format!("REPEATABLE …: {e:?}")))?;
8536 if !w2.eq_ignore_ascii_case("read") {
8537 return Err(self.err(alloc::format!(
8538 "expected READ after REPEATABLE, got {w2:?}"
8539 )));
8540 }
8541 IsolationLevel::RepeatableRead
8542 }
8543 "read" => {
8544 let w2 = self
8545 .expect_ident_like()
8546 .map_err(|e| self.err(alloc::format!("READ …: {e:?}")))?;
8547 match w2.to_ascii_lowercase().as_str() {
8548 "committed" => IsolationLevel::ReadCommitted,
8549 "uncommitted" => IsolationLevel::ReadUncommitted,
8550 other => {
8551 return Err(self.err(alloc::format!(
8552 "expected COMMITTED or UNCOMMITTED after READ, got {other:?}"
8553 )));
8554 }
8555 }
8556 }
8557 other => {
8558 return Err(self.err(alloc::format!("unknown isolation level {other:?}")));
8559 }
8560 };
8561 have_level = true;
8562 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("read")) {
8563 // v7.39 — READ ONLY | READ WRITE. The comment here used to
8564 // read "parsed, not behaviorally honoured", and it was
8565 // accurate: the clause was thrown away, so `BEGIN READ ONLY`
8566 // opened an ordinary read-write transaction and accepted
8567 // every write in it.
8568 self.advance();
8569 match self.peek().clone() {
8570 Token::Ident(s) if s.eq_ignore_ascii_case("only") => {
8571 self.advance();
8572 read_only = Some(true);
8573 }
8574 Token::Ident(s) if s.eq_ignore_ascii_case("write") => {
8575 self.advance();
8576 read_only = Some(false);
8577 }
8578 other => {
8579 return Err(self.err(alloc::format!(
8580 "expected ONLY or WRITE after READ, got {other:?}"
8581 )));
8582 }
8583 }
8584 } else if matches!(self.peek(), Token::Not) {
8585 // NOT DEFERRABLE — `NOT` lexes as a reserved keyword.
8586 self.advance();
8587 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("deferrable")) {
8588 return Err(self.err(alloc::format!(
8589 "expected DEFERRABLE after NOT, got {:?}",
8590 self.peek()
8591 )));
8592 }
8593 self.advance();
8594 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("deferrable"))
8595 {
8596 self.advance();
8597 } else {
8598 break;
8599 }
8600 // Optional comma between modes.
8601 if matches!(self.peek(), Token::Comma) {
8602 self.advance();
8603 }
8604 }
8605 Ok(crate::ast::TransactionModes {
8606 isolation: have_level.then_some(level),
8607 read_only,
8608 })
8609 }
8610
8611 fn parse_wait_after_keyword(&mut self) -> Result<Statement, ParseError> {
8612 // FOR is a v6.1.2-reserved keyword (Token::For). The
8613 // other two are bare idents — they've never needed lexer
8614 // support and we keep it that way.
8615 if !matches!(self.peek(), Token::For) {
8616 return Err(self.err(format!("expected FOR after WAIT, got {:?}", self.peek())));
8617 }
8618 self.advance();
8619 self.expect_keyword_ident("wal")?;
8620 self.expect_keyword_ident("position")?;
8621 let pos = self.expect_u64_literal()?;
8622 let timeout_ms = if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("with"))
8623 {
8624 self.advance();
8625 self.expect_keyword_ident("timeout")?;
8626 Some(self.expect_u64_literal()?)
8627 } else {
8628 None
8629 };
8630 Ok(Statement::WaitForWalPosition { pos, timeout_ms })
8631 }
8632
8633 /// v6.1.7 helper — consume a `Token::Integer` and check it
8634 /// fits `u64`. WAL positions and millisecond timeouts are
8635 /// non-negative.
8636 fn expect_u64_literal(&mut self) -> Result<u64, ParseError> {
8637 match self.advance() {
8638 Token::Integer(n) if n >= 0 => Ok(n as u64),
8639 Token::Integer(n) => Err(ParseError {
8640 message: format!("expected non-negative integer, got {n}"),
8641 token_pos: self.consumed_pos(),
8642 }),
8643 other => Err(ParseError {
8644 message: format!("expected integer literal, got {other:?}"),
8645 token_pos: self.consumed_pos(),
8646 }),
8647 }
8648 }
8649
8650 /// `CREATE USER` body — name + WITH PASSWORD '<pw>' + optional
8651 /// ROLE '<role>' (defaults to readonly). All string slots accept
8652 /// either a quoted ident or a quoted string literal.
8653 /// `CREATE {USER|ROLE} name [WITH] [PASSWORD 'x'] [LOGIN|NOLOGIN]
8654 /// [INHERIT|NOINHERIT] [SUPERUSER|NOSUPERUSER] [ROLE 'admin']`.
8655 ///
8656 /// `is_user` = the statement said USER, which in PG means LOGIN by default.
8657 /// The legacy SPG `ROLE 'readwrite'` clause (the coarse read/write/admin
8658 /// wire role) still parses — it is a different axis from the PG attributes.
8659 /// v7.39 (round 547) — is this ALTER ROLE / DATABASE one of the
8660 /// SET forms? Peeks past the name (and an `IN DATABASE d`) for SET
8661 /// or RESET, so the plain attribute forms keep their old path.
8662 fn peeks_db_role_setting(&self) -> bool {
8663 let mut i = self.pos + 1; // past the object's name
8664 let word = |p: usize| -> Option<String> {
8665 match self.tokens.get(p) {
8666 Some(Token::Ident(s) | Token::QuotedIdent(s)) => Some(s.to_ascii_lowercase()),
8667 Some(Token::In) => Some(String::from("in")),
8668 _ => None,
8669 }
8670 };
8671 if word(i).as_deref() == Some("in") && word(i + 1).as_deref() == Some("database") {
8672 i += 3; // IN DATABASE <name>
8673 }
8674 matches!(word(i).as_deref(), Some("set" | "reset"))
8675 }
8676
8677 fn parse_db_role_setting(&mut self, is_database: bool) -> Result<Statement, ParseError> {
8678 use crate::ast::SetDbRoleSettingStatement;
8679 // `ALTER ROLE ALL SET …` — ALL lexes as a KEYWORD, not an
8680 // identifier, so the ordinary name reader refuses it. Same trap
8681 // as TABLE / INDEX / FULL / DEFAULT before it.
8682 let name = if matches!(self.peek(), Token::All) {
8683 self.advance();
8684 String::from("all")
8685 } else {
8686 self.expect_ident_or_string()?
8687 };
8688 // `ALTER ROLE ALL SET …` is PG's every-role scope (oid 0).
8689 let all = name.eq_ignore_ascii_case("all");
8690 let (mut database, mut role) = if is_database {
8691 (Some(name), None)
8692 } else if all {
8693 (None, None)
8694 } else {
8695 (None, Some(name))
8696 };
8697 if matches!(self.peek(), Token::In) {
8698 self.advance();
8699 self.advance(); // DATABASE
8700 database = Some(self.expect_ident_or_string()?);
8701 }
8702 let resetting = matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("reset"));
8703 self.advance(); // SET | RESET
8704 if resetting && matches!(self.peek(), Token::All) {
8705 self.advance();
8706 self.consume_until_statement_boundary();
8707 return Ok(Statement::SetDbRoleSetting(Box::new(
8708 SetDbRoleSettingStatement {
8709 database,
8710 role,
8711 param: None,
8712 value: None,
8713 },
8714 )));
8715 }
8716 let param = self.expect_ident_like()?;
8717 let value = if resetting {
8718 None
8719 } else {
8720 // `SET p = v` and PG's `SET p TO v` both. TO lexes as a
8721 // KEYWORD, so the ident-only check missed it and consumed
8722 // the word itself as the value — the same trap as ALL, one
8723 // clause over.
8724 if matches!(self.peek(), Token::Eq | Token::To) || self.peek_keyword_ident("to") {
8725 self.advance();
8726 }
8727 Some(self.take_guc_value())
8728 };
8729 self.consume_until_statement_boundary();
8730 Ok(Statement::SetDbRoleSetting(Box::new(
8731 SetDbRoleSettingStatement {
8732 database,
8733 role,
8734 param: Some(param),
8735 value,
8736 },
8737 )))
8738 }
8739
8740 /// The remainder of a `SET <p> = …` clause as PG renders it back:
8741 /// a quoted literal loses its quotes, a bare word or number does not.
8742 fn take_guc_value(&mut self) -> String {
8743 match self.advance() {
8744 Token::String(s) => s,
8745 Token::Integer(n) => format!("{n}"),
8746 Token::Float(f) => format!("{f}"),
8747 Token::Ident(s) | Token::QuotedIdent(s) => s,
8748 other => format!("{other:?}"),
8749 }
8750 }
8751
8752 fn parse_create_user_after_keyword(&mut self, is_user: bool) -> Result<Statement, ParseError> {
8753 let name = self.expect_ident_or_string()?;
8754 if self.peek_keyword_ident("with") {
8755 self.advance();
8756 }
8757 let mut password = String::new();
8758 let mut role = String::new();
8759 let mut login: Option<bool> = None;
8760 let mut inherit: Option<bool> = None;
8761 let mut superuser: Option<bool> = None;
8762 // Not a `while let`: the pattern would borrow `self` across the
8763 // body, which calls `self.advance()` / `self.expect_*` (&mut).
8764 #[allow(clippy::while_let_loop)]
8765 loop {
8766 let (Token::Ident(w) | Token::QuotedIdent(w)) = self.peek() else {
8767 break;
8768 };
8769 match w.to_ascii_lowercase().as_str() {
8770 "password" => {
8771 self.advance();
8772 password = self.expect_string_literal()?;
8773 }
8774 // PG accepts (and pg_dump emits) ENCRYPTED PASSWORD; the value
8775 // is the same slot.
8776 "encrypted" => {
8777 self.advance();
8778 self.expect_keyword_ident("password")?;
8779 password = self.expect_string_literal()?;
8780 }
8781 "login" => {
8782 self.advance();
8783 login = Some(true);
8784 }
8785 "nologin" => {
8786 self.advance();
8787 login = Some(false);
8788 }
8789 "inherit" => {
8790 self.advance();
8791 inherit = Some(true);
8792 }
8793 "noinherit" => {
8794 self.advance();
8795 inherit = Some(false);
8796 }
8797 "superuser" => {
8798 self.advance();
8799 superuser = Some(true);
8800 }
8801 "nosuperuser" => {
8802 self.advance();
8803 superuser = Some(false);
8804 }
8805 // SPG's own coarse wire role: `ROLE 'readwrite'`.
8806 "role" => {
8807 self.advance();
8808 role = self.expect_string_literal()?;
8809 }
8810 // Every other PG role option (CREATEDB, CONNECTION LIMIT n,
8811 // VALID UNTIL '…', CREATEROLE, REPLICATION, BYPASSRLS …) is
8812 // accepted and ignored so a pg_dump role block restores. They
8813 // gate capabilities SPG does not have.
8814 "createdb" | "nocreatedb" | "createrole" | "nocreaterole" | "replication"
8815 | "noreplication" | "bypassrls" | "nobypassrls" => {
8816 self.advance();
8817 }
8818 "connection" => {
8819 self.advance();
8820 self.expect_keyword_ident("limit")?;
8821 self.advance(); // the number
8822 }
8823 "valid" => {
8824 self.advance();
8825 self.expect_keyword_ident("until")?;
8826 self.expect_string_literal()?;
8827 }
8828 _ => break,
8829 }
8830 }
8831 if role.is_empty() {
8832 role = "readonly".to_string();
8833 }
8834 Ok(Statement::CreateUser(crate::ast::CreateUserStatement {
8835 name,
8836 password,
8837 role,
8838 login,
8839 inherit,
8840 superuser,
8841 is_user,
8842 }))
8843 }
8844
8845 /// v7.39 (RLS) — parenthesised policy qualifier `( <expr> )`; caller has
8846 /// consumed the USING / WITH CHECK keyword.
8847 fn parse_paren_expr(&mut self, clause: &str) -> Result<Expr, ParseError> {
8848 if !matches!(self.peek(), Token::LParen) {
8849 return Err(self.err(alloc::format!(
8850 "expected '(' after {clause}, got {:?}",
8851 self.peek()
8852 )));
8853 }
8854 self.advance();
8855 let e = self.parse_expr(0)?;
8856 if !matches!(self.peek(), Token::RParen) {
8857 return Err(self.err(alloc::format!(
8858 "expected ')' to close {clause}, got {:?}",
8859 self.peek()
8860 )));
8861 }
8862 self.advance();
8863 Ok(e)
8864 }
8865
8866 /// v7.39 (RLS) — `TO role [, role]*`; caller has consumed `TO`.
8867 fn parse_policy_roles(&mut self) -> Result<Vec<String>, ParseError> {
8868 let mut roles = Vec::new();
8869 loop {
8870 roles.push(self.expect_ident_like()?);
8871 if matches!(self.peek(), Token::Comma) {
8872 self.advance();
8873 } else {
8874 break;
8875 }
8876 }
8877 Ok(roles)
8878 }
8879
8880 /// v7.39 (RLS) — `CREATE POLICY name ON table [AS {PERMISSIVE|RESTRICTIVE}]
8881 /// [FOR cmd] [TO roles] [USING (expr)] [WITH CHECK (expr)]`. Caller consumed
8882 /// `CREATE POLICY`.
8883 fn parse_create_policy_after_keyword(&mut self) -> Result<Statement, ParseError> {
8884 use crate::ast::PolicyCmd;
8885 let name = self.expect_ident_like()?;
8886 if !matches!(self.peek(), Token::On) {
8887 return Err(self.err(alloc::format!(
8888 "expected ON after CREATE POLICY name, got {:?}",
8889 self.peek()
8890 )));
8891 }
8892 self.advance();
8893 let table = self.expect_ident_like()?;
8894
8895 let mut permissive = true;
8896 if matches!(self.peek(), Token::As) {
8897 self.advance();
8898 let w = self.expect_ident_like()?;
8899 permissive = if w.eq_ignore_ascii_case("permissive") {
8900 true
8901 } else if w.eq_ignore_ascii_case("restrictive") {
8902 false
8903 } else {
8904 return Err(self.err(alloc::format!(
8905 "expected PERMISSIVE or RESTRICTIVE after AS, got {w:?}"
8906 )));
8907 };
8908 }
8909
8910 let mut cmd = PolicyCmd::All;
8911 if matches!(self.peek(), Token::For) {
8912 self.advance();
8913 cmd = self.parse_policy_cmd()?;
8914 }
8915
8916 let mut roles = Vec::new();
8917 if matches!(self.peek(), Token::To) {
8918 self.advance();
8919 roles = self.parse_policy_roles()?;
8920 }
8921
8922 let mut using = None;
8923 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("using"))
8924 {
8925 self.advance();
8926 using = Some(self.parse_paren_expr("USING")?);
8927 }
8928
8929 let mut with_check = None;
8930 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("with"))
8931 {
8932 self.advance();
8933 if !matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("check"))
8934 {
8935 return Err(self.err(alloc::format!(
8936 "expected CHECK after WITH, got {:?}",
8937 self.peek()
8938 )));
8939 }
8940 self.advance();
8941 with_check = Some(self.parse_paren_expr("WITH CHECK")?);
8942 }
8943
8944 // Clause-per-command matrix (PG wording).
8945 match cmd {
8946 PolicyCmd::Insert => {
8947 if using.is_some() {
8948 return Err(self.err("only WITH CHECK expression allowed for INSERT".into()));
8949 }
8950 }
8951 PolicyCmd::Select | PolicyCmd::Delete => {
8952 if with_check.is_some() {
8953 return Err(self.err("WITH CHECK cannot be applied to SELECT or DELETE".into()));
8954 }
8955 }
8956 PolicyCmd::Update | PolicyCmd::All => {}
8957 }
8958
8959 Ok(Statement::CreatePolicy(crate::ast::CreatePolicyStatement {
8960 name,
8961 table,
8962 permissive,
8963 cmd,
8964 roles,
8965 using,
8966 with_check,
8967 }))
8968 }
8969
8970 /// v7.39 (RLS) — the command word after `FOR`.
8971 fn parse_policy_cmd(&mut self) -> Result<crate::ast::PolicyCmd, ParseError> {
8972 use crate::ast::PolicyCmd;
8973 match self.peek().clone() {
8974 Token::All => {
8975 self.advance();
8976 Ok(PolicyCmd::All)
8977 }
8978 Token::Select => {
8979 self.advance();
8980 Ok(PolicyCmd::Select)
8981 }
8982 Token::Insert => {
8983 self.advance();
8984 Ok(PolicyCmd::Insert)
8985 }
8986 Token::Ident(s) if s.eq_ignore_ascii_case("update") => {
8987 self.advance();
8988 Ok(PolicyCmd::Update)
8989 }
8990 Token::Ident(s) if s.eq_ignore_ascii_case("delete") => {
8991 self.advance();
8992 Ok(PolicyCmd::Delete)
8993 }
8994 other => Err(self.err(alloc::format!(
8995 "expected ALL/SELECT/INSERT/UPDATE/DELETE after FOR, got {other:?}"
8996 ))),
8997 }
8998 }
8999
9000 /// v7.39 (RLS) — `ALTER POLICY name ON table { RENAME TO new | [TO roles]
9001 /// [USING (expr)] [WITH CHECK (expr)] }`. Caller consumed `ALTER POLICY`.
9002 fn parse_alter_policy_after_keyword(&mut self) -> Result<Statement, ParseError> {
9003 let name = self.expect_ident_like()?;
9004 if !matches!(self.peek(), Token::On) {
9005 return Err(self.err(alloc::format!(
9006 "expected ON after ALTER POLICY name, got {:?}",
9007 self.peek()
9008 )));
9009 }
9010 self.advance();
9011 let table = self.expect_ident_like()?;
9012
9013 // RENAME TO new
9014 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("rename"))
9015 {
9016 self.advance();
9017 if !matches!(self.peek(), Token::To) {
9018 return Err(self.err(alloc::format!(
9019 "expected TO after RENAME, got {:?}",
9020 self.peek()
9021 )));
9022 }
9023 self.advance();
9024 let new = self.expect_ident_like()?;
9025 return Ok(Statement::AlterPolicy(crate::ast::AlterPolicyStatement {
9026 name,
9027 table,
9028 rename_to: Some(new),
9029 roles: None,
9030 using: None,
9031 with_check: None,
9032 }));
9033 }
9034
9035 let mut roles = None;
9036 if matches!(self.peek(), Token::To) {
9037 self.advance();
9038 roles = Some(self.parse_policy_roles()?);
9039 }
9040 let mut using = None;
9041 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("using"))
9042 {
9043 self.advance();
9044 using = Some(self.parse_paren_expr("USING")?);
9045 }
9046 let mut with_check = None;
9047 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("with"))
9048 {
9049 self.advance();
9050 if !matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("check"))
9051 {
9052 return Err(self.err(alloc::format!(
9053 "expected CHECK after WITH, got {:?}",
9054 self.peek()
9055 )));
9056 }
9057 self.advance();
9058 with_check = Some(self.parse_paren_expr("WITH CHECK")?);
9059 }
9060 Ok(Statement::AlterPolicy(crate::ast::AlterPolicyStatement {
9061 name,
9062 table,
9063 rename_to: None,
9064 roles,
9065 using,
9066 with_check,
9067 }))
9068 }
9069
9070 /// v7.39 (RLS) — `DROP POLICY [IF EXISTS] name ON table`. Caller consumed
9071 /// `DROP POLICY`.
9072 fn parse_drop_policy_after_keyword(&mut self) -> Result<Statement, ParseError> {
9073 let if_exists = self.consume_if_exists();
9074 let name = self.expect_ident_like()?;
9075 if !matches!(self.peek(), Token::On) {
9076 return Err(self.err(alloc::format!(
9077 "expected ON after DROP POLICY name, got {:?}",
9078 self.peek()
9079 )));
9080 }
9081 self.advance();
9082 let table = self.expect_ident_like()?;
9083 Ok(Statement::DropPolicy(crate::ast::DropPolicyStatement {
9084 name,
9085 table,
9086 if_exists,
9087 }))
9088 }
9089}
9090fn wrap_from_leaves(
9091 e: &mut Expr,
9092 names: &[String],
9093 make: &dyn Fn(Expr) -> Expr,
9094 refs: &dyn Fn(&Expr) -> bool,
9095) {
9096 if let Expr::Column(c) = e {
9097 if c.qualifier
9098 .as_deref()
9099 .is_some_and(|q| names.iter().any(|n| n.eq_ignore_ascii_case(q)))
9100 {
9101 let taken = core::mem::replace(e, Expr::Literal(Literal::Null));
9102 *e = make(taken);
9103 }
9104 return;
9105 }
9106 match e {
9107 Expr::Binary { lhs, rhs, .. } => {
9108 wrap_from_leaves(lhs, names, make, refs);
9109 wrap_from_leaves(rhs, names, make, refs);
9110 }
9111 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } | Expr::IsNull { expr, .. } => {
9112 wrap_from_leaves(expr, names, make, refs)
9113 }
9114 Expr::FunctionCall { args, .. } => {
9115 for a in args.iter_mut() {
9116 wrap_from_leaves(a, names, make, refs);
9117 }
9118 }
9119 Expr::Case {
9120 operand,
9121 branches,
9122 else_branch,
9123 } => {
9124 if let Some(o) = operand.as_deref_mut() {
9125 wrap_from_leaves(o, names, make, refs);
9126 }
9127 for (w, t) in branches.iter_mut() {
9128 wrap_from_leaves(w, names, make, refs);
9129 wrap_from_leaves(t, names, make, refs);
9130 }
9131 if let Some(el) = else_branch.as_deref_mut() {
9132 wrap_from_leaves(el, names, make, refs);
9133 }
9134 }
9135 // Compound variants the walk doesn't decompose: keep the
9136 // pre-D.30 behavior — wrap the whole sub-expr if it touches
9137 // a source table, so nothing regresses.
9138 other => {
9139 if refs(other) {
9140 let taken = core::mem::replace(other, Expr::Literal(Literal::Null));
9141 *other = make(taken);
9142 }
9143 }
9144 }
9145}
9146
9147/// v7.39 (round 241) — does this expression reference any of the FROM /
9148/// USING table names (shared by the UPDATE…FROM and DELETE…USING
9149/// lowerings)?
9150fn expr_refs_tables(e: &Expr, names: &[String]) -> bool {
9151 match e {
9152 Expr::Column(c) => c
9153 .qualifier
9154 .as_deref()
9155 .is_some_and(|q| names.iter().any(|n| n.eq_ignore_ascii_case(q))),
9156 Expr::Binary { lhs, rhs, .. } => {
9157 expr_refs_tables(lhs, names) || expr_refs_tables(rhs, names)
9158 }
9159 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } => expr_refs_tables(expr, names),
9160 Expr::FunctionCall { args, .. } => args.iter().any(|a| expr_refs_tables(a, names)),
9161 Expr::Case {
9162 operand,
9163 branches,
9164 else_branch,
9165 } => {
9166 operand
9167 .as_deref()
9168 .is_some_and(|o| expr_refs_tables(o, names))
9169 || branches
9170 .iter()
9171 .any(|(w, t)| expr_refs_tables(w, names) || expr_refs_tables(t, names))
9172 || else_branch
9173 .as_deref()
9174 .is_some_and(|el| expr_refs_tables(el, names))
9175 }
9176 _ => false,
9177 }
9178}
9179
9180impl Parser {
9181 /// v4.4 `UPDATE <table> SET col = expr [, col = expr]* [WHERE cond]`.
9182 /// Caller already consumed the leading `UPDATE` ident.
9183 /// v7.39 (round 420) — does a JOIN clause start here? Used to spot
9184 /// MySQL's multi-table `UPDATE a JOIN b ON …` / `UPDATE a LEFT JOIN b …`
9185 /// after the target name has been read. `JOIN` is a reserved token;
9186 /// the qualifiers are bare idents.
9187 fn peek_is_update_join_start(&self) -> bool {
9188 match self.peek() {
9189 // JOIN and its qualifiers are reserved lexer tokens (the grammar
9190 // dedicates arms to `LEFT [OUTER] JOIN` and friends).
9191 Token::Join
9192 | Token::Inner
9193 | Token::Left
9194 | Token::Right
9195 | Token::Cross
9196 | Token::Full => true,
9197 // NATURAL / STRAIGHT_JOIN arrive as bare idents.
9198 Token::Ident(s) | Token::QuotedIdent(s) => {
9199 matches!(s.to_ascii_lowercase().as_str(), "natural" | "straight_join")
9200 }
9201 _ => false,
9202 }
9203 }
9204
9205 /// v7.39 (round 430) — `SET @x = <expr> [, @y := <expr>]`, MySQL's
9206 /// USER-variable assignment. Its own per-session namespace, an arbitrary
9207 /// expression on the right, and `:=` as a second spelling of `=`.
9208 ///
9209 /// Out-of-line (`inline(never)`): the statement-parse frame it is called
9210 /// from sits on the nesting recursion chain (a CTE body, a subquery),
9211 /// and holding this loop's `Vec` + `String` locals there overflowed the
9212 /// 512 KiB guard (`e2e_in_list_depth::round25_union_cte_search_shape`).
9213 #[inline(never)]
9214 fn parse_set_user_vars(&mut self) -> Result<Statement, ParseError> {
9215 let mut assigns: Vec<(String, Expr)> = Vec::new();
9216 let mut settings: Vec<(String, Expr)> = Vec::new();
9217 loop {
9218 // v7.39 (round 554) — a plain NAME here is a session
9219 // setting, not a user variable. mysqldump writes the two in
9220 // one statement — `SET @OLD_SQL_MODE=@@SQL_MODE,
9221 // SQL_MODE='NO_AUTO_VALUE_ON_ZERO'` saves a value and
9222 // changes it — and this refused the mixture outright, so no
9223 // dump could be restored past its preamble.
9224 if let Token::Ident(name) | Token::QuotedIdent(name) = self.peek().clone() {
9225 self.advance();
9226 if !matches!(self.peek(), Token::Eq | Token::ColonEq) {
9227 return Err(self.err(alloc::format!(
9228 "expected `=` after {name}, got {:?}",
9229 self.peek()
9230 )));
9231 }
9232 self.advance();
9233 let value = self.parse_expr(0)?;
9234 settings.push((name.to_ascii_lowercase(), value));
9235 if matches!(self.peek(), Token::Comma) {
9236 self.advance();
9237 continue;
9238 }
9239 break;
9240 }
9241 let Token::SessionVar(raw) = self.peek().clone() else {
9242 return Err(self.err(alloc::format!(
9243 "expected a user variable after SET, got {:?}",
9244 self.peek()
9245 )));
9246 };
9247 if raw.starts_with("@@") {
9248 return Err(self.err(alloc::string::String::from(
9249 "cannot mix `@@` settings with `@` user variables in one SET",
9250 )));
9251 }
9252 self.advance();
9253 if !matches!(self.peek(), Token::Eq | Token::ColonEq) {
9254 return Err(self.err(alloc::format!(
9255 "expected `=` or `:=` after {raw}, got {:?}",
9256 self.peek()
9257 )));
9258 }
9259 self.advance();
9260 let value = self.parse_expr(0)?;
9261 assigns.push((raw.trim_start_matches('@').to_ascii_lowercase(), value));
9262 if matches!(self.peek(), Token::Comma) {
9263 self.advance();
9264 continue;
9265 }
9266 break;
9267 }
9268 Ok(Statement::SetUserVars(assigns, settings))
9269 }
9270
9271 fn parse_update_after_keyword(&mut self) -> Result<Statement, ParseError> {
9272 // v7.39 (round 646) — `UPDATE ONLY t SET …`. Read as a table
9273 // NAMED `only` until now, which failed on `relation "only" does
9274 // not exist`. The lookahead is what keeps a table actually
9275 // called `only` working: the keyword is only a keyword when a
9276 // TABLE NAME follows it — and `SET` arrives as an identifier
9277 // here, so `UPDATE only SET a = 2` would otherwise take `SET`
9278 // for the table and die on the `=`. Measured by the pin.
9279 let only = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
9280 if s.eq_ignore_ascii_case("only"))
9281 && matches!(
9282 self.tokens.get(self.pos + 1),
9283 Some(Token::Ident(n) | Token::QuotedIdent(n)) if !n.eq_ignore_ascii_case("set")
9284 );
9285 if only {
9286 self.advance();
9287 }
9288 let table = self.expect_ident_like()?;
9289 // v7.39 (round 241) — `UPDATE t [AS] alias SET …`. PG allows the
9290 // bare spelling; a bare identifier that is the SET keyword itself
9291 // is the clause, not an alias.
9292 // v7.39 (round 420) — nor is a bare join qualifier (`LEFT` / `INNER`
9293 // / …) an alias: `UPDATE a LEFT JOIN b …` starts the MySQL
9294 // multi-table form, and swallowing `LEFT` as `a`'s alias made the
9295 // following JOIN a syntax error.
9296 let starts_join = self.mysql_dialect && self.peek_is_update_join_start();
9297 let alias = if matches!(self.peek(), Token::As) {
9298 self.advance();
9299 Some(self.expect_ident_like()?)
9300 } else {
9301 match self.peek() {
9302 Token::Ident(s) | Token::QuotedIdent(s)
9303 if !s.eq_ignore_ascii_case("set") && !starts_join =>
9304 {
9305 let a = s.clone();
9306 self.advance();
9307 Some(a)
9308 }
9309 _ => None,
9310 }
9311 };
9312 // v7.39 (round 420) — MySQL's multi-table UPDATE:
9313 // UPDATE a, b SET a.v = b.v WHERE a.id = b.id
9314 // UPDATE a JOIN b ON a.id = b.id SET a.v = b.v + 1
9315 // UPDATE a LEFT JOIN b ON a.id = b.id SET a.v = COALESCE(b.v, -1)
9316 // The FIRST table is the mutation target and the rest are sources —
9317 // exactly the shape PG spells `UPDATE a SET … FROM b WHERE …`, which
9318 // SPG already lowers onto correlated subqueries. So rewind, let
9319 // `parse_from_clause` read the whole list (it handles aliases, comma
9320 // lists, and every JOIN form), then peel the target off the front.
9321 let (mysql_from, mysql_on, mysql_outer) = if self.mysql_dialect
9322 && (matches!(self.peek(), Token::Comma) || self.peek_is_update_join_start())
9323 {
9324 // NOTE: `advance()` destroys the tokens it returns
9325 // (`mem::replace(.., Eof)`), so re-parsing by rewinding `self.pos`
9326 // is NOT possible — the tail is read forward, once, through the
9327 // same grammar `parse_from_clause` uses after its primary.
9328 let target_qual = alias.clone().unwrap_or_else(|| table.clone());
9329 let mut joins = self.parse_from_joins(&target_qual)?;
9330 if joins.is_empty() {
9331 return Err(self.err(alloc::string::String::from(
9332 "multi-table UPDATE needs at least one source table",
9333 )));
9334 }
9335 let head = joins.remove(0);
9336 // A LEFT join keeps every target row (the unmatched ones see NULL
9337 // on the source side), so it must NOT get the EXISTS row filter
9338 // the inner / comma forms use.
9339 let outer = matches!(head.kind, crate::ast::JoinKind::Left);
9340 let src = FromClause {
9341 primary: head.table,
9342 joins,
9343 };
9344 (Some(src), head.on, outer)
9345 } else {
9346 (None, None, false)
9347 };
9348 self.expect_keyword_ident("set")?;
9349 let mut assignments = Vec::new();
9350 loop {
9351 // `SET (a, b) = (e1, e2)` / `SET (a, b) = (SELECT x, y
9352 // …)` — the parenthesized multi-assignment. Expressions
9353 // assign positionally; a subquery RHS clones per column
9354 // keeping only the Nth projection item.
9355 if matches!(self.peek(), Token::LParen) {
9356 self.advance();
9357 let mut cols = alloc::vec![self.expect_ident_like()?];
9358 while matches!(self.peek(), Token::Comma) {
9359 self.advance();
9360 cols.push(self.expect_ident_like()?);
9361 }
9362 if !matches!(self.peek(), Token::RParen) {
9363 return Err(self.err(format!(
9364 "expected ')' after SET column list, got {:?}",
9365 self.peek()
9366 )));
9367 }
9368 self.advance();
9369 if !matches!(self.peek(), Token::Eq) {
9370 return Err(self.err(format!(
9371 "expected `=` after SET column list, got {:?}",
9372 self.peek()
9373 )));
9374 }
9375 self.advance();
9376 if !matches!(self.peek(), Token::LParen) {
9377 return Err(self.err(format!(
9378 "expected '(' after SET (…) =, got {:?}",
9379 self.peek()
9380 )));
9381 }
9382 self.advance();
9383 if matches!(self.peek(), Token::Select) {
9384 let inner = match self.parse_select_stmt()? {
9385 Statement::Select(s) => s,
9386 other => {
9387 return Err(self.err(alloc::format!(
9388 "expected SELECT in SET (…) = (SELECT …), got {other:?}"
9389 )));
9390 }
9391 };
9392 if !matches!(self.peek(), Token::RParen) {
9393 return Err(self.err(format!(
9394 "expected ')' after SET subquery, got {:?}",
9395 self.peek()
9396 )));
9397 }
9398 self.advance();
9399 if inner.items.len() != cols.len() {
9400 return Err(self.err(alloc::format!(
9401 "SET (…) = (SELECT …) arity mismatch: {} columns, {} items",
9402 cols.len(),
9403 inner.items.len()
9404 )));
9405 }
9406 for (i, col) in cols.into_iter().enumerate() {
9407 let mut sub = inner.clone();
9408 sub.items = alloc::vec![sub.items[i].clone()];
9409 assignments.push((col, Expr::ScalarSubquery(Box::new(sub))));
9410 }
9411 } else {
9412 let mut exprs = alloc::vec![self.parse_expr(0)?];
9413 while matches!(self.peek(), Token::Comma) {
9414 self.advance();
9415 exprs.push(self.parse_expr(0)?);
9416 }
9417 if !matches!(self.peek(), Token::RParen) {
9418 return Err(self.err(format!(
9419 "expected ')' after SET row values, got {:?}",
9420 self.peek()
9421 )));
9422 }
9423 self.advance();
9424 if exprs.len() != cols.len() {
9425 return Err(self.err(alloc::format!(
9426 "SET (…) = (…) arity mismatch: {} columns, {} values",
9427 cols.len(),
9428 exprs.len()
9429 )));
9430 }
9431 for (col, e) in cols.into_iter().zip(exprs) {
9432 assignments.push((col, e));
9433 }
9434 }
9435 if matches!(self.peek(), Token::Comma) {
9436 self.advance();
9437 continue;
9438 }
9439 break;
9440 }
9441 // v7.39 (round 420) — MySQL's multi-table UPDATE qualifies its
9442 // assignment targets (`SET a.v = b.v`). `expect_ident_like`
9443 // SILENTLY strips a `<qual>.` prefix (it exists for PG's
9444 // `public.` dump qualifiers), so the qualifier has to be read off
9445 // the token stream first — otherwise `SET b.v = 888` would write
9446 // to the TARGET table's `v` while naming a source table, a
9447 // silent-wrong. A qualifier naming a SOURCE table means a
9448 // multi-TARGET update — mutating two tables in one statement —
9449 // which SPG does not model, so it is refused loudly.
9450 let set_qual: Option<String> = if mysql_from.is_some()
9451 && matches!(self.tokens.get(self.pos + 1), Some(Token::Dot))
9452 {
9453 match self.peek() {
9454 Token::Ident(s) | Token::QuotedIdent(s) => Some(s.clone()),
9455 _ => None,
9456 }
9457 } else {
9458 None
9459 };
9460 let col = self.expect_ident_like()?;
9461 if let Some(q) = set_qual {
9462 let names_target = q.eq_ignore_ascii_case(&table)
9463 || alias.as_deref().is_some_and(|a| a.eq_ignore_ascii_case(&q));
9464 if !names_target {
9465 return Err(self.err(alloc::format!(
9466 "multi-table UPDATE can only assign to its first table \
9467 ({table}); `{q}.{col}` targets another table"
9468 )));
9469 }
9470 }
9471 // v7.37 D.53 — array element assignment target `SET arr[i] = v`,
9472 // desugared to `arr = __array_assign(arr, i, v)` (mirrors the
9473 // `__column_default` marker lowering just below). PG assigns to the
9474 // i-th (1-based) element, NULL-padding when i exceeds the length.
9475 if matches!(self.peek(), Token::LBracket) {
9476 self.advance();
9477 let index = self.parse_expr(0)?;
9478 // v7.39 (round 257) — the SLICE target `SET arr[lo:hi] = src`
9479 // (and the open `arr[lo:]`), lowered to
9480 // `__array_assign_slice`. Only the single-subscript form
9481 // parsed before, so a slice assignment was a syntax error.
9482 let mut slice_hi: Option<Option<Expr>> = None;
9483 if matches!(self.peek(), Token::Colon) {
9484 self.advance();
9485 slice_hi = Some(if matches!(self.peek(), Token::RBracket) {
9486 None
9487 } else {
9488 Some(self.parse_expr(0)?)
9489 });
9490 }
9491 if !matches!(self.peek(), Token::RBracket) {
9492 return Err(self.err(format!(
9493 "expected `]` after array subscript in UPDATE SET, got {:?}",
9494 self.peek()
9495 )));
9496 }
9497 self.advance();
9498 if !matches!(self.peek(), Token::Eq) {
9499 return Err(self.err(format!(
9500 "expected `=` after array subscript in UPDATE SET, got {:?}",
9501 self.peek()
9502 )));
9503 }
9504 self.advance();
9505 let value = self.parse_expr(0)?;
9506 // PG merges several subscript writes to the same column into one
9507 // array (`SET arr[1]=x, arr[3]=y`), so chain onto any prior
9508 // assignment to `col` rather than each overwriting the original.
9509 let existing = assignments.iter().position(|(c, _)| c == &col);
9510 let base = match existing {
9511 Some(i) => assignments[i].1.clone(),
9512 None => Expr::Column(ColumnName {
9513 qualifier: None,
9514 name: col.clone(),
9515 }),
9516 };
9517 let assigned = match slice_hi {
9518 None => Expr::FunctionCall {
9519 name: "__array_assign".to_string(),
9520 args: alloc::vec![base, index, value],
9521 },
9522 Some(hi) => Expr::FunctionCall {
9523 name: "__array_assign_slice".to_string(),
9524 args: alloc::vec![
9525 base,
9526 index,
9527 hi.unwrap_or(Expr::Literal(crate::ast::Literal::Null)),
9528 value,
9529 ],
9530 },
9531 };
9532 match existing {
9533 Some(i) => assignments[i].1 = assigned,
9534 None => assignments.push((col, assigned)),
9535 }
9536 if matches!(self.peek(), Token::Comma) {
9537 self.advance();
9538 continue;
9539 }
9540 break;
9541 }
9542 if !matches!(self.peek(), Token::Eq) {
9543 return Err(self.err(format!(
9544 "expected `=` after column name in UPDATE SET, got {:?}",
9545 self.peek()
9546 )));
9547 }
9548 self.advance();
9549 // `SET col = DEFAULT` — the column's declared default;
9550 // rides out as a marker call the update executor
9551 // resolves against the schema.
9552 let value = if matches!(self.peek(), Token::Default) {
9553 self.advance();
9554 Expr::FunctionCall {
9555 name: "__column_default".to_string(),
9556 args: Vec::new(),
9557 }
9558 } else {
9559 self.parse_expr(0)?
9560 };
9561 assignments.push((col, value));
9562 if matches!(self.peek(), Token::Comma) {
9563 self.advance();
9564 continue;
9565 }
9566 break;
9567 }
9568 // `UPDATE t SET … FROM src [, …] WHERE cond` — PG's joined
9569 // update. Lowers onto the correlated-subquery machinery:
9570 // the WHERE becomes EXISTS(SELECT 1 FROM src WHERE cond)
9571 // and each assignment that references a FROM-list table
9572 // wraps into a correlated scalar subquery
9573 // (SELECT expr FROM src WHERE cond). Equivalent for the
9574 // unique-join shape (the overwhelmingly common one); a
9575 // multi-match, which PG resolves by arbitrary pick,
9576 // surfaces as a scalar-subquery cardinality error instead
9577 // of a silent arbitrary result.
9578 // v7.39 (round 420) — the MySQL multi-table form supplies the source
9579 // list up front (`UPDATE a, b SET …`) instead of via FROM, so it feeds
9580 // the SAME lowering below. Both spellings together is not legal in
9581 // either dialect.
9582 let from_clause = if let Some(fc) = mysql_from {
9583 if matches!(self.peek(), Token::From) {
9584 return Err(self.err(alloc::string::String::from(
9585 "multi-table UPDATE already names its sources; drop the FROM clause",
9586 )));
9587 }
9588 Some(fc)
9589 } else if matches!(self.peek(), Token::From) {
9590 self.advance();
9591 Some(self.parse_from_clause()?)
9592 } else {
9593 None
9594 };
9595 let where_ = if matches!(self.peek(), Token::Where) {
9596 self.advance();
9597 Some(self.parse_expr(0)?)
9598 } else {
9599 None
9600 };
9601 // v7.39 (round 421, fixing round 420) — the SOURCE subquery's filter
9602 // and the TARGET-row filter are NOT the same predicate once a LEFT
9603 // join is involved:
9604 // * inner / comma / PG's `FROM`: the ON predicate and the WHERE are
9605 // one conjunction, and the whole thing filters target rows via
9606 // EXISTS.
9607 // * LEFT join: only the ON predicate belongs inside the source
9608 // subquery. The WHERE still filters TARGET rows (with source
9609 // columns read through the correlated subquery, which yields NULL
9610 // for an unmatched row — exactly LEFT-join semantics).
9611 // Round 420 folded ON into WHERE unconditionally and then dropped the
9612 // outer filter for the LEFT case, so `UPDATE a LEFT JOIN b ON … SET …
9613 // WHERE a.id > 1` updated EVERY row.
9614 let sub_where = match (mysql_on.clone(), where_.clone()) {
9615 _ if mysql_outer => mysql_on.clone(),
9616 (Some(on), Some(w)) => Some(Expr::Binary {
9617 lhs: Box::new(on),
9618 op: crate::ast::BinOp::And,
9619 rhs: Box::new(w),
9620 }),
9621 (Some(on), None) => Some(on),
9622 (None, w) => w,
9623 };
9624 // v7.39 (round 413) — MySQL `UPDATE … [ORDER BY …] [LIMIT n]`. PG
9625 // has no such clause on UPDATE, so this is accepted only under the
9626 // MySQL dialect; a PG session's `UPDATE … ORDER BY …` still errors.
9627 let update_order_limit = self.parse_mysql_dml_order_limit("UPDATE")?;
9628 let mut returning = self.parse_optional_returning()?;
9629 // v7.39 (round 533) — kept for the engine, which can resolve the
9630 // UNQUALIFIED leaves this lowering has to leave alone.
9631 let from_sources = from_clause.as_ref().map(|fc| {
9632 alloc::boxed::Box::new(crate::ast::UpdateFromSources {
9633 from: fc.clone(),
9634 sub_where: sub_where.clone(),
9635 })
9636 });
9637 let (assignments, where_) = if let Some(fc) = from_clause {
9638 let names: Vec<String> = core::iter::once(&fc.primary)
9639 .chain(fc.joins.iter().map(|j| &j.table))
9640 .flat_map(|t| {
9641 t.alias
9642 .clone()
9643 .into_iter()
9644 .chain(core::iter::once(t.name.clone()))
9645 })
9646 .collect();
9647 let refs_list = |e: &Expr| -> bool {
9648 fn walk(e: &Expr, names: &[String]) -> bool {
9649 match e {
9650 Expr::Column(c) => c
9651 .qualifier
9652 .as_deref()
9653 .is_some_and(|q| names.iter().any(|n| n.eq_ignore_ascii_case(q))),
9654 Expr::Binary { lhs, rhs, .. } => walk(lhs, names) || walk(rhs, names),
9655 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } => walk(expr, names),
9656 Expr::FunctionCall { args, .. } => args.iter().any(|a| walk(a, names)),
9657 Expr::Case {
9658 operand,
9659 branches,
9660 else_branch,
9661 } => {
9662 operand.as_deref().is_some_and(|o| walk(o, names))
9663 || branches
9664 .iter()
9665 .any(|(w, t)| walk(w, names) || walk(t, names))
9666 || else_branch.as_deref().is_some_and(|el| walk(el, names))
9667 }
9668 _ => false,
9669 }
9670 }
9671 walk(e, &names)
9672 };
9673 let sub_select = |items: Vec<SelectItem>| SelectStatement {
9674 locking: None,
9675 ctes: Vec::new(),
9676 distinct: false,
9677 distinct_on: Vec::new(),
9678 items,
9679 from: Some(fc.clone()),
9680 where_: sub_where.clone(),
9681 group_by: None,
9682 group_by_all: false,
9683 having: None,
9684 unions: Vec::new(),
9685 order_by: Vec::new(),
9686 limit: None,
9687 offset: None,
9688 limit_with_ties: false,
9689 window_check_exprs: Vec::new(),
9690 };
9691 // v7.37 D.30 — replace each FROM-qualified column *leaf* in the
9692 // assignment RHS with a correlated scalar subquery, instead of
9693 // wrapping the whole RHS. Wrapping the whole expr moved a target-
9694 // column reference (`SET v = v + u.bonus`, where `v` is the target
9695 // table's column) inside a subquery whose FROM only has the source
9696 // table, so the unqualified `v` resolved against the source and
9697 // errored ColumnNotFound. Leaving target columns in the outer UPDATE
9698 // context — where they belong — fixes it; only the source columns
9699 // (`u.bonus`) become subqueries. A whole-expr fallback covers
9700 // compound variants the leaf-walk doesn't decompose.
9701 let make_subq = |inner: Expr| {
9702 Expr::ScalarSubquery(Box::new(sub_select(alloc::vec![SelectItem::Expr {
9703 expr: inner,
9704 alias: None,
9705 }])))
9706 };
9707 let assignments = assignments
9708 .into_iter()
9709 .map(|(col, mut expr)| {
9710 wrap_from_leaves(&mut expr, &names, &make_subq, &refs_list);
9711 (col, expr)
9712 })
9713 .collect();
9714 let exists = Expr::Exists {
9715 subquery: Box::new(sub_select(alloc::vec![SelectItem::Expr {
9716 expr: Expr::Literal(Literal::Integer(1)),
9717 alias: None,
9718 }])),
9719 negated: false,
9720 };
9721 // v7.39 (round 241) — RETURNING may reference the FROM-list
9722 // tables too (`RETURNING emp.id, dept.name`); the same
9723 // leaf-to-correlated-subquery lowering the assignments get.
9724 // Without it the qualifier died at eval with "unknown table
9725 // qualifier". (RETURNING was parsed before this block — the
9726 // lowering is a pure AST transformation.)
9727 if let Some(items) = returning.as_mut() {
9728 for item in items.iter_mut() {
9729 if let SelectItem::Expr { expr, .. } = item {
9730 wrap_from_leaves(expr, &names, &make_subq, &refs_list);
9731 }
9732 }
9733 }
9734 // v7.39 (round 420, corrected in 421) — a MySQL LEFT JOIN keeps
9735 // EVERY matching target row: it gets no EXISTS filter, but the
9736 // caller's WHERE still applies, with source columns read through
9737 // the correlated subquery (NULL when unmatched — LEFT-join
9738 // semantics). `sub_where` above already excluded the WHERE from
9739 // the source subquery for this case.
9740 if mysql_outer {
9741 let mut outer = where_;
9742 if let Some(w) = outer.as_mut() {
9743 wrap_from_leaves(w, &names, &make_subq, &refs_list);
9744 }
9745 (assignments, outer)
9746 } else {
9747 (assignments, Some(exists))
9748 }
9749 } else {
9750 (assignments, where_)
9751 };
9752 Ok(Statement::Update(crate::ast::UpdateStatement {
9753 ctes: Vec::new(),
9754 table,
9755 only,
9756 alias,
9757 assignments,
9758 from_sources,
9759 where_,
9760 order_limit: update_order_limit,
9761 returning,
9762 }))
9763 }
9764
9765 /// v7.39 (round 432) — MySQL's `[ORDER BY …] [LIMIT n]` tail on a DML
9766 /// statement. UPDATE grew it in round 413 and DELETE in round 432; the
9767 /// clause and its meaning are identical, so both call this rather than
9768 /// keeping two copies that could disagree on, say, whether `LIMIT 0` is
9769 /// legal. PG has no such clause on either statement, so it is read only
9770 /// under the MySQL dialect — a PG session's `DELETE … ORDER BY …` still
9771 /// errors.
9772 ///
9773 /// `#[inline(never)]`: its locals would otherwise land on the statement-
9774 /// parsing recursion frame, which is what tipped the 512 KiB nesting
9775 /// stack in round 430.
9776 #[inline(never)]
9777 fn parse_mysql_dml_order_limit(
9778 &mut self,
9779 what: &str,
9780 ) -> Result<Option<alloc::boxed::Box<crate::ast::DmlOrderLimit>>, ParseError> {
9781 if !self.mysql_dialect {
9782 return Ok(None);
9783 }
9784 let order_by = self.parse_order_by_keys()?;
9785 let limit = if matches!(self.peek(), Token::Limit) {
9786 self.advance();
9787 let tok = self.advance();
9788 let Token::Integer(n) = tok else {
9789 return Err(self.err(alloc::format!(
9790 "expected integer after {what} LIMIT, got {tok:?}"
9791 )));
9792 };
9793 // MySQL rejects the `LIMIT offset, count` form here — only a
9794 // single row count is legal on a DML statement.
9795 if matches!(self.peek(), Token::Comma) {
9796 return Err(self.err(alloc::format!(
9797 "{what} LIMIT takes a row count, not an offset"
9798 )));
9799 }
9800 let n = u32::try_from(n)
9801 .map_err(|_| self.err(alloc::format!("{what} LIMIT out of range: {n}")))?;
9802 Some(n)
9803 } else {
9804 None
9805 };
9806 if order_by.is_empty() && limit.is_none() {
9807 return Ok(None);
9808 }
9809 Ok(Some(alloc::boxed::Box::new(crate::ast::DmlOrderLimit {
9810 order_by,
9811 limit,
9812 })))
9813 }
9814
9815 /// v4.4 `DELETE FROM <table> [WHERE cond]`. Caller already consumed
9816 /// the leading `DELETE` ident.
9817 fn parse_delete_after_keyword(&mut self) -> Result<Statement, ParseError> {
9818 // v7.39 (round 421) — MySQL's multi-table DELETE names its target(s)
9819 // BEFORE the FROM: `DELETE a FROM a JOIN b ON …`. (`DELETE FROM a
9820 // USING a, b WHERE …` — the third MySQL spelling — needs no special
9821 // parse here; it reaches the existing USING path with the target
9822 // repeated in the list, which the source-list peel below handles.)
9823 // More than one name is a multi-TARGET delete, which SPG does not
9824 // model; it is refused rather than half-applied.
9825 let mysql_pre_target: Option<String> =
9826 if self.mysql_dialect && !matches!(self.peek(), Token::From) {
9827 let first = self.expect_ident_like()?;
9828 if matches!(self.peek(), Token::Comma) {
9829 return Err(self.err(alloc::format!(
9830 "multi-table DELETE can only delete from one table; \
9831 `DELETE {first}, …` names several"
9832 )));
9833 }
9834 Some(first)
9835 } else {
9836 None
9837 };
9838 if !matches!(self.peek(), Token::From) {
9839 return Err(self.err(format!("expected FROM after DELETE, got {:?}", self.peek())));
9840 }
9841 self.advance();
9842 // v7.39 (round 646) — `DELETE FROM ONLY t`, same shape and same
9843 // lookahead as the UPDATE spelling.
9844 let only = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
9845 if s.eq_ignore_ascii_case("only"))
9846 && matches!(
9847 self.tokens.get(self.pos + 1),
9848 Some(Token::Ident(_) | Token::QuotedIdent(_))
9849 );
9850 if only {
9851 self.advance();
9852 }
9853 let table = self.expect_ident_like()?;
9854 // v7.39 (round 241) — `DELETE FROM t [AS] alias …`. The bare
9855 // spelling must not swallow the clause keywords that can follow
9856 // the target.
9857 let alias = if matches!(self.peek(), Token::As) {
9858 self.advance();
9859 Some(self.expect_ident_like()?)
9860 } else {
9861 match self.peek() {
9862 Token::Ident(s) | Token::QuotedIdent(s)
9863 if !s.eq_ignore_ascii_case("using") && !s.eq_ignore_ascii_case("returning") =>
9864 {
9865 let a = s.clone();
9866 self.advance();
9867 Some(a)
9868 }
9869 _ => None,
9870 }
9871 };
9872 // v7.39 (round 421) — MySQL's multi-table DELETE source list, read
9873 // through the SAME join grammar the FROM clause uses (see the
9874 // `advance()`-destroys-tokens note on `parse_from_joins`).
9875 let mut mysql_on: Option<Expr> = None;
9876 let mut mysql_outer = false;
9877 let mysql_using = if mysql_pre_target.is_some()
9878 && (matches!(self.peek(), Token::Comma) || self.peek_is_update_join_start())
9879 {
9880 let target_qual = alias.clone().unwrap_or_else(|| table.clone());
9881 let mut joins = self.parse_from_joins(&target_qual)?;
9882 if joins.is_empty() {
9883 return Err(self.err(alloc::string::String::from(
9884 "multi-table DELETE needs at least one source table",
9885 )));
9886 }
9887 let head = joins.remove(0);
9888 mysql_outer = matches!(head.kind, crate::ast::JoinKind::Left);
9889 mysql_on = head.on;
9890 Some(FromClause {
9891 primary: head.table,
9892 joins,
9893 })
9894 } else {
9895 None
9896 };
9897 // The pre-FROM target must be the table the FROM names (or its
9898 // alias) — `DELETE b FROM a JOIN b …` would delete from a table that
9899 // is not the scan target.
9900 if let Some(t) = &mysql_pre_target {
9901 let names_target = t.eq_ignore_ascii_case(&table)
9902 || alias.as_deref().is_some_and(|a| a.eq_ignore_ascii_case(t));
9903 if !names_target {
9904 return Err(self.err(alloc::format!(
9905 "DELETE target `{t}` is not the first table in the FROM clause ({table})"
9906 )));
9907 }
9908 }
9909 // `DELETE FROM t USING src [, …] WHERE cond` — PG's joined
9910 // delete. Same lowering as UPDATE … FROM: the WHERE
9911 // becomes EXISTS(SELECT 1 FROM src WHERE cond), driven per
9912 // target row by the correlated machinery.
9913 let using_clause = if let Some(fc) = mysql_using {
9914 Some(fc)
9915 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
9916 self.advance();
9917 let mut fc = self.parse_from_clause()?;
9918 // v7.39 (round 421) — MySQL's `DELETE FROM a USING a, b WHERE …`
9919 // repeats the TARGET as the first USING entry (PG's spelling
9920 // lists only the extra sources). Peel it so the source subquery
9921 // does not re-scan — and shadow — the target table.
9922 let primary_is_target =
9923 fc.primary.name.eq_ignore_ascii_case(&table) && fc.primary.alias.is_none();
9924 if self.mysql_dialect && primary_is_target && !fc.joins.is_empty() {
9925 let head = fc.joins.remove(0);
9926 mysql_outer = matches!(head.kind, crate::ast::JoinKind::Left);
9927 mysql_on = head.on;
9928 fc = FromClause {
9929 primary: head.table,
9930 joins: fc.joins,
9931 };
9932 }
9933 Some(fc)
9934 } else {
9935 None
9936 };
9937 let where_ = if matches!(self.peek(), Token::Where) {
9938 self.advance();
9939 Some(self.parse_expr(0)?)
9940 } else {
9941 None
9942 };
9943 // v7.39 (round 432) — MySQL's `DELETE … [ORDER BY …] [LIMIT n]`,
9944 // read before RETURNING (MariaDB's own extension trails the LIMIT).
9945 let delete_order_limit = self.parse_mysql_dml_order_limit("DELETE")?;
9946 let mut returning = self.parse_optional_returning()?;
9947 let where_ = if let Some(fc) = using_clause {
9948 // v7.39 (round 241) — same RETURNING lowering as UPDATE…FROM:
9949 // a USING-table reference in RETURNING becomes a correlated
9950 // scalar subquery over the USING list.
9951 let names: Vec<String> = core::iter::once(&fc.primary)
9952 .chain(fc.joins.iter().map(|j| &j.table))
9953 .flat_map(|t| {
9954 t.alias
9955 .clone()
9956 .into_iter()
9957 .chain(core::iter::once(t.name.clone()))
9958 })
9959 .collect();
9960 // v7.39 (round 421) — same ON / WHERE split as UPDATE: a LEFT
9961 // join filters the SOURCE subquery on the ON predicate alone and
9962 // leaves the WHERE filtering TARGET rows (so the anti-join idiom
9963 // `LEFT JOIN b ON … WHERE b.id IS NULL` deletes the unmatched
9964 // rows); every other form folds ON and WHERE into one EXISTS.
9965 let sub_where = match (mysql_on.clone(), where_.clone()) {
9966 _ if mysql_outer => mysql_on.clone(),
9967 (Some(on), Some(w)) => Some(Expr::Binary {
9968 lhs: Box::new(on),
9969 op: crate::ast::BinOp::And,
9970 rhs: Box::new(w),
9971 }),
9972 (Some(on), None) => Some(on),
9973 (None, w) => w,
9974 };
9975 let exists_where = sub_where.clone();
9976 let sub_fc = fc.clone();
9977 let make_subq = move |leaf: Expr| -> Expr {
9978 Expr::ScalarSubquery(Box::new(SelectStatement {
9979 locking: None,
9980 ctes: Vec::new(),
9981 distinct: false,
9982 distinct_on: Vec::new(),
9983 items: alloc::vec![SelectItem::Expr {
9984 expr: leaf,
9985 alias: None,
9986 }],
9987 from: Some(sub_fc.clone()),
9988 where_: sub_where.clone(),
9989 group_by: None,
9990 group_by_all: false,
9991 having: None,
9992 unions: Vec::new(),
9993 order_by: Vec::new(),
9994 limit: None,
9995 offset: None,
9996 limit_with_ties: false,
9997 window_check_exprs: Vec::new(),
9998 }))
9999 };
10000 let refs = |e: &Expr| expr_refs_tables(e, &names);
10001 if let Some(items) = returning.as_mut() {
10002 for item in items.iter_mut() {
10003 if let SelectItem::Expr { expr, .. } = item {
10004 wrap_from_leaves(expr, &names, &make_subq, &refs);
10005 }
10006 }
10007 }
10008 // A LEFT join deletes the target rows the WHERE selects, reading
10009 // source columns through the correlated subquery (NULL when
10010 // unmatched); no EXISTS row filter.
10011 if mysql_outer {
10012 let mut outer = where_;
10013 if let Some(w) = outer.as_mut() {
10014 wrap_from_leaves(w, &names, &make_subq, &refs);
10015 }
10016 outer
10017 } else {
10018 Some(Expr::Exists {
10019 subquery: Box::new(SelectStatement {
10020 locking: None,
10021 ctes: Vec::new(),
10022 distinct: false,
10023 distinct_on: Vec::new(),
10024 items: alloc::vec![SelectItem::Expr {
10025 expr: Expr::Literal(Literal::Integer(1)),
10026 alias: None,
10027 }],
10028 from: Some(fc),
10029 where_: exists_where,
10030 group_by: None,
10031 group_by_all: false,
10032 having: None,
10033 unions: Vec::new(),
10034 order_by: Vec::new(),
10035 limit: None,
10036 offset: None,
10037 limit_with_ties: false,
10038 window_check_exprs: Vec::new(),
10039 }),
10040 negated: false,
10041 })
10042 }
10043 } else {
10044 where_
10045 };
10046 Ok(Statement::Delete(crate::ast::DeleteStatement {
10047 ctes: Vec::new(),
10048 table,
10049 only,
10050 alias,
10051 where_,
10052 order_limit: delete_order_limit,
10053 returning,
10054 }))
10055 }
10056
10057 /// v7.17.0 Phase 3.P0-42 — parse `MERGE INTO <target> [alias]
10058 /// USING <source> [alias] ON <expr> WHEN [NOT] MATCHED [AND
10059 /// <expr>] THEN <action> [WHEN …]` after the leading `MERGE`
10060 /// keyword. v7.17 surface:
10061 /// * source: table reference (subquery source is a follow-up)
10062 /// * actions: UPDATE SET / DELETE / DO NOTHING (matched);
10063 /// INSERT (cols) VALUES (vals) / DO NOTHING (not matched)
10064 /// * AND-conditioned WHEN clauses; clauses tried in declaration
10065 /// order
10066 fn parse_merge_after_keyword(&mut self) -> Result<Statement, ParseError> {
10067 // INTO
10068 let is_into_kw = matches!(self.peek(), Token::Into)
10069 || matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("into"));
10070 if !is_into_kw {
10071 return Err(self.err(format!("expected INTO after MERGE, got {:?}", self.peek())));
10072 }
10073 self.advance();
10074 let target = self.expect_ident_like()?;
10075 // Optional alias — bare ident before USING.
10076 let target_alias = match self.peek() {
10077 Token::Ident(s) | Token::QuotedIdent(s) if !s.eq_ignore_ascii_case("using") => {
10078 Some(self.expect_ident_like()?)
10079 }
10080 _ => None,
10081 };
10082 // USING
10083 let is_using_kw = matches!(
10084 self.peek(),
10085 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("using")
10086 );
10087 if !is_using_kw {
10088 return Err(self.err(format!(
10089 "expected USING after MERGE INTO target, got {:?}",
10090 self.peek()
10091 )));
10092 }
10093 self.advance();
10094 // v7.37 D.44 — `USING (SELECT …) alias` subquery source, or `USING
10095 // <table> [alias]`. PG requires an alias after a subquery source.
10096 let (source, source_select) = if matches!(self.peek(), Token::LParen) {
10097 self.advance(); // (
10098 // v7.39 (round 768, F31-D5) — `USING (VALUES …)`: the same
10099 // constant-SELECT lowering the derived-table parser uses
10100 // (PG deletes through this form; it was a parse error).
10101 let inner = if matches!(self.peek(), Token::Values) {
10102 self.advance(); // VALUES
10103 Statement::Select(self.parse_values_rows_body()?)
10104 } else {
10105 self.parse_select_stmt()?
10106 };
10107 match self.advance() {
10108 Token::RParen => {}
10109 other => {
10110 return Err(self.err(format!(
10111 "expected ')' after MERGE USING subquery, got {other:?}"
10112 )));
10113 }
10114 }
10115 let Statement::Select(sub) = inner else {
10116 return Err(self.err("MERGE USING subquery must be a SELECT".into()));
10117 };
10118 (String::new(), Some(Box::new(sub)))
10119 } else {
10120 (self.expect_ident_like()?, None)
10121 };
10122 let source_alias = match self.peek() {
10123 Token::Ident(s) | Token::QuotedIdent(s)
10124 if !s.eq_ignore_ascii_case("on") && !s.eq_ignore_ascii_case("as") =>
10125 {
10126 Some(self.expect_ident_like()?)
10127 }
10128 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("as") => {
10129 self.advance(); // AS
10130 Some(self.expect_ident_like()?)
10131 }
10132 _ => None,
10133 };
10134 // v7.39 (round 768, F31-D5) — optional positional column-alias
10135 // list after the source alias (`s(id, v)`).
10136 let mut source_column_aliases: Vec<String> = Vec::new();
10137 if source_alias.is_some() && matches!(self.peek(), Token::LParen) {
10138 self.advance();
10139 loop {
10140 source_column_aliases.push(self.expect_ident_like()?);
10141 match self.peek() {
10142 Token::Comma => {
10143 self.advance();
10144 }
10145 Token::RParen => {
10146 self.advance();
10147 break;
10148 }
10149 other => {
10150 return Err(self.err(format!(
10151 "expected ',' or ')' in MERGE source column list, got {other:?}"
10152 )));
10153 }
10154 }
10155 }
10156 }
10157 if source_select.is_some() && source_alias.is_none() {
10158 return Err(self.err("MERGE USING (subquery) requires an alias".into()));
10159 }
10160 // ON
10161 if !matches!(self.peek(), Token::On) {
10162 return Err(self.err(format!(
10163 "expected ON after MERGE … USING source, got {:?}",
10164 self.peek()
10165 )));
10166 }
10167 self.advance();
10168 let on = self.parse_expr(0)?;
10169 // One or more WHEN clauses.
10170 let mut clauses: Vec<crate::ast::MergeWhenClause> = Vec::new();
10171 loop {
10172 let is_when_kw = matches!(
10173 self.peek(),
10174 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("when")
10175 );
10176 if !is_when_kw {
10177 break;
10178 }
10179 self.advance(); // WHEN
10180 // [NOT] MATCHED
10181 let matched = if matches!(self.peek(), Token::Not) {
10182 self.advance();
10183 crate::ast::MergeMatched::NotMatched
10184 } else {
10185 crate::ast::MergeMatched::Matched
10186 };
10187 let is_matched_kw = matches!(
10188 self.peek(),
10189 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("matched")
10190 );
10191 if !is_matched_kw {
10192 return Err(self.err(format!(
10193 "expected MATCHED in WHEN clause, got {:?}",
10194 self.peek()
10195 )));
10196 }
10197 self.advance();
10198 // v7.39 (round 146, PG17) — `NOT MATCHED [BY TARGET | BY SOURCE]`.
10199 // BY TARGET is the default (a synonym); BY SOURCE flips the clause
10200 // to fire for target rows no source row matches.
10201 let mut matched = matched;
10202 if matches!(matched, crate::ast::MergeMatched::NotMatched) && self.peek_is_by() {
10203 self.advance();
10204 match self.peek() {
10205 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("source") => {
10206 self.advance();
10207 matched = crate::ast::MergeMatched::NotMatchedBySource;
10208 }
10209 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("target") => {
10210 self.advance();
10211 }
10212 other => {
10213 return Err(self.err(format!(
10214 "expected SOURCE or TARGET after NOT MATCHED BY, got {other:?}"
10215 )));
10216 }
10217 }
10218 }
10219 // Optional AND <expr>
10220 let condition = if matches!(self.peek(), Token::And) {
10221 self.advance();
10222 Some(self.parse_expr(0)?)
10223 } else {
10224 None
10225 };
10226 // THEN
10227 let is_then_kw = matches!(
10228 self.peek(),
10229 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("then")
10230 );
10231 if !is_then_kw {
10232 return Err(self.err(format!(
10233 "expected THEN in WHEN clause, got {:?}",
10234 self.peek()
10235 )));
10236 }
10237 self.advance();
10238 // Action: INSERT / UPDATE / DELETE / DO NOTHING
10239 let action = match self.peek().clone() {
10240 Token::Insert => {
10241 self.advance();
10242 // v7.39 (read01 round 123) — the `(cols)` list is OPTIONAL,
10243 // exactly like a plain INSERT: `WHEN NOT MATCHED THEN INSERT
10244 // VALUES (…)` omits it and fills every column in declaration
10245 // order. PG accepts this; SPG used to require the list.
10246 let mut columns: Vec<String> = Vec::new();
10247 if matches!(self.peek(), Token::LParen) {
10248 self.advance();
10249 loop {
10250 columns.push(self.expect_ident_like()?);
10251 if matches!(self.peek(), Token::Comma) {
10252 self.advance();
10253 continue;
10254 }
10255 break;
10256 }
10257 if !matches!(self.peek(), Token::RParen) {
10258 return Err(self.err(format!(
10259 "expected ')' after INSERT column list, got {:?}",
10260 self.peek()
10261 )));
10262 }
10263 self.advance();
10264 }
10265 // VALUES (...)
10266 if !matches!(self.peek(), Token::Values) {
10267 return Err(self.err(format!(
10268 "expected VALUES in MERGE INSERT, got {:?}",
10269 self.peek()
10270 )));
10271 }
10272 self.advance();
10273 if !matches!(self.peek(), Token::LParen) {
10274 return Err(self.err(format!(
10275 "expected '(' after VALUES in MERGE INSERT, got {:?}",
10276 self.peek()
10277 )));
10278 }
10279 self.advance();
10280 let mut values: Vec<crate::ast::Expr> = Vec::new();
10281 loop {
10282 values.push(self.parse_expr(0)?);
10283 if matches!(self.peek(), Token::Comma) {
10284 self.advance();
10285 continue;
10286 }
10287 break;
10288 }
10289 if !matches!(self.peek(), Token::RParen) {
10290 return Err(self.err(format!(
10291 "expected ')' after MERGE INSERT values, got {:?}",
10292 self.peek()
10293 )));
10294 }
10295 self.advance();
10296 // Empty column list = positional into every column, so the
10297 // count is checked against the table arity at execution.
10298 if !columns.is_empty() && columns.len() != values.len() {
10299 return Err(self.err(format!(
10300 "MERGE INSERT column count ({}) ≠ value count ({})",
10301 columns.len(),
10302 values.len()
10303 )));
10304 }
10305 crate::ast::MergeAction::Insert { columns, values }
10306 }
10307 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("update") => {
10308 self.advance();
10309 // SET
10310 let is_set_kw = matches!(
10311 self.peek(),
10312 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("set")
10313 );
10314 if !is_set_kw {
10315 return Err(self.err(format!(
10316 "expected SET after UPDATE in MERGE, got {:?}",
10317 self.peek()
10318 )));
10319 }
10320 self.advance();
10321 let mut assignments: Vec<(String, crate::ast::Expr)> = Vec::new();
10322 loop {
10323 let col = self.expect_ident_like()?;
10324 if !matches!(self.peek(), Token::Eq) {
10325 return Err(self.err(format!(
10326 "expected '=' in MERGE UPDATE assignment, got {:?}",
10327 self.peek()
10328 )));
10329 }
10330 self.advance();
10331 let expr = self.parse_expr(0)?;
10332 assignments.push((col, expr));
10333 if matches!(self.peek(), Token::Comma) {
10334 self.advance();
10335 continue;
10336 }
10337 break;
10338 }
10339 crate::ast::MergeAction::Update { assignments }
10340 }
10341 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("delete") => {
10342 self.advance();
10343 crate::ast::MergeAction::Delete
10344 }
10345 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("do") => {
10346 self.advance();
10347 let is_nothing_kw = matches!(
10348 self.peek(),
10349 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("nothing")
10350 );
10351 if !is_nothing_kw {
10352 return Err(self.err(format!(
10353 "expected NOTHING after DO in MERGE clause, got {:?}",
10354 self.peek()
10355 )));
10356 }
10357 self.advance();
10358 crate::ast::MergeAction::DoNothing
10359 }
10360 other => {
10361 return Err(self.err(format!(
10362 "expected INSERT / UPDATE / DELETE / DO NOTHING in MERGE clause, got {other:?}"
10363 )));
10364 }
10365 };
10366 // PG's grammar simply has no INSERT production under BY SOURCE
10367 // (a target row already exists there) — same syntax error.
10368 if matches!(matched, crate::ast::MergeMatched::NotMatchedBySource)
10369 && matches!(action, crate::ast::MergeAction::Insert { .. })
10370 {
10371 return Err(self.err(String::from("syntax error at or near \"INSERT\"")));
10372 }
10373 clauses.push(crate::ast::MergeWhenClause {
10374 matched,
10375 condition,
10376 action,
10377 });
10378 }
10379 if clauses.is_empty() {
10380 return Err(self.err(String::from("MERGE requires at least one WHEN clause")));
10381 }
10382 // v7.38 (read01 U-merge) — PG rejects a WHEN clause that follows an
10383 // unconditional (no `AND`) WHEN of the same match kind: it could
10384 // never fire. Check per match kind in clause order.
10385 let mut seen_unconditional_matched = false;
10386 let mut seen_unconditional_not_matched = false;
10387 let mut seen_unconditional_by_source = false;
10388 for c in &clauses {
10389 let seen = match c.matched {
10390 crate::ast::MergeMatched::Matched => &mut seen_unconditional_matched,
10391 crate::ast::MergeMatched::NotMatched => &mut seen_unconditional_not_matched,
10392 crate::ast::MergeMatched::NotMatchedBySource => &mut seen_unconditional_by_source,
10393 };
10394 if *seen {
10395 return Err(self.err(String::from(
10396 "unreachable WHEN clause specified after unconditional WHEN clause",
10397 )));
10398 }
10399 if c.condition.is_none() {
10400 *seen = true;
10401 }
10402 }
10403 // v7.39 (round 130) — optional trailing `RETURNING <projection>` (PG17+).
10404 let returning = self.parse_optional_returning()?;
10405 Ok(Statement::Merge(crate::ast::MergeStatement {
10406 // Attached by `parse_with_cte_then_select` when the MERGE
10407 // heads a WITH clause (round 149).
10408 ctes: Vec::new(),
10409 target,
10410 target_alias,
10411 source,
10412 source_alias,
10413 source_select,
10414 source_column_aliases,
10415 on,
10416 clauses,
10417 returning,
10418 }))
10419 }
10420
10421 /// v7.9.4 — parse the optional trailing `RETURNING <projection>`
10422 /// clause on INSERT / UPDATE / DELETE. Same projection grammar
10423 /// as SELECT, so `RETURNING *`, `RETURNING col`,
10424 /// `RETURNING expr AS alias`, and `RETURNING a, b, c` all work.
10425 fn parse_optional_returning(
10426 &mut self,
10427 ) -> Result<Option<Vec<crate::ast::SelectItem>>, ParseError> {
10428 let is_returning_kw = matches!(
10429 self.peek(),
10430 Token::Ident(s) if s.eq_ignore_ascii_case("returning")
10431 );
10432 if !is_returning_kw {
10433 return Ok(None);
10434 }
10435 self.advance();
10436 let mut items = Vec::new();
10437 loop {
10438 items.push(self.parse_select_item()?);
10439 if matches!(self.peek(), Token::Comma) {
10440 self.advance();
10441 continue;
10442 }
10443 break;
10444 }
10445 Ok(Some(items))
10446 }
10447
10448 /// v6.0.4 — parse the tail of an ALTER statement after the
10449 /// leading `ALTER` keyword has been consumed. Only one form is
10450 /// supported in v6.0.4:
10451 ///
10452 /// ```text
10453 /// ALTER INDEX <name> REBUILD [WITH (encoding = <enc>)]
10454 /// ```
10455 fn parse_alter_after_keyword(&mut self) -> Result<Statement, ParseError> {
10456 // ALTER INDEX <name> ... | ALTER TABLE <name> SET hot_tier_bytes = <n>
10457 // v7.14.0 — `ALTER TABLE ONLY` modifier (PG partition-
10458 // exclusion) is accepted by stripping the `ONLY` keyword
10459 // before the table parse.
10460 // v7.14.0 — `ALTER SEQUENCE / ALTER VIEW / ALTER OWNER`
10461 // and the long PG-dump tail are accepted as no-ops.
10462 match self.advance() {
10463 Token::Index => {}
10464 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("index") => {}
10465 // v6.7.2 — ALTER TABLE t SET hot_tier_bytes = X
10466 // v7.14.0 — ALTER TABLE ONLY t … strip the `ONLY`.
10467 Token::Table => {
10468 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("only")) {
10469 self.advance();
10470 }
10471 return self.parse_alter_table_after_keyword();
10472 }
10473 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("policy") => {
10474 return self.parse_alter_policy_after_keyword();
10475 }
10476 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("table") => {
10477 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("only")) {
10478 self.advance();
10479 }
10480 return self.parse_alter_table_after_keyword();
10481 }
10482 // v7.17.0 — ALTER SEQUENCE name <options>. Moved out
10483 // of the silent-noop tail.
10484 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("sequence") => {
10485 return self.parse_alter_sequence_after_keyword();
10486 }
10487 // v7.37 D.55 — ALTER TYPE name ADD VALUE [IF NOT EXISTS] 'label'
10488 // [{BEFORE | AFTER} 'existing']. Real enum evolution; other ALTER
10489 // TYPE forms (RENAME / OWNER / SET SCHEMA) still no-op below.
10490 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("type") => {
10491 // NB: the match arm consumed `TYPE` via self.advance(); the
10492 // cursor is now at the type name — do NOT advance again.
10493 let type_name = self.expect_ident_like()?;
10494 let is_add_value = matches!(self.peek(), Token::Ident(a) if a.eq_ignore_ascii_case("add"))
10495 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(v)) if v.eq_ignore_ascii_case("value"));
10496 if is_add_value {
10497 self.advance(); // ADD
10498 self.advance(); // VALUE
10499 // `IF NOT EXISTS` — NOT lexes as the keyword `Token::Not`,
10500 // IF/EXISTS as identifiers.
10501 let if_not_exists = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if"))
10502 {
10503 let n1 = self.tokens.get(self.pos + 1);
10504 let n2 = self.tokens.get(self.pos + 2);
10505 if matches!(n1, Some(Token::Not))
10506 && matches!(n2, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("exists"))
10507 {
10508 self.advance();
10509 self.advance();
10510 self.advance();
10511 true
10512 } else {
10513 false
10514 }
10515 } else {
10516 false
10517 };
10518 let label = self.expect_string_literal()?;
10519 let position = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("before") || s.eq_ignore_ascii_case("after"))
10520 {
10521 let is_before = matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("before"));
10522 self.advance();
10523 let anchor = self.expect_string_literal()?;
10524 Some((is_before, anchor))
10525 } else {
10526 None
10527 };
10528 return Ok(Statement::AlterTypeAddValue {
10529 type_name,
10530 label,
10531 if_not_exists,
10532 position,
10533 });
10534 }
10535 // v7.39 (read01 round 49) — `RENAME VALUE 'old' TO 'new'`.
10536 // Used to fall into the no-op tail below: accepted, silently
10537 // ignored. `RENAME TO <newtype>` keeps falling through.
10538 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("rename"))
10539 && matches!(
10540 self.tokens.get(self.pos + 1),
10541 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("value")
10542 )
10543 {
10544 self.advance(); // RENAME
10545 self.advance(); // VALUE
10546 let old = self.expect_string_literal()?;
10547 if matches!(self.peek(), Token::To) {
10548 self.advance();
10549 } else {
10550 self.expect_keyword_ident("to")?;
10551 }
10552 let new = self.expect_string_literal()?;
10553 return Ok(Statement::AlterTypeRenameValue {
10554 type_name,
10555 old,
10556 new,
10557 });
10558 }
10559 // Other ALTER TYPE forms — the ACTION stays a no-op
10560 // (pg_dump tail), but v7.39 (round 708) the NAME is
10561 // validated: `ALTER TYPE nosuch RENAME TO x` reported
10562 // success for a type that does not exist.
10563 self.consume_until_statement_boundary();
10564 return Ok(Statement::ValidateOnly {
10565 kind: crate::ast::ValidateOnlyKind::TypeName,
10566 names: alloc::vec![type_name],
10567 });
10568 }
10569 // v7.14.0 — ALTER VIEW / ALTER FUNCTION /
10570 // ALTER DOMAIN / ALTER DATABASE / ALTER USER / ALTER
10571 // ROLE / ALTER SCHEMA / ALTER OWNER / ALTER DEFAULT
10572 // PRIVILEGES — accept as no-op so pg_dump's tail loads.
10573 // v7.17.0 NOTE: ALTER SEQUENCE moved out (above).
10574 // v7.39 (round 260) — ALTER DOMAIN is REAL now, so it leaves the
10575 // pg_dump no-op list below: every form used to report success
10576 // and change nothing, which is worse than refusing outright
10577 // (a migration dropping a constraint kept rejecting data).
10578 // NOTE: the enclosing `match self.advance()` already consumed
10579 // the DOMAIN keyword, so the name is next.
10580 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("domain") => {
10581 return self.parse_alter_domain_after_keyword();
10582 }
10583 // v7.39 (round 547) — `ALTER ROLE|USER <r> [IN DATABASE <d>]
10584 // SET|RESET …` and `ALTER DATABASE <d> SET|RESET …`. These
10585 // used to fall into the pg_dump no-op tail below, so a DBA
10586 // setting a per-role default was told it worked and nothing
10587 // happened. Intercepted here, BEFORE that tail.
10588 // v7.39 (round 695) — `ALTER SYSTEM SET <name> = …` / `RESET
10589 // <name>` / `RESET ALL`. Same reason the ROLE / DATABASE
10590 // interception below exists: swallowed with the no-op tail, an
10591 // unknown parameter name was ACCEPTED where PG18 answers
10592 // `unrecognized configuration parameter`. SPG applies nothing
10593 // either way — there is no postgresql.auto.conf — but it now
10594 // says so about a name it does not know.
10595 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("system") => {
10596 // NOTE: the scrutinee is `self.advance()`, so SYSTEM is
10597 // already consumed here. An extra advance eats the SET and
10598 // the parameter name is never seen — which is exactly the
10599 // bug a panic in this branch disproved: the branch WAS on
10600 // the path, the reading of it was wrong.
10601 let mut parameter = None;
10602 // SET <name> … | RESET <name> | RESET ALL
10603 if matches!(self.peek(), Token::Ident(k)
10604 if k.eq_ignore_ascii_case("set") || k.eq_ignore_ascii_case("reset"))
10605 {
10606 self.advance();
10607 if let Token::Ident(n) | Token::QuotedIdent(n) = self.peek().clone()
10608 && !n.eq_ignore_ascii_case("all")
10609 {
10610 self.advance();
10611 // A dotted GUC (`plpgsql.check_asserts`) is two
10612 // tokens; keep the whole name.
10613 let mut full = n;
10614 while matches!(self.peek(), Token::Dot) {
10615 self.advance();
10616 if let Token::Ident(t) | Token::QuotedIdent(t) = self.advance() {
10617 full.push('.');
10618 full.push_str(&t);
10619 }
10620 }
10621 parameter = Some(full);
10622 }
10623 }
10624 self.consume_until_statement_boundary();
10625 return Ok(Statement::AlterSystem { parameter });
10626 }
10627 Token::Ident(s) | Token::QuotedIdent(s)
10628 if matches!(
10629 s.to_ascii_lowercase().as_str(),
10630 "role" | "user" | "database"
10631 ) && self.peeks_db_role_setting() =>
10632 {
10633 let is_database = s.eq_ignore_ascii_case("database");
10634 return self.parse_db_role_setting(is_database);
10635 }
10636 // v7.39 (round 708) — `ALTER ROLE|USER <name> [WITH attrs…]`
10637 // (the non-SET forms; SET/RESET took the branch above). The
10638 // attributes still no-op — recorded, and the ignored PASSWORD
10639 // is ledgered as its own follow-up — but the ROLE is validated:
10640 // any name was accepted for a role that does not exist.
10641 Token::Ident(s) | Token::QuotedIdent(s)
10642 if s.eq_ignore_ascii_case("role") || s.eq_ignore_ascii_case("user") =>
10643 {
10644 // NB: the enclosing `match self.advance()` already consumed
10645 // ROLE/USER — the round-695 trap, hit again in this round's
10646 // first draft (the name was eaten and WITH parsed as the
10647 // role). The cursor is at the name.
10648 let name = self.expect_ident_or_string()?;
10649 // v7.39 (round 750) — scan the attribute tail for
10650 // PASSWORD. Everything else stays a recorded no-op, but
10651 // a dropped credential rotation is a SECURITY bug:
10652 // `ALTER USER x PASSWORD 'new'` answered ALTER ROLE and
10653 // changed nothing, so the old password kept working.
10654 // ENCRYPTED/UNENCRYPTED are PG-noise prefixes; `PASSWORD
10655 // NULL` clears the credential.
10656 let mut password: Option<Option<String>> = None;
10657 loop {
10658 match self.peek() {
10659 Token::Semicolon | Token::Eof => break,
10660 Token::Ident(w) if w.eq_ignore_ascii_case("password") => {
10661 self.advance();
10662 match self.advance() {
10663 Token::String(p) => password = Some(Some(p)),
10664 Token::Null => password = Some(None),
10665 Token::Ident(n) if n.eq_ignore_ascii_case("null") => {
10666 password = Some(None);
10667 }
10668 other => {
10669 return Err(self.err(alloc::format!(
10670 "expected password string or NULL after PASSWORD, got {other:?}"
10671 )));
10672 }
10673 }
10674 }
10675 _ => {
10676 self.advance();
10677 }
10678 }
10679 }
10680 if name.eq_ignore_ascii_case("all") {
10681 // `ALTER ROLE ALL …` names every role; nothing to check.
10682 return Ok(Statement::Empty);
10683 }
10684 if let Some(pw) = password {
10685 return Ok(Statement::AlterRolePassword { name, password: pw });
10686 }
10687 return Ok(Statement::ValidateOnly {
10688 kind: crate::ast::ValidateOnlyKind::RoleName,
10689 names: alloc::vec![name],
10690 });
10691 }
10692 // v7.39 (round 709) — ALTER COLLATION / TEXT SEARCH
10693 // CONFIGURATION / EVENT TRIGGER / LARGE OBJECT leave the no-op
10694 // list far enough to validate the NAME; the actions still no-op.
10695 // (TEXT SEARCH DICTIONARY / PARSER / TEMPLATE stay noise: SPG
10696 // models none of them and their dumps are rare.)
10697 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("collation") => {
10698 let name = self.expect_ident_or_string()?;
10699 self.consume_until_statement_boundary();
10700 return Ok(Statement::ValidateOnly {
10701 kind: crate::ast::ValidateOnlyKind::CollationName,
10702 names: alloc::vec![name],
10703 });
10704 }
10705 Token::Ident(s) | Token::QuotedIdent(s)
10706 if s.eq_ignore_ascii_case("text")
10707 && matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("search"))
10708 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(k)) if k.eq_ignore_ascii_case("configuration")) =>
10709 {
10710 self.advance(); // SEARCH
10711 self.advance(); // CONFIGURATION
10712 let name = self.expect_ident_like()?;
10713 self.consume_until_statement_boundary();
10714 return Ok(Statement::ValidateOnly {
10715 kind: crate::ast::ValidateOnlyKind::TsConfigName,
10716 names: alloc::vec![name],
10717 });
10718 }
10719 Token::Ident(s) | Token::QuotedIdent(s)
10720 if s.eq_ignore_ascii_case("event")
10721 && matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("trigger")) =>
10722 {
10723 self.advance(); // TRIGGER
10724 let name = self.expect_ident_like()?;
10725 self.consume_until_statement_boundary();
10726 return Ok(Statement::ValidateOnly {
10727 kind: crate::ast::ValidateOnlyKind::EventTriggerName,
10728 names: alloc::vec![name],
10729 });
10730 }
10731 Token::Ident(s) | Token::QuotedIdent(s)
10732 if s.eq_ignore_ascii_case("large")
10733 && matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("object")) =>
10734 {
10735 self.advance(); // OBJECT
10736 let oid = match self.advance() {
10737 Token::Integer(n) => alloc::format!("{n}"),
10738 other => {
10739 return Err(
10740 self.err(alloc::format!("expected large object oid, got {other:?}"))
10741 );
10742 }
10743 };
10744 self.consume_until_statement_boundary();
10745 return Ok(Statement::ValidateOnly {
10746 kind: crate::ast::ValidateOnlyKind::LargeObjectOid,
10747 names: alloc::vec![oid],
10748 });
10749 }
10750 // v7.39 (round 708) — `ALTER AGGREGATE name(args) …`. Same
10751 // argument-list parse as DROP AGGREGATE (round 707); the
10752 // action no-ops, the existence check is real.
10753 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("aggregate") => {
10754 // Same round-695 trap as above: AGGREGATE is already
10755 // consumed; the cursor is at the name.
10756 let name = self.expect_ident_like()?;
10757 let mut names = alloc::vec![name];
10758 if matches!(self.peek(), Token::LParen) {
10759 self.advance();
10760 loop {
10761 match self.peek().clone() {
10762 Token::RParen => {
10763 self.advance();
10764 break;
10765 }
10766 Token::Star => {
10767 self.advance();
10768 names.push(String::from("*"));
10769 }
10770 Token::Comma => {
10771 self.advance();
10772 }
10773 _ => {
10774 let mut t = self.expect_ident_like()?;
10775 while let Token::Ident(nx) = self.peek() {
10776 let nx = nx.clone();
10777 self.advance();
10778 t.push(' ');
10779 t.push_str(&nx);
10780 }
10781 names.push(t);
10782 }
10783 }
10784 }
10785 }
10786 self.consume_until_statement_boundary();
10787 return Ok(Statement::ValidateOnly {
10788 kind: crate::ast::ValidateOnlyKind::AggregateName,
10789 names,
10790 });
10791 }
10792 Token::Ident(s) | Token::QuotedIdent(s)
10793 if matches!(
10794 s.to_ascii_lowercase().as_str(),
10795 "view"
10796 | "function"
10797 | "database"
10798 | "schema"
10799 | "owner"
10800 | "default"
10801 | "extension"
10802 | "materialized"
10803 | "publication"
10804 | "subscription"
10805 // v7.37.17 (17.6 siblings) — additional ALTER
10806 // targets pg_dump / pg_dumpall / operator DB
10807 // migration scripts commonly emit. SPG has
10808 // no matching machinery for any of these; the
10809 // parser accepts + Empty-returns so pg_dump
10810 // tail statements don't stall.
10811 | "tablespace"
10812 | "language"
10813 | "operator"
10814 | "conversion"
10815 | "statistics"
10816 | "server"
10817 | "foreign"
10818 // `text` stays for TEXT SEARCH DICTIONARY / PARSER
10819 // / TEMPLATE (CONFIGURATION intercepted above).
10820 | "text"
10821 ) =>
10822 {
10823 self.consume_until_statement_boundary();
10824 return Ok(Statement::Empty);
10825 }
10826 other => {
10827 return Err(self.err(format!(
10828 "expected INDEX / TABLE / SEQUENCE / VIEW / FUNCTION / TYPE / OWNER / etc \
10829 after ALTER, got {other:?}"
10830 )));
10831 }
10832 }
10833 // v7.16.2 — optional `IF EXISTS` after ALTER INDEX
10834 // (mailrs migrate-042 ships these). The presence of an
10835 // IF EXISTS makes the subsequent name lookup tolerate
10836 // a missing index — engine returns CommandOk no-op.
10837 let if_exists = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if")) {
10838 let next = self.tokens.get(self.pos + 1);
10839 if matches!(next, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("exists")) {
10840 self.advance();
10841 self.advance();
10842 true
10843 } else {
10844 false
10845 }
10846 } else {
10847 false
10848 };
10849 let name = self.expect_ident_like()?;
10850 // v7.16.2 — RENAME TO new_name shape (mailrs migrate-042).
10851 // Detect BEFORE the REBUILD path so the existing REBUILD
10852 // arm stays untouched.
10853 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("rename")) {
10854 self.advance();
10855 if matches!(self.peek(), Token::To) {
10856 self.advance();
10857 } else {
10858 self.expect_keyword_ident("to")?;
10859 }
10860 let new = self.expect_ident_like()?;
10861 return Ok(Statement::AlterIndex(crate::ast::AlterIndexStatement {
10862 name,
10863 target: crate::ast::AlterIndexTarget::Rename { new, if_exists },
10864 }));
10865 }
10866 // v7.39 (round 710) — SET ( … ) / RESET ( … ) storage parameters.
10867 // A syntax error before; the index is validated, the params no-op.
10868 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("reset"))
10869 || (matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("set"))
10870 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen)))
10871 {
10872 self.consume_until_statement_boundary();
10873 return Ok(Statement::AlterIndex(crate::ast::AlterIndexStatement {
10874 name,
10875 target: crate::ast::AlterIndexTarget::StorageParams,
10876 }));
10877 }
10878 // REBUILD
10879 self.expect_keyword_ident("rebuild")?;
10880 // Optional: WITH (encoding = <enc>)
10881 let encoding = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with")) {
10882 self.advance();
10883 if !matches!(self.peek(), Token::LParen) {
10884 return Err(self.err(format!(
10885 "expected '(' after WITH in ALTER INDEX REBUILD, got {:?}",
10886 self.peek()
10887 )));
10888 }
10889 self.advance();
10890 self.expect_keyword_ident("encoding")?;
10891 if !matches!(self.peek(), Token::Eq) {
10892 return Err(self.err(format!(
10893 "expected '=' after encoding in ALTER INDEX REBUILD, got {:?}",
10894 self.peek()
10895 )));
10896 }
10897 self.advance();
10898 let enc_ident = match self.advance() {
10899 Token::Ident(s) | Token::QuotedIdent(s) => s,
10900 other => {
10901 return Err(self.err(format!("expected encoding name after =, got {other:?}")));
10902 }
10903 };
10904 let enc = match enc_ident.to_ascii_lowercase().as_str() {
10905 "f32" => VecEncoding::F32,
10906 "sq8" => VecEncoding::Sq8,
10907 "half" => VecEncoding::F16,
10908 other => {
10909 return Err(self.err(format!(
10910 "unknown vector encoding {other:?} in ALTER INDEX REBUILD; supported: F32, SQ8, HALF"
10911 )));
10912 }
10913 };
10914 if !matches!(self.peek(), Token::RParen) {
10915 return Err(self.err(format!(
10916 "expected ')' after encoding value, got {:?}",
10917 self.peek()
10918 )));
10919 }
10920 self.advance();
10921 Some(enc)
10922 } else {
10923 None
10924 };
10925 Ok(Statement::AlterIndex(crate::ast::AlterIndexStatement {
10926 name,
10927 target: crate::ast::AlterIndexTarget::Rebuild { encoding },
10928 }))
10929 }
10930
10931 /// v6.7.2 — `ALTER TABLE <name> SET hot_tier_bytes = <n>`. The
10932 /// only `SET` form currently supported; future v6.7.x can add
10933 /// more SET subjects without changing the dispatch shape.
10934 /// v7.13.2 — mailrs round-6 S1: accepts comma-separated
10935 /// subactions. Single-subaction shape stays a 1-element vec.
10936 fn parse_alter_table_after_keyword(&mut self) -> Result<Statement, ParseError> {
10937 let table_name = self.expect_ident_like()?;
10938 let mut targets: Vec<crate::ast::AlterTableTarget> = Vec::new();
10939 loop {
10940 let subaction = self.parse_alter_table_subaction()?;
10941 // ADD COLUMN with inline REFERENCES emits both an
10942 // AddColumn and an AddForeignKey subaction; the
10943 // helper returns 1 or 2 items.
10944 targets.extend(subaction);
10945 if matches!(self.peek(), Token::Comma) {
10946 self.advance();
10947 continue;
10948 }
10949 break;
10950 }
10951 Ok(Statement::AlterTable(crate::ast::AlterTableStatement {
10952 name: table_name,
10953 targets,
10954 }))
10955 }
10956
10957 /// Parse one ALTER TABLE subaction. Returns a Vec because
10958 /// inline `REFERENCES` on `ADD COLUMN` produces both an
10959 /// AddColumn and an AddForeignKey entry (mailrs round-6 S3).
10960 /// v7.39.9 — MySQL's `FIRST` / `AFTER <col>` trailer on ADD /
10961 /// MODIFY / CHANGE COLUMN. Absent is the PostgreSQL form, which
10962 /// appends.
10963 fn parse_column_position(&mut self) -> Option<crate::ast::ColumnPosition> {
10964 match self.peek() {
10965 Token::Ident(s) if s.eq_ignore_ascii_case("first") => {
10966 self.advance();
10967 Some(crate::ast::ColumnPosition::First)
10968 }
10969 Token::Ident(s) if s.eq_ignore_ascii_case("after") => {
10970 self.advance();
10971 let name = self.expect_ident_like().ok()?;
10972 Some(crate::ast::ColumnPosition::After(name))
10973 }
10974 _ => None,
10975 }
10976 }
10977
10978 fn parse_alter_table_subaction(
10979 &mut self,
10980 ) -> Result<Vec<crate::ast::AlterTableTarget>, ParseError> {
10981 match self.peek() {
10982 // v7.39.9 — MySQL's own ALTER TABLE vocabulary. Each one is
10983 // a statement a real migration emits and SPG answered 1064
10984 // for; measured against MySQL 9.7.2, one at a time, beside
10985 // the published image.
10986 Token::Ident(s)
10987 if s.eq_ignore_ascii_case("modify") || s.eq_ignore_ascii_case("change") =>
10988 {
10989 let changing = s.eq_ignore_ascii_case("change");
10990 self.advance();
10991 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("column")) {
10992 self.advance();
10993 }
10994 // `parse_column_def_with_fk` reads the NAME itself, so
10995 // `MODIFY` hands it the column and `CHANGE` eats the old
10996 // name first and lets it read the new one.
10997 let old_name = if changing {
10998 Some(self.expect_ident_like()?)
10999 } else {
11000 None
11001 };
11002 let (definition, _fk) = self.parse_column_def_with_fk()?;
11003 let column = old_name.clone().unwrap_or_else(|| definition.name.clone());
11004 let rename_to = if changing {
11005 Some(definition.name.clone())
11006 } else {
11007 None
11008 };
11009 let position = self.parse_column_position();
11010 Ok(alloc::vec![crate::ast::AlterTableTarget::ModifyColumn {
11011 column,
11012 rename_to,
11013 definition,
11014 position,
11015 }])
11016 }
11017 Token::Ident(s) if s.eq_ignore_ascii_case("auto_increment") => {
11018 self.advance();
11019 if matches!(self.peek(), Token::Eq) {
11020 self.advance();
11021 }
11022 let n = self.expect_u64_literal()?;
11023 Ok(alloc::vec![
11024 crate::ast::AlterTableTarget::SetTableAutoIncrement(
11025 i64::try_from(n).unwrap_or(i64::MAX)
11026 )
11027 ])
11028 }
11029 Token::Ident(s) if s.eq_ignore_ascii_case("engine") => {
11030 self.advance();
11031 if matches!(self.peek(), Token::Eq) {
11032 self.advance();
11033 }
11034 // v7.39.10 — as WRITTEN, the way `CREATE TABLE`'s ENGINE
11035 // clause has kept it since v7.39.3. The lexer folds a
11036 // bare identifier, and MySQL names the engine back
11037 // exactly: measured, `ALTER TABLE f1 ENGINE=NoSuchEng`
11038 // answers `Unknown storage engine 'NoSuchEng'` there and
11039 // answered `'nosucheng'` here — the one thing that
11040 // message is for is telling the operator which word in
11041 // their migration was wrong.
11042 let at = self.pos;
11043 let name = self.expect_ident_like()?;
11044 let written = self
11045 .source_span(at, at)
11046 .map(|raw| raw.trim().trim_matches('`').trim_matches('\''))
11047 .filter(|raw| raw.eq_ignore_ascii_case(&name))
11048 .map(alloc::string::String::from);
11049 Ok(alloc::vec![crate::ast::AlterTableTarget::SetEngine(
11050 written.unwrap_or(name)
11051 )])
11052 }
11053 Token::Ident(s) if s.eq_ignore_ascii_case("convert") => {
11054 self.advance();
11055 // CONVERT TO CHARACTER SET <cs> [COLLATE <c>]
11056 if matches!(self.peek(), Token::To) {
11057 self.advance();
11058 }
11059 let kw = self.expect_ident_like()?;
11060 if !kw.eq_ignore_ascii_case("character") {
11061 return Err(self.err("expected CHARACTER after CONVERT TO".into()));
11062 }
11063 let set_kw = self.expect_ident_like()?;
11064 if !set_kw.eq_ignore_ascii_case("set") {
11065 return Err(self.err("expected SET after CHARACTER".into()));
11066 }
11067 let charset = self.expect_ident_like()?;
11068 let collate =
11069 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("collate")) {
11070 self.advance();
11071 Some(self.expect_ident_like()?)
11072 } else {
11073 None
11074 };
11075 Ok(alloc::vec![
11076 crate::ast::AlterTableTarget::ConvertToCharacterSet { charset, collate }
11077 ])
11078 }
11079 Token::Ident(s) if s.eq_ignore_ascii_case("set") => {
11080 self.advance();
11081 // v7.37.18 (18.7-18.15) — SET ( option = value, … )
11082 // storage parameters: paren-prefixed; consume.
11083 if matches!(self.peek(), Token::LParen) {
11084 self.consume_until_statement_boundary();
11085 return Ok(Vec::new());
11086 }
11087 let setting = self.expect_ident_like()?;
11088 if setting.eq_ignore_ascii_case("hot_tier_bytes") {
11089 if !matches!(self.peek(), Token::Eq) {
11090 return Err(self.err(alloc::format!(
11091 "expected '=' after hot_tier_bytes, got {:?}",
11092 self.peek()
11093 )));
11094 }
11095 self.advance();
11096 let n = self.expect_u64_literal()?;
11097 return Ok(alloc::vec![crate::ast::AlterTableTarget::SetHotTierBytes(n)]);
11098 }
11099 // v7.37.18 (18.7 / 18.8 / 18.11 / 18.13 / 18.14) —
11100 // accept-and-no-op for ALTER TABLE SET <subject>
11101 // forms that pg_dump emits but SPG either treats
11102 // as N/A (single-tenant, single-owner, no shared
11103 // tablespaces) or accepts the dump-side declaration
11104 // without runtime effect:
11105 // SET SCHEMA <name> (18.11)
11106 // SET TABLESPACE <name> (18.8)
11107 // SET LOGGED / UNLOGGED (18.7 alt-form)
11108 // SET WITHOUT CLUSTER (18.13)
11109 // SET WITHOUT OIDS (PG legacy)
11110 // SET (option = value, …) (storage parameters)
11111 // SET REPLICA IDENTITY {…} (18.14)
11112 if setting.eq_ignore_ascii_case("schema")
11113 || setting.eq_ignore_ascii_case("tablespace")
11114 || setting.eq_ignore_ascii_case("logged")
11115 || setting.eq_ignore_ascii_case("unlogged")
11116 || setting.eq_ignore_ascii_case("without")
11117 {
11118 self.consume_until_statement_boundary();
11119 return Ok(Vec::new());
11120 }
11121 if setting.eq_ignore_ascii_case("replica") {
11122 // SET REPLICA IDENTITY {DEFAULT|FULL|NOTHING|USING INDEX <name>}
11123 self.consume_until_statement_boundary();
11124 return Ok(Vec::new());
11125 }
11126 // SET (option=value, …) — storage parameters.
11127 if matches!(self.peek(), Token::LParen) {
11128 self.consume_until_statement_boundary();
11129 return Ok(Vec::new());
11130 }
11131 Err(self.err(alloc::format!(
11132 "ALTER TABLE SET: unknown setting {setting:?}; supported: \
11133 hot_tier_bytes / SCHEMA / TABLESPACE / LOGGED / UNLOGGED / \
11134 WITHOUT CLUSTER / WITHOUT OIDS / REPLICA IDENTITY / (storage_params)"
11135 )))
11136 }
11137 // v7.39 (round 647) — `ALTER TABLE c INHERIT p`. Carried now,
11138 // not ignored: round 645 gave SPG the inheritance the
11139 // v7.37.18 no-op said it did not have.
11140 Token::Ident(s) if s.eq_ignore_ascii_case("inherit") => {
11141 self.advance();
11142 let parent = self.expect_ident_like()?;
11143 self.consume_until_statement_boundary();
11144 Ok(alloc::vec![crate::ast::AlterTableTarget::Inherit {
11145 parent,
11146 detach: false
11147 }])
11148 }
11149 // `NO INHERIT <parent>`. Guarded to NOT match `NO FORCE ROW
11150 // LEVEL SECURITY`, which has its own RLS arm below — without
11151 // the guard this swallowed NO FORCE as a no-op.
11152 Token::Ident(s)
11153 if s.eq_ignore_ascii_case("no")
11154 && !matches!(
11155 self.tokens.get(self.pos + 1),
11156 Some(Token::Ident(t)) if t.eq_ignore_ascii_case("force")
11157 ) =>
11158 {
11159 self.advance();
11160 if matches!(self.peek(), Token::Ident(k) | Token::QuotedIdent(k)
11161 if k.eq_ignore_ascii_case("inherit"))
11162 {
11163 self.advance();
11164 let parent = self.expect_ident_like()?;
11165 self.consume_until_statement_boundary();
11166 return Ok(alloc::vec![crate::ast::AlterTableTarget::Inherit {
11167 parent,
11168 detach: true
11169 }]);
11170 }
11171 self.consume_until_statement_boundary();
11172 Ok(Vec::new())
11173 }
11174 // v7.37.18 (18.10) — ALTER TABLE OWNER TO <user>. SPG is
11175 // single-owner, so there is still nothing to record.
11176 //
11177 // v7.39 (round 652) — but the name now reaches the engine,
11178 // which refuses a role that does not exist as PG does. The
11179 // no-op was swallowing the whole statement, so a dump naming
11180 // a role this server never heard of restored clean and left
11181 // the table owned by whoever ran the restore.
11182 Token::Ident(s) if s.eq_ignore_ascii_case("owner") => {
11183 self.advance();
11184 if matches!(self.peek(), Token::To) {
11185 self.advance();
11186 }
11187 let role = self.expect_ident_like()?;
11188 Ok(alloc::vec![crate::ast::AlterTableTarget::OwnerTo {
11189 role
11190 }])
11191 }
11192 // v7.37.18 (18.13) — ALTER TABLE CLUSTER ON <index>.
11193 // PG sets a hint; SPG doesn't have clustered storage, so the
11194 // hint itself stays a no-op.
11195 //
11196 // v7.39 (round 652) — the index name is checked now. PG
11197 // errors on one that does not exist, and swallowing that let
11198 // a typo'd CLUSTER ON pass silently.
11199 Token::Ident(s) if s.eq_ignore_ascii_case("cluster") => {
11200 self.advance();
11201 // `ON` is a reserved token, not an ident.
11202 if !matches!(self.peek(), Token::On) {
11203 return Err(self.err(alloc::format!(
11204 "expected ON after CLUSTER, got {:?}",
11205 self.peek()
11206 )));
11207 }
11208 self.advance();
11209 let index = self.expect_ident_like()?;
11210 Ok(alloc::vec![crate::ast::AlterTableTarget::ClusterOn {
11211 index: Some(index)
11212 }])
11213 }
11214 // v7.39 (read01 round 49) — ALTER TABLE REPLICA IDENTITY
11215 // { DEFAULT | FULL | NOTHING | USING INDEX <name> }. PG records
11216 // what a logical decoder puts in the old-tuple image; SPG's
11217 // replication is SQL-text, so there is nothing to record.
11218 // Accept-and-no-op (it used to be a parse error).
11219 Token::Ident(s) if s.eq_ignore_ascii_case("replica") => {
11220 self.advance();
11221 // v7.39 (round 710) — `REPLICA IDENTITY USING INDEX <i>`
11222 // validates the index; DEFAULT / FULL / NOTHING stay no-op.
11223 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("identity"))
11224 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(k)) if k.eq_ignore_ascii_case("using"))
11225 {
11226 self.advance(); // IDENTITY
11227 self.advance(); // USING
11228 if matches!(self.peek(), Token::Index)
11229 || matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("index"))
11230 {
11231 self.advance();
11232 }
11233 let index = self.expect_ident_like()?;
11234 self.consume_until_statement_boundary();
11235 return Ok(alloc::vec![
11236 crate::ast::AlterTableTarget::ReplicaIdentityUsingIndex { index }
11237 ]);
11238 }
11239 self.consume_until_statement_boundary();
11240 Ok(Vec::new())
11241 }
11242 // v7.37.18 (18.15) — ALTER TABLE VALIDATE CONSTRAINT <name>.
11243 //
11244 // v7.39 (round 652) — it used to consume the statement and
11245 // return nothing, on the stated theory that SPG validated at
11246 // ADD CONSTRAINT time so there was never anything left to
11247 // validate. Measured against PG18, ADD CONSTRAINT did not
11248 // scan the existing rows at all — the comment described a
11249 // property SPG did not have, which is why nobody looked. Both
11250 // halves are real now: ADD scans unless told NOT VALID, and
11251 // this scans what NOT VALID skipped.
11252 Token::Ident(s) if s.eq_ignore_ascii_case("validate") => {
11253 self.advance();
11254 self.expect_keyword_ident("constraint")?;
11255 let name = self.expect_ident_like()?;
11256 Ok(alloc::vec![
11257 crate::ast::AlterTableTarget::ValidateConstraint { name }
11258 ])
11259 }
11260 // v7.37.18 (18.18) — RESET ( option [, …] ). Inverse of
11261 // SET (option = value, …). PG uses it to clear per-table
11262 // storage params like fillfactor or autovacuum_*. SPG
11263 // engine-manages those parameters; accept-and-no-op.
11264 Token::Ident(s) if s.eq_ignore_ascii_case("reset") => {
11265 self.advance();
11266 self.consume_until_statement_boundary();
11267 Ok(Vec::new())
11268 }
11269 // v7.37.18 (18.18) — OF <type_name> / NOT OF. Composite-
11270 // type-of binding (PG 9.0+). SPG composite types
11271 // (v7.37.5 ζ-B sub-commit) follow CREATE TYPE; ALTER
11272 // TABLE OF is rare and inverse of CREATE TABLE OF.
11273 // Accept-and-no-op until a customer dump round-trips it.
11274 Token::Ident(s) if s.eq_ignore_ascii_case("of") => {
11275 self.advance();
11276 // v7.39 (round 710) — the type name is validated now.
11277 let type_name = self.expect_ident_like()?;
11278 self.consume_until_statement_boundary();
11279 Ok(alloc::vec![crate::ast::AlterTableTarget::OfType {
11280 type_name
11281 }])
11282 }
11283 // v7.37.18 (18.18) — `NOT OF` lexes NOT as Token::Not
11284 // (reserved keyword) rather than Token::Ident("not"),
11285 // so it needs its own arm. Accept-and-no-op same as OF.
11286 Token::Not => {
11287 self.advance();
11288 self.consume_until_statement_boundary();
11289 Ok(Vec::new())
11290 }
11291 // v7.39 (RLS) — FORCE ROW LEVEL SECURITY (sets relforcerowsecurity).
11292 Token::Ident(s) if s.eq_ignore_ascii_case("force") => {
11293 self.advance();
11294 self.expect_row_level_security()?;
11295 Ok(alloc::vec![crate::ast::AlterTableTarget::SetRowSecurity {
11296 enabled: None,
11297 force: Some(true),
11298 }])
11299 }
11300 // v7.39 (RLS) — NO FORCE ROW LEVEL SECURITY.
11301 Token::Ident(s)
11302 if s.eq_ignore_ascii_case("no")
11303 && matches!(
11304 self.tokens.get(self.pos + 1),
11305 Some(Token::Ident(t)) if t.eq_ignore_ascii_case("force")
11306 ) =>
11307 {
11308 self.advance(); // NO
11309 self.advance(); // FORCE
11310 self.expect_row_level_security()?;
11311 Ok(alloc::vec![crate::ast::AlterTableTarget::SetRowSecurity {
11312 enabled: None,
11313 force: Some(false),
11314 }])
11315 }
11316 // v7.39 (RLS) — ENABLE/DISABLE ROW LEVEL SECURITY
11317 // (sets relrowsecurity). The guard requires the next token to be
11318 // `ROW` so the ENABLE/DISABLE TRIGGER arm still matches its case.
11319 Token::Ident(s)
11320 if (s.eq_ignore_ascii_case("enable") || s.eq_ignore_ascii_case("disable"))
11321 && matches!(
11322 self.tokens.get(self.pos + 1),
11323 Some(Token::Ident(t)) if t.eq_ignore_ascii_case("row")
11324 ) =>
11325 {
11326 let enabled = s.eq_ignore_ascii_case("enable");
11327 self.advance(); // ENABLE/DISABLE
11328 self.expect_row_level_security()?;
11329 Ok(alloc::vec![crate::ast::AlterTableTarget::SetRowSecurity {
11330 enabled: Some(enabled),
11331 force: None,
11332 }])
11333 }
11334 Token::Ident(s) if s.eq_ignore_ascii_case("add") => {
11335 self.advance();
11336 // v7.39 (round 431) — MySQL's `ALTER TABLE t ADD [UNIQUE]
11337 // {INDEX|KEY} [name] (cols)`, which every ORM migration
11338 // emits. The same grammar CREATE TABLE already accepts
11339 // inline (`KEY idx (a)`, prefix lengths and all), so it goes
11340 // through the SAME parser — an ALTER-only copy would be a
11341 // second place for the two to drift.
11342 if self.peek_mysql_inline_key_start() {
11343 return Ok(match self.parse_mysql_inline_key()? {
11344 Some(c) => {
11345 alloc::vec![crate::ast::AlterTableTarget::AddTableConstraint(c)]
11346 }
11347 // FULLTEXT / SPATIAL parse and are accepted as a
11348 // no-op here exactly as they are inline.
11349 None => Vec::new(),
11350 });
11351 }
11352 // v7.14.0 — ADD CONSTRAINT <name> { FOREIGN KEY |
11353 // PRIMARY KEY | UNIQUE | CHECK }. pg_dump emits
11354 // PRIMARY KEY this way; mysqldump emits both.
11355 // Peek-only dispatch (no advance) — `advance()`
11356 // destructively replaces consumed tokens with Eof,
11357 // so saved-pos restore would land on Eofs.
11358 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint"))
11359 {
11360 // The next-but-one ident is the constraint
11361 // name; the one after THAT is the kind.
11362 let kind_pos = self.pos + 2;
11363 let kind = self.tokens.get(kind_pos).cloned();
11364 if matches!(&kind, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("foreign"))
11365 {
11366 let fk = self.parse_table_level_fk()?;
11367 return Ok(alloc::vec![
11368 crate::ast::AlterTableTarget::AddForeignKey(fk)
11369 ]);
11370 }
11371 if matches!(&kind, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("primary"))
11372 {
11373 self.advance(); // CONSTRAINT
11374 // v7.39 (read01 round 48) — keep the name; the engine
11375 // stores it now instead of dropping it on the floor.
11376 let con_name = self.expect_ident_like()?;
11377 self.advance(); // PRIMARY
11378 self.expect_keyword_ident("key")?;
11379 let cols = self.parse_paren_ident_list("PRIMARY KEY")?;
11380 // v7.39 (round 711) — the ALTER form carries the
11381 // timing too (pg_dump writes it here).
11382 let (deferrable, initially_deferred) =
11383 self.consume_deferrable_clauses_timed()?;
11384 return Ok(alloc::vec![
11385 crate::ast::AlterTableTarget::AddTableConstraint(
11386 crate::ast::TableConstraint::PrimaryKey {
11387 name: Some(con_name),
11388 columns: cols,
11389 deferrable,
11390 initially_deferred,
11391 }
11392 )
11393 ]);
11394 }
11395 if matches!(&kind, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("unique"))
11396 {
11397 self.advance(); // CONSTRAINT
11398 // v7.39 (read01 round 48) — keep the name.
11399 let con_name = self.expect_ident_like()?;
11400 // v7.22 (mailrs round-13 gap 6) — delegate so
11401 // the optional `NULLS [NOT] DISTINCT` modifier
11402 // parses here too (pg_dump emits the ALTER
11403 // form; semantics enforced by the engine
11404 // since v7.13).
11405 let mut uc = self.parse_table_level_unique()?;
11406 if let crate::ast::TableConstraint::Unique { name, .. } = &mut uc {
11407 *name = Some(con_name);
11408 }
11409 return Ok(alloc::vec![
11410 crate::ast::AlterTableTarget::AddTableConstraint(uc)
11411 ]);
11412 }
11413 if matches!(&kind, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("check"))
11414 {
11415 self.advance(); // CONSTRAINT
11416 // v7.39 (read01 round 48) — keep the name.
11417 let con_name = self.expect_ident_like()?;
11418 self.advance(); // CHECK
11419 if !matches!(self.peek(), Token::LParen) {
11420 return Err(self.err(alloc::format!(
11421 "expected '(' after CHECK, got {:?}", self.peek()
11422 )));
11423 }
11424 self.advance();
11425 let expr = self.parse_expr(0)?;
11426 if matches!(self.peek(), Token::RParen) {
11427 self.advance();
11428 }
11429 let not_valid = self.parse_not_valid_suffix();
11430 return Ok(alloc::vec![
11431 crate::ast::AlterTableTarget::AddTableConstraint(
11432 crate::ast::TableConstraint::Check {
11433 name: Some(con_name),
11434 expr,
11435 not_valid,
11436 }
11437 )
11438 ]);
11439 }
11440 // v7.39 (round 211) — ADD CONSTRAINT <name> EXCLUDE
11441 // [USING <am>] (<col> WITH <op>[, …]). pg_dump emits
11442 // exclusion constraints via this ALTER form.
11443 if matches!(&kind, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("exclude"))
11444 {
11445 self.advance(); // CONSTRAINT
11446 let con_name = self.expect_ident_like()?;
11447 let mut ex = self.parse_table_level_exclude()?;
11448 if let crate::ast::TableConstraint::Exclude { name, .. } = &mut ex {
11449 *name = Some(con_name);
11450 }
11451 return Ok(alloc::vec![
11452 crate::ast::AlterTableTarget::AddTableConstraint(ex)
11453 ]);
11454 }
11455 // Unknown kind — fall through to FK path which
11456 // produces a descriptive parse error.
11457 }
11458 let is_fk = matches!(
11459 self.peek(),
11460 Token::Ident(s) if s.eq_ignore_ascii_case("constraint")
11461 || s.eq_ignore_ascii_case("foreign")
11462 );
11463 if is_fk {
11464 let fk = self.parse_table_level_fk()?;
11465 return Ok(alloc::vec![crate::ast::AlterTableTarget::AddForeignKey(fk)]);
11466 }
11467 // v7.14.0 — bare ADD PRIMARY KEY / UNIQUE / CHECK
11468 // (no CONSTRAINT prefix) — same dispatch.
11469 match self.peek().clone() {
11470 Token::Ident(s) if s.eq_ignore_ascii_case("primary") => {
11471 self.advance();
11472 self.expect_keyword_ident("key")?;
11473 let cols = self.parse_paren_ident_list("PRIMARY KEY")?;
11474 let (deferrable, initially_deferred) =
11475 self.consume_deferrable_clauses_timed()?;
11476 return Ok(alloc::vec![
11477 crate::ast::AlterTableTarget::AddTableConstraint(
11478 crate::ast::TableConstraint::PrimaryKey {
11479 name: None,
11480 columns: cols,
11481 deferrable,
11482 initially_deferred,
11483 }
11484 )
11485 ]);
11486 }
11487 Token::Ident(s) if s.eq_ignore_ascii_case("unique") => {
11488 // v7.22 — delegate (NULLS [NOT] DISTINCT).
11489 let uc = self.parse_table_level_unique()?;
11490 return Ok(alloc::vec![
11491 crate::ast::AlterTableTarget::AddTableConstraint(uc)
11492 ]);
11493 }
11494 // v7.39 (round 652) — bare ADD CHECK (no CONSTRAINT
11495 // prefix). The other three bare forms were here and
11496 // this one was not, so it fell through to the column
11497 // path and came back as "unexpected reserved keyword
11498 // 'check' at start of column definition".
11499 _ if self.peek_table_level_check_start() => {
11500 let chk = self.parse_table_level_check()?;
11501 let not_valid = self.parse_not_valid_suffix();
11502 let crate::ast::TableConstraint::Check { expr, .. } = chk else {
11503 unreachable!("parse_table_level_check returns Check")
11504 };
11505 return Ok(alloc::vec![
11506 crate::ast::AlterTableTarget::AddTableConstraint(
11507 crate::ast::TableConstraint::Check {
11508 name: None,
11509 expr,
11510 not_valid,
11511 }
11512 )
11513 ]);
11514 }
11515 // v7.39 (round 211) — bare ADD EXCLUDE (no CONSTRAINT prefix).
11516 Token::Ident(s) if s.eq_ignore_ascii_case("exclude") => {
11517 let ex = self.parse_table_level_exclude()?;
11518 return Ok(alloc::vec![
11519 crate::ast::AlterTableTarget::AddTableConstraint(ex)
11520 ]);
11521 }
11522 _ => {}
11523 }
11524 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("column")) {
11525 self.advance();
11526 }
11527 let mut if_not_exists = false;
11528 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if")) {
11529 self.advance();
11530 if !matches!(self.peek(), Token::Not) {
11531 return Err(self.err(alloc::format!(
11532 "expected NOT after IF in ALTER TABLE ADD COLUMN, got {:?}",
11533 self.peek()
11534 )));
11535 }
11536 self.advance();
11537 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("exists")) {
11538 return Err(self.err(alloc::format!(
11539 "expected EXISTS after IF NOT in ALTER TABLE ADD COLUMN, got {:?}",
11540 self.peek()
11541 )));
11542 }
11543 self.advance();
11544 if_not_exists = true;
11545 }
11546 // v7.13.2 — mailrs round-6 S3: `ADD COLUMN col TYPE
11547 // REFERENCES other(col) [ON DELETE …]`. parse_column_def
11548 // returns ColumnDef + an optional inline FK.
11549 let (column, col_level_fk) = self.parse_column_def_with_fk()?;
11550 let col_name = column.name.clone();
11551 // v7.39.9 — MySQL says where the column goes.
11552 let position = self.parse_column_position();
11553 let mut out = alloc::vec![crate::ast::AlterTableTarget::AddColumn {
11554 column,
11555 if_not_exists,
11556 position,
11557 }];
11558 if let Some(mut fk) = col_level_fk {
11559 if fk.columns.is_empty() {
11560 fk.columns.push(col_name);
11561 }
11562 out.push(crate::ast::AlterTableTarget::AddForeignKey(fk));
11563 }
11564 Ok(out)
11565 }
11566 Token::Drop => {
11567 self.advance();
11568 // v7.13.3 — dispatch on the next token. mailrs round-7
11569 // S8 closed DROP COLUMN; round-6 S7 closed
11570 // DROP CONSTRAINT. Both share IF EXISTS / CASCADE /
11571 // RESTRICT modifiers.
11572 // DROP CONSTRAINT [IF EXISTS] <name> [CASCADE|RESTRICT]
11573 // DROP [COLUMN] [IF EXISTS] <col> [CASCADE|RESTRICT]
11574 let subject = match self.peek() {
11575 Token::Ident(s) if s.eq_ignore_ascii_case("constraint") => {
11576 self.advance();
11577 "constraint"
11578 }
11579 Token::Ident(s) if s.eq_ignore_ascii_case("column") => {
11580 self.advance();
11581 "column"
11582 }
11583 // v7.39 (round 431) — MySQL `DROP {INDEX|KEY} name`.
11584 // `INDEX` lexes as the reserved Token::Index, so it is
11585 // unambiguous. `KEY` is a plain ident, and PG allows a
11586 // column literally named "key", so only read it as the
11587 // keyword when a name follows it.
11588 Token::Index => {
11589 self.advance();
11590 "index"
11591 }
11592 Token::Ident(s)
11593 if s.eq_ignore_ascii_case("key")
11594 && matches!(
11595 self.tokens.get(self.pos + 1),
11596 Some(Token::Ident(_) | Token::QuotedIdent(_))
11597 ) =>
11598 {
11599 self.advance();
11600 "index"
11601 }
11602 // PG-canonical bare `DROP <col>` without COLUMN
11603 // keyword is also valid; treat any other ident
11604 // as the column name.
11605 Token::Ident(_) | Token::QuotedIdent(_) => "column",
11606 other => {
11607 return Err(self.err(alloc::format!(
11608 "expected COLUMN / CONSTRAINT after DROP in ALTER TABLE, got {other:?}"
11609 )));
11610 }
11611 };
11612 let mut if_exists = false;
11613 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if")) {
11614 let n1 = self.tokens.get(self.pos + 1);
11615 if matches!(n1, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("exists")) {
11616 self.advance();
11617 self.advance();
11618 if_exists = true;
11619 }
11620 }
11621 let name = self.expect_ident_like()?;
11622 let mut cascade = false;
11623 if matches!(
11624 self.peek(),
11625 Token::Ident(s) if s.eq_ignore_ascii_case("cascade")
11626 || s.eq_ignore_ascii_case("restrict")
11627 ) {
11628 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("cascade"))
11629 {
11630 cascade = true;
11631 }
11632 self.advance();
11633 }
11634 if subject == "index" {
11635 Ok(alloc::vec![crate::ast::AlterTableTarget::DropIndex {
11636 name,
11637 if_exists,
11638 }])
11639 } else if subject == "constraint" {
11640 Ok(alloc::vec![crate::ast::AlterTableTarget::DropForeignKey {
11641 name,
11642 if_exists,
11643 }])
11644 } else {
11645 Ok(alloc::vec![crate::ast::AlterTableTarget::DropColumn {
11646 column: name,
11647 if_exists,
11648 cascade,
11649 }])
11650 }
11651 }
11652 Token::Ident(s) if s.eq_ignore_ascii_case("alter") => {
11653 self.advance();
11654 // v7.37.18 (18.16) — `ALTER TABLE … ALTER CONSTRAINT
11655 // <name> {DEFERRABLE|NOT DEFERRABLE} [INITIALLY
11656 // {IMMEDIATE|DEFERRED}]`. SPG enforces constraints
11657 // immediately; accept-and-no-op.
11658 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint")) {
11659 self.advance();
11660 self.consume_until_statement_boundary();
11661 return Ok(Vec::new());
11662 }
11663 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("column")) {
11664 self.advance();
11665 }
11666 let col_name = self.expect_ident_like()?;
11667 match self.peek() {
11668 Token::Ident(s) if s.eq_ignore_ascii_case("type") => {
11669 self.advance();
11670 }
11671 // v7.14.0 — pg_dump emits BIGSERIAL via
11672 // `ALTER TABLE … ALTER COLUMN id SET DEFAULT
11673 // nextval('seq')` (the sequence is created
11674 // separately). SPG's BIGSERIAL already uses
11675 // AUTO_INCREMENT; accept SET DEFAULT / DROP
11676 // DEFAULT / SET NOT NULL / DROP NOT NULL as
11677 // engine no-ops by consuming the tail.
11678 Token::Ident(s) if s.eq_ignore_ascii_case("set") => {
11679 // v7.22 (round-13 T2) — `SET DEFAULT
11680 // nextval('…')` is how pg_dump spells a
11681 // SERIAL column (plain integer in CREATE
11682 // TABLE + this ALTER). It used to be
11683 // swallowed as a no-op, which silently
11684 // STRIPPED auto-increment from imported
11685 // schemas — the first post-import INSERT
11686 // without an explicit id then violated NOT
11687 // NULL. Lower it to the auto-increment
11688 // marker instead.
11689 let is_default_nextval =
11690 matches!(self.tokens.get(self.pos + 1), Some(Token::Default))
11691 && matches!(
11692 self.tokens.get(self.pos + 2),
11693 Some(Token::Ident(f)) if f.eq_ignore_ascii_case("nextval")
11694 );
11695 if is_default_nextval {
11696 let seq_name = self.scan_sequence_name_until_boundary();
11697 return Ok(alloc::vec![
11698 crate::ast::AlterTableTarget::SetColumnAutoIncrement {
11699 column: col_name,
11700 seq_name,
11701 }
11702 ]);
11703 }
11704 // v7.37.18 (18.1 + 18.2) — proper lowering.
11705 self.advance(); // consume "set"
11706 match self.peek().clone() {
11707 Token::Default => {
11708 self.advance();
11709 let default_expr = self.parse_expr(0)?;
11710 return Ok(alloc::vec![
11711 crate::ast::AlterTableTarget::AlterColumnSetDefault {
11712 column: col_name,
11713 default_expr,
11714 }
11715 ]);
11716 }
11717 Token::Not => {
11718 self.advance();
11719 if !matches!(self.peek(), Token::Null) {
11720 return Err(self.err(alloc::format!(
11721 "expected NULL after ALTER COLUMN SET NOT, got {:?}",
11722 self.peek()
11723 )));
11724 }
11725 self.advance();
11726 return Ok(alloc::vec![
11727 crate::ast::AlterTableTarget::AlterColumnSetNotNull {
11728 column: col_name,
11729 }
11730 ]);
11731 }
11732 // `SET EXPRESSION AS (expr)` (PG 17) — change a
11733 // stored generated column's expression and
11734 // recompute existing rows.
11735 Token::Ident(s) if s.eq_ignore_ascii_case("expression") => {
11736 self.advance(); // EXPRESSION
11737 if matches!(self.peek(), Token::As) {
11738 self.advance();
11739 }
11740 let expr = self.parse_expr(0)?;
11741 return Ok(alloc::vec![
11742 crate::ast::AlterTableTarget::AlterColumnSetExpression {
11743 column: col_name,
11744 expr,
11745 }
11746 ]);
11747 }
11748 other => {
11749 // Other SET subjects (STATISTICS,
11750 // STORAGE, COMPRESSION, …) stay no-ops —
11751 // storage hints with no SPG semantics.
11752 let _ = other;
11753 self.consume_until_statement_boundary();
11754 return Ok(Vec::new());
11755 }
11756 }
11757 }
11758 Token::Ident(s) if s.eq_ignore_ascii_case("drop") => {
11759 self.advance(); // consume "drop"
11760 return self.parse_alter_column_drop_tail(col_name);
11761 }
11762 Token::Drop => {
11763 self.advance(); // consume Drop token
11764 return self.parse_alter_column_drop_tail(col_name);
11765 }
11766 Token::Ident(s) if s.eq_ignore_ascii_case("add") => {
11767 // v7.22 (round-13 T2) — `ALTER COLUMN c ADD
11768 // GENERATED { ALWAYS | BY DEFAULT } AS
11769 // IDENTITY ( … )`: pg_dump's spelling for
11770 // identity columns. Same auto-increment
11771 // lowering as the nextval default; the
11772 // sequence options inside the parens are
11773 // no-ops under SPG's max+1 semantics.
11774 let is_generated = matches!(
11775 self.tokens.get(self.pos + 1),
11776 Some(Token::Ident(g)) if g.eq_ignore_ascii_case("generated")
11777 );
11778 if !is_generated {
11779 return Err(self.err(alloc::format!(
11780 "expected GENERATED after ALTER COLUMN {col_name} ADD, got {:?}",
11781 self.tokens.get(self.pos + 1)
11782 )));
11783 }
11784 let seq_name = self.scan_sequence_name_until_boundary();
11785 return Ok(alloc::vec![
11786 crate::ast::AlterTableTarget::SetColumnAutoIncrement {
11787 column: col_name,
11788 seq_name,
11789 }
11790 ]);
11791 }
11792 // v7.39 (round 220) — `RESTART [WITH n]` on an identity
11793 // column: floor the next allocated value at n (bare
11794 // RESTART = restart from the start value, 1).
11795 Token::Ident(s) if s.eq_ignore_ascii_case("restart") => {
11796 self.advance();
11797 let with = if matches!(self.peek(), Token::Ident(w) if w.eq_ignore_ascii_case("with"))
11798 {
11799 self.advance();
11800 let neg = if matches!(self.peek(), Token::Minus) {
11801 self.advance();
11802 true
11803 } else {
11804 false
11805 };
11806 match self.advance() {
11807 Token::Integer(v) => Some(if neg { -v } else { v }),
11808 other => {
11809 return Err(self.err(alloc::format!(
11810 "expected integer after RESTART WITH, got {other:?}"
11811 )));
11812 }
11813 }
11814 } else {
11815 None
11816 };
11817 return Ok(alloc::vec![
11818 crate::ast::AlterTableTarget::AlterColumnRestart {
11819 column: col_name,
11820 with,
11821 }
11822 ]);
11823 }
11824 other => {
11825 return Err(self.err(alloc::format!(
11826 "expected TYPE / SET / DROP / ADD after ALTER COLUMN <name>, got {other:?}"
11827 )));
11828 }
11829 }
11830 // v7.39 (round 713) — the type parser has consumed a
11831 // trailing `COLLATE <name>` since Phase 2.5, and
11832 // `parse_column_type_name` discarded it: `ALTER COLUMN t
11833 // TYPE text COLLATE "C"` parsed clean and changed
11834 // nothing. Keep the clause; the engine re-collates.
11835 let (new_type, _, _, _, coll, coll_explicit, coll_name, _, _, _, _, _, _, _) =
11836 self.parse_type_with_implied_flags()?;
11837 let collation = if coll_explicit {
11838 coll_name.map(|n| (coll, n))
11839 } else {
11840 None
11841 };
11842 let using = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using"))
11843 {
11844 self.advance();
11845 Some(self.parse_expr(0)?)
11846 } else {
11847 None
11848 };
11849 Ok(alloc::vec![crate::ast::AlterTableTarget::AlterColumnType {
11850 column: col_name,
11851 new_type,
11852 using,
11853 collation,
11854 }])
11855 }
11856 // v7.15.0 — `ALTER TABLE t RENAME [COLUMN] old TO new`.
11857 // PG also supports `RENAME TO new_table` for table-name
11858 // rename; that surface is deferred (pg_dump never emits
11859 // it). If the first post-RENAME ident is `TO`, the user
11860 // is asking for table rename — error with a clear
11861 // message rather than misparsing `TO` as a column name.
11862 Token::Ident(s) if s.eq_ignore_ascii_case("rename") => {
11863 self.advance();
11864 // v7.16.2 — `ALTER TABLE t RENAME TO new_table`
11865 // table-name rename (mailrs round-10 A.5 — used
11866 // by migrate-042's `RENAME TO email_contacts`).
11867 // `TO` lexes as Token::To.
11868 if matches!(self.peek(), Token::To)
11869 || matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("to"))
11870 {
11871 self.advance();
11872 let new = self.expect_ident_like()?;
11873 return Ok(alloc::vec![crate::ast::AlterTableTarget::RenameTable {
11874 new,
11875 }]);
11876 }
11877 // v7.39 (read01 round 48) — `RENAME CONSTRAINT old TO new`.
11878 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint")) {
11879 self.advance();
11880 let old = self.expect_ident_like()?;
11881 if matches!(self.peek(), Token::To) {
11882 self.advance();
11883 } else {
11884 self.expect_keyword_ident("to")?;
11885 }
11886 let new = self.expect_ident_like()?;
11887 return Ok(alloc::vec![
11888 crate::ast::AlterTableTarget::RenameConstraint { old, new }
11889 ]);
11890 }
11891 // v7.39.9 — MySQL's `RENAME {INDEX|KEY} old TO new`.
11892 // PostgreSQL renames an index with its own top-level
11893 // `ALTER INDEX`, so this spelling had nowhere to go and
11894 // answered 1064; MySQL 9.7.2 parses it and answers 1176
11895 // when the key is not there.
11896 if matches!(self.peek(), Token::Index)
11897 || matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("key"))
11898 {
11899 self.advance();
11900 let old = self.expect_ident_like()?;
11901 if matches!(self.peek(), Token::To) {
11902 self.advance();
11903 } else {
11904 self.expect_keyword_ident("to")?;
11905 }
11906 let new = self.expect_ident_like()?;
11907 return Ok(alloc::vec![crate::ast::AlterTableTarget::RenameIndex {
11908 old,
11909 new,
11910 }]);
11911 }
11912 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("column")) {
11913 self.advance();
11914 }
11915 let old = self.expect_ident_like()?;
11916 // `TO` is a reserved keyword token; accept both
11917 // Token::To and Token::Ident("to") for consistency.
11918 if matches!(self.peek(), Token::To) {
11919 self.advance();
11920 } else {
11921 self.expect_keyword_ident("to")?;
11922 }
11923 let new = self.expect_ident_like()?;
11924 Ok(alloc::vec![crate::ast::AlterTableTarget::RenameColumn {
11925 old,
11926 new,
11927 }])
11928 }
11929 // v7.16.1 — `ALTER TABLE t { ENABLE | DISABLE } TRIGGER
11930 // { ALL | <name> }`. pg_dump --disable-triggers wraps
11931 // every data block with these. Real disable semantics —
11932 // not no-op — because reload correctness assumes the
11933 // triggers don't fire (rows already carry their
11934 // computed values from prod).
11935 Token::Ident(s)
11936 if s.eq_ignore_ascii_case("enable") || s.eq_ignore_ascii_case("disable") =>
11937 {
11938 let enabled = s.eq_ignore_ascii_case("enable");
11939 self.advance();
11940 // PG also accepts ENABLE/DISABLE { REPLICA | ALWAYS }
11941 // TRIGGER … and ENABLE/DISABLE RULE / ROW LEVEL
11942 // SECURITY. v7.16.1 only matches TRIGGER (mailrs's
11943 // pg_dump output) — anything else falls through to
11944 // the catch-all error below.
11945 // v7.22 (round-13 T3) — mysqldump wraps every data
11946 // section in `/*!40000 ALTER TABLE t DISABLE KEYS */`
11947 // + ENABLE KEYS (a MyISAM index-rebuild hint). SPG
11948 // maintains indexes incrementally — engine no-op.
11949 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("keys")) {
11950 self.advance();
11951 return Ok(Vec::new());
11952 }
11953 // v7.37.18 (18.12) — ENABLE/DISABLE ALWAYS TRIGGER
11954 // and ENABLE/DISABLE REPLICA TRIGGER. PG uses these
11955 // to gate triggers on session_replication_role; SPG
11956 // has no replica role, so the prefix is consumed and
11957 // treated identically to the plain ENABLE/DISABLE
11958 // TRIGGER form.
11959 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("always"))
11960 || matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("replica"))
11961 {
11962 self.advance();
11963 }
11964 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("trigger")) {
11965 return Err(self.err(alloc::format!(
11966 "expected TRIGGER after {}, got {:?}",
11967 if enabled { "ENABLE" } else { "DISABLE" },
11968 self.peek()
11969 )));
11970 }
11971 self.advance();
11972 // `ALL` lexes as Token::All (reserved); also
11973 // accept Token::Ident("all") for symmetry.
11974 // v7.37.18 (18.12) — USER / REPLICA / ALWAYS post-
11975 // TRIGGER selectors. USER (= all user triggers) is
11976 // semantically ALL here; REPLICA / ALWAYS gate on
11977 // session_replication_role which SPG doesn't track.
11978 // All map to TriggerSelector::All.
11979 let which = if matches!(self.peek(), Token::All)
11980 || matches!(self.peek(), Token::Ident(s)
11981 if s.eq_ignore_ascii_case("all")
11982 || s.eq_ignore_ascii_case("user")
11983 || s.eq_ignore_ascii_case("replica")
11984 || s.eq_ignore_ascii_case("always"))
11985 {
11986 self.advance();
11987 crate::ast::TriggerSelector::All
11988 } else {
11989 let name = self.expect_ident_like()?;
11990 crate::ast::TriggerSelector::Named(name)
11991 };
11992 Ok(alloc::vec![crate::ast::AlterTableTarget::SetTriggerEnabled {
11993 which,
11994 enabled,
11995 }])
11996 }
11997 // v7.37.16 (16.3) — ATTACH PARTITION child <bounds>
11998 Token::Ident(s) if s.eq_ignore_ascii_case("attach") => {
11999 self.advance();
12000 if !matches!(self.peek(), Token::Partition)
12001 && !matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12002 if s.eq_ignore_ascii_case("partition"))
12003 {
12004 return Err(self.err(alloc::format!(
12005 "expected PARTITION after ATTACH, got {:?}",
12006 self.peek()
12007 )));
12008 }
12009 self.advance();
12010 let child = self.expect_ident_like()?;
12011 let bounds = self.parse_partition_bounds_tail()?;
12012 Ok(alloc::vec![
12013 crate::ast::AlterTableTarget::AttachPartition { child, bounds }
12014 ])
12015 }
12016 // v7.37.16 (16.4 + 16.5) — DETACH PARTITION child [CONCURRENTLY] [FINALIZE]
12017 Token::Ident(s) if s.eq_ignore_ascii_case("detach") => {
12018 self.advance();
12019 if !matches!(self.peek(), Token::Partition)
12020 && !matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12021 if s.eq_ignore_ascii_case("partition"))
12022 {
12023 return Err(self.err(alloc::format!(
12024 "expected PARTITION after DETACH, got {:?}",
12025 self.peek()
12026 )));
12027 }
12028 self.advance();
12029 let child = self.expect_ident_like()?;
12030 let mut concurrently = false;
12031 let mut finalize = false;
12032 loop {
12033 match self.peek().clone() {
12034 Token::Ident(s) | Token::QuotedIdent(s)
12035 if s.eq_ignore_ascii_case("concurrently") =>
12036 {
12037 self.advance();
12038 concurrently = true;
12039 }
12040 Token::Ident(s) | Token::QuotedIdent(s)
12041 if s.eq_ignore_ascii_case("finalize") =>
12042 {
12043 self.advance();
12044 finalize = true;
12045 }
12046 _ => break,
12047 }
12048 }
12049 Ok(alloc::vec![crate::ast::AlterTableTarget::DetachPartition {
12050 child,
12051 concurrently,
12052 finalize,
12053 }])
12054 }
12055 other => Err(self.err(alloc::format!(
12056 "expected SET / ADD / DROP / ALTER / RENAME / ENABLE / DISABLE / ATTACH / DETACH in ALTER TABLE, got {other:?}"
12057 ))),
12058 }
12059 }
12060
12061 /// v7.37.16 (16.3) — parse the `FOR VALUES …` / `DEFAULT`
12062 /// tail used by both CREATE TABLE … PARTITION OF and ALTER
12063 /// TABLE … ATTACH PARTITION. Shares the same grammar as
12064 /// `parse_partition_of_tail`'s bounds branch.
12065 /// v7.37.18 (18.1 + 18.2) — parse the tail of `ALTER COLUMN
12066 /// col DROP …`. Accepts `DROP DEFAULT` and `DROP NOT NULL`,
12067 /// lowering each to the respective AlterTableTarget. Any
12068 /// other DROP subject (IDENTITY, EXPRESSION, etc.) stays a
12069 /// no-op via consume_until_statement_boundary.
12070 fn parse_alter_column_drop_tail(
12071 &mut self,
12072 col_name: String,
12073 ) -> Result<Vec<crate::ast::AlterTableTarget>, ParseError> {
12074 match self.peek().clone() {
12075 Token::Default => {
12076 self.advance();
12077 Ok(alloc::vec![
12078 crate::ast::AlterTableTarget::AlterColumnDropDefault { column: col_name }
12079 ])
12080 }
12081 Token::Not => {
12082 self.advance();
12083 if !matches!(self.peek(), Token::Null) {
12084 return Err(self.err(alloc::format!(
12085 "expected NULL after ALTER COLUMN DROP NOT, got {:?}",
12086 self.peek()
12087 )));
12088 }
12089 self.advance();
12090 Ok(alloc::vec![
12091 crate::ast::AlterTableTarget::AlterColumnDropNotNull { column: col_name }
12092 ])
12093 }
12094 // `DROP EXPRESSION [IF EXISTS]` — de-generate a stored
12095 // generated column into a plain column.
12096 Token::Ident(s) if s.eq_ignore_ascii_case("expression") => {
12097 self.advance();
12098 // v7.39 (round 187, U10) — IF EXISTS was consumed but
12099 // dropped, so the engine still errored on a plain
12100 // column; PG's semantics are NOTICE + skip.
12101 let mut if_exists = false;
12102 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if")) {
12103 self.advance();
12104 if matches!(self.peek(), Token::Ident(e) if e.eq_ignore_ascii_case("exists")) {
12105 self.advance();
12106 if_exists = true;
12107 }
12108 }
12109 Ok(alloc::vec![
12110 crate::ast::AlterTableTarget::AlterColumnDropExpression {
12111 column: col_name,
12112 if_exists,
12113 }
12114 ])
12115 }
12116 // v7.38 (read01, T28) — `DROP IDENTITY [IF EXISTS]` — de-generate an
12117 // identity column into a plain column.
12118 Token::Ident(s) if s.eq_ignore_ascii_case("identity") => {
12119 self.advance();
12120 let mut if_exists = false;
12121 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if")) {
12122 self.advance();
12123 if matches!(self.peek(), Token::Ident(e) if e.eq_ignore_ascii_case("exists")) {
12124 self.advance();
12125 if_exists = true;
12126 }
12127 }
12128 Ok(alloc::vec![
12129 crate::ast::AlterTableTarget::AlterColumnDropIdentity {
12130 column: col_name,
12131 if_exists,
12132 }
12133 ])
12134 }
12135 _ => {
12136 self.consume_until_statement_boundary();
12137 Ok(Vec::new())
12138 }
12139 }
12140 }
12141
12142 /// Parse the optional trailer of `COPY … TO STDOUT`: nothing (text
12143 /// format, no header), the modern `[WITH] ( opt [, opt]* )` list, or
12144 /// the legacy space-separated `[WITH] CSV|TEXT [HEADER] [DELIMITER
12145 /// 'c'] [NULL 'str'] [QUOTE 'c']` spelling.
12146 fn parse_copy_to_options(&mut self) -> Result<crate::ast::CopyOptions, ParseError> {
12147 let mut opts = crate::ast::CopyOptions::default();
12148 if matches!(self.peek(), Token::Eof | Token::Semicolon) {
12149 return Ok(opts);
12150 }
12151 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with")) {
12152 self.advance();
12153 }
12154 if matches!(self.peek(), Token::LParen) {
12155 self.advance();
12156 loop {
12157 self.parse_one_copy_option(&mut opts)?;
12158 match self.peek() {
12159 Token::Comma => {
12160 self.advance();
12161 }
12162 Token::RParen => {
12163 self.advance();
12164 break;
12165 }
12166 other => {
12167 return Err(self.err(alloc::format!(
12168 "expected ',' or ')' in COPY options, got {other:?}"
12169 )));
12170 }
12171 }
12172 }
12173 } else {
12174 while !matches!(self.peek(), Token::Eof | Token::Semicolon) {
12175 self.parse_one_copy_option(&mut opts)?;
12176 }
12177 }
12178 if !matches!(self.peek(), Token::Eof | Token::Semicolon) {
12179 return Err(self.err(alloc::format!(
12180 "unexpected token after COPY options: {:?}",
12181 self.peek()
12182 )));
12183 }
12184 Ok(opts)
12185 }
12186
12187 fn parse_one_copy_option(
12188 &mut self,
12189 opts: &mut crate::ast::CopyOptions,
12190 ) -> Result<(), ParseError> {
12191 use crate::ast::CopyFormat;
12192 // The option keyword. NULL lexes as its own token; the rest are
12193 // bare identifiers.
12194 let kw = match self.advance() {
12195 Token::Null => alloc::string::String::from("NULL"),
12196 Token::Ident(s) => s.to_uppercase(),
12197 other => {
12198 return Err(self.err(alloc::format!(
12199 "expected a COPY option keyword, got {other:?}"
12200 )));
12201 }
12202 };
12203 match kw.as_str() {
12204 "FORMAT" => {
12205 let fmt = self.expect_ident_like()?;
12206 match fmt.to_ascii_uppercase().as_str() {
12207 "CSV" => opts.format = CopyFormat::Csv,
12208 "TEXT" => opts.format = CopyFormat::Text,
12209 other => {
12210 return Err(self.err(alloc::format!(
12211 "COPY format \"{}\" not recognized",
12212 other.to_ascii_lowercase()
12213 )));
12214 }
12215 }
12216 }
12217 // Legacy bare format keywords.
12218 "CSV" => opts.format = CopyFormat::Csv,
12219 "TEXT" => opts.format = CopyFormat::Text,
12220 "HEADER" => {
12221 opts.header = match self.peek() {
12222 Token::True => {
12223 self.advance();
12224 true
12225 }
12226 Token::False => {
12227 self.advance();
12228 false
12229 }
12230 Token::Ident(s) if s.eq_ignore_ascii_case("on") => {
12231 self.advance();
12232 true
12233 }
12234 Token::Ident(s) if s.eq_ignore_ascii_case("off") => {
12235 self.advance();
12236 false
12237 }
12238 // Bare HEADER (no boolean) means HEADER true.
12239 _ => true,
12240 };
12241 }
12242 // r1066 (7.38 S5.1) — pgbench 14+ loads with
12243 // `COPY … WITH (FREEZE ON)`. The hint's PG effect is
12244 // vacuum bookkeeping on a freshly created/truncated
12245 // table; SPG's per-statement visibility makes it a
12246 // faithful no-op, and rejecting it aborted `pgbench -i`
12247 // against the drop-in. Accept ON/OFF/bare, change nothing.
12248 "FREEZE" => match self.peek() {
12249 Token::True | Token::False => {
12250 self.advance();
12251 }
12252 Token::Ident(s)
12253 if s.eq_ignore_ascii_case("on") || s.eq_ignore_ascii_case("off") =>
12254 {
12255 self.advance();
12256 }
12257 _ => {}
12258 },
12259 "DELIMITER" | "QUOTE" | "ESCAPE" => {
12260 let s = match self.advance() {
12261 Token::String(s) => s,
12262 other => {
12263 return Err(self.err(alloc::format!(
12264 "COPY {kw} expects a single-character string, got {other:?}"
12265 )));
12266 }
12267 };
12268 // v7.39 (round 247) — PG's wording (0A000), keyword in
12269 // lowercase: "COPY delimiter must be a single one-byte
12270 // character".
12271 let one_byte_err = || {
12272 self.err(alloc::format!(
12273 "COPY {} must be a single one-byte character",
12274 kw.to_ascii_lowercase()
12275 ))
12276 };
12277 let mut chars = s.chars();
12278 let c = chars.next().ok_or_else(one_byte_err)?;
12279 if chars.next().is_some() || c.len_utf8() != 1 {
12280 return Err(one_byte_err());
12281 }
12282 match kw.as_str() {
12283 "DELIMITER" => opts.delimiter = Some(c),
12284 "QUOTE" => opts.quote = Some(c),
12285 _ => opts.escape = Some(c),
12286 }
12287 }
12288 // v7.39 (round 247) — `FORCE_QUOTE (col, …)` / `FORCE_QUOTE *`.
12289 "FORCE_QUOTE" => {
12290 if matches!(self.peek(), Token::Star) {
12291 self.advance();
12292 opts.force_quote = Some(Vec::new());
12293 } else {
12294 if !matches!(self.peek(), Token::LParen) {
12295 return Err(self.err(alloc::format!(
12296 "expected '(' or '*' after FORCE_QUOTE, got {:?}",
12297 self.peek()
12298 )));
12299 }
12300 self.advance();
12301 let mut cols = Vec::new();
12302 loop {
12303 cols.push(self.expect_ident_like()?);
12304 match self.peek() {
12305 Token::Comma => {
12306 self.advance();
12307 }
12308 Token::RParen => {
12309 self.advance();
12310 break;
12311 }
12312 other => {
12313 return Err(self.err(alloc::format!(
12314 "expected ',' or ')' in FORCE_QUOTE list, got {other:?}"
12315 )));
12316 }
12317 }
12318 }
12319 opts.force_quote = Some(cols);
12320 }
12321 }
12322 "NULL" => {
12323 opts.null_str = Some(match self.advance() {
12324 Token::String(s) => s,
12325 other => {
12326 return Err(self.err(alloc::format!(
12327 "COPY NULL expects a quoted string, got {other:?}"
12328 )));
12329 }
12330 });
12331 }
12332 // v7.39 (round 265) — the two CSV FROM-side column lists. Same
12333 // grammar as FORCE_QUOTE; PG accepts `*` for FORCE_NOT_NULL /
12334 // FORCE_NULL too.
12335 "FORCE_NOT_NULL" | "FORCE_NULL" => {
12336 let cols = self.parse_copy_column_list(&kw)?;
12337 if kw == "FORCE_NOT_NULL" {
12338 opts.force_not_null = Some(cols);
12339 } else {
12340 opts.force_null = Some(cols);
12341 }
12342 }
12343 other => {
12344 // PG's wording, lowercased option name.
12345 return Err(self.err(alloc::format!(
12346 "option \"{}\" not recognized",
12347 other.to_ascii_lowercase()
12348 )));
12349 }
12350 }
12351 Ok(())
12352 }
12353
12354 /// v7.39 (round 265) — `( col, … )` or `*` after a COPY column-list
12355 /// option (FORCE_QUOTE / FORCE_NOT_NULL / FORCE_NULL). An empty vec
12356 /// is the `*` spelling.
12357 fn parse_copy_column_list(&mut self, kw: &str) -> Result<Vec<String>, ParseError> {
12358 if matches!(self.peek(), Token::Star) {
12359 self.advance();
12360 return Ok(Vec::new());
12361 }
12362 if !matches!(self.peek(), Token::LParen) {
12363 return Err(self.err(alloc::format!(
12364 "expected '(' or '*' after {kw}, got {:?}",
12365 self.peek()
12366 )));
12367 }
12368 self.advance();
12369 let mut cols = Vec::new();
12370 loop {
12371 cols.push(self.expect_ident_like()?);
12372 match self.peek() {
12373 Token::Comma => {
12374 self.advance();
12375 }
12376 Token::RParen => {
12377 self.advance();
12378 break;
12379 }
12380 other => {
12381 return Err(self.err(alloc::format!(
12382 "expected ',' or ')' in {kw} list, got {other:?}"
12383 )));
12384 }
12385 }
12386 }
12387 Ok(cols)
12388 }
12389
12390 fn parse_partition_bounds_tail(
12391 &mut self,
12392 ) -> Result<crate::ast::PartitionOfBoundsAst, ParseError> {
12393 use crate::ast::PartitionOfBoundsAst;
12394 match self.peek() {
12395 Token::Default => {
12396 self.advance();
12397 Ok(PartitionOfBoundsAst::Default)
12398 }
12399 Token::For => {
12400 self.advance();
12401 if !matches!(self.peek(), Token::Values) {
12402 return Err(
12403 self.err(format!("expected VALUES after FOR, got {:?}", self.peek()))
12404 );
12405 }
12406 self.advance();
12407 let want_with = matches!(
12408 self.peek(),
12409 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("with")
12410 );
12411 if want_with {
12412 self.advance();
12413 if !matches!(self.peek(), Token::LParen) {
12414 return Err(self.err(format!(
12415 "expected '(' after FOR VALUES WITH, got {:?}",
12416 self.peek()
12417 )));
12418 }
12419 self.advance();
12420 let (mut modulus, mut remainder): (Option<u32>, Option<u32>) = (None, None);
12421 loop {
12422 let key = self.expect_ident_like()?;
12423 let n = match self.peek().clone() {
12424 Token::Integer(v) if u32::try_from(v).is_ok() => {
12425 self.advance();
12426 v as u32
12427 }
12428 other => {
12429 return Err(self.err(format!(
12430 "FOR VALUES WITH: expected unsigned integer literal, got {other:?}"
12431 )));
12432 }
12433 };
12434 match key.to_ascii_uppercase().as_str() {
12435 "MODULUS" => modulus = Some(n),
12436 "REMAINDER" => remainder = Some(n),
12437 other => {
12438 return Err(self.err(format!(
12439 "FOR VALUES WITH: unknown key {other:?}; \
12440 expected MODULUS or REMAINDER"
12441 )));
12442 }
12443 }
12444 match self.peek() {
12445 Token::Comma => {
12446 self.advance();
12447 }
12448 Token::RParen => {
12449 self.advance();
12450 break;
12451 }
12452 other => {
12453 return Err(self.err(format!(
12454 "expected ',' or ')' in FOR VALUES WITH list, got {other:?}"
12455 )));
12456 }
12457 }
12458 }
12459 let modulus = modulus
12460 .ok_or_else(|| self.err("FOR VALUES WITH: missing MODULUS".to_string()))?;
12461 let remainder = remainder.ok_or_else(|| {
12462 self.err("FOR VALUES WITH: missing REMAINDER".to_string())
12463 })?;
12464 if modulus == 0 {
12465 return Err(self.err("FOR VALUES WITH: MODULUS must be > 0".to_string()));
12466 }
12467 if remainder >= modulus {
12468 return Err(self.err(format!(
12469 "FOR VALUES WITH: REMAINDER ({remainder}) must be < MODULUS ({modulus})"
12470 )));
12471 }
12472 return Ok(PartitionOfBoundsAst::Hash { modulus, remainder });
12473 }
12474 match self.peek() {
12475 Token::From => {
12476 self.advance();
12477 let lower = Box::new(self.parse_partition_bound_expr()?);
12478 if !matches!(self.peek(), Token::To) {
12479 return Err(self.err(format!(
12480 "expected TO after FROM (...), got {:?}",
12481 self.peek()
12482 )));
12483 }
12484 self.advance();
12485 let upper = Box::new(self.parse_partition_bound_expr()?);
12486 Ok(PartitionOfBoundsAst::Range { lower, upper })
12487 }
12488 Token::In => {
12489 self.advance();
12490 if !matches!(self.peek(), Token::LParen) {
12491 return Err(self.err(format!(
12492 "expected '(' after FOR VALUES IN, got {:?}",
12493 self.peek()
12494 )));
12495 }
12496 self.advance();
12497 let mut values = Vec::new();
12498 loop {
12499 values.push(self.parse_expr(0)?);
12500 match self.peek() {
12501 Token::Comma => {
12502 self.advance();
12503 }
12504 Token::RParen => {
12505 self.advance();
12506 break;
12507 }
12508 other => {
12509 return Err(self.err(format!(
12510 "expected ',' or ')' in FOR VALUES IN list, got {other:?}"
12511 )));
12512 }
12513 }
12514 }
12515 if values.is_empty() {
12516 return Err(
12517 self.err("FOR VALUES IN requires at least one literal".to_string())
12518 );
12519 }
12520 Ok(PartitionOfBoundsAst::List { values })
12521 }
12522 other => Err(self.err(format!(
12523 "expected FROM / IN / WITH after FOR VALUES, got {other:?}"
12524 ))),
12525 }
12526 }
12527 other => Err(self.err(format!(
12528 "expected DEFAULT or FOR VALUES after ATTACH PARTITION child, got {other:?}"
12529 ))),
12530 }
12531 }
12532
12533 /// v7.16.2 — peek for `information_schema.<tbl>` /
12534 /// `pg_catalog.<tbl>` triples and, if matched, consume all
12535 /// three tokens + return a synthetic table name the engine's
12536 /// SELECT path recognises as a virtual view. Returns `None`
12537 /// when the head doesn't look like a meta-qualified name.
12538 /// Used by `parse_table_ref` to bypass the
12539 /// `expect_ident_like` schema-strip for these specific PG
12540 /// meta schemas (mailrs round-10 A.3).
12541 fn try_peek_meta_qualified(&mut self) -> Option<(String, String)> {
12542 // Extract the schema name. Must be a plain ident token.
12543 let schema = match self.tokens.get(self.pos) {
12544 Some(Token::Ident(s) | Token::QuotedIdent(s)) => s.clone(),
12545 _ => return None,
12546 };
12547 // Dot.
12548 if !matches!(self.tokens.get(self.pos + 1), Some(Token::Dot)) {
12549 return None;
12550 }
12551 // The table-side ident may lex as a reserved keyword
12552 // (e.g. `Token::Tables`). Tolerate the common ones via a
12553 // helper that reads the trailing token's underlying name.
12554 let tbl = match self.tokens.get(self.pos + 2)? {
12555 Token::Ident(t) | Token::QuotedIdent(t) => t.clone(),
12556 Token::Tables => "tables".to_string(),
12557 // Other PG meta table names that may collide with
12558 // reserved keywords land here as needed.
12559 _ => return None,
12560 };
12561 // Strip the `pg_` prefix from `pg_catalog.pg_class`-style
12562 // names so the synthetic name doesn't double-prefix
12563 // (`__spg_pg_class`, not `__spg_pg_pg_class`).
12564 let (prefix, normalised) = if schema.eq_ignore_ascii_case("information_schema") {
12565 ("__spg_info_", tbl.to_ascii_lowercase())
12566 } else if schema.eq_ignore_ascii_case("pg_catalog") {
12567 // v7.39 (round 541) — only the catalogs SPG actually
12568 // synthesises are rewritten, which is what the BARE path
12569 // has always checked. Anything else keeps its own name and
12570 // takes the ordinary route: `pg_stat_activity` and friends
12571 // resolve through meta_view_result, and a name that is no
12572 // catalog at all gets PG's "relation does not exist"
12573 // instead of a message about a view SPG cannot materialise.
12574 let lowered = tbl.to_ascii_lowercase();
12575 if !SYNTHESISED_PG_CATALOGS.contains(&lowered.as_str()) {
12576 self.advance(); // schema
12577 self.advance(); // dot
12578 self.advance(); // tbl
12579 return Some((lowered.clone(), lowered));
12580 }
12581 let bare = lowered
12582 .strip_prefix("pg_")
12583 .map(alloc::string::String::from)
12584 .unwrap_or(lowered);
12585 ("__spg_pg_", bare)
12586 } else if schema.eq_ignore_ascii_case("mysql") {
12587 // v7.17.0 Phase 3.P0-65 — MySQL system schema
12588 // (`mysql.user`, `mysql.db`). Same synthetic-name
12589 // shape as pg_catalog.
12590 ("__spg_mysql_", tbl.to_ascii_lowercase())
12591 } else {
12592 return None;
12593 };
12594 self.advance(); // schema
12595 self.advance(); // dot
12596 self.advance(); // tbl
12597 Some((
12598 alloc::format!("{prefix}{normalised}"),
12599 tbl.to_ascii_lowercase(),
12600 ))
12601 }
12602
12603 /// Unqualified PG meta-table names (`FROM pg_extension`, `FROM
12604 /// pg_class`) resolve the same way: PG puts `pg_catalog` at the
12605 /// implicit front of every search_path, so a bare reference to a
12606 /// known catalog table always means the catalog table. Only the
12607 /// names the engine actually synthesises are recognised — any
12608 /// other `pg_*` ident stays a user table (mailrs embed round-12).
12609 fn try_peek_meta_bare(&mut self) -> Option<(String, String)> {
12610 // v7.38 (read01 P3.21) — every catalog view SPG synthesises
12611 // (`__spg_pg_*`) is bare-resolvable, matching PG's implicit
12612 // `pg_catalog` at the front of every search_path. (pg_stat_activity
12613 // / pg_stat_statements / pg_locks / pg_statio_user_tables route
12614 // through the meta_view_result path instead, and already resolve
12615 // bare — they must NOT be listed here or the __spg_ rewrite would
12616 // mis-target them.)
12617 const PG_META_TABLES: &[&str] = SYNTHESISED_PG_CATALOGS;
12618 let name = match self.tokens.get(self.pos) {
12619 Some(Token::Ident(s)) => s.to_ascii_lowercase(),
12620 _ => return None,
12621 };
12622 // A following dot means this ident is a schema qualifier,
12623 // not a table name — let the qualified path handle it.
12624 if matches!(self.tokens.get(self.pos + 1), Some(Token::Dot)) {
12625 return None;
12626 }
12627 if !PG_META_TABLES.contains(&name.as_str()) {
12628 return None;
12629 }
12630 self.advance();
12631 let bare = name.strip_prefix("pg_").unwrap_or(&name);
12632 Some((alloc::format!("__spg_pg_{bare}"), name.clone()))
12633 }
12634
12635 /// Consume a bare ident if its lowercase matches `kw`, else err.
12636 /// v7.39 (read01 round 57) — is the next token this bare keyword-ident?
12637 /// Peeks only; the caller advances.
12638 fn peek_keyword_ident(&self, kw: &str) -> bool {
12639 matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case(kw))
12640 }
12641
12642 fn expect_keyword_ident(&mut self, kw: &str) -> Result<(), ParseError> {
12643 match self.advance() {
12644 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case(kw) => Ok(()),
12645 other => Err(ParseError {
12646 message: format!("expected {kw:?}, got {other:?}"),
12647 token_pos: self.consumed_pos(),
12648 }),
12649 }
12650 }
12651
12652 /// Accept either a quoted identifier (`"foo"`) or a quoted string
12653 /// literal (`'foo'`) — same shape used by CREATE USER for the
12654 /// username slot.
12655 fn expect_ident_or_string(&mut self) -> Result<String, ParseError> {
12656 match self.advance() {
12657 Token::Ident(s) | Token::QuotedIdent(s) | Token::String(s) => Ok(s),
12658 other => Err(ParseError {
12659 message: format!("expected identifier or string, got {other:?}"),
12660 token_pos: self.consumed_pos(),
12661 }),
12662 }
12663 }
12664
12665 fn expect_string_literal(&mut self) -> Result<String, ParseError> {
12666 match self.advance() {
12667 Token::String(s) => Ok(s),
12668 other => Err(ParseError {
12669 message: format!("expected quoted string, got {other:?}"),
12670 token_pos: self.consumed_pos(),
12671 }),
12672 }
12673 }
12674
12675 fn parse_select_stmt(&mut self) -> Result<Statement, ParseError> {
12676 // v7.30.2 (mailrs round-25 ask 2) — derived tables /
12677 // subqueries recurse through here without passing
12678 // parse_expr; share the same nesting budget.
12679 self.enter_nested()?;
12680 let r = self.parse_select_stmt_inner();
12681 self.nest_depth -= 1;
12682 r
12683 }
12684
12685 fn parse_select_stmt_inner(&mut self) -> Result<Statement, ParseError> {
12686 // Caller dispatches on Token::Select; the inner helper handles
12687 // the rest. ORDER BY / LIMIT bind at this top level; UNION peers
12688 // get a fresh bare-select parse and may not have their own ORDER
12689 // BY / LIMIT.
12690 let mut head = self.parse_bare_select()?;
12691 let into = self.pending_select_into.take();
12692 self.parse_setop_chain_into(&mut head)?;
12693 self.parse_select_tail_into(&mut head)?;
12694 // v7.38.19 — `SELECT … INTO t` lowers to the SAME node as
12695 // `CREATE TABLE t AS SELECT …`, which is what a comment in
12696 // `ast.rs` has claimed since v7.38 and what only CTAS actually
12697 // did. The tail (ORDER BY / LIMIT) is parsed first so it belongs
12698 // to the body, as it does in PostgreSQL.
12699 if let Some((name, temporary)) = into {
12700 return Ok(Statement::CreateMaterializedView(
12701 crate::ast::CreateMaterializedViewStatement {
12702 temporary,
12703 name,
12704 if_not_exists: false,
12705 columns: Vec::new(),
12706 body: head,
12707 with_data: true,
12708 as_plain_table: true,
12709 },
12710 ));
12711 }
12712 Ok(Statement::Select(head))
12713 }
12714
12715 /// v7.37.17 (17.6 siblings) — the three SQL set operations
12716 /// share the peer chain: UNION [ALL], EXCEPT [ALL] (a reserved
12717 /// token), and INTERSECT [ALL] (a bare ident — it was never
12718 /// reserved in SPG's lexer). PG precedence: INTERSECT binds
12719 /// tighter than UNION / EXCEPT — the executor folds the chain
12720 /// left-to-right, which is already correct for LEADING
12721 /// intersects; an INTERSECT pair that FOLLOWS a union/except
12722 /// pair nests into that previous peer, so A UNION B INTERSECT C
12723 /// = A ∪ (B ∩ C). Shared by the top level and parenthesized
12724 /// groups.
12725 fn parse_setop_chain_into(&mut self, head: &mut SelectStatement) -> Result<(), ParseError> {
12726 // A parenthesized group arrives with its own (already
12727 // regrouped) unions on `head`; only the pairs THIS chain
12728 // appends participate in the precedence regroup below —
12729 // nesting an outer INTERSECT into a group-internal peer
12730 // would dissolve the explicit grouping.
12731 let boundary = head.unions.len();
12732 loop {
12733 let base = match self.peek() {
12734 Token::Union => UnionKind::Distinct,
12735 Token::Except => UnionKind::Except,
12736 Token::Ident(s) if s.eq_ignore_ascii_case("intersect") => UnionKind::Intersect,
12737 _ => break,
12738 };
12739 self.advance();
12740 let kind = if matches!(self.peek(), Token::All) {
12741 self.advance();
12742 match base {
12743 UnionKind::Distinct => UnionKind::All,
12744 UnionKind::Except => UnionKind::ExceptAll,
12745 _ => UnionKind::IntersectAll,
12746 }
12747 } else {
12748 base
12749 };
12750 let peer = self.parse_bare_select()?;
12751 head.unions.push((kind, peer));
12752 }
12753 let mut pairs = core::mem::take(&mut head.unions);
12754 let tail = pairs.split_off(boundary);
12755 let mut regrouped: Vec<(UnionKind, SelectStatement)> = pairs;
12756 for (kind, peer) in tail {
12757 let is_intersect = matches!(kind, UnionKind::Intersect | UnionKind::IntersectAll);
12758 // An intersect nests into the previous element of THIS
12759 // chain only; with no new previous element it stays at
12760 // the outer level (the left fold applies it to the
12761 // whole head, group included).
12762 match (
12763 is_intersect,
12764 regrouped.len() > boundary,
12765 regrouped.last_mut(),
12766 ) {
12767 (true, true, Some((_, prev))) => prev.unions.push((kind, peer)),
12768 _ => regrouped.push((kind, peer)),
12769 }
12770 }
12771 head.unions = regrouped;
12772 Ok(())
12773 }
12774
12775 /// v7.37.17 (17.6 siblings) — the shared SELECT tail: ORDER BY /
12776 /// LIMIT / OFFSET / FETCH FIRST / FOR-lock clauses. Extracted so
12777 /// the top-level bare VALUES statement reuses it verbatim.
12778 /// v6.4.0 — parse an optional `ORDER BY <expr> [ASC|DESC] [NULLS …], …`
12779 /// clause into its key list (empty when no `ORDER BY` follows). Extracted
12780 /// (v7.39 round 135) so the grouping-set path can parse ORDER BY early,
12781 /// where the grouping-set universe is still in scope.
12782 fn parse_order_by_keys(&mut self) -> Result<Vec<OrderBy>, ParseError> {
12783 if !matches!(self.peek(), Token::Order) {
12784 return Ok(Vec::new());
12785 }
12786 self.advance();
12787 if !self.peek_is_by() {
12788 return Err(self.err(format!("expected BY after ORDER, got {:?}", self.peek())));
12789 }
12790 self.advance();
12791 let mut keys = Vec::new();
12792 loop {
12793 // v7.39 (round 691) — save/restore, the discipline this parser
12794 // already uses around `pending_sample_preds`, so a subquery inside
12795 // a key neither inherits nor leaks the channel.
12796 let saved_flag = core::mem::replace(&mut self.in_order_by_key, true);
12797 let saved_coll = self.order_key_collation.take();
12798 let parsed = self.parse_expr(0);
12799 self.in_order_by_key = saved_flag;
12800 let collation = core::mem::replace(&mut self.order_key_collation, saved_coll);
12801 let expr = parsed?;
12802 let desc = if matches!(self.peek(), Token::Desc) {
12803 self.advance();
12804 true
12805 } else if matches!(self.peek(), Token::Asc) {
12806 self.advance();
12807 false
12808 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
12809 // `ORDER BY x USING <op>` — PG's operator-class spelling. SPG has
12810 // one ordering per type, so the btree comparison operators map
12811 // onto it: < / <= are ASC, > / >= are DESC. Any other operator
12812 // would need a custom operator class — honest error.
12813 self.advance();
12814 match self.advance() {
12815 Token::Lt | Token::LtEq => false,
12816 Token::Gt | Token::GtEq => true,
12817 other => {
12818 return Err(self.err(alloc::format!(
12819 "ORDER BY USING supports the btree comparison \
12820 operators (< <= > >=); got {other:?}"
12821 )));
12822 }
12823 }
12824 } else {
12825 false
12826 };
12827 // v7.24 (round-16 A) — explicit NULLS FIRST/LAST.
12828 let nulls_first = self.parse_optional_nulls_placement()?;
12829 keys.push(OrderBy {
12830 expr,
12831 desc,
12832 nulls_first,
12833 collation,
12834 });
12835 if matches!(self.peek(), Token::Comma) {
12836 self.advance();
12837 } else {
12838 break;
12839 }
12840 }
12841 Ok(keys)
12842 }
12843
12844 fn parse_select_tail_into(&mut self, head: &mut SelectStatement) -> Result<(), ParseError> {
12845 // v7.39 (round 135) — a grouping-set query may have already parsed +
12846 // rewritten its ORDER BY (to reference synthetic grouping columns); if
12847 // no ORDER BY token is present, keep that pre-set order_by rather than
12848 // clobbering it with an empty list.
12849 let parsed_keys = self.parse_order_by_keys()?;
12850 head.order_by = if parsed_keys.is_empty() {
12851 core::mem::take(&mut head.order_by)
12852 } else {
12853 parsed_keys
12854 };
12855 // v7.39 (round 314, V39) — the row-count clauses come in EITHER
12856 // order. PG's grammar takes a limit clause and an offset clause
12857 // as an unordered pair, so `OFFSET 2 LIMIT 3` means exactly what
12858 // `LIMIT 3 OFFSET 2` does (measured: same rows). This used to
12859 // parse them in a fixed LIMIT-then-OFFSET sequence, so the other
12860 // spelling died on `expected end of input, got Limit`.
12861 //
12862 // Each may appear at most once, and LIMIT and FETCH FIRST are
12863 // two spellings of the same clause — PG rejects `LIMIT 1 LIMIT 2`,
12864 // `OFFSET 1 OFFSET 2` and `LIMIT 2 FETCH FIRST 3 ROWS ONLY` alike.
12865 // A second one is left unconsumed here, which the caller reports
12866 // as trailing input rather than silently taking the last.
12867 let mut saw_limit = false;
12868 let mut saw_offset = false;
12869 loop {
12870 if !saw_limit && matches!(self.peek(), Token::Limit) {
12871 self.advance();
12872 // v7.17.0 Phase 5.1 — `LIMIT NULL` / `LIMIT ALL` are
12873 // PG synonyms for "no limit". Treat both as None
12874 // (no head.limit set) so the engine's existing
12875 // unlimited-result path takes over. Reject was the
12876 // pre-5.1 behaviour and broke pg_dump-flavoured
12877 // tooling that occasionally emits LIMIT NULL.
12878 if self.consume_limit_unbounded_sentinel() {
12879 head.limit = None;
12880 } else {
12881 let first = self.parse_limit_expr("LIMIT")?;
12882 // MySQL `LIMIT offset, count` — the first number is
12883 // the offset when a comma follows.
12884 if matches!(self.peek(), Token::Comma) {
12885 self.advance();
12886 let count = self.parse_limit_expr("LIMIT")?;
12887 head.offset = Some(first);
12888 saw_offset = true;
12889 head.limit = Some(count);
12890 } else {
12891 head.limit = Some(first);
12892 }
12893 }
12894 saw_limit = true;
12895 continue;
12896 }
12897 if !saw_offset && matches!(self.peek(), Token::Offset) {
12898 self.advance();
12899 // PG also accepts an optional `ROW` / `ROWS` trailer
12900 // after the offset value (`OFFSET 10 ROWS`). The
12901 // FETCH-FIRST branch below relies on the same.
12902 let off = self.parse_limit_expr("OFFSET")?;
12903 self.consume_optional_rows_keyword();
12904 head.offset = Some(off);
12905 saw_offset = true;
12906 continue;
12907 }
12908 // v7.17.0 Phase 5.1 — `FETCH FIRST <int|$N> ROWS ONLY` is
12909 // the SQL-standard alias for LIMIT. PG accepts both
12910 // spellings interchangeably; pg_dump emits FETCH FIRST in
12911 // newer versions. We map it onto `head.limit` so the
12912 // engine path is unified.
12913 if !saw_limit
12914 && matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12915 if s.eq_ignore_ascii_case("fetch"))
12916 {
12917 self.advance(); // FETCH
12918 // `FIRST` or `NEXT` (both legal per SQL standard).
12919 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12920 if s.eq_ignore_ascii_case("first") || s.eq_ignore_ascii_case("next"))
12921 {
12922 self.advance();
12923 }
12924 // Count (optional in the bare `FETCH FIRST ROW ONLY` —
12925 // implicit 1 — but we always consume one if present).
12926 let count = if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12927 if s.eq_ignore_ascii_case("row") || s.eq_ignore_ascii_case("rows"))
12928 {
12929 // Bare `FETCH FIRST ROW ONLY` = LIMIT 1.
12930 crate::ast::LimitExpr::Literal(1)
12931 } else {
12932 self.parse_limit_expr("FETCH FIRST")?
12933 };
12934 // Eat `ROW` / `ROWS` if not already consumed above.
12935 self.consume_optional_rows_keyword();
12936 // Optional `ONLY` (the spec form) — or the SQL:2008
12937 // `WITH TIES` form. v7.17.0 Phase 3.P0-49: the executor
12938 // now honours WITH TIES by extending past the LIMIT
12939 // truncation point through every row that shares the
12940 // last-kept row's ORDER BY key.
12941 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12942 if s.eq_ignore_ascii_case("only"))
12943 {
12944 self.advance();
12945 } else if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12946 if s.eq_ignore_ascii_case("with"))
12947 {
12948 self.advance(); // WITH
12949 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12950 if s.eq_ignore_ascii_case("ties"))
12951 {
12952 self.advance();
12953 head.limit_with_ties = true;
12954 }
12955 }
12956 head.limit = Some(count);
12957 saw_limit = true;
12958 continue;
12959 }
12960 break;
12961 }
12962 // v7.17.0 Phase 3.4 — trailing row-lock clauses:
12963 // FOR { UPDATE | NO KEY UPDATE | SHARE | KEY SHARE }
12964 // [ OF table_name [, …] ]
12965 // [ NOWAIT | SKIP LOCKED ]
12966 // Multiple FOR clauses may stack (PG: `FOR UPDATE OF t1
12967 // FOR SHARE OF t2`). SPG is a single-writer engine — every
12968 // SELECT already returns a consistent snapshot — so these
12969 // are accept-and-discard: the parser absorbs them so
12970 // mailrs / Rails / Django code paths that emit `SELECT
12971 // … FOR UPDATE` for advisory pessimistic locking load
12972 // without a parser error. The on-disk locking model is
12973 // unchanged; callers that rely on FOR UPDATE for read-
12974 // through-write ordering still get the right answer
12975 // because SPG serialises writes anyway.
12976 head.locking = self
12977 .consume_optional_for_lock_clauses()
12978 .map(alloc::boxed::Box::new);
12979 Ok(())
12980 }
12981
12982 /// v7.17.0 Phase 3.4 — eat zero or more `FOR { UPDATE | NO KEY
12983 /// UPDATE | SHARE | KEY SHARE } [ OF tbl[, …] ] [ NOWAIT | SKIP
12984 /// LOCKED ]` trailers. Each clause is fully accepted and
12985 /// discarded — SPG's single-writer model already satisfies the
12986 /// callers' implicit ordering requirement. Stops at the first
12987 /// token that isn't `FOR`.
12988 fn consume_optional_for_lock_clauses(&mut self) -> Option<crate::ast::LockingClause> {
12989 // v7.39 (round 293, E3 Phase 1) — the clause is REPORTED now,
12990 // not discarded. PG keeps the strongest of several clauses; the
12991 // policy of the last one wins, which is what this loop records.
12992 let mut seen: Option<crate::ast::LockingClause> = None;
12993 while matches!(self.peek(), Token::For) {
12994 // v7.37.14 (A2.5-stub) — record that this query asked
12995 // for a row lock the parser is about to silently
12996 // discard. Operators surface the count via
12997 // `spg_sql::silent_for_update_count()` so they can
12998 // gauge how much of the workload depends on advisory
12999 // FOR UPDATE / FOR SHARE / FOR KEY SHARE semantics
13000 // before v7.37.15's per-row tuple locking lands.
13001 crate::record_silent_for_update_clause();
13002 self.advance(); // FOR
13003 // `NO KEY` prefix (PG) — `NO` is reserved-keyword-shaped
13004 // (`Token::Not` isn't it; PG `NO` lexes as Token::Ident).
13005 let mut no_key = false;
13006 let mut key = false;
13007 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13008 if s.eq_ignore_ascii_case("no"))
13009 {
13010 self.advance(); // NO
13011 no_key = true;
13012 // The next ident should be KEY but be generous;
13013 // anything followed by UPDATE/SHARE is accepted.
13014 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13015 if s.eq_ignore_ascii_case("key"))
13016 {
13017 self.advance(); // KEY
13018 }
13019 }
13020 // `KEY` prefix (PG `FOR KEY SHARE`).
13021 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13022 if s.eq_ignore_ascii_case("key"))
13023 {
13024 self.advance(); // KEY
13025 key = true;
13026 }
13027 // Lock-strength keyword: UPDATE / SHARE. Required, but
13028 // we're lenient — an unexpected token here just bails
13029 // (we already consumed FOR; caller's downstream
13030 // dispatch will error if anything actually depends on
13031 // the trailing tokens).
13032 let is_update = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13033 if s.eq_ignore_ascii_case("update"));
13034 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13035 if s.eq_ignore_ascii_case("update") || s.eq_ignore_ascii_case("share"))
13036 {
13037 self.advance();
13038 use crate::ast::LockStrength as LS;
13039 let strength = match (is_update, no_key, key) {
13040 (true, true, _) => LS::NoKeyUpdate,
13041 (true, _, _) => LS::Update,
13042 (false, _, true) => LS::KeyShare,
13043 (false, _, _) => LS::Share,
13044 };
13045 seen = Some(crate::ast::LockingClause {
13046 strength,
13047 of_tables: alloc::vec::Vec::new(),
13048 policy: crate::ast::LockWait::Wait,
13049 });
13050 } else {
13051 // FOR by itself (or `FOR KEY` with nothing after) —
13052 // give up on the lock-clause path. We've already
13053 // advanced past FOR; further attempts to parse
13054 // here would clobber state.
13055 return seen;
13056 }
13057 // Optional `OF tbl[, tbl …]`. mailrs emits this when
13058 // joining and locking only a subset of tables.
13059 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13060 if s.eq_ignore_ascii_case("of"))
13061 {
13062 self.advance(); // OF
13063 #[allow(clippy::while_let_loop)]
13064 loop {
13065 match self.peek() {
13066 Token::Ident(_) | Token::QuotedIdent(_) => {
13067 // v7.39 (round 294) — the name is CAPTURED now: PG
13068 // validates it against the FROM clause, and an
13069 // uncaptured list silently means "lock everything".
13070 let mut nm = match self.advance() {
13071 Token::Ident(n) | Token::QuotedIdent(n) => n,
13072 _ => alloc::string::String::new(),
13073 };
13074 // Optional schema-qualified `schema.table`.
13075 if matches!(self.peek(), Token::Dot) {
13076 self.advance();
13077 if let Token::Ident(n) | Token::QuotedIdent(n) = self.peek().clone()
13078 {
13079 self.advance();
13080 nm = n;
13081 }
13082 }
13083 if let Some(c) = seen.as_mut() {
13084 c.of_tables.push(nm);
13085 }
13086 }
13087 _ => break,
13088 }
13089 if matches!(self.peek(), Token::Comma) {
13090 self.advance();
13091 } else {
13092 break;
13093 }
13094 }
13095 }
13096 // Optional `NOWAIT` | `SKIP LOCKED`.
13097 match self.peek().clone() {
13098 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("nowait") => {
13099 self.advance();
13100 if let Some(c) = seen.as_mut() {
13101 c.policy = crate::ast::LockWait::NoWait;
13102 }
13103 }
13104 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("skip") => {
13105 self.advance(); // SKIP
13106 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13107 if s.eq_ignore_ascii_case("locked"))
13108 {
13109 self.advance(); // LOCKED
13110 if let Some(c) = seen.as_mut() {
13111 c.policy = crate::ast::LockWait::SkipLocked;
13112 }
13113 }
13114 }
13115 _ => {}
13116 }
13117 // Loop: PG allows multiple FOR clauses chained.
13118 }
13119 seen
13120 }
13121
13122 /// v7.9.24 — accept `LIMIT <int>` or `LIMIT $N`. mailrs H2.
13123 /// Bind value gets resolved during prepared-statement Execute;
13124 /// the Pratt expression parser would over-accept here (e.g.
13125 /// `LIMIT 5 + 5`), so we narrowly accept only the two PG forms.
13126 /// v7.17.0 Phase 5.1 — consume the `LIMIT NULL` / `LIMIT ALL`
13127 /// sentinel tokens (PG synonyms for "no limit"). Returns true
13128 /// when one was consumed; caller skips the regular
13129 /// limit-value parse and leaves `head.limit` at None.
13130 fn consume_limit_unbounded_sentinel(&mut self) -> bool {
13131 if matches!(self.peek(), Token::Null) {
13132 self.advance();
13133 return true;
13134 }
13135 if matches!(self.peek(), Token::All) {
13136 self.advance();
13137 return true;
13138 }
13139 false
13140 }
13141
13142 /// v7.17.0 Phase 5.1 — eat an optional trailing `ROW` / `ROWS`
13143 /// keyword after a LIMIT / OFFSET / FETCH FIRST value, the
13144 /// SQL-standard shape. No-op when missing.
13145 fn consume_optional_rows_keyword(&mut self) {
13146 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13147 if s.eq_ignore_ascii_case("row") || s.eq_ignore_ascii_case("rows"))
13148 {
13149 self.advance();
13150 }
13151 }
13152
13153 /// v7.39 (round 284) — `LIMIT` / `OFFSET` over a general expression.
13154 ///
13155 /// PG's row-count clause takes an `a_expr`, so `LIMIT 1+1` and
13156 /// `OFFSET 2+3` are legal; only `FETCH FIRST` is restricted to a
13157 /// constant, which is why that spelling keeps the token path below.
13158 ///
13159 /// Constants are folded here rather than carried into the tree: the
13160 /// 15+ execution paths that read the row count go through
13161 /// `limit_literal()`, which answers `Option<u32>` — and `None` there
13162 /// means "no limit". A clause the engine could not resolve would
13163 /// therefore return the WHOLE table instead of failing. Folding at
13164 /// parse time keeps that impossible; a non-constant clause is still
13165 /// a clean error (recorded residual — closing it wants a resolution
13166 /// pre-pass on the simple-query path, where `substitute_placeholders`
13167 /// does not run).
13168 fn parse_limit_expr(&mut self, label: &str) -> Result<crate::ast::LimitExpr, ParseError> {
13169 // PG restricts FETCH FIRST to a constant or a PARENTHESISED
13170 // expression: `FETCH FIRST 1+1 ROWS ONLY` is a syntax error, but
13171 // `FETCH FIRST (1+1) ROWS ONLY` and `FETCH FIRST (SELECT 3) ROWS
13172 // ONLY` both work (its grammar takes a c_expr). Both measured
13173 // against PG 18.4 in round 305.
13174 if label == "FETCH FIRST" && !matches!(self.peek(), Token::LParen) {
13175 return self.parse_limit_constant(label);
13176 }
13177 // One pass, no rewind: `advance()` takes each token by
13178 // `mem::replace`, so a consumed token reads back as Eof and this
13179 // parser cannot backtrack. Everything — bare literal included —
13180 // is therefore folded from the parsed expression rather than
13181 // re-read from the token stream.
13182 let start = self.pos;
13183 let e = self.parse_expr(0)?;
13184 if let crate::ast::Expr::Placeholder(n) = e {
13185 return Ok(crate::ast::LimitExpr::Placeholder(n));
13186 }
13187 let neg_label = if label == "OFFSET" { "OFFSET" } else { "LIMIT" };
13188 match fold_limit_constant(&e) {
13189 Some(Ok(v)) if v < 0 => Err(ParseError {
13190 message: alloc::format!("{neg_label} must not be negative"),
13191 token_pos: start,
13192 }),
13193 Some(Ok(v)) => u32::try_from(v)
13194 .map(crate::ast::LimitExpr::Literal)
13195 .map_err(|_| ParseError {
13196 message: alloc::format!("{label} value too large: {v}"),
13197 token_pos: start,
13198 }),
13199 Some(Err(message)) => Err(ParseError {
13200 message: message.replace("{L}", neg_label),
13201 token_pos: start,
13202 }),
13203 // v7.39 (round 305, V23) — not foldable at parse time
13204 // (`LIMIT (SELECT 4)`, `LIMIT greatest(2,3)`). Carry the
13205 // expression; the engine evaluates it once before dispatch.
13206 None => Ok(crate::ast::LimitExpr::Expr(alloc::boxed::Box::new(e))),
13207 }
13208 }
13209
13210 fn parse_limit_constant(&mut self, label: &str) -> Result<crate::ast::LimitExpr, ParseError> {
13211 // v7.39 (round 239) — PG's row-count clause takes a bigint with its
13212 // coercion rules, not just an integer token: a NUMERIC rounds half
13213 // away from zero (`LIMIT 2.5` keeps 3 rows), a negative count is
13214 // refused with PG's wording ("LIMIT must not be negative", 2201W /
13215 // 2201X — FETCH FIRST shares LIMIT's), and a string coerces by its
13216 // content, failing as an input-syntax error on the value. General
13217 // expressions (`LIMIT 1+1`) stay unsupported — a recorded residual;
13218 // they need an Expr-carrying LimitExpr variant.
13219 let neg_label = if label == "OFFSET" { "OFFSET" } else { "LIMIT" };
13220 let err_at = |message: alloc::string::String, pos: usize| ParseError {
13221 message,
13222 token_pos: pos,
13223 };
13224 match self.advance() {
13225 Token::Integer(n) if n >= 0 => u32::try_from(n)
13226 .map(crate::ast::LimitExpr::Literal)
13227 .map_err(|_| ParseError {
13228 message: alloc::format!("{label} value too large: {n}"),
13229 token_pos: self.consumed_pos(),
13230 }),
13231 Token::Integer(_) => Err(err_at(
13232 alloc::format!("{neg_label} must not be negative"),
13233 self.pos.saturating_sub(1),
13234 )),
13235 Token::Numeric(t) => {
13236 let pos = self.pos.saturating_sub(1);
13237 let v: f64 = t.parse().map_err(|_| {
13238 err_at(
13239 alloc::format!("invalid input syntax for type bigint: \"{t}\""),
13240 pos,
13241 )
13242 })?;
13243 if v < 0.0 {
13244 return Err(err_at(
13245 alloc::format!("{neg_label} must not be negative"),
13246 pos,
13247 ));
13248 }
13249 // Round half away from zero — PG's numeric→bigint cast.
13250 // (no_std: no f64::round; v is non-negative, so truncating
13251 // v + 0.5 is the same thing.)
13252 #[allow(clippy::cast_possible_truncation, clippy::cast_sign_loss)]
13253 let rounded = (v + 0.5) as u64;
13254 u32::try_from(rounded)
13255 .map(crate::ast::LimitExpr::Literal)
13256 .map_err(|_| err_at(alloc::format!("{label} value too large: {t}"), pos))
13257 }
13258 Token::Minus => {
13259 let pos = self.pos.saturating_sub(1);
13260 match self.peek() {
13261 Token::Integer(_) | Token::Numeric(_) => {
13262 self.advance();
13263 Err(err_at(
13264 alloc::format!("{neg_label} must not be negative"),
13265 pos,
13266 ))
13267 }
13268 other => Err(err_at(
13269 alloc::format!(
13270 "expected non-negative integer or $N placeholder after {label}, got {other:?}"
13271 ),
13272 pos,
13273 )),
13274 }
13275 }
13276 Token::String(t) => {
13277 let pos = self.pos.saturating_sub(1);
13278 match t.trim().parse::<i64>() {
13279 Ok(n) if n < 0 => Err(err_at(
13280 alloc::format!("{neg_label} must not be negative"),
13281 pos,
13282 )),
13283 Ok(n) => u32::try_from(n)
13284 .map(crate::ast::LimitExpr::Literal)
13285 .map_err(|_| err_at(alloc::format!("{label} value too large: {t}"), pos)),
13286 Err(_) => Err(err_at(
13287 alloc::format!("invalid input syntax for type bigint: \"{t}\""),
13288 pos,
13289 )),
13290 }
13291 }
13292 Token::Placeholder(n) => Ok(crate::ast::LimitExpr::Placeholder(n)),
13293 other => Err(ParseError {
13294 message: alloc::format!(
13295 "expected non-negative integer or $N placeholder after {label}, got {other:?}"
13296 ),
13297 token_pos: self.consumed_pos(),
13298 }),
13299 }
13300 }
13301
13302 /// Parse one SELECT block without ORDER BY / LIMIT / UNION chaining —
13303 /// just `[DISTINCT] items [FROM] [WHERE] [GROUP BY]`. Returned with
13304 /// `unions` empty and `order_by` / `limit` `None`; the top-level
13305 /// `parse_select_stmt` is responsible for filling those in.
13306 /// v7.37.17 (17.6 siblings) — rewrite every `grouping(keys…)`
13307 /// call in the expression tree to the per-set integer bitmask
13308 /// (PG semantics: one bit per argument, MSB first; 1 = the key
13309 /// is dropped in this grouping set). Runs during the ROLLUP /
13310 /// CUBE / GROUPING SETS expansion, where the set is known.
13311 /// v7.39 (round 135) — collect the distinct `grouping(...)` calls appearing
13312 /// anywhere in `expr` (an ORDER BY key), without recursing into their args.
13313 fn collect_grouping_calls(expr: &Expr, out: &mut Vec<Expr>) {
13314 if let Expr::FunctionCall { name, .. } = expr
13315 && name.eq_ignore_ascii_case("grouping")
13316 {
13317 if !out.iter().any(|e| e == expr) {
13318 out.push(expr.clone());
13319 }
13320 return;
13321 }
13322 match expr {
13323 Expr::Binary { lhs, rhs, .. } => {
13324 Self::collect_grouping_calls(lhs, out);
13325 Self::collect_grouping_calls(rhs, out);
13326 }
13327 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } | Expr::IsNull { expr, .. } => {
13328 Self::collect_grouping_calls(expr, out)
13329 }
13330 Expr::FunctionCall { args, .. } => {
13331 for a in args {
13332 Self::collect_grouping_calls(a, out);
13333 }
13334 }
13335 Expr::Case {
13336 operand,
13337 branches,
13338 else_branch,
13339 } => {
13340 if let Some(o) = operand {
13341 Self::collect_grouping_calls(o, out);
13342 }
13343 for (c, v) in branches {
13344 Self::collect_grouping_calls(c, out);
13345 Self::collect_grouping_calls(v, out);
13346 }
13347 if let Some(x) = else_branch {
13348 Self::collect_grouping_calls(x, out);
13349 }
13350 }
13351 _ => {}
13352 }
13353 }
13354
13355 /// v7.39 (round 135) — replace each `grouping(...)` call in `expr` equal to
13356 /// `grp_exprs[k]` with a reference to the synthetic ordering column
13357 /// `__grp_ord_k` (injected per grouping-set branch).
13358 fn rewrite_grouping_to_col(expr: &mut Expr, grp_exprs: &[Expr]) {
13359 if let Expr::FunctionCall { name, .. } = expr
13360 && name.eq_ignore_ascii_case("grouping")
13361 {
13362 if let Some(k) = grp_exprs.iter().position(|e| e == expr) {
13363 *expr = Expr::Column(crate::ast::ColumnName {
13364 qualifier: None,
13365 name: alloc::format!("__grp_ord_{k}"),
13366 });
13367 }
13368 return;
13369 }
13370 match expr {
13371 Expr::Binary { lhs, rhs, .. } => {
13372 Self::rewrite_grouping_to_col(lhs, grp_exprs);
13373 Self::rewrite_grouping_to_col(rhs, grp_exprs);
13374 }
13375 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } | Expr::IsNull { expr, .. } => {
13376 Self::rewrite_grouping_to_col(expr, grp_exprs)
13377 }
13378 Expr::FunctionCall { args, .. } => {
13379 for a in args {
13380 Self::rewrite_grouping_to_col(a, grp_exprs);
13381 }
13382 }
13383 Expr::Case {
13384 operand,
13385 branches,
13386 else_branch,
13387 } => {
13388 if let Some(o) = operand {
13389 Self::rewrite_grouping_to_col(o, grp_exprs);
13390 }
13391 for (c, v) in branches {
13392 Self::rewrite_grouping_to_col(c, grp_exprs);
13393 Self::rewrite_grouping_to_col(v, grp_exprs);
13394 }
13395 if let Some(x) = else_branch {
13396 Self::rewrite_grouping_to_col(x, grp_exprs);
13397 }
13398 }
13399 _ => {}
13400 }
13401 }
13402
13403 /// v7.39 (round 242) — one grouping element of PG's GROUP BY grammar,
13404 /// as the list of key sets it contributes. A bare expression is one
13405 /// single-key set; `ROLLUP (u1, …, un)` the n+1 unit-prefixes (largest
13406 /// first); `CUBE` every unit-subset (largest first); `GROUPING SETS`
13407 /// the concatenation of its items' sets, where an item is itself an
13408 /// element, a parenthesized key list, or the empty set `()`. A
13409 /// ROLLUP/CUBE member in parentheses is a composite UNIT: its keys
13410 /// move together.
13411 fn parse_grouping_element(&mut self) -> Result<Vec<Vec<Expr>>, ParseError> {
13412 let is_kw = |t: &Token, kw: &str| matches!(t, Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case(kw));
13413 // ROLLUP ( … ) / CUBE ( … )
13414 if (is_kw(self.peek(), "rollup") || is_kw(self.peek(), "cube"))
13415 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
13416 {
13417 let is_cube = is_kw(self.peek(), "cube");
13418 self.advance(); // ROLLUP / CUBE
13419 self.advance(); // (
13420 let mut units: Vec<Vec<Expr>> = Vec::new();
13421 loop {
13422 if matches!(self.peek(), Token::LParen) {
13423 // Composite unit: (a, b) rolls up as one.
13424 self.advance();
13425 let mut unit = Vec::new();
13426 if !matches!(self.peek(), Token::RParen) {
13427 loop {
13428 unit.push(self.parse_expr(0)?);
13429 match self.peek() {
13430 Token::Comma => {
13431 self.advance();
13432 }
13433 Token::RParen => break,
13434 other => {
13435 return Err(self.err(format!(
13436 "expected ',' or ')' in grouping unit, got {other:?}"
13437 )));
13438 }
13439 }
13440 }
13441 }
13442 self.advance(); // )
13443 units.push(unit);
13444 } else {
13445 units.push(alloc::vec![self.parse_expr(0)?]);
13446 }
13447 match self.peek() {
13448 Token::Comma => {
13449 self.advance();
13450 }
13451 Token::RParen => break,
13452 other => {
13453 return Err(self.err(format!(
13454 "expected ',' or ')' in grouping list, got {other:?}"
13455 )));
13456 }
13457 }
13458 }
13459 self.advance(); // )
13460 let flatten = |unit_sel: &[bool]| -> Vec<Expr> {
13461 units
13462 .iter()
13463 .zip(unit_sel.iter())
13464 .filter(|(_, keep)| **keep)
13465 .flat_map(|(u, _)| u.iter().cloned())
13466 .collect()
13467 };
13468 let n = units.len();
13469 if is_cube {
13470 let mut subsets: Vec<Vec<bool>> = (0..(1u32 << n))
13471 .map(|mask| (0..n).map(|i| mask & (1 << i) != 0).collect())
13472 .collect();
13473 subsets.sort_by_key(|sel| core::cmp::Reverse(sel.iter().filter(|b| **b).count()));
13474 return Ok(subsets.iter().map(|sel| flatten(sel)).collect());
13475 }
13476 return Ok((0..=n)
13477 .rev()
13478 .map(|keep| {
13479 let sel: Vec<bool> = (0..n).map(|i| i < keep).collect();
13480 flatten(&sel)
13481 })
13482 .collect());
13483 }
13484 // GROUPING SETS ( item [, item]* )
13485 if is_kw(self.peek(), "grouping")
13486 && matches!(self.tokens.get(self.pos + 1), Some(t) if is_kw(t, "sets"))
13487 {
13488 self.advance(); // GROUPING
13489 self.advance(); // SETS
13490 if !matches!(self.peek(), Token::LParen) {
13491 return Err(self.err(format!(
13492 "expected '(' after GROUPING SETS, got {:?}",
13493 self.peek()
13494 )));
13495 }
13496 self.advance(); // outer (
13497 let mut sets: Vec<Vec<Expr>> = Vec::new();
13498 loop {
13499 if matches!(self.peek(), Token::LParen) {
13500 // A parenthesized key list (or the empty set).
13501 self.advance();
13502 let mut set = Vec::new();
13503 if !matches!(self.peek(), Token::RParen) {
13504 loop {
13505 set.push(self.parse_expr(0)?);
13506 match self.peek() {
13507 Token::Comma => {
13508 self.advance();
13509 }
13510 Token::RParen => break,
13511 other => {
13512 return Err(self.err(format!(
13513 "expected ',' or ')' in grouping set, got {other:?}"
13514 )));
13515 }
13516 }
13517 }
13518 }
13519 self.advance(); // )
13520 sets.push(set);
13521 } else {
13522 // A nested element: ROLLUP/CUBE/GROUPING SETS or a
13523 // bare expression.
13524 sets.extend(self.parse_grouping_element()?);
13525 }
13526 match self.peek() {
13527 Token::Comma => {
13528 self.advance();
13529 }
13530 Token::RParen => break,
13531 other => {
13532 return Err(self.err(format!(
13533 "expected ',' or ')' after a grouping set, got {other:?}"
13534 )));
13535 }
13536 }
13537 }
13538 self.advance(); // outer )
13539 return Ok(sets);
13540 }
13541 Ok(alloc::vec![alloc::vec![self.parse_expr(0)?]])
13542 }
13543
13544 fn substitute_grouping_calls(expr: &mut Expr, dropped: &[Expr]) {
13545 // v7.38 (read01) — a reference to a key that is dropped in this grouping
13546 // set evaluates to NULL, at any depth. Previously only a *top-level*
13547 // select item equal to a dropped key was nullified, so a key nested in
13548 // an expression (`COALESCE(g,'TOTAL')`, `g || sum(v)`) survived as a raw
13549 // column and failed to resolve against the set's synthetic schema.
13550 if dropped.iter().any(|d| d == expr) {
13551 *expr = Expr::Literal(Literal::Null);
13552 return;
13553 }
13554 if let Expr::FunctionCall { name, args } = expr
13555 && name.eq_ignore_ascii_case("grouping")
13556 {
13557 let mut mask: i64 = 0;
13558 for a in args.iter() {
13559 mask <<= 1;
13560 if dropped.iter().any(|d| d == a) {
13561 mask |= 1;
13562 }
13563 }
13564 // v7.39 (round 242) — wrapped in a cast, NOT a bare integer
13565 // literal: a bare integer in a select item is indistinguishable
13566 // from a positional reference once `ORDER BY 1` substitutes the
13567 // item back in, and the round-232 position check then read the
13568 // mask value as an out-of-range position. The cast changes
13569 // nothing semantically (grouping() is integer).
13570 *expr = Expr::Cast {
13571 expr: alloc::boxed::Box::new(Expr::Literal(Literal::Integer(mask))),
13572 target: crate::ast::CastTarget::Int,
13573 };
13574 return;
13575 }
13576 // Generic recursion over the common expression shapes the
13577 // SELECT list uses; anything without child expressions is
13578 // left alone.
13579 match expr {
13580 Expr::FunctionCall { args, .. } => {
13581 for a in args {
13582 Self::substitute_grouping_calls(a, dropped);
13583 }
13584 }
13585 Expr::Binary { lhs, rhs, .. } => {
13586 Self::substitute_grouping_calls(lhs, dropped);
13587 Self::substitute_grouping_calls(rhs, dropped);
13588 }
13589 Expr::Unary { expr: inner, .. } => {
13590 Self::substitute_grouping_calls(inner, dropped);
13591 }
13592 Expr::Cast { expr: inner, .. } => {
13593 Self::substitute_grouping_calls(inner, dropped);
13594 }
13595 Expr::Case {
13596 operand,
13597 branches,
13598 else_branch,
13599 } => {
13600 if let Some(op) = operand {
13601 Self::substitute_grouping_calls(op, dropped);
13602 }
13603 for (w, t) in branches {
13604 Self::substitute_grouping_calls(w, dropped);
13605 Self::substitute_grouping_calls(t, dropped);
13606 }
13607 if let Some(e) = else_branch {
13608 Self::substitute_grouping_calls(e, dropped);
13609 }
13610 }
13611 // v7.38 (read01) — recurse into the remaining child-bearing shapes
13612 // so a dropped key nested in `IS NULL` / `LIKE` / `IN (…)` / `EXTRACT`
13613 // / a subscript / `ANY`/`ALL` is nullified too (`CASE WHEN g IS NULL
13614 // …` is the canonical rollup-total label idiom).
13615 Expr::IsNull { expr: inner, .. } => Self::substitute_grouping_calls(inner, dropped),
13616 Expr::Like { expr, pattern, .. } => {
13617 Self::substitute_grouping_calls(expr, dropped);
13618 Self::substitute_grouping_calls(pattern, dropped);
13619 }
13620 Expr::InList { expr, list, .. } => {
13621 Self::substitute_grouping_calls(expr, dropped);
13622 for item in list {
13623 Self::substitute_grouping_calls(item, dropped);
13624 }
13625 }
13626 Expr::Extract { source, .. } => Self::substitute_grouping_calls(source, dropped),
13627 Expr::Array(items) => {
13628 for item in items {
13629 Self::substitute_grouping_calls(item, dropped);
13630 }
13631 }
13632 Expr::ArraySubscript { target, index } => {
13633 Self::substitute_grouping_calls(target, dropped);
13634 Self::substitute_grouping_calls(index, dropped);
13635 }
13636 Expr::ArraySlice { target, lo, hi } => {
13637 Self::substitute_grouping_calls(target, dropped);
13638 if let Some(lo) = lo {
13639 Self::substitute_grouping_calls(lo, dropped);
13640 }
13641 if let Some(hi) = hi {
13642 Self::substitute_grouping_calls(hi, dropped);
13643 }
13644 }
13645 Expr::AnyAll { expr, array, .. } => {
13646 Self::substitute_grouping_calls(expr, dropped);
13647 Self::substitute_grouping_calls(array, dropped);
13648 }
13649 _ => {}
13650 }
13651 }
13652
13653 fn parse_bare_select(&mut self) -> Result<SelectStatement, ParseError> {
13654 // v7.37.17 (17.6 siblings) — parenthesized set-operation
13655 // group: `( <select chain> )` usable anywhere a query block
13656 // is (head or peer of an outer chain). The group's own
13657 // unions ride the returned SelectStatement; the executor's
13658 // nested-peer recursion runs them.
13659 if matches!(self.peek(), Token::LParen)
13660 && matches!(
13661 self.tokens.get(self.pos + 1),
13662 Some(Token::Select | Token::LParen | Token::Values)
13663 )
13664 {
13665 self.advance(); // (
13666 self.enter_nested()?;
13667 // v7.37 D.20 — a group whose head is a VALUES list:
13668 // `(VALUES (1),(2)) UNION (VALUES (3))`. Parse the VALUES body,
13669 // otherwise recurse into a nested SELECT/group head.
13670 let mut head = (if matches!(self.peek(), Token::Values) {
13671 self.advance(); // VALUES
13672 self.parse_values_rows_body()
13673 } else {
13674 self.parse_bare_select()
13675 })
13676 .and_then(|mut h| {
13677 self.parse_setop_chain_into(&mut h)?;
13678 Ok(h)
13679 });
13680 self.nest_depth -= 1;
13681 let mut head = match &mut head {
13682 Ok(h) => core::mem::take(h),
13683 Err(_) => return head,
13684 };
13685 // v7.37.17 (17.6 siblings) — group-internal tail:
13686 // `(A UNION B ORDER BY 1 LIMIT 5)`. Parse it into the
13687 // group head, then wrap the group as a derived table
13688 // (SELECT * FROM (group)) so the outer chain / outer
13689 // tail can't clobber the group's own ordering or limit.
13690 let has_tail = matches!(self.peek(), Token::Order | Token::Limit | Token::Offset)
13691 || matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13692 if s.eq_ignore_ascii_case("fetch"));
13693 if has_tail {
13694 self.parse_select_tail_into(&mut head)?;
13695 head = SelectStatement {
13696 locking: None,
13697 ctes: Vec::new(),
13698 distinct: false,
13699 distinct_on: Vec::new(),
13700 items: alloc::vec![SelectItem::Wildcard],
13701 from: Some(FromClause {
13702 primary: TableRef {
13703 name: "subquery".to_string(),
13704 alias: None,
13705 only: false,
13706 as_of_segment: None,
13707 unnest_expr: None,
13708 unnest_column_aliases: Vec::new(),
13709 with_ordinality: false,
13710 generate_series_args: None,
13711 lateral_subquery: Some(Box::new(head)),
13712 jsonb_each_text_arg: None,
13713 table_fn_call: None,
13714 rows_from: None,
13715 json_table: None,
13716 scalar_fn_item: false,
13717 },
13718 joins: Vec::new(),
13719 }),
13720 where_: None,
13721 group_by: None,
13722 group_by_all: false,
13723 having: None,
13724 unions: Vec::new(),
13725 order_by: Vec::new(),
13726 limit: None,
13727 offset: None,
13728 limit_with_ties: false,
13729 window_check_exprs: Vec::new(),
13730 };
13731 }
13732 if !matches!(self.peek(), Token::RParen) {
13733 return Err(self.err(format!(
13734 "expected ')' after parenthesized query group, got {:?}",
13735 self.peek()
13736 )));
13737 }
13738 self.advance();
13739 return Ok(head);
13740 }
13741 // `TABLE name` shorthand as a query block — valid anywhere
13742 // a SELECT head is (set-op peers included).
13743 if matches!(self.peek(), Token::Table)
13744 && matches!(
13745 self.tokens.get(self.pos + 1),
13746 Some(Token::Ident(_) | Token::QuotedIdent(_))
13747 )
13748 {
13749 return self.parse_table_shorthand();
13750 }
13751 if !matches!(self.peek(), Token::Select) {
13752 return Err(self.err(format!(
13753 "expected SELECT to start a query block, got {:?}",
13754 self.peek()
13755 )));
13756 }
13757 self.advance();
13758 // v7.39.9 — MySQL's `SELECT STRAIGHT_JOIN …` join-order hint.
13759 //
13760 // It sits where `DISTINCT` sits and tells the optimiser to join
13761 // in the written order. SPG plans its own joins, so the hint is
13762 // accepted and not acted on — but it has to PARSE, because as a
13763 // bare identifier it became a column: measured on the published
13764 // image, `SELECT STRAIGHT_JOIN a FROM t` answered `Unknown
13765 // column 'straight_join' in 'field list'` where MySQL 9.7.2
13766 // returns the rows. Only in this position, which is the only one
13767 // MySQL accepts either — a trailing `STRAIGHT_JOIN` is its 1064.
13768 if self.mysql_dialect
13769 && matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("straight_join"))
13770 {
13771 self.advance();
13772 }
13773 let distinct = if matches!(self.peek(), Token::Distinct) {
13774 self.advance();
13775 true
13776 } else {
13777 false
13778 };
13779 // v7.37.17 (17.6 siblings) — `DISTINCT ON (expr [, …])`:
13780 // keep the first row (per ORDER BY) of each group the
13781 // expressions define. Django's .distinct('field') shape.
13782 let distinct_on: Vec<Expr> = if distinct && matches!(self.peek(), Token::On) {
13783 self.advance(); // ON
13784 if !matches!(self.peek(), Token::LParen) {
13785 return Err(self.err(format!(
13786 "expected '(' after DISTINCT ON, got {:?}",
13787 self.peek()
13788 )));
13789 }
13790 self.advance();
13791 let mut exprs = Vec::new();
13792 loop {
13793 exprs.push(self.parse_expr(0)?);
13794 match self.peek() {
13795 Token::Comma => {
13796 self.advance();
13797 }
13798 Token::RParen => break,
13799 other => {
13800 return Err(self.err(format!(
13801 "expected ',' or ')' in DISTINCT ON list, got {other:?}"
13802 )));
13803 }
13804 }
13805 }
13806 self.advance(); // )
13807 exprs
13808 } else {
13809 Vec::new()
13810 };
13811 let mut items = self.parse_select_list()?;
13812 // v7.38.19 — `SELECT … INTO <table>`, PostgreSQL's other spelling
13813 // of CTAS. It sits exactly here in PG's grammar, right after the
13814 // target list.
13815 //
13816 // A comment in `ast.rs` has said since v7.38 that CTAS and
13817 // `SELECT INTO` lower to the same node. Only CTAS ever did:
13818 // `SELECT i INTO t FROM src` answered `syntax error at or near
13819 // "INTO"`, which the differential found while measuring what
13820 // PostgreSQL tags each of the five materialising forms with. A
13821 // comment describing a capability the code does not have is the
13822 // defect this version has been finding all day, and this is the
13823 // one it found in the parser.
13824 //
13825 // `INTO` is captured rather than consumed here: the name has to
13826 // travel out of a function that returns a `SelectStatement`, and
13827 // the caller lowers the whole thing to the CTAS node.
13828 if matches!(self.peek(), Token::Into) {
13829 self.advance();
13830 // `TEMP` / `TEMPORARY` / `UNLOGGED` / `TABLE` are modifiers on
13831 // the target, not part of its name. SPG has one storage
13832 // class, so `UNLOGGED` is accepted and means nothing, which
13833 // is what it already means on `CREATE TABLE`.
13834 let mut temporary = false;
13835 loop {
13836 match self.peek().clone() {
13837 Token::Ident(w) | Token::QuotedIdent(w)
13838 if w.eq_ignore_ascii_case("temp")
13839 || w.eq_ignore_ascii_case("temporary") =>
13840 {
13841 temporary = true;
13842 self.advance();
13843 }
13844 Token::Ident(w) | Token::QuotedIdent(w)
13845 if w.eq_ignore_ascii_case("unlogged") =>
13846 {
13847 self.advance();
13848 }
13849 Token::Table => {
13850 self.advance();
13851 }
13852 _ => break,
13853 }
13854 }
13855 let name = match self.peek().clone() {
13856 Token::Ident(w) | Token::QuotedIdent(w) => {
13857 self.advance();
13858 w
13859 }
13860 other => {
13861 return Err(self.err(alloc::format!(
13862 "expected a table name after SELECT … INTO, got {other:?}"
13863 )));
13864 }
13865 };
13866 self.pending_select_into = Some((name, temporary));
13867 }
13868 // Scope the TABLESAMPLE lowering channel to this SELECT:
13869 // stash whatever an enclosing select accumulated, collect
13870 // our own FROM's predicates, restore after the combine.
13871 let enclosing_sample_preds = core::mem::take(&mut self.pending_sample_preds);
13872 let mut from = if matches!(self.peek(), Token::From) {
13873 self.advance();
13874 Some(self.parse_from_clause()?)
13875 } else {
13876 None
13877 };
13878 // v7.37 D.22 — a set-returning function in the projection with no FROM
13879 // (`SELECT unnest(arr)`, `SELECT 'x', generate_series(a,b)`) expands to
13880 // rows. Move the first SRF projection item to a FROM-position derived
13881 // table and replace it in the projection with a reference to its output
13882 // column; sibling scalar columns repeat per SRF row. PG names the output
13883 // column after the function (or its AS alias). Reuses the FROM-SRF
13884 // machinery. Only fires with no FROM — mixed SRF over a real FROM already
13885 // works via the targetlist-SRF path.
13886 // v7.39 (read01 round 69) — lower `(f(args)).*`. With no FROM it IS
13887 // `SELECT * FROM f(args)` — the record's fields become the columns, which
13888 // is exactly what the function's own row shape already is. Anywhere else
13889 // (per outer row, or beside other items) it would need a real record-typed
13890 // projection, so it says so rather than answering something else.
13891 if let [
13892 SelectItem::Expr {
13893 expr: Expr::FunctionCall { name, args },
13894 ..
13895 },
13896 ] = items.as_slice()
13897 && name == "__record_expand"
13898 {
13899 let Some(Expr::FunctionCall {
13900 name: inner_name,
13901 args: inner_args,
13902 }) = args.first()
13903 else {
13904 return Err(self.err(
13905 "(<expr>).* expands a function's record — it needs a function call".into(),
13906 ));
13907 };
13908 if from.is_some() {
13909 return Err(self.err(
13910 "(<fn>).* over a FROM clause is not supported — call the function in FROM"
13911 .into(),
13912 ));
13913 }
13914 let fn_ref = TableRef {
13915 name: inner_name.clone(),
13916 alias: None,
13917 only: false,
13918 as_of_segment: None,
13919 unnest_expr: None,
13920 unnest_column_aliases: Vec::new(),
13921 with_ordinality: false,
13922 generate_series_args: None,
13923 lateral_subquery: None,
13924 jsonb_each_text_arg: None,
13925 table_fn_call: Some(Box::new((
13926 inner_name.to_ascii_lowercase(),
13927 inner_args.clone(),
13928 ))),
13929 rows_from: None,
13930 json_table: None,
13931 scalar_fn_item: false,
13932 };
13933 items = alloc::vec![SelectItem::Wildcard];
13934 from = Some(FromClause {
13935 primary: fn_ref,
13936 joins: Vec::new(),
13937 });
13938 }
13939 // v7.39 (read01 round 74) — `(f(args)).*` beside other items, or over a
13940 // FROM, keeps its marker: the ENGINE lowers it, because naming the
13941 // record's fields takes the catalog. It becomes a LATERAL of the same
13942 // function plus one item per declared column — the machinery rounds 65
13943 // and 69 already built.
13944 // v7.39 (read01 round 67) — the lift moves ONE SRF into FROM. With two
13945 // (`SELECT generate_series(1,3), generate_series(10,11)`) PG runs them in
13946 // LOCKSTEP, padding the shorter with NULLs — a shape the lift cannot
13947 // express, since the lifted one becomes a scan and the other would
13948 // expand per its rows (a cross product, not a zip). So when the
13949 // projection holds more than one top-level function call, the lift steps
13950 // aside and the engine's target-list expansion takes the whole list.
13951 let fn_call_items = items
13952 .iter()
13953 .filter(|it| {
13954 matches!(
13955 it,
13956 SelectItem::Expr {
13957 expr: Expr::FunctionCall { .. },
13958 ..
13959 }
13960 )
13961 })
13962 .count();
13963 if from.is_none() && fn_call_items <= 1 {
13964 let mut found: Option<(usize, TableRef, String)> = None;
13965 for (i, item) in items.iter().enumerate() {
13966 if let SelectItem::Expr {
13967 expr: Expr::FunctionCall { name, args },
13968 alias,
13969 } = item
13970 {
13971 let lname = name.to_ascii_lowercase();
13972 let colname = alias.clone().unwrap_or_else(|| lname.clone());
13973 let (unnest, gs) = match lname.as_str() {
13974 "unnest" if args.len() == 1 => (Some(Box::new(args[0].clone())), None),
13975 "generate_series" if (2..=3).contains(&args.len()) => {
13976 (None, Some(args.clone()))
13977 }
13978 // v7.38 (read01) — generate_subscripts(arr, dim) in a
13979 // no-FROM projection yields the 1-based subscripts, i.e.
13980 // generate_series(1, array_length(arr, dim)); an invalid
13981 // dimension makes array_length NULL → 0 rows, as in PG.
13982 "generate_subscripts" if args.len() == 2 => (
13983 None,
13984 Some(alloc::vec![
13985 Expr::Literal(Literal::Integer(1)),
13986 Expr::FunctionCall {
13987 name: "array_length".to_string(),
13988 args: args.clone(),
13989 },
13990 ]),
13991 ),
13992 // v7.38 (read01, T-srf) — string_to_table / regexp_split_to_table
13993 // in a no-FROM projection unnest their *_to_array form.
13994 "string_to_table" | "regexp_split_to_table" => {
13995 let array_fn = if lname == "string_to_table" {
13996 "string_to_array"
13997 } else {
13998 "regexp_split_to_array"
13999 };
14000 (
14001 Some(Box::new(Expr::FunctionCall {
14002 name: array_fn.to_string(),
14003 args: args.clone(),
14004 })),
14005 None,
14006 )
14007 }
14008 // v7.38 (read01, T15) — jsonb/json_array_elements[_text] in
14009 // a no-FROM projection expand per element. The scalar form
14010 // returns the elements as a TEXT array, so unnest over the
14011 // same call materialises one row each (same rewrite the
14012 // FROM-clause form uses).
14013 "jsonb_array_elements"
14014 | "json_array_elements"
14015 | "jsonb_array_elements_text"
14016 | "json_array_elements_text"
14017 if args.len() == 1 =>
14018 {
14019 (
14020 Some(Box::new(Expr::FunctionCall {
14021 name: lname.clone(),
14022 args: args.clone(),
14023 })),
14024 None,
14025 )
14026 }
14027 // v7.38 (read01, T15) — jsonb/json_path_query(doc, path)
14028 // in a no-FROM projection expands per match (scalar form
14029 // returns the matches as a TEXT array → unnest).
14030 "jsonb_path_query" | "json_path_query" if args.len() == 2 => (
14031 Some(Box::new(Expr::FunctionCall {
14032 name: lname.clone(),
14033 args: args.clone(),
14034 })),
14035 None,
14036 ),
14037 _ => continue,
14038 };
14039 found = Some((
14040 i,
14041 TableRef {
14042 name: colname.clone(),
14043 alias: Some(colname.clone()),
14044 only: false,
14045 as_of_segment: None,
14046 unnest_expr: unnest,
14047 unnest_column_aliases: alloc::vec![colname.clone()],
14048 with_ordinality: false,
14049 generate_series_args: gs,
14050 lateral_subquery: None,
14051 jsonb_each_text_arg: None,
14052 table_fn_call: None,
14053 rows_from: None,
14054 json_table: None,
14055 scalar_fn_item: false,
14056 },
14057 colname,
14058 ));
14059 break;
14060 }
14061 }
14062 if let Some((idx, tref, colname)) = found {
14063 from = Some(FromClause {
14064 primary: tref,
14065 joins: Vec::new(),
14066 });
14067 items[idx] = SelectItem::Expr {
14068 expr: Expr::Column(ColumnName {
14069 qualifier: None,
14070 name: colname.clone(),
14071 }),
14072 alias: Some(colname),
14073 };
14074 }
14075 }
14076 let sample_preds = core::mem::take(&mut self.pending_sample_preds);
14077 let where_ = if matches!(self.peek(), Token::Where) {
14078 self.advance();
14079 Some(self.parse_expr(0)?)
14080 } else {
14081 None
14082 };
14083 let where_ = sample_preds.into_iter().fold(where_, |acc, pred| {
14084 Some(match acc {
14085 Some(w) => Expr::Binary {
14086 lhs: Box::new(pred),
14087 op: crate::ast::BinOp::And,
14088 rhs: Box::new(w),
14089 },
14090 None => pred,
14091 })
14092 });
14093 self.pending_sample_preds = enclosing_sample_preds;
14094 let mut group_by_all = false;
14095 // v7.37.17 (17.6 siblings) — ROLLUP / CUBE / GROUPING SETS
14096 // share one expansion: `grouping_sets` lists the key subsets
14097 // (first = primary, assigned to stmt.group_by; the rest
14098 // become UNION ALL peers), `grouping_universe` is the full
14099 // key list used to compute each peer's dropped keys.
14100 let mut grouping_sets: Vec<Vec<Expr>> = Vec::new();
14101 let mut grouping_universe: Vec<Expr> = Vec::new();
14102 // v7.39 (round 472) — did the GROUP BY end in MySQL's `WITH ROLLUP`?
14103 // A BOOL, not the key list: this frame is the statement parser's, and
14104 // round 430 measured that a `Vec` local here is enough on its own to
14105 // tip the 512 KiB nesting guard. The keys are recoverable from
14106 // `grouping_universe`, which a rollup fills with exactly them.
14107 let mut mysql_rollup = false;
14108 let group_by = if matches!(self.peek(), Token::Group) {
14109 self.advance();
14110 if !self.peek_is_by() {
14111 return Err(self.err(format!("expected BY after GROUP, got {:?}", self.peek())));
14112 }
14113 self.advance();
14114 // v6.4.1 — `GROUP BY ALL` shortcut. Planner expands to
14115 // every non-aggregate SELECT-list item later.
14116 if matches!(self.peek(), Token::All) {
14117 self.advance();
14118 group_by_all = true;
14119 None
14120 } else {
14121 // v7.39 (round 242) — PG's general grouping-element grammar:
14122 // GROUP BY [DISTINCT] element [, element]*, where an element
14123 // is a bare expression, ROLLUP (…), CUBE (…) or GROUPING
14124 // SETS (…) — mixed freely. Each element yields a list of
14125 // key sets; the query's grouping sets are the CARTESIAN
14126 // PRODUCT of the elements' lists (so `a, ROLLUP (b)` is
14127 // {(a,b),(a)}), and DISTINCT drops duplicate sets by
14128 // content. ROLLUP/CUBE members may be composite
14129 // (`ROLLUP ((a, b))` moves a and b as one unit), and a
14130 // GROUPING SETS item may itself be a ROLLUP/CUBE. The old
14131 // parser handled only a lone ROLLUP/CUBE/GS as the whole
14132 // clause.
14133 let distinct_sets = if matches!(self.peek(), Token::Distinct) {
14134 self.advance();
14135 true
14136 } else {
14137 false
14138 };
14139 let mut element_sets: Vec<Vec<Vec<Expr>>> = Vec::new();
14140 loop {
14141 element_sets.push(self.parse_grouping_element()?);
14142 if matches!(self.peek(), Token::Comma) {
14143 self.advance();
14144 } else {
14145 break;
14146 }
14147 }
14148 let mut total: Vec<Vec<Expr>> = alloc::vec![Vec::new()];
14149 for el in &element_sets {
14150 let mut next: Vec<Vec<Expr>> = Vec::new();
14151 for base in &total {
14152 for set in el {
14153 let mut merged = base.clone();
14154 for k in set {
14155 if !merged.iter().any(|m| m == k) {
14156 merged.push(k.clone());
14157 }
14158 }
14159 next.push(merged);
14160 }
14161 }
14162 total = next;
14163 }
14164 // v7.39 (round 472) — MySQL spells a rollup as a SUFFIX:
14165 // `GROUP BY a, b WITH ROLLUP` is PG's `GROUP BY ROLLUP(a, b)`.
14166 // The keys and the aggregates come out identical; the ROW
14167 // ORDER does not, and that is the part a report depends on.
14168 // MySQL interleaves each group's subtotal right after its
14169 // own rows (east/a, east/b, east/NULL, west/a, …, NULL/NULL)
14170 // where the union-of-grouping-sets expansion emits every
14171 // leaf first and then every subtotal. MariaDB REFUSES an
14172 // ORDER BY next to ROLLUP (1221), so a client cannot fix the
14173 // order itself — measured on MariaDB 11 and MySQL 9.7, which
14174 // agree on the order and disagree only on whether ORDER BY
14175 // is allowed (MySQL allows it; SPG allows it too, since
14176 // refusing would break the clients that can write it).
14177 if self.mysql_dialect
14178 && matches!(self.peek(), Token::Ident(w) if w.eq_ignore_ascii_case("with"))
14179 && matches!(
14180 self.tokens.get(self.pos + 1),
14181 Some(Token::Ident(r)) if r.eq_ignore_ascii_case("rollup")
14182 )
14183 {
14184 self.advance(); // WITH
14185 self.advance(); // ROLLUP
14186 let keys = total.into_iter().next().unwrap_or_default();
14187 mysql_rollup = true;
14188 // n+1 prefixes, largest first — the same expansion
14189 // `ROLLUP (…)` produces.
14190 total = (0..=keys.len()).rev().map(|n| keys[..n].to_vec()).collect();
14191 }
14192 if distinct_sets {
14193 let mut seen: Vec<Vec<String>> = Vec::new();
14194 total.retain(|set| {
14195 let mut key: Vec<String> =
14196 set.iter().map(|e| alloc::format!("{e}")).collect();
14197 key.sort();
14198 if seen.contains(&key) {
14199 false
14200 } else {
14201 seen.push(key);
14202 true
14203 }
14204 });
14205 }
14206 if total.len() > 1 {
14207 let mut universe: Vec<Expr> = Vec::new();
14208 for set in &total {
14209 for k in set {
14210 if !universe.iter().any(|u| u == k) {
14211 universe.push(k.clone());
14212 }
14213 }
14214 }
14215 grouping_universe = universe;
14216 let primary = total[0].clone();
14217 grouping_sets = total;
14218 Some(primary)
14219 } else {
14220 // One set (a plain GROUP BY list, or a single-set
14221 // spelling like GROUPING SETS ((a, b))). An EMPTY
14222 // single set — GROUPING SETS (()) — stays
14223 // `Some(vec![])`: the grand-total group, which must
14224 // run the aggregate path.
14225 Some(total.into_iter().next().unwrap_or_default())
14226 }
14227 }
14228 } else {
14229 None
14230 };
14231 let having = if matches!(self.peek(), Token::Having) {
14232 self.advance();
14233 Some(self.parse_expr(0)?)
14234 } else {
14235 None
14236 };
14237 // `WINDOW w AS ( <window-def> ) [, ...]` — named windows.
14238 // OVER w parsed to a marker above; inline each definition
14239 // into the referencing WindowFunction nodes.
14240 let mut window_defs: Vec<(String, WindowDef)> = Vec::new();
14241 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("window")) {
14242 self.advance();
14243 loop {
14244 let wname = self.expect_ident_like()?;
14245 if !matches!(self.peek(), Token::As) {
14246 return Err(self.err(format!(
14247 "expected AS after WINDOW {wname}, got {:?}",
14248 self.peek()
14249 )));
14250 }
14251 self.advance();
14252 // v7.39 (round 229) — PG rejects a redefinition outright.
14253 if window_defs
14254 .iter()
14255 .any(|(n, _)| n.eq_ignore_ascii_case(&wname))
14256 {
14257 return Err(self.err(alloc::format!("window \"{wname}\" is already defined")));
14258 }
14259 let def = self.parse_over_clause()?;
14260 // A definition may itself copy an earlier one
14261 // (`WINDOW w1 AS (PARTITION BY g), w2 AS (w1 ORDER BY v)`),
14262 // so resolve it against the defs already in scope. Same
14263 // copy rules as an `OVER (w1 …)` in the select list.
14264 let mut probe = Expr::WindowFunction {
14265 name: String::new(),
14266 args: Vec::new(),
14267 partition_by: def.0,
14268 order_by: def.1,
14269 frame: def.2,
14270 null_treatment: crate::ast::NullTreatment::Respect,
14271 filter: None,
14272 };
14273 Self::substitute_named_windows(&mut probe, &window_defs)
14274 .map_err(|m| self.err(m))?;
14275 let Expr::WindowFunction {
14276 partition_by,
14277 order_by,
14278 frame,
14279 ..
14280 } = probe
14281 else {
14282 unreachable!("probe is a WindowFunction")
14283 };
14284 window_defs.push((wname, (partition_by, order_by, frame)));
14285 if matches!(self.peek(), Token::Comma) {
14286 self.advance();
14287 continue;
14288 }
14289 break;
14290 }
14291 }
14292 // v7.39 (round 705) — which definitions did anything reference?
14293 // The ones nothing did used to be dropped here, unexamined, so
14294 // `WINDOW w AS (ORDER BY nosuch)` succeeded — PG analyses every
14295 // definition whether referenced or not. Their key expressions ride
14296 // out on the statement for the engine to resolve.
14297 let mut window_refs: Vec<String> = Vec::new();
14298 if !window_defs.is_empty() {
14299 for it in &items {
14300 if let SelectItem::Expr { expr, .. } = it {
14301 Self::collect_named_window_refs(expr, &mut window_refs);
14302 }
14303 }
14304 }
14305 let window_check_exprs: Vec<Expr> = window_defs
14306 .iter()
14307 .filter(|(n, _)| !window_refs.iter().any(|r| r.eq_ignore_ascii_case(n)))
14308 .flat_map(|(_, (partition, order, _))| {
14309 partition
14310 .iter()
14311 .cloned()
14312 .chain(order.iter().map(|(e, _, _)| e.clone()))
14313 })
14314 .collect();
14315 if !window_defs.is_empty()
14316 || items
14317 .iter()
14318 .any(|it| matches!(it, SelectItem::Expr { expr, .. } if Self::expr_has_named_window(expr)))
14319 {
14320 for it in &mut items {
14321 if let SelectItem::Expr { expr, .. } = it {
14322 Self::substitute_named_windows(expr, &window_defs)
14323 .map_err(|m| self.err(m))?;
14324 }
14325 }
14326 }
14327 // `GROUP BY 1` — positional keys substitute with the Nth
14328 // select item's expression (same contract ORDER BY has had
14329 // since v6.x). Out-of-range positions error.
14330 let group_by = match group_by {
14331 Some(mut keys) => {
14332 for k in &mut keys {
14333 if let Expr::Literal(Literal::Integer(n)) = k {
14334 let idx = *n;
14335 if idx < 1 || idx as usize > items.len() {
14336 return Err(self.err(alloc::format!(
14337 "GROUP BY position {idx} is not in select list"
14338 )));
14339 }
14340 match &items[(idx - 1) as usize] {
14341 SelectItem::Expr { expr, .. } => *k = expr.clone(),
14342 SelectItem::Wildcard | SelectItem::QualifiedWildcard(_) => {
14343 return Err(self.err(alloc::format!(
14344 "GROUP BY position {idx} references a wildcard item"
14345 )));
14346 }
14347 }
14348 }
14349 }
14350 Some(keys)
14351 }
14352 None => None,
14353 };
14354 let mut stmt = SelectStatement {
14355 locking: None,
14356 ctes: Vec::new(),
14357 distinct,
14358 distinct_on,
14359 items,
14360 from,
14361 where_,
14362 group_by,
14363 group_by_all,
14364 having,
14365 unions: Vec::new(),
14366 order_by: Vec::new(),
14367 limit: None,
14368 offset: None,
14369 limit_with_ties: false,
14370 window_check_exprs,
14371 };
14372 // Grouping expansion (ROLLUP / CUBE / GROUPING SETS): the
14373 // first set is the primary (already on stmt.group_by); each
14374 // further set becomes a UNION ALL peer with its dropped
14375 // keys (universe minus the set) replaced by NULL literals
14376 // in the peer's items and group_by. PG-legal: non-grouped
14377 // select items must be group keys or aggregates, so a
14378 // dropped key's occurrences in the projection are exactly
14379 // the ones to nullify.
14380 // v7.39 (round 242) — grouping() OUTSIDE an expansion: PG allows it
14381 // over a plain GROUP BY (every argument must be a group key; the
14382 // mask is then 0) and rejects anything else with 42803. SPG's
14383 // rewrite only ran during the ROLLUP/CUBE expansion, so a plain
14384 // `SELECT grouping(a) … GROUP BY a` died at eval with "unknown
14385 // function `grouping`".
14386 if grouping_sets.len() <= 1 {
14387 let keys: Vec<Expr> = stmt.group_by.clone().unwrap_or_default();
14388 let mut calls: Vec<Expr> = Vec::new();
14389 for item in &stmt.items {
14390 if let SelectItem::Expr { expr, .. } = item {
14391 Self::collect_grouping_calls(expr, &mut calls);
14392 }
14393 }
14394 if let Some(h) = &stmt.having {
14395 Self::collect_grouping_calls(h, &mut calls);
14396 }
14397 for call in &calls {
14398 let Expr::FunctionCall { args, .. } = call else {
14399 continue;
14400 };
14401 for a in args {
14402 if !keys.iter().any(|k| k == a) {
14403 return Err(self.err(
14404 "arguments to GROUPING must be grouping expressions of the associated query level"
14405 .to_string(),
14406 ));
14407 }
14408 }
14409 }
14410 if !calls.is_empty() {
14411 for item in &mut stmt.items {
14412 if let SelectItem::Expr { expr, .. } = item {
14413 Self::substitute_grouping_calls(expr, &[]);
14414 }
14415 }
14416 if let Some(h) = &mut stmt.having {
14417 Self::substitute_grouping_calls(h, &[]);
14418 }
14419 }
14420 }
14421 if grouping_sets.len() > 1 {
14422 // The primary set's own dropped keys nullify in the
14423 // HEAD's projection too (GROUPING SETS's first set may
14424 // omit keys other sets use).
14425 let primary = grouping_sets[0].clone();
14426 let head_dropped: Vec<Expr> = grouping_universe
14427 .iter()
14428 .filter(|u| !primary.iter().any(|k| k == *u))
14429 .cloned()
14430 .collect();
14431 for set in grouping_sets.iter().skip(1) {
14432 let mut peer = stmt.clone();
14433 peer.unions = Vec::new();
14434 let dropped: Vec<&Expr> = grouping_universe
14435 .iter()
14436 .filter(|u| !set.iter().any(|k| k == *u))
14437 .collect();
14438 // Empty set = grand-total group: `Some(vec![])` forces
14439 // the aggregate path (one collapsed row) instead of a
14440 // per-row passthrough. See the primary-set note above.
14441 peer.group_by = Some(set.clone());
14442 let dropped_owned: Vec<Expr> = dropped.iter().map(|d| (*d).clone()).collect();
14443 for item in &mut peer.items {
14444 if let SelectItem::Expr { expr, alias } = item {
14445 if dropped.iter().any(|d| *d == expr) {
14446 // v7.39 — keep the dropped key's name on the
14447 // NULL literal so the UNION output column
14448 // (and any top-level ORDER BY on it) still
14449 // resolves.
14450 if alias.is_none()
14451 && let Expr::Column(c) = &expr
14452 {
14453 *alias = Some(c.name.clone());
14454 }
14455 *expr = Expr::Literal(Literal::Null);
14456 } else {
14457 Self::substitute_grouping_calls(expr, &dropped_owned);
14458 }
14459 }
14460 }
14461 if let Some(h) = &mut peer.having {
14462 Self::substitute_grouping_calls(h, &dropped_owned);
14463 }
14464 stmt.unions.push((UnionKind::All, peer));
14465 }
14466 for item in &mut stmt.items {
14467 if let SelectItem::Expr { expr, alias } = item {
14468 if head_dropped.iter().any(|d| d == expr) {
14469 if alias.is_none()
14470 && let Expr::Column(c) = &expr
14471 {
14472 *alias = Some(c.name.clone());
14473 }
14474 *expr = Expr::Literal(Literal::Null);
14475 } else {
14476 Self::substitute_grouping_calls(expr, &head_dropped);
14477 }
14478 }
14479 }
14480 if let Some(h) = &mut stmt.having {
14481 Self::substitute_grouping_calls(h, &head_dropped);
14482 }
14483 // v7.39 (round 135) — GROUPING() in ORDER BY. Parse the ORDER BY now
14484 // (while `grouping_universe` / the per-branch sets are in scope). For
14485 // each grouping() call in it, inject a per-branch hidden column
14486 // `__grp_ord_K` carrying that branch's mask into the head + every
14487 // peer, and rewrite the ORDER BY to reference it. `parse_select_tail_into`
14488 // preserves this pre-set order_by; the engine strips `__grp_ord_*`
14489 // from the final output. A standalone grouping-set query has ORDER BY
14490 // (not an explicit set-op) next, so consuming it here is safe.
14491 // v7.39 (round 472) — absent the client's own ORDER BY, a MySQL
14492 // rollup carries the hierarchical order: sort by the grouping
14493 // keys with the rolled-up NULLs last, which is exactly the
14494 // interleaving both oracles emit. A client's own ORDER BY wins,
14495 // which is what MySQL does (MariaDB refuses to let one be
14496 // written at all).
14497 // The synthesised keys have to travel the SAME path a written
14498 // ORDER BY does: the block below is what turns a `grouping()`
14499 // call into the per-branch `__grp_ord_K` column the engine can
14500 // actually sort on. Bypassing it left a bare `grouping(text)`
14501 // for the evaluator to reject.
14502 let synthesised_or_parsed: Vec<OrderBy> = if matches!(self.peek(), Token::Order) {
14503 self.parse_order_by_keys()?
14504 } else if mysql_rollup {
14505 Self::mysql_rollup_order(&grouping_universe)
14506 } else {
14507 Vec::new()
14508 };
14509 if !synthesised_or_parsed.is_empty() {
14510 let mut order_keys = synthesised_or_parsed;
14511 let mut grp_exprs: Vec<Expr> = Vec::new();
14512 for ob in &order_keys {
14513 Self::collect_grouping_calls(&ob.expr, &mut grp_exprs);
14514 }
14515 for (k, gexpr) in grp_exprs.iter().enumerate() {
14516 let colname = alloc::format!("__grp_ord_{k}");
14517 // Head branch (primary set) uses `head_dropped`.
14518 let mut he = gexpr.clone();
14519 Self::substitute_grouping_calls(&mut he, &head_dropped);
14520 stmt.items.push(SelectItem::Expr {
14521 expr: he,
14522 alias: Some(colname.clone()),
14523 });
14524 // Each peer `stmt.unions[i]` corresponds to `grouping_sets[i+1]`.
14525 for (i, (_, peer)) in stmt.unions.iter_mut().enumerate() {
14526 let set = &grouping_sets[i + 1];
14527 let dropped: Vec<Expr> = grouping_universe
14528 .iter()
14529 .filter(|u| !set.iter().any(|k| k == *u))
14530 .cloned()
14531 .collect();
14532 let mut pe = gexpr.clone();
14533 Self::substitute_grouping_calls(&mut pe, &dropped);
14534 peer.items.push(SelectItem::Expr {
14535 expr: pe,
14536 alias: Some(colname.clone()),
14537 });
14538 }
14539 }
14540 for ob in &mut order_keys {
14541 Self::rewrite_grouping_to_col(&mut ob.expr, &grp_exprs);
14542 }
14543 stmt.order_by = order_keys;
14544 }
14545 }
14546 Ok(stmt)
14547 }
14548
14549 /// v7.39 (round 472) — the row order MySQL's `WITH ROLLUP` promises,
14550 /// as ORDER BY keys.
14551 ///
14552 /// Per key: the rollup marker, then the key. Sorting on the key alone
14553 /// is not enough, and a table with a NULL in it says why — MariaDB puts
14554 /// the DATA-NULL group where a plain GROUP BY puts it (first) and only
14555 /// the ROLLUP-introduced NULL last, and both print as NULL.
14556 /// `GROUPING(k)` is the one thing that tells them apart: 0 for every
14557 /// real group including the data-NULL one, 1 only for the row the
14558 /// rollup added. Measured on MariaDB 11 — `('a',1),(NULL,2),('b',3)`
14559 /// rolls up to NULL|2, a|1, b|3, NULL|6.
14560 ///
14561 /// `#[inline(never)]`: its locals must not join the statement parser's
14562 /// frame, which round 430 measured sitting against the nesting guard.
14563 #[inline(never)]
14564 fn mysql_rollup_order(keys: &[Expr]) -> Vec<OrderBy> {
14565 let mut out: Vec<OrderBy> = Vec::with_capacity(keys.len() * 2);
14566 for e in keys {
14567 out.push(OrderBy {
14568 expr: Expr::FunctionCall {
14569 name: "grouping".into(),
14570 args: alloc::vec![e.clone()],
14571 },
14572 desc: false,
14573 nulls_first: None,
14574 collation: None,
14575 });
14576 out.push(OrderBy {
14577 expr: e.clone(),
14578 desc: false,
14579 // MySQL orders NULL first on an ascending key.
14580 nulls_first: Some(true),
14581 collation: None,
14582 });
14583 }
14584 out
14585 }
14586
14587 /// v7.39 (round 535) — `REINDEX [(opts)] { INDEX | TABLE | SCHEMA |
14588 /// DATABASE | SYSTEM } [CONCURRENTLY] [<name>]`.
14589 #[inline(never)]
14590 fn parse_reindex_tail(&mut self) -> Result<Statement, ParseError> {
14591 use crate::ast::MaintainKind;
14592 self.skip_paren_option_list();
14593 let kind = match self.peek() {
14594 // `TABLE` and `INDEX` lex as keywords, not identifiers.
14595 Token::Table | Token::Index => {
14596 self.advance();
14597 MaintainKind::ReindexRelation
14598 }
14599 Token::Ident(s) | Token::QuotedIdent(s) => match s.to_ascii_lowercase().as_str() {
14600 "index" | "table" => {
14601 self.advance();
14602 MaintainKind::ReindexRelation
14603 }
14604 "schema" => {
14605 self.advance();
14606 MaintainKind::ReindexSchema
14607 }
14608 "system" | "database" => {
14609 self.advance();
14610 MaintainKind::Whole
14611 }
14612 // PG requires the object type; anything else is the
14613 // caller's problem, not something to swallow.
14614 _ => MaintainKind::ReindexRelation,
14615 },
14616 _ => MaintainKind::Whole,
14617 };
14618 // PG bars `REINDEX … CONCURRENTLY` inside a transaction block and
14619 // allows the plain form, so the modifier is recorded rather than
14620 // skipped. It still has no effect on how the reindex runs.
14621 let mut concurrently = false;
14622 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("concurrently")) {
14623 self.advance();
14624 concurrently = true;
14625 }
14626 let target = self.take_optional_maintain_name();
14627 self.consume_until_statement_boundary();
14628 Ok(Statement::Maintain {
14629 kind,
14630 concurrently,
14631 target,
14632 })
14633 }
14634
14635 /// v7.39 (round 535) — `CLUSTER [VERBOSE] [<table> [USING <index>]]`
14636 /// and `CLUSTER [VERBOSE] <index> ON <table>`.
14637 #[inline(never)]
14638 fn parse_cluster_tail(&mut self) -> Result<Statement, ParseError> {
14639 use crate::ast::MaintainKind;
14640 self.skip_paren_option_list();
14641 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("verbose")) {
14642 self.advance();
14643 }
14644 let target = self.take_optional_maintain_name();
14645 self.consume_until_statement_boundary();
14646 Ok(Statement::Maintain {
14647 kind: if target.is_some() {
14648 MaintainKind::ClusterRelation
14649 } else {
14650 MaintainKind::Whole
14651 },
14652 // CLUSTER has no CONCURRENTLY form, and PG runs it inside a
14653 // transaction block quite happily (measured).
14654 concurrently: false,
14655 target,
14656 })
14657 }
14658
14659 /// The next token as a relation / schema name, when there is one.
14660 fn take_optional_maintain_name(&mut self) -> Option<alloc::string::String> {
14661 match self.peek() {
14662 Token::Ident(_) | Token::QuotedIdent(_) => match self.advance() {
14663 Token::Ident(n) | Token::QuotedIdent(n) => Some(n),
14664 _ => None,
14665 },
14666 _ => None,
14667 }
14668 }
14669
14670 /// A parenthesised option list, absorbed.
14671 fn skip_paren_option_list(&mut self) {
14672 if !matches!(self.peek(), Token::LParen) {
14673 return;
14674 }
14675 let mut depth = 0usize;
14676 loop {
14677 match self.advance() {
14678 Token::LParen => depth += 1,
14679 Token::RParen => {
14680 depth -= 1;
14681 if depth == 0 {
14682 return;
14683 }
14684 }
14685 Token::Eof => return,
14686 _ => {}
14687 }
14688 }
14689 }
14690
14691 /// v7.39 (round 531) — the `LIKE` clause inside a CREATE TABLE
14692 /// column list.
14693 ///
14694 /// PG's option names are COMMENTS / COMPRESSION / CONSTRAINTS /
14695 /// DEFAULTS / GENERATED / IDENTITY / INDEXES / STATISTICS / STORAGE
14696 /// / ALL. The three that describe physical storage have no meaning
14697 /// here, so they parse and change nothing rather than making a
14698 /// dump that mentions them fail to load.
14699 ///
14700 /// `#[inline(never)]` because the CREATE TABLE frame sits on the
14701 /// parse chain the nesting sentinel is tuned against.
14702 #[inline(never)]
14703 fn parse_create_table_like(&mut self, at: usize) -> Result<crate::ast::LikeSpec, ParseError> {
14704 self.advance(); // LIKE
14705 let source = self.expect_ident_like()?;
14706 let mut options = crate::ast::LikeOptions::default();
14707 loop {
14708 let including = match self.peek() {
14709 Token::Ident(s) if s.eq_ignore_ascii_case("including") => true,
14710 Token::Ident(s) if s.eq_ignore_ascii_case("excluding") => false,
14711 _ => break,
14712 };
14713 self.advance();
14714 // `ALL` lexes as its own keyword, not an identifier.
14715 let opt = if matches!(self.peek(), Token::All) {
14716 self.advance();
14717 alloc::string::String::from("all")
14718 } else {
14719 self.expect_ident_like()?
14720 };
14721 let set = |o: &mut crate::ast::LikeOptions, on: bool| {
14722 o.defaults = on;
14723 o.constraints = on;
14724 o.identity = on;
14725 o.generated = on;
14726 o.indexes = on;
14727 o.comments = on;
14728 };
14729 match opt.to_ascii_lowercase().as_str() {
14730 "all" => set(&mut options, including),
14731 "defaults" => options.defaults = including,
14732 "constraints" => options.constraints = including,
14733 "identity" => options.identity = including,
14734 "generated" => options.generated = including,
14735 "indexes" => options.indexes = including,
14736 "comments" => options.comments = including,
14737 // No storage model to copy into.
14738 "storage" | "statistics" | "compression" => {}
14739 other => {
14740 return Err(self.err(alloc::format!("unrecognized LIKE option {other:?}")));
14741 }
14742 }
14743 }
14744 Ok(crate::ast::LikeSpec {
14745 source,
14746 at,
14747 options,
14748 })
14749 }
14750
14751 fn parse_create_table_stmt_after_create(&mut self) -> Result<Statement, ParseError> {
14752 // Caller already consumed CREATE; we're sitting on TABLE.
14753 debug_assert!(matches!(self.peek(), Token::Table));
14754 self.advance();
14755 let if_not_exists = self.consume_if_not_exists();
14756 let name = self.expect_ident_like()?;
14757 // v7.37.6-B — `CREATE TABLE c PARTITION OF parent <bounds>`
14758 // child shape has no column list; the child inherits its
14759 // columns from the parent at engine-DDL time. Detect it
14760 // before the `(` requirement below.
14761 if matches!(self.peek(), Token::Partition)
14762 && Self::tokens_match_ident_ci(self.tokens.get(self.pos + 1), "of")
14763 {
14764 self.advance(); // PARTITION
14765 self.advance(); // of
14766 let partition_of = self.parse_partition_of_tail()?;
14767 return Ok(Statement::CreateTable(CreateTableStatement {
14768 temporary: false,
14769 name,
14770 engine: None,
14771 columns: Vec::new(),
14772 like_specs: Vec::new(),
14773 inherits: Vec::new(),
14774 if_not_exists,
14775 foreign_keys: Vec::new(),
14776 table_constraints: Vec::new(),
14777 partition_by: None,
14778 partition_of: Some(partition_of),
14779 }));
14780 }
14781 // v7.38 (read01 P6.49) — CTAS: `CREATE TABLE name AS <select>`. Reuses
14782 // the materialized-view materialisation path (run the SELECT, infer the
14783 // column types, create + populate the table) but marks the node so the
14784 // executor creates a plain table without a mat-view registry entry.
14785 if matches!(self.peek(), Token::As) {
14786 self.advance();
14787 let body_stmt = self.parse_select_stmt()?;
14788 let Statement::Select(body) = body_stmt else {
14789 return Err(self.err(format!(
14790 "CREATE TABLE {name:?} AS body must be a SELECT, got {body_stmt:?}"
14791 )));
14792 };
14793 let with_data = self.parse_optional_with_data(true)?;
14794 return Ok(Statement::CreateMaterializedView(
14795 crate::ast::CreateMaterializedViewStatement {
14796 temporary: false,
14797 name,
14798 if_not_exists,
14799 columns: Vec::new(),
14800 body,
14801 with_data,
14802 as_plain_table: true,
14803 },
14804 ));
14805 }
14806 if !matches!(self.peek(), Token::LParen) {
14807 return Err(self.err(format!(
14808 "expected '(' after table name, got {:?}",
14809 self.peek()
14810 )));
14811 }
14812 self.advance();
14813 let mut columns = Vec::new();
14814 let mut foreign_keys: Vec<ForeignKeyConstraint> = Vec::new();
14815 let mut table_constraints: Vec<crate::ast::TableConstraint> = Vec::new();
14816 let mut like_specs: Vec<crate::ast::LikeSpec> = Vec::new();
14817 loop {
14818 // v7.39 (round 621) — `CREATE TABLE c () INHERITS (p)`, the empty
14819 // column list. It is how a child that adds nothing of its own is
14820 // written, and this loop demanded at least one entry: `syntax
14821 // error at or near ")"`. The child takes the parent's columns,
14822 // which the INHERITS clause already arranges.
14823 if columns.is_empty() && matches!(self.peek(), Token::RParen) {
14824 self.advance();
14825 break;
14826 }
14827 // v7.6.0 / v7.9.18 — distinguish table-level constraint
14828 // clauses from column definitions. Constraints start
14829 // with `CONSTRAINT <name> …`, `FOREIGN KEY (…)`,
14830 // `PRIMARY KEY (…)`, or `UNIQUE (…)`. Anything else is
14831 // a column.
14832 if self.peek_table_level_pk_start() {
14833 table_constraints.push(self.parse_table_level_primary_key()?);
14834 } else if matches!(self.peek(), Token::Like) {
14835 // v7.39 (round 531) — `LIKE <table> [ {INCLUDING|EXCLUDING}
14836 // <opt> ]*`. The source table's shape lives in the catalog,
14837 // so this records the clause and the engine expands it.
14838 like_specs.push(self.parse_create_table_like(columns.len())?);
14839 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("exclude")) {
14840 // v7.39 (round 210) — bare `EXCLUDE [USING m] (col WITH op, …)`.
14841 table_constraints.push(self.parse_table_level_exclude()?);
14842 } else if self.peek_table_level_unique_start() {
14843 table_constraints.push(self.parse_table_level_unique()?);
14844 } else if self.peek_table_level_check_start() {
14845 // v7.13.0 — table-level CHECK (mailrs round-5 G3).
14846 table_constraints.push(self.parse_table_level_check()?);
14847 } else if self.peek_mysql_inline_key_start() {
14848 // v7.14.0 — mysqldump emits inline `KEY name (cols)`,
14849 // `INDEX name (cols)`, `UNIQUE KEY name (cols)`,
14850 // `FULLTEXT KEY name (cols)`, `SPATIAL KEY name (cols)`
14851 // inside the column list. Skip name + paren list;
14852 // for UNIQUE KEY, register as a UC.
14853 if let Some(uc) = self.parse_mysql_inline_key()? {
14854 table_constraints.push(uc);
14855 }
14856 } else if let Some(kind) = self.peek_named_table_constraint_kind() {
14857 // v7.22 (mailrs round-13 gap 5) — `CONSTRAINT <name>
14858 // { CHECK | UNIQUE | PRIMARY KEY }`: every pg_dump'd
14859 // CHECK is named, and the named-CONSTRAINT arm used
14860 // to accept FOREIGN KEY only. The name is accepted
14861 // and discarded — same handling as every other SPG
14862 // constraint name.
14863 self.advance(); // CONSTRAINT
14864 // v7.39 (read01 round 48) — the name is kept now: the schema
14865 // stores it, so DROP / RENAME CONSTRAINT can find it.
14866 let con_name = self.expect_ident_like()?;
14867 let mut tc = match kind {
14868 NamedTableConstraintKind::Check => self.parse_table_level_check()?,
14869 NamedTableConstraintKind::Unique => self.parse_table_level_unique()?,
14870 NamedTableConstraintKind::PrimaryKey => self.parse_table_level_primary_key()?,
14871 NamedTableConstraintKind::Exclude => self.parse_table_level_exclude()?,
14872 };
14873 match &mut tc {
14874 crate::ast::TableConstraint::Check { name, .. }
14875 | crate::ast::TableConstraint::Unique { name, .. }
14876 | crate::ast::TableConstraint::PrimaryKey { name, .. }
14877 | crate::ast::TableConstraint::Exclude { name, .. } => {
14878 *name = Some(con_name);
14879 }
14880 _ => {}
14881 }
14882 table_constraints.push(tc);
14883 } else if self.peek_constraint_or_fk_start() {
14884 foreign_keys.push(self.parse_table_level_fk()?);
14885 } else {
14886 let (col, col_level_fk) = self.parse_column_def_with_fk()?;
14887 // v7.13.0 — fold inline UNIQUE / CHECK column
14888 // constraints into table-level entries so the
14889 // engine path stays uniform.
14890 if col.is_unique {
14891 table_constraints.push(crate::ast::TableConstraint::Unique {
14892 name: None,
14893 columns: alloc::vec![col.name.clone()],
14894 nulls_not_distinct: col.unique_nulls_not_distinct,
14895 deferrable: col.constraint_deferrable,
14896 initially_deferred: col.constraint_initially_deferred,
14897 });
14898 }
14899 if let Some(check_expr) = col.check.clone() {
14900 table_constraints.push(crate::ast::TableConstraint::Check {
14901 name: None,
14902 expr: check_expr,
14903 not_valid: false,
14904 });
14905 }
14906 columns.push(col);
14907 if let Some(fk) = col_level_fk {
14908 foreign_keys.push(fk);
14909 }
14910 }
14911 match self.peek() {
14912 Token::Comma => {
14913 self.advance();
14914 }
14915 Token::RParen => {
14916 self.advance();
14917 break;
14918 }
14919 other => {
14920 return Err(
14921 self.err(format!("expected ',' or ')' in column list, got {other:?}"))
14922 );
14923 }
14924 }
14925 }
14926 // v7.39 (round 531) — a `LIKE` clause brings its own columns, so
14927 // `CREATE TABLE k (LIKE t)` is a complete definition even though
14928 // nothing is written between the parentheses.
14929 // v7.39 (round 621) — a table with NO columns is legal: PG creates it
14930 // and `INSERT … DEFAULT VALUES` puts a row in it. This refused, so the
14931 // empty parentheses were a parse error in their own right — quite apart
14932 // from `CREATE TABLE c () INHERITS (p)`, which needs table inheritance
14933 // SPG does not have (filed separately).
14934 let _ = &like_specs;
14935 // v7.39 (round 645) — `INHERITS (p1, p2)`, PG table inheritance.
14936 // It sits between the column list and the MySQL table options,
14937 // and it was a syntax error until this round.
14938 let mut inherits: Vec<String> = Vec::new();
14939 if matches!(self.peek(), Token::Ident(k) | Token::QuotedIdent(k)
14940 if k.eq_ignore_ascii_case("inherits"))
14941 {
14942 self.advance();
14943 if !matches!(self.peek(), Token::LParen) {
14944 return Err(self.err(alloc::format!(
14945 "expected ( after INHERITS, got {:?}",
14946 self.peek()
14947 )));
14948 }
14949 self.advance();
14950 loop {
14951 inherits.push(self.expect_ident_like()?);
14952 if matches!(self.peek(), Token::Comma) {
14953 self.advance();
14954 continue;
14955 }
14956 break;
14957 }
14958 if !matches!(self.peek(), Token::RParen) {
14959 return Err(self.err(alloc::format!(
14960 "expected ) closing INHERITS, got {:?}",
14961 self.peek()
14962 )));
14963 }
14964 self.advance();
14965 }
14966 // v7.14.0 — consume MySQL/MariaDB table options after the
14967 // closing `)`. mysqldump emits things like
14968 // `ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_unicode_ci
14969 // AUTO_INCREMENT=42 ROW_FORMAT=DYNAMIC COMMENT='blog posts'`.
14970 // SPG accepts all forms as no-ops (each option is
14971 // `<ident> [=] <ident-or-string>` separated by whitespace).
14972 let engine = self.consume_mysql_table_options();
14973 // v7.38 (read01 P6.55) — PG storage parameters `WITH (opt=val, …)`.
14974 // SPG has no per-table reloptions, so accept and ignore them so a
14975 // pg_dump `CREATE TABLE … WITH (fillfactor=70, …)` restores cleanly.
14976 self.consume_with_reloptions();
14977 // v7.37.6-B — declarative-partition-parent suffix
14978 // (`PARTITION BY RANGE (key_col)`) sits after the column
14979 // list + MySQL table-options. v7.37.6-B only accepts RANGE
14980 // and locks the key column at one ident; the engine then
14981 // verifies the column type is TIMESTAMPTZ.
14982 let partition_by = if matches!(self.peek(), Token::Partition) {
14983 self.advance(); // PARTITION
14984 if !self.peek_is_by() {
14985 return Err(self.err(format!(
14986 "expected BY after PARTITION, got {:?}",
14987 self.peek()
14988 )));
14989 }
14990 self.advance();
14991 Some(self.parse_partition_by_tail()?)
14992 } else {
14993 None
14994 };
14995 Ok(Statement::CreateTable(CreateTableStatement {
14996 temporary: false,
14997 name,
14998 engine,
14999 columns,
15000 like_specs,
15001 inherits,
15002 if_not_exists,
15003 foreign_keys,
15004 table_constraints,
15005 partition_by,
15006 partition_of: None,
15007 }))
15008 }
15009
15010 /// v7.37.6-B — case-insensitive ident match helper for the
15011 /// `PARTITION OF` / `MINVALUE` / `MAXVALUE` keywords. They lex
15012 /// as `Token::Ident("of"/"minvalue"/"maxvalue")` because we
15013 /// didn't burn a global keyword slot for each (see the
15014 /// `Token::Partition` doc-comment in `lexer.rs`).
15015 fn tokens_match_ident_ci(t: Option<&Token>, want: &str) -> bool {
15016 matches!(t, Some(Token::Ident(s) | Token::QuotedIdent(s)) if s.eq_ignore_ascii_case(want))
15017 }
15018
15019 /// v7.37.6-B — after `PARTITION BY`, expect `RANGE (key_col [, ...])`.
15020 /// v7.37.16 (16.1/16.2) — extended to LIST + HASH.
15021 fn parse_partition_by_tail(&mut self) -> Result<crate::ast::PartitionBySpec, ParseError> {
15022 use crate::ast::{PartitionBySpec, PartitionKindAst};
15023 let kind = match self.peek() {
15024 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("range") => {
15025 self.advance();
15026 PartitionKindAst::Range
15027 }
15028 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("list") => {
15029 self.advance();
15030 PartitionKindAst::List
15031 }
15032 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("hash") => {
15033 self.advance();
15034 PartitionKindAst::Hash
15035 }
15036 other => {
15037 return Err(self.err(format!(
15038 "PARTITION BY: expected RANGE / LIST / HASH, got {other:?}"
15039 )));
15040 }
15041 };
15042 if !matches!(self.peek(), Token::LParen) {
15043 return Err(self.err(format!(
15044 "expected '(' after PARTITION BY <strategy>, got {:?}",
15045 self.peek()
15046 )));
15047 }
15048 self.advance();
15049 let mut key_columns = Vec::new();
15050 loop {
15051 key_columns.push(self.expect_ident_like()?);
15052 match self.peek() {
15053 Token::Comma => {
15054 self.advance();
15055 }
15056 Token::RParen => {
15057 self.advance();
15058 break;
15059 }
15060 other => {
15061 return Err(self.err(format!(
15062 "expected ',' or ')' in PARTITION BY key list, got {other:?}"
15063 )));
15064 }
15065 }
15066 }
15067 if key_columns.is_empty() {
15068 return Err(self.err("PARTITION BY requires at least one key column".to_string()));
15069 }
15070 Ok(PartitionBySpec { kind, key_columns })
15071 }
15072
15073 /// v7.37.6-B — after `PARTITION OF`, expect
15074 /// <parent> FOR VALUES FROM ( <expr> ) TO ( <expr> )
15075 /// or
15076 /// <parent> DEFAULT
15077 fn parse_partition_of_tail(&mut self) -> Result<crate::ast::PartitionOfSpec, ParseError> {
15078 use crate::ast::{PartitionOfBoundsAst, PartitionOfSpec};
15079 let parent_name = self.expect_ident_like()?;
15080 // v7.37.6-B rejects an explicit column list — the child
15081 // inherits from the parent. mailrs round-7 taught us that
15082 // CREATE TABLE-side schema reconciliation hides drift, so
15083 // we surface this as a parse error rather than silently
15084 // ignoring user columns.
15085 if matches!(self.peek(), Token::LParen) {
15086 return Err(self.err(
15087 "CREATE TABLE … PARTITION OF parent: explicit column list not supported \
15088 at v7.37.6-B; the child inherits its columns from the parent"
15089 .to_string(),
15090 ));
15091 }
15092 let bounds = match self.peek() {
15093 Token::Default => {
15094 self.advance();
15095 PartitionOfBoundsAst::Default
15096 }
15097 Token::For => {
15098 self.advance();
15099 if !matches!(self.peek(), Token::Values) {
15100 return Err(
15101 self.err(format!("expected VALUES after FOR, got {:?}", self.peek()))
15102 );
15103 }
15104 self.advance();
15105 // WITH is not a reserved Token in the lexer — it lexes
15106 // as Token::Ident("with"). Disambiguate manually.
15107 let want_with = matches!(
15108 self.peek(),
15109 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("with")
15110 );
15111 if want_with {
15112 self.advance();
15113 if !matches!(self.peek(), Token::LParen) {
15114 return Err(self.err(format!(
15115 "expected '(' after FOR VALUES WITH, got {:?}",
15116 self.peek()
15117 )));
15118 }
15119 self.advance();
15120 let (mut modulus, mut remainder): (Option<u32>, Option<u32>) = (None, None);
15121 loop {
15122 let key = self.expect_ident_like()?;
15123 let n = match self.peek().clone() {
15124 Token::Integer(v) if u32::try_from(v).is_ok() => {
15125 self.advance();
15126 v as u32
15127 }
15128 other => {
15129 return Err(self.err(format!(
15130 "FOR VALUES WITH: expected unsigned integer literal, got {other:?}"
15131 )));
15132 }
15133 };
15134 match key.to_ascii_uppercase().as_str() {
15135 "MODULUS" => modulus = Some(n),
15136 "REMAINDER" => remainder = Some(n),
15137 other => {
15138 return Err(self.err(format!(
15139 "FOR VALUES WITH: unknown key {other:?}; \
15140 expected MODULUS or REMAINDER"
15141 )));
15142 }
15143 }
15144 match self.peek() {
15145 Token::Comma => {
15146 self.advance();
15147 }
15148 Token::RParen => {
15149 self.advance();
15150 break;
15151 }
15152 other => {
15153 return Err(self.err(format!(
15154 "expected ',' or ')' in FOR VALUES WITH list, got {other:?}"
15155 )));
15156 }
15157 }
15158 }
15159 let modulus = modulus
15160 .ok_or_else(|| self.err("FOR VALUES WITH: missing MODULUS".to_string()))?;
15161 let remainder = remainder.ok_or_else(|| {
15162 self.err("FOR VALUES WITH: missing REMAINDER".to_string())
15163 })?;
15164 if modulus == 0 {
15165 return Err(self.err("FOR VALUES WITH: MODULUS must be > 0".to_string()));
15166 }
15167 if remainder >= modulus {
15168 return Err(self.err(format!(
15169 "FOR VALUES WITH: REMAINDER ({remainder}) \
15170 must be < MODULUS ({modulus})"
15171 )));
15172 }
15173 PartitionOfBoundsAst::Hash { modulus, remainder }
15174 } else {
15175 match self.peek() {
15176 Token::From => {
15177 self.advance();
15178 let lower = Box::new(self.parse_partition_bound_expr()?);
15179 if !matches!(self.peek(), Token::To) {
15180 return Err(self.err(format!(
15181 "expected TO after FROM (...), got {:?}",
15182 self.peek()
15183 )));
15184 }
15185 self.advance();
15186 let upper = Box::new(self.parse_partition_bound_expr()?);
15187 PartitionOfBoundsAst::Range { lower, upper }
15188 }
15189 // v7.37.16 (16.1) — FOR VALUES IN (lit [, lit, …])
15190 Token::In => {
15191 self.advance();
15192 if !matches!(self.peek(), Token::LParen) {
15193 return Err(self.err(format!(
15194 "expected '(' after FOR VALUES IN, got {:?}",
15195 self.peek()
15196 )));
15197 }
15198 self.advance();
15199 let mut values = Vec::new();
15200 loop {
15201 values.push(self.parse_expr(0)?);
15202 match self.peek() {
15203 Token::Comma => {
15204 self.advance();
15205 }
15206 Token::RParen => {
15207 self.advance();
15208 break;
15209 }
15210 other => {
15211 return Err(self.err(format!(
15212 "expected ',' or ')' in FOR VALUES IN list, got {other:?}"
15213 )));
15214 }
15215 }
15216 }
15217 if values.is_empty() {
15218 return Err(self.err(
15219 "FOR VALUES IN requires at least one literal".to_string(),
15220 ));
15221 }
15222 PartitionOfBoundsAst::List { values }
15223 }
15224 other => {
15225 return Err(self.err(format!(
15226 "expected FROM / IN / WITH after FOR VALUES, got {other:?}"
15227 )));
15228 }
15229 }
15230 }
15231 }
15232 other => {
15233 return Err(self.err(format!(
15234 "expected FOR VALUES or DEFAULT after PARTITION OF parent, got {other:?}"
15235 )));
15236 }
15237 };
15238 Ok(PartitionOfSpec {
15239 parent_name,
15240 bounds,
15241 })
15242 }
15243
15244 /// v7.37.6-B — a single `( <expr> )` bound. `MINVALUE` /
15245 /// `MAXVALUE` lex as Ident; rewrite them into FunctionCall
15246 /// markers (no-arg builtins) so the engine resolves them
15247 /// against [`spg_storage::PartitionBound::{MinValue, MaxValue}`].
15248 fn parse_partition_bound_expr(&mut self) -> Result<crate::ast::Expr, ParseError> {
15249 if !matches!(self.peek(), Token::LParen) {
15250 return Err(self.err(format!(
15251 "expected '(' before partition bound, got {:?}",
15252 self.peek()
15253 )));
15254 }
15255 self.advance();
15256 let expr = match self.peek() {
15257 Token::Ident(s) | Token::QuotedIdent(s)
15258 if s.eq_ignore_ascii_case("minvalue") || s.eq_ignore_ascii_case("maxvalue") =>
15259 {
15260 let name = s.to_ascii_uppercase();
15261 self.advance();
15262 crate::ast::Expr::FunctionCall {
15263 name,
15264 args: Vec::new(),
15265 }
15266 }
15267 _ => self.parse_expr(0)?,
15268 };
15269 if !matches!(self.peek(), Token::RParen) {
15270 return Err(self.err(format!(
15271 "expected ')' after partition bound, got {:?}",
15272 self.peek()
15273 )));
15274 }
15275 self.advance();
15276 Ok(expr)
15277 }
15278
15279 /// v7.14.0 — true when the next tokens look like an inline
15280 /// MySQL index declaration: KEY / INDEX / UNIQUE KEY /
15281 /// UNIQUE INDEX / FULLTEXT [KEY|INDEX] / SPATIAL [KEY|INDEX]
15282 /// — each followed by an optional name + `(...)`. Critical:
15283 /// a column NAMED `key` / `index` (PG accepts as ident) must
15284 /// NOT be mistaken for the KEY constraint shape. We disambig
15285 /// by requiring the keyword to be followed by either `(` or
15286 /// `<ident> (`.
15287 fn peek_mysql_inline_key_start(&self) -> bool {
15288 let cur = self.peek();
15289 // Shapes:
15290 // KEY (cols)
15291 // KEY name (cols)
15292 // INDEX (cols)
15293 // INDEX name (cols)
15294 // UNIQUE KEY [name] (cols)
15295 // UNIQUE INDEX [name] (cols)
15296 // FULLTEXT [KEY|INDEX] [name] (cols)
15297 // SPATIAL [KEY|INDEX] [name] (cols)
15298 let after_keyword_followed_by_paren_or_ident_paren = |skip: usize| -> bool {
15299 // tokens at skip = the position AFTER the index-form
15300 // keywords (KEY/INDEX) have been consumed.
15301 match self.tokens.get(skip) {
15302 Some(Token::LParen) => true,
15303 Some(Token::Ident(_) | Token::QuotedIdent(_)) => {
15304 matches!(self.tokens.get(skip + 1), Some(Token::LParen))
15305 }
15306 _ => false,
15307 }
15308 };
15309 // `INDEX` lexes as Token::Index (reserved), not as
15310 // Token::Ident("index"). Both shapes count as a KEY/INDEX
15311 // start; the peek helper below handles either.
15312 let is_key_or_index_tok = |t: &Token| -> bool {
15313 matches!(t, Token::Index)
15314 || matches!(t, Token::Ident(s) if s.eq_ignore_ascii_case("key") || s.eq_ignore_ascii_case("index"))
15315 };
15316 match cur {
15317 Token::Index => after_keyword_followed_by_paren_or_ident_paren(self.pos + 1),
15318 Token::Ident(s) if s.eq_ignore_ascii_case("key") || s.eq_ignore_ascii_case("index") => {
15319 after_keyword_followed_by_paren_or_ident_paren(self.pos + 1)
15320 }
15321 Token::Ident(s)
15322 if s.eq_ignore_ascii_case("fulltext") || s.eq_ignore_ascii_case("spatial") =>
15323 {
15324 let nxt = self.tokens.get(self.pos + 1);
15325 let after_after = if nxt.is_some_and(is_key_or_index_tok) {
15326 self.pos + 2
15327 } else {
15328 self.pos + 1
15329 };
15330 after_keyword_followed_by_paren_or_ident_paren(after_after)
15331 }
15332 Token::Ident(s) if s.eq_ignore_ascii_case("unique") => {
15333 let nxt = self.tokens.get(self.pos + 1);
15334 if !nxt.is_some_and(is_key_or_index_tok) {
15335 return false;
15336 }
15337 after_keyword_followed_by_paren_or_ident_paren(self.pos + 2)
15338 }
15339 _ => false,
15340 }
15341 }
15342
15343 /// v7.14.0 — parse the MySQL inline KEY/INDEX form. Returns
15344 /// Some(TableConstraint::Unique) for UNIQUE KEY (so SPG
15345 /// enforces uniqueness on INSERT). v7.15.0: plain KEY/INDEX
15346 /// returns Some(TableConstraint::Index) so the engine builds
15347 /// a real BTree index on the leading column (mysqldump
15348 /// `KEY idx_posts_author (author_id)` shape).
15349 /// FULLTEXT / SPATIAL still return None — accepted-as-no-op
15350 /// (the storage layer has no matching AM).
15351 fn parse_mysql_inline_key(
15352 &mut self,
15353 ) -> Result<Option<crate::ast::TableConstraint>, ParseError> {
15354 // Detect UNIQUE prefix.
15355 let is_unique = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("unique"))
15356 {
15357 self.advance();
15358 true
15359 } else {
15360 false
15361 };
15362 // Consume FULLTEXT / SPATIAL prefix and record which one
15363 // it was. v7.17.0 Phase 2.2 — FULLTEXT routes through a
15364 // dedicated TableConstraint variant so the engine can
15365 // build a tsvector-GIN; SPATIAL still has no matching
15366 // AM, so it falls back to accept-as-no-op.
15367 let mut is_fulltext = false;
15368 let mut is_spatial = false;
15369 if let Token::Ident(s) = self.peek().clone() {
15370 if s.eq_ignore_ascii_case("fulltext") {
15371 self.advance();
15372 is_fulltext = true;
15373 } else if s.eq_ignore_ascii_case("spatial") {
15374 self.advance();
15375 is_spatial = true;
15376 }
15377 }
15378 // KEY / INDEX keyword. `INDEX` lexes as Token::Index
15379 // (reserved); accept either token shape.
15380 match self.peek() {
15381 Token::Index => {
15382 self.advance();
15383 }
15384 Token::Ident(s) if s.eq_ignore_ascii_case("key") || s.eq_ignore_ascii_case("index") => {
15385 self.advance();
15386 }
15387 other => {
15388 return Err(self.err(alloc::format!(
15389 "expected KEY/INDEX in inline index declaration, got {other:?}"
15390 )));
15391 }
15392 }
15393 // Optional index name (an ident before the `(`).
15394 // v7.15.0 — capture the name when present so the engine
15395 // builds the secondary index under the user's chosen
15396 // name (matches mysqldump's `KEY idx_x (col)` shape).
15397 let mut idx_name: Option<String> = None;
15398 if matches!(self.peek(), Token::Ident(_) | Token::QuotedIdent(_))
15399 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
15400 {
15401 if let Token::Ident(s) | Token::QuotedIdent(s) = self.advance() {
15402 idx_name = Some(s);
15403 }
15404 }
15405 // Optional `USING BTREE` / `USING HASH` (MySQL).
15406 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
15407 self.advance();
15408 if matches!(self.peek(), Token::Ident(_) | Token::QuotedIdent(_)) {
15409 self.advance();
15410 }
15411 }
15412 // Required column list `(col [, col]*)`.
15413 if !matches!(self.peek(), Token::LParen) {
15414 return Err(self.err(alloc::format!(
15415 "expected '(' in inline KEY/INDEX, got {:?}",
15416 self.peek()
15417 )));
15418 }
15419 self.advance();
15420 let mut cols: Vec<String> = Vec::new();
15421 while let Token::Ident(s) | Token::QuotedIdent(s) = self.peek().clone() {
15422 self.advance();
15423 cols.push(s);
15424 // Skip optional `(length)` per-column prefix.
15425 if matches!(self.peek(), Token::LParen) {
15426 let mut depth = 1usize;
15427 self.advance();
15428 while depth > 0 {
15429 match self.peek() {
15430 Token::LParen => depth += 1,
15431 Token::RParen => depth -= 1,
15432 Token::Eof => break,
15433 _ => {}
15434 }
15435 self.advance();
15436 }
15437 }
15438 // Skip optional ASC / DESC.
15439 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("asc") || s.eq_ignore_ascii_case("desc"))
15440 || matches!(self.peek(), Token::Asc | Token::Desc)
15441 {
15442 self.advance();
15443 }
15444 if matches!(self.peek(), Token::Comma) {
15445 self.advance();
15446 continue;
15447 }
15448 break;
15449 }
15450 if matches!(self.peek(), Token::RParen) {
15451 self.advance();
15452 }
15453 // Trailing options on the inline index — comment / etc.
15454 // Skip until comma or `)`.
15455 while !matches!(self.peek(), Token::Comma | Token::RParen | Token::Eof) {
15456 self.advance();
15457 }
15458 if cols.is_empty() {
15459 return Ok(None);
15460 }
15461 if is_unique {
15462 // Carry the captured idx_name on UNIQUE too so future
15463 // engine work can name the underlying BTree
15464 // accordingly; today the unique-constraint installer
15465 // synthesises the name itself, but Display round-trip
15466 // benefits from preserving it.
15467 Ok(Some(crate::ast::TableConstraint::Unique {
15468 name: idx_name,
15469 columns: cols,
15470 nulls_not_distinct: false,
15471 // MySQL inline UNIQUE KEY has no deferral vocabulary.
15472 deferrable: false,
15473 initially_deferred: false,
15474 }))
15475 } else if is_fulltext {
15476 // v7.17.0 Phase 2.2 — MySQL `FULLTEXT KEY` now
15477 // routes through `TableConstraint::FulltextIndex`;
15478 // the engine builds a tsvector-GIN over each named
15479 // column so MATCH AGAINST gets a real inverted
15480 // index instead of a silently-dropped declaration.
15481 Ok(Some(crate::ast::TableConstraint::FulltextIndex {
15482 name: idx_name,
15483 columns: cols,
15484 }))
15485 } else if is_spatial {
15486 // SPG has no native SPATIAL AM. Accept-as-no-op
15487 // (declaration is parsed, but no index is built).
15488 Ok(None)
15489 } else {
15490 // v7.15.0 — plain KEY / INDEX builds a real BTree
15491 // secondary index.
15492 Ok(Some(crate::ast::TableConstraint::Index {
15493 name: idx_name,
15494 columns: cols,
15495 }))
15496 }
15497 }
15498
15499 /// v7.14.0 — consume MySQL/MariaDB table-options tail after
15500 /// the closing `)`: ENGINE=..., DEFAULT CHARSET=...,
15501 /// COLLATE=..., AUTO_INCREMENT=N, ROW_FORMAT=..., COMMENT='...'
15502 /// (in any order, separated by whitespace).
15503 /// v7.38 (read01 P6.55) — consume and discard a PG `WITH (opt=val, …)`
15504 /// storage-parameter clause on CREATE TABLE. SPG has no per-table
15505 /// reloptions; accepting them keeps pg_dump restores working. `WITH` is a
15506 /// bare ident here, and only the parenthesised form is reloptions (so this
15507 /// never eats a `WITH DATA` / `WITH CHECK OPTION` trailer).
15508 fn consume_with_reloptions(&mut self) {
15509 let is_with = matches!(
15510 self.peek(),
15511 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("with")
15512 );
15513 if !is_with || !matches!(self.tokens.get(self.pos + 1), Some(Token::LParen)) {
15514 return;
15515 }
15516 self.advance(); // WITH
15517 self.advance(); // (
15518 let mut depth = 1u32;
15519 while depth > 0 && !matches!(self.peek(), Token::Eof) {
15520 match self.peek() {
15521 Token::LParen => depth += 1,
15522 Token::RParen => depth -= 1,
15523 _ => {}
15524 }
15525 self.advance();
15526 }
15527 }
15528
15529 /// v7.39 — returns the `ENGINE=` name, which used to be consumed and
15530 /// dropped with everything else here. The rest of the MySQL table
15531 /// options genuinely have no meaning for SPG's storage; the engine
15532 /// name does, because MySQL REFUSES one it does not know and a dump
15533 /// with a typo in it should not quietly become a table.
15534 fn consume_mysql_table_options(&mut self) -> Option<alloc::string::String> {
15535 let mut engine: Option<alloc::string::String> = None;
15536 loop {
15537 // Heuristic: a table option is an ident (or `DEFAULT`
15538 // reserved keyword) followed by `=` and an
15539 // ident / string / integer.
15540 let name_lc = match self.peek().clone() {
15541 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
15542 Token::Default => alloc::string::String::from("default"),
15543 _ => break,
15544 };
15545 let known = matches!(
15546 name_lc.as_str(),
15547 "engine"
15548 | "default"
15549 | "charset"
15550 | "collate"
15551 | "auto_increment"
15552 | "row_format"
15553 | "comment"
15554 | "pack_keys"
15555 | "stats_persistent"
15556 | "stats_auto_recalc"
15557 | "stats_sample_pages"
15558 | "key_block_size"
15559 | "tablespace"
15560 | "min_rows"
15561 | "max_rows"
15562 | "checksum"
15563 | "delay_key_write"
15564 | "insert_method"
15565 | "data"
15566 | "index"
15567 | "encryption"
15568 | "compression"
15569 );
15570 if !known {
15571 break;
15572 }
15573 self.advance(); // option name
15574 // `DEFAULT` optional prefix is followed by `CHARSET` /
15575 // `COLLATE`; consume the next ident too.
15576 if name_lc == "default" {
15577 if matches!(self.peek(), Token::Ident(_) | Token::QuotedIdent(_)) {
15578 self.advance();
15579 }
15580 }
15581 if matches!(self.peek(), Token::Eq) {
15582 self.advance();
15583 }
15584 match self.peek().clone() {
15585 Token::Ident(v) | Token::QuotedIdent(v) | Token::String(v) => {
15586 if name_lc == "engine" {
15587 // v7.39.3 — as WRITTEN. MySQL 9.7.2 refuses an
15588 // engine it does not know and names it back
15589 // exactly: `Unknown storage engine 'NoSuchEng'`,
15590 // measured. The lexer folds a bare identifier, so
15591 // the message quoted a name the dump did not
15592 // contain, which is the one thing that message is
15593 // for. Guarded the same way the column spelling
15594 // is: the span runs to the next token, so what
15595 // comes back has to be the same word.
15596 let written = self
15597 .source_span(self.pos, self.pos)
15598 .map(|raw| raw.trim().trim_matches('`').trim_matches('\''))
15599 .filter(|raw| raw.eq_ignore_ascii_case(&v))
15600 .map(alloc::string::String::from);
15601 engine = Some(written.unwrap_or(v));
15602 }
15603 self.advance();
15604 }
15605 Token::Integer(_) => {
15606 self.advance();
15607 }
15608 _ => {}
15609 }
15610 }
15611 engine
15612 }
15613
15614 /// v7.9.18 — true when the next tokens are `PRIMARY KEY (…)`.
15615 /// PRIMARY and KEY are bare idents; we look-ahead 2 to be
15616 /// sure (otherwise a column literally named `primary` would
15617 /// be mistaken).
15618 fn peek_table_level_pk_start(&self) -> bool {
15619 let cur = self.peek();
15620 let nxt = self.tokens.get(self.pos + 1);
15621 let nxt2 = self.tokens.get(self.pos + 2);
15622 let is_primary = matches!(cur, Token::Ident(s) if s.eq_ignore_ascii_case("primary"));
15623 let is_key = matches!(nxt, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("key"));
15624 let is_lparen = matches!(nxt2, Some(Token::LParen));
15625 is_primary && is_key && is_lparen
15626 }
15627
15628 /// v7.9.18 — true when the next tokens are `UNIQUE (…)`.
15629 /// v7.13.0 — also matches `UNIQUE NULLS [NOT] DISTINCT (…)`
15630 /// (mailrs round-5 G10).
15631 fn peek_table_level_unique_start(&self) -> bool {
15632 let cur = self.peek();
15633 let is_unique = matches!(cur, Token::Ident(s) if s.eq_ignore_ascii_case("unique"));
15634 if !is_unique {
15635 return false;
15636 }
15637 let n1 = self.tokens.get(self.pos + 1);
15638 // Plain `UNIQUE (…)`.
15639 if matches!(n1, Some(Token::LParen)) {
15640 return true;
15641 }
15642 // `UNIQUE NULLS [NOT] DISTINCT (…)`.
15643 let is_nulls = matches!(n1, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("nulls"));
15644 if !is_nulls {
15645 return false;
15646 }
15647 let n2 = self.tokens.get(self.pos + 2);
15648 let n3 = self.tokens.get(self.pos + 3);
15649 let n4 = self.tokens.get(self.pos + 4);
15650 // `UNIQUE NULLS DISTINCT (…)` — 4 tokens before `(`.
15651 if matches!(n2, Some(Token::Distinct)) && matches!(n3, Some(Token::LParen)) {
15652 return true;
15653 }
15654 // `UNIQUE NULLS NOT DISTINCT (…)` — 5 tokens before `(`.
15655 if matches!(n2, Some(Token::Not))
15656 && matches!(n3, Some(Token::Distinct))
15657 && matches!(n4, Some(Token::LParen))
15658 {
15659 return true;
15660 }
15661 false
15662 }
15663
15664 fn parse_table_level_primary_key(&mut self) -> Result<crate::ast::TableConstraint, ParseError> {
15665 self.advance(); // PRIMARY
15666 self.advance(); // KEY
15667 let columns = self.parse_paren_ident_list("PRIMARY KEY")?;
15668 // v7.39 (round 711) — the trailer's values are CARRIED now; round
15669 // 621 consumed and dropped them (the storing half of F08).
15670 let (deferrable, initially_deferred) = self.consume_deferrable_clauses_timed()?;
15671 Ok(crate::ast::TableConstraint::PrimaryKey {
15672 name: None,
15673 columns,
15674 deferrable,
15675 initially_deferred,
15676 })
15677 }
15678
15679 fn parse_table_level_unique(&mut self) -> Result<crate::ast::TableConstraint, ParseError> {
15680 self.advance(); // UNIQUE
15681 // v7.13.0 — optional `NULLS NOT DISTINCT` modifier
15682 // (mailrs round-5 G10, PG 15+ surface). Default behaviour
15683 // is `NULLS DISTINCT` per the SQL standard.
15684 let mut nulls_not_distinct = false;
15685 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("nulls")) {
15686 let n1 = self.tokens.get(self.pos + 1);
15687 let n2 = self.tokens.get(self.pos + 2);
15688 let is_not = matches!(n1, Some(Token::Not));
15689 let is_distinct = matches!(n2, Some(Token::Distinct));
15690 if is_not && is_distinct {
15691 self.advance(); // NULLS
15692 self.advance(); // NOT
15693 self.advance(); // DISTINCT
15694 nulls_not_distinct = true;
15695 } else if matches!(n1, Some(Token::Distinct)) {
15696 self.advance(); // NULLS
15697 self.advance(); // DISTINCT
15698 }
15699 }
15700 let columns = self.parse_paren_ident_list("UNIQUE")?;
15701 let (deferrable, initially_deferred) = self.consume_deferrable_clauses_timed()?;
15702 Ok(crate::ast::TableConstraint::Unique {
15703 name: None,
15704 columns,
15705 nulls_not_distinct,
15706 deferrable,
15707 initially_deferred,
15708 })
15709 }
15710
15711 /// v7.13.0 — table-level `CHECK (<expr>)` constraint
15712 /// (mailrs round-5 G3). Consumes `CHECK` then a parenthesised
15713 /// expression.
15714 /// v7.39 (round 210) — `EXCLUDE [USING <method>] ( <col> WITH <op>
15715 /// [, <col> WITH <op>]* ) [WHERE (...)]`. The operator is read as a
15716 /// standalone token spelling (`&&`, `=`, `@>`, `<@`, `&<`, `&>`).
15717 /// v7.39 (round 652) — the optional `NOT VALID` suffix on a
15718 /// constraint added by ALTER TABLE. `NOT` alone is not enough to
15719 /// commit: `NOT` starts no other suffix here, but reading both
15720 /// tokens before advancing keeps the caller's error message intact
15721 /// if someone writes `NOT NULL` by mistake.
15722 fn parse_not_valid_suffix(&mut self) -> bool {
15723 if !matches!(self.peek(), Token::Not) {
15724 return false;
15725 }
15726 if !matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("valid"))
15727 {
15728 return false;
15729 }
15730 self.advance();
15731 self.advance();
15732 true
15733 }
15734
15735 fn parse_table_level_exclude(&mut self) -> Result<crate::ast::TableConstraint, ParseError> {
15736 self.advance(); // EXCLUDE
15737 // Optional `USING <method>`.
15738 let mut method = None;
15739 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
15740 self.advance();
15741 method = Some(match self.advance() {
15742 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
15743 other => {
15744 return Err(self.err(alloc::format!(
15745 "expected index method after USING, got {other:?}"
15746 )));
15747 }
15748 });
15749 }
15750 if !matches!(self.peek(), Token::LParen) {
15751 return Err(self.err(alloc::format!(
15752 "expected '(' after EXCLUDE, got {:?}",
15753 self.peek()
15754 )));
15755 }
15756 self.advance();
15757 let mut elements: Vec<(String, String)> = Vec::new();
15758 loop {
15759 let col = match self.advance() {
15760 Token::Ident(s) | Token::QuotedIdent(s) => s,
15761 other => {
15762 return Err(self.err(alloc::format!(
15763 "expected column name in EXCLUDE, got {other:?}"
15764 )));
15765 }
15766 };
15767 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with")) {
15768 return Err(self.err(alloc::format!(
15769 "expected WITH after EXCLUDE column, got {:?}",
15770 self.peek()
15771 )));
15772 }
15773 self.advance();
15774 let op = match self.advance() {
15775 Token::InetOverlap => String::from("&&"),
15776 Token::Intersects => String::from("?#"),
15777 Token::IsBelow => String::from("<^"),
15778 Token::IsAbove => String::from(">^"),
15779 Token::PatternLt => String::from("~<~"),
15780 Token::PatternLtEq => String::from("~<=~"),
15781 Token::PatternGt => String::from("~>~"),
15782 Token::PatternGtEq => String::from("~>=~"),
15783 Token::TsMatchOld => String::from("@@@"),
15784 Token::Eq => String::from("="),
15785 Token::JsonContains => String::from("@>"),
15786 Token::JsonContainedBy => String::from("<@"),
15787 Token::OverLeft => String::from("&<"),
15788 Token::OverRight => String::from("&>"),
15789 other => {
15790 return Err(self.err(alloc::format!(
15791 "unsupported EXCLUDE operator {other:?} (SPG supports &&, =, @>, <@, &<, &>)"
15792 )));
15793 }
15794 };
15795 elements.push((col, op));
15796 if matches!(self.peek(), Token::Comma) {
15797 self.advance();
15798 continue;
15799 }
15800 break;
15801 }
15802 if !matches!(self.peek(), Token::RParen) {
15803 return Err(self.err(alloc::format!(
15804 "expected ')' to close EXCLUDE, got {:?}",
15805 self.peek()
15806 )));
15807 }
15808 self.advance();
15809 Ok(crate::ast::TableConstraint::Exclude {
15810 name: None,
15811 method,
15812 elements,
15813 })
15814 }
15815
15816 fn parse_table_level_check(&mut self) -> Result<crate::ast::TableConstraint, ParseError> {
15817 self.advance(); // CHECK
15818 if !matches!(self.peek(), Token::LParen) {
15819 return Err(self.err(alloc::format!(
15820 "expected '(' after CHECK, got {:?}",
15821 self.peek()
15822 )));
15823 }
15824 self.advance();
15825 let expr = self.parse_expr(0)?;
15826 if !matches!(self.peek(), Token::RParen) {
15827 return Err(self.err(alloc::format!(
15828 "expected ')' to close CHECK predicate, got {:?}",
15829 self.peek()
15830 )));
15831 }
15832 self.advance();
15833 // A CHECK written inside CREATE TABLE cannot be NOT VALID: there
15834 // are no existing rows for PG to skip, so it rejects the suffix.
15835 Ok(crate::ast::TableConstraint::Check {
15836 name: None,
15837 expr,
15838 not_valid: false,
15839 })
15840 }
15841
15842 /// v7.13.0 — `true` when the next token is `CHECK` (a bare ident).
15843 fn peek_table_level_check_start(&self) -> bool {
15844 matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("check"))
15845 }
15846
15847 /// v7.22 (round-13 gap 5) — `Some(kind)` when the next tokens are
15848 /// `CONSTRAINT <name> { CHECK | UNIQUE | PRIMARY }`. FOREIGN stays
15849 /// on the dedicated FK path (`parse_table_level_fk` consumes its
15850 /// own CONSTRAINT prefix).
15851 fn peek_named_table_constraint_kind(&self) -> Option<NamedTableConstraintKind> {
15852 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint")) {
15853 return None;
15854 }
15855 // tokens[pos+1] is the constraint name (any ident-like);
15856 // tokens[pos+2] is the kind keyword.
15857 match self.tokens.get(self.pos + 2) {
15858 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("check") => {
15859 Some(NamedTableConstraintKind::Check)
15860 }
15861 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("unique") => {
15862 Some(NamedTableConstraintKind::Unique)
15863 }
15864 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("primary") => {
15865 Some(NamedTableConstraintKind::PrimaryKey)
15866 }
15867 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("exclude") => {
15868 Some(NamedTableConstraintKind::Exclude)
15869 }
15870 _ => None,
15871 }
15872 }
15873
15874 fn parse_paren_ident_list(&mut self, ctx: &str) -> Result<Vec<String>, ParseError> {
15875 if !matches!(self.peek(), Token::LParen) {
15876 return Err(self.err(alloc::format!(
15877 "expected '(' after {ctx}, got {:?}",
15878 self.peek()
15879 )));
15880 }
15881 self.advance();
15882 let mut out = Vec::new();
15883 loop {
15884 out.push(self.expect_ident_like()?);
15885 match self.peek() {
15886 Token::Comma => {
15887 self.advance();
15888 }
15889 Token::RParen => {
15890 self.advance();
15891 break;
15892 }
15893 other => {
15894 return Err(self.err(alloc::format!(
15895 "expected ',' or ')' in {ctx} list, got {other:?}"
15896 )));
15897 }
15898 }
15899 }
15900 if out.is_empty() {
15901 return Err(self.err(alloc::format!("{ctx} requires at least one column")));
15902 }
15903 Ok(out)
15904 }
15905
15906 /// v7.6.0 — true when the next tokens are `CONSTRAINT <name>
15907 /// FOREIGN KEY` or bare `FOREIGN KEY`. Both introduce a
15908 /// table-level FK; a column def never starts with either keyword
15909 /// (column names are not in this reserved set).
15910 fn peek_constraint_or_fk_start(&self) -> bool {
15911 let is_constraint_kw = matches!(
15912 self.peek(),
15913 Token::Ident(s) if s.eq_ignore_ascii_case("constraint")
15914 );
15915 let is_foreign_kw = matches!(
15916 self.peek(),
15917 Token::Ident(s) if s.eq_ignore_ascii_case("foreign")
15918 );
15919 is_constraint_kw || is_foreign_kw
15920 }
15921
15922 /// v7.6.0 — parse a table-level FK clause:
15923 /// `[CONSTRAINT <name>] FOREIGN KEY (<col>[,<col>]*) REFERENCES
15924 /// <tbl> [(<pcol>[,<pcol>]*)] [ON DELETE <action>] [ON UPDATE <action>]`.
15925 fn parse_table_level_fk(&mut self) -> Result<ForeignKeyConstraint, ParseError> {
15926 let mut name: Option<String> = None;
15927 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint")) {
15928 self.advance();
15929 name = Some(self.expect_ident_like()?);
15930 }
15931 // `FOREIGN`
15932 match self.advance() {
15933 Token::Ident(s) if s.eq_ignore_ascii_case("foreign") => {}
15934 other => return Err(self.err(format!("expected FOREIGN, got {other:?}"))),
15935 }
15936 // `KEY`
15937 match self.advance() {
15938 Token::Ident(s) if s.eq_ignore_ascii_case("key") => {}
15939 other => return Err(self.err(format!("expected KEY after FOREIGN, got {other:?}"))),
15940 }
15941 // `(col, col, ...)`
15942 if !matches!(self.peek(), Token::LParen) {
15943 return Err(self.err(format!(
15944 "expected '(' after FOREIGN KEY, got {:?}",
15945 self.peek()
15946 )));
15947 }
15948 self.advance();
15949 let mut columns = Vec::new();
15950 loop {
15951 columns.push(self.expect_ident_like()?);
15952 match self.peek() {
15953 Token::Comma => {
15954 self.advance();
15955 }
15956 Token::RParen => {
15957 self.advance();
15958 break;
15959 }
15960 other => {
15961 return Err(self.err(format!(
15962 "expected ',' or ')' in FK column list, got {other:?}"
15963 )));
15964 }
15965 }
15966 }
15967 if columns.is_empty() {
15968 return Err(self.err("FOREIGN KEY requires at least one column".into()));
15969 }
15970 let (
15971 parent_table,
15972 parent_columns,
15973 on_delete,
15974 on_update,
15975 match_type,
15976 deferrable,
15977 initially_deferred,
15978 ) = self.parse_references_tail(columns.len())?;
15979 Ok(ForeignKeyConstraint {
15980 name,
15981 columns,
15982 parent_table,
15983 parent_columns,
15984 on_delete,
15985 on_update,
15986 match_type,
15987 deferrable,
15988 initially_deferred,
15989 })
15990 }
15991
15992 /// v7.6.0 — parse the tail `REFERENCES <tbl> [(<pcol>...)] [ON
15993 /// DELETE <action>] [ON UPDATE <action>]`. `expected_arity` is
15994 /// the local column count, used to default the parent column
15995 /// list when omitted (SQL spec: parent's PK is implied).
15996 fn parse_references_tail(
15997 &mut self,
15998 expected_arity: usize,
15999 ) -> Result<
16000 (
16001 String,
16002 Vec<String>,
16003 FkAction,
16004 FkAction,
16005 crate::ast::MatchType,
16006 // v7.39 (round 288) — deferrable, initially_deferred.
16007 bool,
16008 bool,
16009 ),
16010 ParseError,
16011 > {
16012 match self.advance() {
16013 Token::Ident(s) if s.eq_ignore_ascii_case("references") => {}
16014 other => return Err(self.err(format!("expected REFERENCES, got {other:?}"))),
16015 }
16016 let parent_table = self.expect_ident_like()?;
16017 let mut parent_columns: Vec<String> = Vec::new();
16018 if matches!(self.peek(), Token::LParen) {
16019 self.advance();
16020 loop {
16021 parent_columns.push(self.expect_ident_like()?);
16022 match self.peek() {
16023 Token::Comma => {
16024 self.advance();
16025 }
16026 Token::RParen => {
16027 self.advance();
16028 break;
16029 }
16030 other => {
16031 return Err(self.err(format!(
16032 "expected ',' or ')' in REFERENCES column list, got {other:?}"
16033 )));
16034 }
16035 }
16036 }
16037 }
16038 if !parent_columns.is_empty() && parent_columns.len() != expected_arity {
16039 return Err(self.err(format!(
16040 "FK arity mismatch: {} local column(s) vs {} parent column(s)",
16041 expected_arity,
16042 parent_columns.len()
16043 )));
16044 }
16045 // Optional `MATCH {SIMPLE | FULL | PARTIAL}`. PG's grammar puts
16046 // it between the referenced column list and the ON / DEFERRABLE
16047 // trailers. SPG implements MATCH SIMPLE semantics (the FK check
16048 // is skipped when any referencing column is NULL), so SIMPLE —
16049 // the default, and the only spelling pg_dump emits — is accepted
16050 // as a no-op. MATCH FULL / MATCH PARTIAL need the per-FK
16051 // mixed-NULL rule, which is not wired yet; reject them honestly
16052 // rather than silently applying SIMPLE (PG itself errors on
16053 // MATCH PARTIAL as "not yet implemented").
16054 let mut match_type = crate::ast::MatchType::Simple;
16055 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("match")) {
16056 self.advance();
16057 // `FULL` is a reserved keyword token (FULL OUTER JOIN);
16058 // SIMPLE / PARTIAL arrive as bare identifiers.
16059 let kind = match self.advance() {
16060 Token::Full => "FULL".to_string(),
16061 Token::Ident(s) => s.to_uppercase(),
16062 other => {
16063 return Err(self.err(format!(
16064 "expected FULL, PARTIAL or SIMPLE after MATCH, got {other:?}"
16065 )));
16066 }
16067 };
16068 match kind.as_str() {
16069 "SIMPLE" => {} // Default — match_type stays Simple.
16070 // v7.38 (read01, T29) — MATCH FULL: the check is skipped only
16071 // when ALL referencing columns are NULL; a mixed-NULL key errors.
16072 "FULL" => match_type = crate::ast::MatchType::Full,
16073 "PARTIAL" => {
16074 return Err(self.err("MATCH PARTIAL not yet implemented".to_string()));
16075 }
16076 _ => {
16077 return Err(self.err(format!(
16078 "expected FULL, PARTIAL or SIMPLE after MATCH, got {kind}"
16079 )));
16080 }
16081 }
16082 }
16083 // v7.6.7 / v7.17.0 Phase 3.1 — interleave `[NOT] DEFERRABLE
16084 // [INITIALLY {DEFERRED | IMMEDIATE}]` and `ON DELETE
16085 // <action>` / `ON UPDATE <action>` in either order. PG /
16086 // pg_dump emits the timing clause AFTER the ON clauses
16087 // (`ON DELETE CASCADE DEFERRABLE INITIALLY DEFERRED`),
16088 // but the SQL spec allows either order. We loop over
16089 // every possible trailer and dispatch on the next token,
16090 // stopping when nothing matches. Phase 3.1 changes the
16091 // bare DEFERRABLE form from hard-error to accept-as-
16092 // immediate; SPG is single-writer with no deferred-
16093 // constraint window so the runtime semantics are always
16094 // immediate even when INITIALLY DEFERRED is requested.
16095 // PG's default referential action (no ON DELETE / ON UPDATE
16096 // clause) is NO ACTION, not RESTRICT — the two enforce
16097 // identically in SPG (single-writer, no deferred window; see the
16098 // shared match arm in constraints.rs) but information_schema.
16099 // referential_constraints must report NO ACTION to match PG.
16100 let mut on_delete = FkAction::NoAction;
16101 let mut on_update = FkAction::NoAction;
16102 let mut seen_on_delete = false;
16103 let mut seen_on_update = false;
16104 let mut deferrable = false;
16105 let mut initially_deferred = false;
16106 loop {
16107 // DEFERRABLE / NOT DEFERRABLE / INITIALLY shapes.
16108 let before = self.pos;
16109 let (d, idef) = self.consume_deferrable_clauses_timed()?;
16110 if self.pos != before {
16111 deferrable = d;
16112 initially_deferred = idef;
16113 continue;
16114 }
16115 // ON DELETE / ON UPDATE.
16116 if !matches!(self.peek(), Token::On) {
16117 break;
16118 }
16119 self.advance();
16120 let which = self.advance();
16121 let action = self.parse_fk_action()?;
16122 match which {
16123 Token::Ident(ref s) if s.eq_ignore_ascii_case("delete") => {
16124 if seen_on_delete {
16125 return Err(self.err("ON DELETE specified twice".into()));
16126 }
16127 seen_on_delete = true;
16128 on_delete = action;
16129 }
16130 Token::Ident(ref s) if s.eq_ignore_ascii_case("update") => {
16131 if seen_on_update {
16132 return Err(self.err("ON UPDATE specified twice".into()));
16133 }
16134 seen_on_update = true;
16135 on_update = action;
16136 }
16137 other => {
16138 return Err(
16139 self.err(format!("expected DELETE or UPDATE after ON, got {other:?}"))
16140 );
16141 }
16142 }
16143 }
16144 Ok((
16145 parent_table,
16146 parent_columns,
16147 on_delete,
16148 on_update,
16149 match_type,
16150 deferrable,
16151 initially_deferred,
16152 ))
16153 }
16154
16155 /// v7.6.0 — parse `CASCADE | RESTRICT | SET NULL | SET DEFAULT |
16156 /// NO ACTION`.
16157 fn parse_fk_action(&mut self) -> Result<FkAction, ParseError> {
16158 match self.advance() {
16159 Token::Ident(s) if s.eq_ignore_ascii_case("cascade") => Ok(FkAction::Cascade),
16160 Token::Ident(s) if s.eq_ignore_ascii_case("restrict") => Ok(FkAction::Restrict),
16161 Token::Ident(s) if s.eq_ignore_ascii_case("set") => match self.advance() {
16162 Token::Null => Ok(FkAction::SetNull),
16163 Token::Default => Ok(FkAction::SetDefault),
16164 other => Err(self.err(format!(
16165 "expected NULL or DEFAULT after SET in FK action, got {other:?}"
16166 ))),
16167 },
16168 Token::Ident(s) if s.eq_ignore_ascii_case("no") => match self.advance() {
16169 Token::Ident(s) if s.eq_ignore_ascii_case("action") => Ok(FkAction::NoAction),
16170 other => Err(self.err(format!(
16171 "expected ACTION after NO in FK action, got {other:?}"
16172 ))),
16173 },
16174 other => Err(self.err(format!(
16175 "expected CASCADE | RESTRICT | SET NULL | SET DEFAULT | NO ACTION, got {other:?}"
16176 ))),
16177 }
16178 }
16179
16180 /// Recognise the optional `IF NOT EXISTS` prefix shared by `CREATE
16181 /// TABLE` and `CREATE INDEX`. Returns `true` if consumed.
16182 fn consume_if_not_exists(&mut self) -> bool {
16183 // `IF` arrives as a bare Ident (we don't reserve it because it
16184 // also appears mid-expression in PG, though we don't support
16185 // those forms yet).
16186 let looks_like_if = matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if"));
16187 if !looks_like_if {
16188 return false;
16189 }
16190 // Peek one ahead before committing: only consume IF when it's
16191 // actually `IF NOT EXISTS`.
16192 if !matches!(self.tokens.get(self.pos + 1), Some(Token::Not)) {
16193 return false;
16194 }
16195 if !matches!(
16196 self.tokens.get(self.pos + 2),
16197 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("exists")
16198 ) {
16199 return false;
16200 }
16201 self.advance(); // IF
16202 self.advance(); // NOT
16203 self.advance(); // EXISTS
16204 true
16205 }
16206
16207 /// v7.12.4 — `IF EXISTS` modifier for DROP statements.
16208 /// Consumes IF EXISTS as a pair; returns false otherwise
16209 /// without consuming any tokens.
16210 /// v7.39 (RLS) — consume the `ROW LEVEL SECURITY` keyword triple after
16211 /// ENABLE/DISABLE/FORCE/NO FORCE.
16212 fn expect_row_level_security(&mut self) -> Result<(), ParseError> {
16213 for kw in ["row", "level", "security"] {
16214 if !matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case(kw))
16215 {
16216 return Err(self.err(alloc::format!(
16217 "expected {} in ROW LEVEL SECURITY, got {:?}",
16218 kw.to_ascii_uppercase(),
16219 self.peek()
16220 )));
16221 }
16222 self.advance();
16223 }
16224 Ok(())
16225 }
16226
16227 fn consume_if_exists(&mut self) -> bool {
16228 let looks_like_if = matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if"));
16229 if !looks_like_if {
16230 return false;
16231 }
16232 if !matches!(
16233 self.tokens.get(self.pos + 1),
16234 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("exists")
16235 ) {
16236 return false;
16237 }
16238 self.advance(); // IF
16239 self.advance(); // EXISTS
16240 true
16241 }
16242
16243 /// v7.9.14 — consume `ASC | DESC | NULLS FIRST | NULLS LAST`
16244 /// qualifiers after an index column ref. ASC / DESC are
16245 /// reserved tokens; NULLS / FIRST / LAST are bare idents.
16246 /// We accept and discard them since single-column BTree
16247 /// stores rows in natural key order today.
16248 /// v7.24 (round-16 A) — `NULLS FIRST` / `NULLS LAST` after an
16249 /// ORDER BY key. Returns None when absent.
16250 fn parse_optional_nulls_placement(&mut self) -> Result<Option<bool>, ParseError> {
16251 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("nulls")) {
16252 return Ok(None);
16253 }
16254 self.advance();
16255 match self.advance() {
16256 Token::Ident(s) if s.eq_ignore_ascii_case("first") => Ok(Some(true)),
16257 Token::Ident(s) if s.eq_ignore_ascii_case("last") => Ok(Some(false)),
16258 other => Err(self.err(alloc::format!(
16259 "expected FIRST or LAST after NULLS, got {other:?}"
16260 ))),
16261 }
16262 }
16263
16264 /// v7.39 (round 537) — the per-column ordering clause, REPORTED now
16265 /// rather than discarded.
16266 ///
16267 /// SPG's index does not scan in a direction — column ordering is
16268 /// intrinsic to the storage — but `pg_indexes.indexdef` is a
16269 /// reproduction of the DDL, and dropping the clause meant
16270 /// `CREATE INDEX i ON t (a DESC NULLS LAST)` read back as `(a)`. A
16271 /// dump lost it, and a schema diff saw drift on every run.
16272 fn consume_optional_index_column_qualifiers(&mut self) -> crate::ast::IndexColumnOrder {
16273 let mut order = crate::ast::IndexColumnOrder::default();
16274 loop {
16275 match self.peek() {
16276 Token::Asc => {
16277 self.advance();
16278 }
16279 Token::Desc => {
16280 order.descending = true;
16281 self.advance();
16282 }
16283 Token::Ident(s) if s.eq_ignore_ascii_case("nulls") => {
16284 let look = self.tokens.get(self.pos + 1);
16285 if matches!(
16286 look,
16287 Some(Token::Ident(k)) if k.eq_ignore_ascii_case("first")
16288 || k.eq_ignore_ascii_case("last")
16289 ) {
16290 self.advance();
16291 order.nulls_first = Some(matches!(
16292 self.advance(),
16293 Token::Ident(k) if k.eq_ignore_ascii_case("first")
16294 ));
16295 } else {
16296 break;
16297 }
16298 }
16299 _ => break,
16300 }
16301 }
16302 order
16303 }
16304
16305 fn parse_create_index_stmt_after_create(
16306 &mut self,
16307 is_unique: bool,
16308 ) -> Result<Statement, ParseError> {
16309 // Caller consumed CREATE (and the optional UNIQUE); we're on INDEX.
16310 debug_assert!(matches!(self.peek(), Token::Index));
16311 self.advance();
16312 // v7.37.17 (17.6 partial) — CONCURRENTLY noise word (PG 8.2+).
16313 // SPG's CREATE INDEX is synchronous end-to-end today (real
16314 // CONCURRENTLY variant with restartable scans queues with
16315 // v7.39 indexes epic), so the modifier has no runtime effect
16316 // — same accept-and-no-op shape as v7.37.16.5 DETACH
16317 // PARTITION CONCURRENTLY and v7.37.19.8 REFRESH MATERIALIZED
16318 // VIEW CONCURRENTLY.
16319 let mut concurrently = false;
16320 if matches!(
16321 self.peek(),
16322 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("concurrently")
16323 ) {
16324 self.advance();
16325 concurrently = true;
16326 }
16327 let if_not_exists = self.consume_if_not_exists();
16328 // v7.39 (read01 round 93) — the index name is optional (PG since
16329 // forever): `CREATE INDEX ON t (a)` lets the server pick a name.
16330 // When the token after `[IF NOT EXISTS]` is already `ON`, no name
16331 // was given; leave it empty and the engine derives a PG-style
16332 // `<table>_<cols>_idx` name at CREATE time (with collision counter).
16333 let name = if matches!(self.peek(), Token::On) {
16334 String::new()
16335 } else {
16336 self.expect_ident_like()?
16337 };
16338 if !matches!(self.peek(), Token::On) {
16339 return Err(self.err(format!(
16340 "expected ON after CREATE INDEX <name>, got {:?}",
16341 self.peek()
16342 )));
16343 }
16344 self.advance();
16345 let table = self.expect_ident_like()?;
16346 // Optional `USING <method>` — only recognised method in v2.0 is
16347 // `hnsw` (a single-layer NSW graph for kNN). `USING` is the bare
16348 // ident `using` (we don't promote it to a reserved keyword
16349 // because it isn't reserved anywhere else in our SQL surface).
16350 let mut method_name: Option<String> = None;
16351 let method = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
16352 self.advance();
16353 let m = self.expect_ident_like()?;
16354 method_name = Some(m.to_ascii_lowercase());
16355 match m.to_ascii_lowercase().as_str() {
16356 "hnsw" => IndexMethod::Hnsw,
16357 "btree" => IndexMethod::BTree,
16358 "brin" => IndexMethod::Brin,
16359 // v7.12.3 — real GIN inverted index over `tsvector`.
16360 // v7.9.26b's `USING gin` → BTree silent fallback is
16361 // gone; the engine validates that the indexed column
16362 // is `tsvector` at CREATE INDEX time.
16363 "gin" => IndexMethod::Gin,
16364 // v7.9.26b — PG `pg_dump` emits `USING gist` /
16365 // `USING spgist` / `USING hash` for their built-in
16366 // AMs that SPG doesn't have a matching
16367 // implementation for; degrade to BTree on the
16368 // leading column so the schema loads + the index
16369 // catalogue stays consistent. Operator pays the
16370 // planner cost only for the queries that would have
16371 // used the specialised AM.
16372 "gist" | "spgist" | "hash" => IndexMethod::BTree,
16373 // v7.11.3 — pgvector ships both `ivfflat` and
16374 // `hnsw`. Customers shouldn't have to choose
16375 // their on-disk index method based on what SPG
16376 // implements; accept `ivfflat` as a synonym for
16377 // `hnsw` so PG schemas using either method drop
16378 // in. The vector distance op (`<->` / `<#>` /
16379 // `<=>`) at query time still picks the metric.
16380 "ivfflat" => IndexMethod::Hnsw,
16381 other => {
16382 return Err(self.err(alloc::format!(
16383 "unknown index method {other:?}; supported: hnsw, btree, brin, gin (gist/spgist/hash accepted as BTree fallback)"
16384 )));
16385 }
16386 }
16387 } else {
16388 IndexMethod::BTree
16389 };
16390 if !matches!(self.peek(), Token::LParen) {
16391 return Err(self.err(format!(
16392 "expected '(' before indexed column, got {:?}",
16393 self.peek()
16394 )));
16395 }
16396 self.advance();
16397 // v6.8.2 — accept either a bare column ident (legacy) or
16398 // an expression `fn(col, …)` for expression indexes.
16399 // Distinguish by peeking the token *after* the current
16400 // ident: `ident )` is the legacy column-only path;
16401 // anything else triggers the Pratt expression parser.
16402 // (`advance()` uses `mem::replace` to nil out the current
16403 // slot, so we can't save+rewind cleanly — peek-ahead via
16404 // direct index avoids the mutation.)
16405 let mut opclass: Option<String> = None;
16406 let mut key_collation: Option<String> = None;
16407 let (column, expression): (String, Option<Expr>) = match self.peek().clone() {
16408 // Single column with `)` immediately after — fast path.
16409 // v7.9.29 — also: bare column followed by `,` (the
16410 // multi-column form `(a, b, c)`). Without this branch
16411 // the leading ident gets pulled into `parse_expr`
16412 // which then sets `expression = Some(Column(a))` and
16413 // breaks Display round-trip on the multi-column shape.
16414 Token::Ident(s) | Token::QuotedIdent(s)
16415 if matches!(
16416 self.tokens.get(self.pos + 1),
16417 Some(Token::RParen | Token::Comma)
16418 ) =>
16419 {
16420 self.advance();
16421 (s, None)
16422 }
16423 // v7.9.22 — single column followed by a pgvector
16424 // opclass ident: `(col vector_cosine_ops)`. mailrs G5.
16425 // v7.15.0 — capture the opclass instead of discarding
16426 // it so the engine can dispatch (e.g. `gin_trgm_ops`
16427 // → real trigram-shingle GIN over a TEXT column).
16428 // Vector/HNSW opclasses still take their distance
16429 // metric from the query operator (`<->` / `<#>` /
16430 // `<=>`), so for those callers the opclass stays
16431 // informational.
16432 // v7.22 (mailrs round-13 gap 7) — pg_dump qualifies the
16433 // opclass: `(embedding public.vector_cosine_ops)`. Strip
16434 // the schema and dispatch on the bare opclass, the same
16435 // treatment table/type names get.
16436 Token::Ident(s) | Token::QuotedIdent(s)
16437 if matches!(
16438 self.tokens.get(self.pos + 1),
16439 Some(Token::Ident(_) | Token::QuotedIdent(_))
16440 ) && matches!(self.tokens.get(self.pos + 2), Some(Token::Dot))
16441 && matches!(
16442 self.tokens.get(self.pos + 3),
16443 Some(Token::Ident(op) | Token::QuotedIdent(op))
16444 if is_vector_opclass_name(op)
16445 ) =>
16446 {
16447 self.advance(); // column name
16448 self.advance(); // schema qualifier
16449 self.advance(); // dot
16450 let op_tok = self.advance();
16451 if let Token::Ident(op) | Token::QuotedIdent(op) = op_tok {
16452 opclass = Some(op.to_ascii_lowercase());
16453 }
16454 (s, None)
16455 }
16456 // r1038 — an operator class is recognised by its POSITION, not
16457 // by a list of names. It used to be `is_vector_opclass_name`,
16458 // so `USING gin (doc jsonb_path_ops)` — ordinary PG, and what
16459 // sentori's migration wrote — was a syntax error while
16460 // `USING gin (doc)` parsed. Anything sitting between a column
16461 // name and a `,` `)` ASC DESC NULLS COLLATE is an opclass;
16462 // two bare identifiers in a row are not valid there otherwise.
16463 Token::Ident(s) | Token::QuotedIdent(s)
16464 if matches!(
16465 self.tokens.get(self.pos + 1),
16466 Some(Token::Ident(op) | Token::QuotedIdent(op))
16467 if is_vector_opclass_name(op) || Self::opclass_position_follows(
16468 self.tokens.get(self.pos + 2)
16469 )
16470 ) =>
16471 {
16472 self.advance(); // column name
16473 // Capture the opclass token, lower-cased for
16474 // case-insensitive engine dispatch.
16475 let op_tok = self.advance();
16476 if let Token::Ident(op) | Token::QuotedIdent(op) = op_tok {
16477 opclass = Some(op.to_ascii_lowercase());
16478 }
16479 (s, None)
16480 }
16481 Token::Ident(_) | Token::QuotedIdent(_) => {
16482 // v7.39 (round 538) — an explicit COLLATE on the key,
16483 // read by LOOKAHEAD because `parse_expr` absorbs the
16484 // clause as a no-op (SPG orders text by bytes, which is
16485 // the C collation, so it changes nothing to honour). PG
16486 // still PRINTS it: an explicitly written `"C"` and the
16487 // collation a column inherits are different collation
16488 // OBJECTS even where they sort identically, which is why
16489 // `(a COLLATE "C")` shows on a C-collation database too.
16490 if matches!(
16491 self.tokens.get(self.pos + 1),
16492 Some(Token::Ident(w)) if w.eq_ignore_ascii_case("collate")
16493 ) {
16494 key_collation = match self.tokens.get(self.pos + 2) {
16495 Some(Token::Ident(n) | Token::QuotedIdent(n) | Token::String(n)) => {
16496 Some(n.clone())
16497 }
16498 _ => None,
16499 };
16500 }
16501 // v7.39.2 — the clause is read by the LOOKAHEAD above and
16502 // belongs to the KEY, not to the expression. Since
16503 // `COLLATE` became a node, letting `parse_expr` build one
16504 // here put the collation in twice and the key deparsed as
16505 // `(c COLLATE "C" COLLATE "C")`. The ORDER-BY-key channel
16506 // is the same idea and already exists, so this borrows it:
16507 // absorb into the side channel, and the key's own
16508 // lookahead is what carries it.
16509 // v7.39.2 — and the key can only CARRY the byte-order
16510 // spellings. Absorbing into the side channel accepts any
16511 // name, so suppressing the node here without this check
16512 // silently accepted `(name COLLATE "en_US")`, which SPG's
16513 // index cannot honour — a refusal that was doing real
16514 // work, removed by the suppression and put back here.
16515 if let Some(name) = &key_collation {
16516 let lc = name.to_ascii_lowercase();
16517 let byte_order = matches!(
16518 lc.as_str(),
16519 "c" | "posix" | "default" | "ucs_basic" | "pg_c_utf8"
16520 );
16521 let mysql_ok = self.mysql_dialect
16522 && (lc.ends_with("_ci")
16523 || lc.ends_with("_bin")
16524 || lc == "binary"
16525 || matches!(lc.as_str(), "case_insensitive" | "nocase"));
16526 if !byte_order && !mysql_ok {
16527 return Err(self.err(alloc::format!(
16528 "COLLATE {name:?} is not supported in this position: an index \
16529 key carries the byte-order spellings only. Declare it on the \
16530 column (`x text COLLATE {name:?}`) instead"
16531 )));
16532 }
16533 }
16534 let saved_key_ctx = self.in_order_by_key;
16535 self.in_order_by_key = true;
16536 let key_expr = self.parse_expr(0);
16537 self.in_order_by_key = saved_key_ctx;
16538 let key_expr = key_expr?;
16539 let primary = extract_first_column(&key_expr).ok_or_else(|| {
16540 self.err("expression index key must reference at least one column".into())
16541 })?;
16542 (primary, Some(key_expr))
16543 }
16544 // v7.37.43-T4 — parenthesised expression index key
16545 // `CREATE INDEX … ON t ((payload->'bundle'->>'id'))`.
16546 // PG's CREATE INDEX requires the expression to be in
16547 // its own parens to disambiguate function calls from
16548 // column lists, so this `LParen` is the inner open-paren
16549 // of an expression key. parse_expr handles the recursive
16550 // descent and consumes the matching `RParen`.
16551 Token::LParen => {
16552 let key_expr = self.parse_expr(0)?;
16553 let primary = extract_first_column(&key_expr).ok_or_else(|| {
16554 self.err("expression index key must reference at least one column".into())
16555 })?;
16556 (primary, Some(key_expr))
16557 }
16558 other => {
16559 return Err(self.err(format!(
16560 "expected column ident or expression, got {other:?}"
16561 )));
16562 }
16563 };
16564 // v7.9.14 — accept extra comma-separated columns inside
16565 // the index key parens (`CREATE INDEX … (a, b, c)`).
16566 // mailrs F2.
16567 //
16568 // v7.39.11 — each extra column's `ASC` / `DESC` / `NULLS FIRST`
16569 // / `NULLS LAST` is KEPT. It used to be parsed and dropped on
16570 // the floor, so `CREATE INDEX i ON t (a, b DESC)` read back from
16571 // `pg_get_indexdef` as `(a, b)`: a dump lost the clause and a
16572 // schema diff saw drift on every run. Reported by sentori
16573 // against 7.39.10, and the same defect round 537 fixed for the
16574 // LEADING column, in the loop right beside it.
16575 let mut extra_columns: Vec<String> = Vec::new();
16576 let mut extra_orders: Vec<crate::ast::IndexColumnOrder> = Vec::new();
16577 // The leading column may also have ASC/DESC after it — and that
16578 // one is the column SPG indexes, so its clause is kept.
16579 let key_order = self.consume_optional_index_column_qualifiers();
16580 while matches!(self.peek(), Token::Comma) {
16581 self.advance();
16582 let extra = self.expect_ident_like()?;
16583 extra_orders.push(self.consume_optional_index_column_qualifiers());
16584 extra_columns.push(extra);
16585 }
16586 if !matches!(self.peek(), Token::RParen) {
16587 return Err(self.err(format!(
16588 "expected ')' after indexed column / expression, got {:?}",
16589 self.peek()
16590 )));
16591 }
16592 self.advance();
16593 // v6.8.0 — optional `INCLUDE (col1, col2, …)` clause for
16594 // index-only-scan annotation. Bare ident (not a reserved
16595 // keyword) so we test by case-insensitive string match.
16596 let included_columns = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("include"))
16597 {
16598 self.advance();
16599 if !matches!(self.peek(), Token::LParen) {
16600 return Err(self.err(format!("expected '(' after INCLUDE, got {:?}", self.peek())));
16601 }
16602 self.advance();
16603 let mut cols = Vec::new();
16604 loop {
16605 cols.push(self.expect_ident_like()?);
16606 match self.peek() {
16607 Token::Comma => {
16608 self.advance();
16609 }
16610 Token::RParen => {
16611 self.advance();
16612 break;
16613 }
16614 other => {
16615 return Err(self.err(format!(
16616 "expected ',' or ')' in INCLUDE list, got {other:?}"
16617 )));
16618 }
16619 }
16620 }
16621 cols
16622 } else {
16623 Vec::new()
16624 };
16625 // v7.11.3 — accept and discard PG `WITH (k = v, ...)` index
16626 // storage parameters. pgvector emits `WITH (lists = N)` for
16627 // ivfflat and `WITH (m = N, ef_construction = M)` for hnsw;
16628 // SPG's HNSW picks its own parameters today (tunable via
16629 // env vars), so the WITH clause is informational and dropped.
16630 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with")) {
16631 self.advance();
16632 if !matches!(self.peek(), Token::LParen) {
16633 return Err(self.err(format!(
16634 "expected '(' after WITH in CREATE INDEX, got {:?}",
16635 self.peek()
16636 )));
16637 }
16638 self.advance();
16639 loop {
16640 if matches!(self.peek(), Token::RParen) {
16641 self.advance();
16642 break;
16643 }
16644 // Drain `key = value` or bare `key` tokens.
16645 let _ = self.advance(); // key
16646 if matches!(self.peek(), Token::Eq) {
16647 self.advance();
16648 let _ = self.advance(); // value (int / string / ident)
16649 }
16650 match self.peek() {
16651 Token::Comma => {
16652 self.advance();
16653 }
16654 Token::RParen => {
16655 self.advance();
16656 break;
16657 }
16658 other => {
16659 return Err(self.err(format!(
16660 "expected ',' or ')' in WITH (…) clause, got {other:?}"
16661 )));
16662 }
16663 }
16664 }
16665 }
16666 // v7.39 (read01 round 52) — optional `NULLS [NOT] DISTINCT` (PG 15+),
16667 // which sits between the key list and the WHERE clause.
16668 let mut nulls_not_distinct = false;
16669 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("nulls")) {
16670 let n1 = self.tokens.get(self.pos + 1);
16671 let n2 = self.tokens.get(self.pos + 2);
16672 if matches!(n1, Some(Token::Not)) && matches!(n2, Some(Token::Distinct)) {
16673 self.advance(); // NULLS
16674 self.advance(); // NOT
16675 self.advance(); // DISTINCT
16676 nulls_not_distinct = true;
16677 } else if matches!(n1, Some(Token::Distinct)) {
16678 self.advance(); // NULLS
16679 self.advance(); // DISTINCT
16680 }
16681 }
16682 // v6.8.1 — optional `WHERE <expr>` partial-index predicate.
16683 let partial_predicate = if matches!(self.peek(), Token::Where) {
16684 self.advance();
16685 Some(self.parse_expr(0)?)
16686 } else {
16687 None
16688 };
16689 // v7.9.29 — UNIQUE on a vector index (HNSW) makes no
16690 // sense: uniqueness over an ANN structure has no clean
16691 // semantics. Reject early. (BRIN UNIQUE is similarly
16692 // meaningless — block both.)
16693 if is_unique && !matches!(method, IndexMethod::BTree) {
16694 return Err(self.err(alloc::format!(
16695 "UNIQUE is only supported on BTree indexes, got USING {:?}",
16696 method
16697 )));
16698 }
16699 Ok(Statement::CreateIndex(CreateIndexStatement {
16700 concurrently,
16701 name,
16702 key_order,
16703 key_collation,
16704 table,
16705 column,
16706 nulls_not_distinct,
16707 method,
16708 if_not_exists,
16709 included_columns,
16710 partial_predicate,
16711 extra_columns: extra_columns.clone(),
16712 extra_orders: extra_orders.clone(),
16713 expression,
16714 is_unique,
16715 opclass,
16716 method_name,
16717 }))
16718 }
16719
16720 /// v7.6.0 — wraps `parse_column_def` and consumes an optional
16721 /// column-level `REFERENCES ...` clause. The trailing FK is
16722 /// normalised into table-level shape (single-element columns +
16723 /// parent_columns) so the engine sees one uniform constraint list.
16724 fn parse_column_def_with_fk(
16725 &mut self,
16726 ) -> Result<(ColumnDef, Option<ForeignKeyConstraint>), ParseError> {
16727 let col = self.parse_column_def()?;
16728 // v7.39 (round 308, V29) — an explicitly named inline FK:
16729 // `col INT CONSTRAINT fk_a REFERENCES tbl(pcol)`. The column-def
16730 // loop leaves this spelling intact precisely so the name can be
16731 // kept here; PG reports it in violation messages and matches it
16732 // in `SET CONSTRAINTS`.
16733 let declared_name = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint"))
16734 {
16735 self.advance();
16736 Some(self.expect_ident_like()?)
16737 } else {
16738 None
16739 };
16740 // Inline form: `col INT REFERENCES tbl(pcol) [ON DELETE ...] [ON UPDATE ...]`.
16741 let inline_references = matches!(
16742 self.peek(),
16743 Token::Ident(s) if s.eq_ignore_ascii_case("references")
16744 );
16745 if !inline_references {
16746 return Ok((col, None));
16747 }
16748 let (
16749 parent_table,
16750 parent_columns,
16751 on_delete,
16752 on_update,
16753 match_type,
16754 deferrable,
16755 initially_deferred,
16756 ) = self.parse_references_tail(1)?;
16757 let fk = ForeignKeyConstraint {
16758 name: declared_name,
16759 columns: vec![col.name.clone()],
16760 parent_table,
16761 parent_columns,
16762 on_delete,
16763 on_update,
16764 match_type,
16765 deferrable,
16766 initially_deferred,
16767 };
16768 Ok((col, Some(fk)))
16769 }
16770
16771 /// v7.13.0 — parse a column type (consuming the type ident and
16772 /// any trailing parameters / `[]`), without surrounding column
16773 /// constraints. Used by ALTER COLUMN TYPE (mailrs round-5 G8).
16774 /// Returns the resolved `ColumnTypeName` plus implied
16775 /// `(auto_increment, not_null)` flags from PG SERIAL family
16776 /// shorthands — callers that don't expect those (ALTER COLUMN
16777 /// TYPE) can discard them.
16778 fn parse_column_type_name(&mut self) -> Result<ColumnTypeName, ParseError> {
16779 let (ty, _, _, _, _, _, _, _, _, _, _, _, _, _) = self.parse_type_with_implied_flags()?;
16780 Ok(ty)
16781 }
16782
16783 #[allow(clippy::type_complexity)]
16784 fn parse_type_with_implied_flags(
16785 &mut self,
16786 ) -> Result<
16787 (
16788 ColumnTypeName,
16789 bool,
16790 bool,
16791 Option<String>,
16792 Collation,
16793 // v7.39 (round 370, M4 P4a) — was `COLLATE` written explicitly?
16794 bool,
16795 // v7.39 (round 676) — the collation NAME as written, which the
16796 // `Collation` enum above cannot carry.
16797 Option<String>,
16798 bool,
16799 // v7.17.0 Phase 3.P0-36 — MySQL inline ENUM variant
16800 // list captured at type-parse time. None for all
16801 // non-ENUM types.
16802 Option<Vec<String>>,
16803 // v7.17.0 Phase 3.P0-37 — MySQL inline SET variant
16804 // list. Distinct from ENUM (subset semantics).
16805 Option<Vec<String>>,
16806 // v7.39 (round 386, epic P1) — declared TINYINT / MEDIUMINT
16807 // width, lost when the type collapses to SmallInt / Int.
16808 Option<MysqlIntWidth>,
16809 // v7.39 (round 424) — declared fractional-seconds precision of a
16810 // MySQL temporal column (bare spelling = 0). None under PG.
16811 Option<u8>,
16812 // v7.39.2 — written `TIMESTAMP` rather than `DATETIME`. The
16813 // two are different types on MySQL and SPG stores both as
16814 // `Timestamp`, so the spelling has to travel separately or
16815 // a dump silently rewrites the column.
16816 bool,
16817 // v7.39.3 — a MySQL `FLOAT(m,d)` / `DOUBLE(m,d)` pair. Not a
16818 // display hint: it rounds on write.
16819 Option<(u8, u8)>,
16820 ),
16821 ParseError,
16822 > {
16823 let mut ty_ident = match self.advance() {
16824 Token::Ident(s) => s,
16825 // v7.37.5 β-P2 — `INTERVAL` lexes as a reserved keyword
16826 // (Token::Interval) since v7.9.25 to drive the `INTERVAL
16827 // '<span>'` literal grammar. As a column type it lands
16828 // here directly; downstream resolution still uses the
16829 // canonical lowercase string.
16830 Token::Interval => "interval".to_string(),
16831 other => {
16832 return Err(ParseError {
16833 message: format!("expected column type, got {other:?}"),
16834 token_pos: self.consumed_pos(),
16835 });
16836 }
16837 };
16838 // v7.22 (mailrs round-13 gap 4) — schema-qualified type names:
16839 // pg_dump qualifies extension types (`public.vector(1024)`).
16840 // SPG is single-namespace; drop the schema and resolve the
16841 // bare type — same treatment table names already get.
16842 while matches!(self.peek(), Token::Dot) {
16843 self.advance();
16844 ty_ident = self.expect_ident_like()?;
16845 }
16846 let mut implied_auto_increment = false;
16847 let mut implied_not_null = false;
16848 let mut user_type_ref: Option<String> = None;
16849 // v7.17.0 Phase 3.P0-36 — MySQL inline ENUM('a','b','c')
16850 // value list, captured here and bubbled up through the
16851 // ColumnDef so the engine can attach it to the column
16852 // schema (and validate INSERT cells against it).
16853 let mut inline_enum_variants: Option<Vec<String>> = None;
16854 // v7.17.0 Phase 3.P0-37 — MySQL inline SET variant list.
16855 let mut inline_set_variants: Option<Vec<String>> = None;
16856 // v7.39 (round 386, type-fidelity epic P1) — the declared MySQL
16857 // narrow-int width (TINYINT / MEDIUMINT), captured before the type
16858 // collapses to SmallInt / Int. Only under the MySQL dialect.
16859 let mut mysql_int_width: Option<MysqlIntWidth> = None;
16860 // v7.39 (round 424) — the declared fractional-seconds precision of a
16861 // MySQL temporal column. Set by the temporal arms below; stays None
16862 // for PG (whose temporal columns keep full microseconds).
16863 let mut mysql_fsp: Option<u8> = None;
16864 let mut mysql_declared_timestamp = false;
16865 let mut mysql_float_md: Option<(u8, u8)> = None;
16866 let mut ty = match ty_ident.as_str() {
16867 // PG SERIAL family. Implies NOT NULL + AUTO_INCREMENT.
16868 "smallserial" | "serial2" => {
16869 implied_auto_increment = true;
16870 implied_not_null = true;
16871 ColumnTypeName::SmallInt
16872 }
16873 "serial" | "serial4" => {
16874 implied_auto_increment = true;
16875 implied_not_null = true;
16876 ColumnTypeName::Int
16877 }
16878 "bigserial" | "serial8" => {
16879 implied_auto_increment = true;
16880 implied_not_null = true;
16881 ColumnTypeName::BigInt
16882 }
16883 // MySQL flavours we accept by aliasing to the closest SPG
16884 // type. TINYINT covers MySQL's i8 — held inside SMALLINT
16885 // since SPG doesn't have a dedicated i8. MEDIUMINT (MySQL
16886 // 24-bit) → INT. UNSIGNED modifiers are consumed below
16887 // without semantic effect.
16888 // v7.38 (read01 P4.19-sibling) — `int2` / `int4` / `int8` are
16889 // PG's internal type names; pg_dump and hand-written PG schemas
16890 // use them interchangeably with smallint / int / bigint (the cast
16891 // path already accepted them, only the column grammar didn't).
16892 "smallint" | "int2" => {
16893 // v7.14.0 — MySQL display-width on integers
16894 // (`SMALLINT(5)`, `INT(11)`, `BIGINT(20)`). The
16895 // parenthesised number is purely cosmetic — it
16896 // doesn't change storage. Accept + discard.
16897 self.consume_optional_paren_size();
16898 ColumnTypeName::SmallInt
16899 }
16900 // v7.17.0 Phase 4.3 — MySQL `TINYINT(1)` is the
16901 // canonical encoding for BOOLEAN. Every MySQL driver
16902 // (JDBC `tinyInt1isBit=true`, PHP `mysql_field_type`,
16903 // .NET `MySqlConnection`, sqlx) maps it to bit. Pre-
16904 // 4.3 SPG classified TINYINT(1) as SmallInt, which
16905 // gave the customer i16-shaped values where the app
16906 // expected bool — a Tier-A silent type drift on
16907 // mysqldump restores. Now: `TINYINT(1)` → Bool;
16908 // `TINYINT` (no width) and `TINYINT(N)` for N ≠ 1
16909 // stay SmallInt (the legacy width-agnostic path).
16910 "tinyint" => {
16911 let width = self.peek_optional_paren_size_value();
16912 self.consume_optional_paren_size();
16913 if width == Some(1) {
16914 ColumnTypeName::Bool
16915 } else {
16916 // v7.39 (round 386, epic P1) — TINYINT is i8; record the
16917 // lost width so the write path can enforce -128..127.
16918 if self.mysql_dialect {
16919 mysql_int_width = Some(MysqlIntWidth::Tiny);
16920 }
16921 ColumnTypeName::SmallInt
16922 }
16923 }
16924 "mediumint" => {
16925 self.consume_optional_paren_size();
16926 // v7.39 (round 386, epic P1) — MEDIUMINT is 24-bit; record it.
16927 if self.mysql_dialect {
16928 mysql_int_width = Some(MysqlIntWidth::Medium);
16929 }
16930 ColumnTypeName::Int
16931 }
16932 "int" | "integer" | "int4" => {
16933 self.consume_optional_paren_size();
16934 ColumnTypeName::Int
16935 }
16936 "bigint" | "int8" => {
16937 self.consume_optional_paren_size();
16938 ColumnTypeName::BigInt
16939 }
16940 // v7.13.0 — `DOUBLE PRECISION` (PG canonical spelling)
16941 // (mailrs round-5 G6). Consume the optional `PRECISION`
16942 // tail when the type keyword was `double` / `DOUBLE`.
16943 //
16944 // v7.39 (round 269) — REAL is 32-bit, not "the same as our
16945 // FLOAT". `FLOAT(p)` picks the width the way PG does:
16946 // p in 1..=24 is real, 25..=53 is double precision, and
16947 // anything else is an error.
16948 "float" | "double" | "real" => {
16949 if ty_ident.eq_ignore_ascii_case("double")
16950 && matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("precision"))
16951 {
16952 self.advance();
16953 }
16954 if ty_ident.eq_ignore_ascii_case("real") {
16955 // v7.39 (round 274) — the two dialects genuinely
16956 // disagree: PG's REAL is 4-byte, MySQL's REAL is a
16957 // synonym for DOUBLE (8-byte). Round 269 made REAL
16958 // 32-bit globally and thereby narrowed the stored
16959 // precision of every MySQL REAL column.
16960 if self.mysql_dialect {
16961 ColumnTypeName::Float
16962 } else {
16963 ColumnTypeName::Real
16964 }
16965 } else if self.mysql_dialect
16966 && matches!(self.peek(), Token::LParen)
16967 && self.peek_paren_has_comma()
16968 {
16969 // v7.39 (round 360) — MySQL's `FLOAT(m,d)` / `DOUBLE(m,d)`
16970 // display form (`FLOAT(10,2)`), which PG has no
16971 // equivalent of. It was `syntax error at or near ","`,
16972 // so the whole CREATE failed.
16973 //
16974 // v7.39.2 — the guard said `float` while the comment
16975 // said both, so `DOUBLE(10,2)` — which every legacy
16976 // MySQL schema uses for money — still failed the
16977 // whole CREATE with `syntax error at or near "("`.
16978 // Measured on 9.7.2: both forms are accepted, and the
16979 // digits are NOT a display hint, they round on write
16980 // (3.14159265358979 into either stores 3.14). The
16981 // rounding is recorded as a residual; accepting the
16982 // syntax and keeping the width is the half this
16983 // change makes.
16984 // v7.39.3 — keep the pair. The digits are not a
16985 // display hint: MySQL 9.7.2 ROUNDS on write and
16986 // refuses a value wider than `m` (errno 1264), so a
16987 // column declared for money held more precision here
16988 // than its schema said.
16989 let (m, d) = self.parse_optional_numeric_params()?;
16990 mysql_float_md = Some((
16991 u8::try_from(m).unwrap_or(u8::MAX),
16992 u8::try_from(d.max(0)).unwrap_or(u8::MAX),
16993 ));
16994 if ty_ident.eq_ignore_ascii_case("float") {
16995 ColumnTypeName::Real
16996 } else {
16997 ColumnTypeName::Float
16998 }
16999 } else if ty_ident.eq_ignore_ascii_case("float")
17000 && matches!(self.peek(), Token::LParen)
17001 {
17002 // PG words the two bounds differently, and
17003 // parse_paren_size already rejects a zero.
17004 let p = self.parse_paren_size("FLOAT")?;
17005 if p > 53 {
17006 return Err(self.err(String::from(
17007 "precision for type float must be less than 54 bits",
17008 )));
17009 }
17010 if p <= 24 {
17011 ColumnTypeName::Real
17012 } else {
17013 ColumnTypeName::Float
17014 }
17015 } else if ty_ident.eq_ignore_ascii_case("float") && self.mysql_dialect {
17016 // v7.39.2 — MySQL's bare FLOAT is FOUR bytes; PG's is
17017 // eight (it is `float8`'s spelling there). SPG used
17018 // PG's for both, so a MySQL FLOAT column silently
17019 // kept more precision than MySQL does — measured,
17020 // 3.14159265358979 comes back as 3.14159 there and
17021 // came back whole here — and reported itself as
17022 // `double` to every reflection.
17023 //
17024 // This is the mirror of the REAL split above: the
17025 // two dialects disagree about which spelling means
17026 // which width, and one of them was already honoured.
17027 ColumnTypeName::Real
17028 } else {
17029 ColumnTypeName::Float
17030 }
17031 }
17032 // v7.13.0 — `FLOAT8` (PG short form) maps the same as FLOAT.
17033 "float4" => ColumnTypeName::Real,
17034 "float8" => ColumnTypeName::Float,
17035 "text" => ColumnTypeName::Text,
17036 // v7.39 (round 360) — MySQL's sized TEXT and BLOB families.
17037 // `LONGTEXT`, `BLOB` and `VARBINARY` appear in nearly every
17038 // real MySQL schema and NONE of them existed: the CREATE
17039 // failed outright with `type "blob" does not exist`, so the
17040 // table was never made. The sizes differ only in MySQL's
17041 // maximum length, which SPG does not cap, so they collapse
17042 // onto TEXT and BYTEA the way the unsized spellings do.
17043 "tinytext" | "mediumtext" | "longtext" => ColumnTypeName::Text,
17044 "blob" | "tinyblob" | "mediumblob" | "longblob" => ColumnTypeName::Bytes,
17045 // `VARBINARY(n)` / `BINARY(n)` — a length that SPG does not
17046 // enforce, consumed so the declaration parses.
17047 "varbinary" | "binary" => {
17048 self.consume_optional_paren_size();
17049 ColumnTypeName::Bytes
17050 }
17051 "name" => ColumnTypeName::Name,
17052 "xid" => ColumnTypeName::Xid,
17053 "oid" => ColumnTypeName::Oid,
17054 "xid8" => ColumnTypeName::Xid8,
17055 "bool" | "boolean" => ColumnTypeName::Bool,
17056 // v7.39 (round 620) — an UNBOUNDED `varchar` is the same type as
17057 // an unbounded `character varying`, which the arm below has always
17058 // read as text. Only the short spelling demanded a length, so
17059 // `CREATE TABLE t (x VARCHAR)` — as ordinary a line of DDL as
17060 // there is — failed on `VARCHAR type requires (N)` while the long
17061 // spelling of the same thing was accepted. The same asymmetry
17062 // round 613 closed on the CAST side, here on the DDL side.
17063 "varchar" => {
17064 if matches!(self.peek(), Token::LParen) {
17065 ColumnTypeName::Varchar(self.parse_paren_size("VARCHAR")?)
17066 } else {
17067 ColumnTypeName::Text
17068 }
17069 }
17070 // v7.39 (bpchar epic) — bare `char` = char(1), same as bare
17071 // `character` below (SQL standard).
17072 "char" => {
17073 if matches!(self.peek(), Token::LParen) {
17074 ColumnTypeName::Char(self.parse_paren_size("CHAR")?)
17075 } else {
17076 ColumnTypeName::Char(1)
17077 }
17078 }
17079 // pg_dump's canonical spellings: `character varying(n)` = varchar,
17080 // `character(n)` = char, bare `character` = char(1). Unbounded
17081 // `character varying` maps to text.
17082 "character" => {
17083 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("varying")) {
17084 self.advance();
17085 if matches!(self.peek(), Token::LParen) {
17086 ColumnTypeName::Varchar(self.parse_paren_size("VARCHAR")?)
17087 } else {
17088 ColumnTypeName::Text
17089 }
17090 } else if matches!(self.peek(), Token::LParen) {
17091 ColumnTypeName::Char(self.parse_paren_size("CHAR")?)
17092 } else {
17093 ColumnTypeName::Char(1)
17094 }
17095 }
17096 "vector" => {
17097 let dim = self.parse_paren_size("VECTOR")?;
17098 let encoding = self.parse_optional_vector_encoding()?;
17099 ColumnTypeName::Vector { dim, encoding }
17100 }
17101 // v7.39 (round 345, M5) — `DECIMAL` and `DEC` are the SQL
17102 // standard's own spellings of NUMERIC, and PG 18.4 accepts both
17103 // (measured: `DECIMAL(10,2)` and `DEC(5,1)` both report as
17104 // `numeric`). Only `NUMERIC` parsed, so `CREATE TABLE t (a
17105 // DECIMAL(10,2))` — how nearly every money column is written,
17106 // in either dialect — was a syntax error and the table was
17107 // never created. `FIXED` is MySQL's alias alone, so it is
17108 // taken only in that dialect.
17109 "numeric" | "decimal" | "dec" => {
17110 let (precision, scale) = self.parse_optional_numeric_params()?;
17111 ColumnTypeName::Numeric(precision, scale)
17112 }
17113 "fixed" if self.mysql_dialect => {
17114 let (precision, scale) = self.parse_optional_numeric_params()?;
17115 ColumnTypeName::Numeric(precision, scale)
17116 }
17117 "date" => ColumnTypeName::Date,
17118 // MySQL's `DATETIME` is the same domain as standard
17119 // `TIMESTAMP` — accept both spellings.
17120 "timestamp" | "datetime" => {
17121 // pg_dump emits `TIMESTAMP(6) WITH TIME ZONE` — the optional
17122 // fractional-seconds precision comes BEFORE the `WITH/WITHOUT
17123 // TIME ZONE` clause, so consume it first.
17124 // v7.39 (round 424) — under MySQL the precision is SEMANTIC
17125 // (it truncates on write and pads on render), so capture it;
17126 // a bare spelling means precision 0 there. PG stores µs always
17127 // and keeps `None`.
17128 let n = self.take_optional_paren_size();
17129 if self.mysql_dialect {
17130 mysql_fsp = Some(n.unwrap_or(0).min(6));
17131 // v7.39.2 — remember WHICH spelling was written.
17132 // MySQL and MariaDB keep `timestamp` and `datetime`
17133 // apart everywhere a client can read the type back,
17134 // and SPG reported `datetime` for both — so a dump
17135 // and reload silently changed the column's declared
17136 // type, and MySQL's TIMESTAMP is not DATETIME (a
17137 // different range, and UTC conversion on the way in
17138 // and out).
17139 mysql_declared_timestamp = ty_ident.eq_ignore_ascii_case("timestamp");
17140 }
17141 // v7.14.0 — PG canonical `TIMESTAMP WITH TIME ZONE`
17142 // / `TIMESTAMP WITHOUT TIME ZONE`. pg_dump emits
17143 // the full form. SPG canonicalises:
17144 // - WITH TIME ZONE → Timestamptz
17145 // - WITHOUT TIME ZONE → Timestamp
17146 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with"))
17147 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("time"))
17148 && matches!(self.tokens.get(self.pos + 2), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("zone"))
17149 {
17150 self.advance(); // WITH
17151 self.advance(); // TIME
17152 self.advance(); // ZONE
17153 ColumnTypeName::Timestamptz
17154 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("without"))
17155 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("time"))
17156 && matches!(self.tokens.get(self.pos + 2), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("zone"))
17157 {
17158 self.advance(); // WITHOUT
17159 self.advance(); // TIME
17160 self.advance(); // ZONE
17161 ColumnTypeName::Timestamp
17162 } else {
17163 // A second `(precision)` cannot legally follow, but the
17164 // old grammar tolerated it; keep that tolerance.
17165 self.consume_optional_paren_size();
17166 ColumnTypeName::Timestamp
17167 }
17168 }
17169 // v7.9.2 — `TIMESTAMPTZ` and full PG spelling
17170 // `TIMESTAMP WITH TIME ZONE`. Same storage as TIMESTAMP;
17171 // only PG-wire OID differs.
17172 "timestamptz" => {
17173 self.consume_optional_paren_size();
17174 ColumnTypeName::Timestamptz
17175 }
17176 // v4.9: JSON / JSONB. Stored as raw text — no parse-time
17177 // validation. We accept the JSONB spelling too because
17178 // most PG clients default to it; SPG doesn't distinguish
17179 // the two (no path-operator perf advantage to model).
17180 "json" => ColumnTypeName::Json,
17181 "jsonb" => ColumnTypeName::Jsonb,
17182 // v7.10.4 — PG `BYTEA` and the SPG `BYTES` alias both
17183 // surface here. Same storage shape; mapping happens at
17184 // the engine side via the ColumnTypeName → DataType
17185 // resolver. Literal forms are handled at coerce_value
17186 // time so the lexer stays untouched.
17187 "bytea" | "bytes" => ColumnTypeName::Bytes,
17188 // v7.17.0 Phase 7 — PG network address types
17189 // v7.17.0 had a Text-backed fallback here for
17190 // `inet` / `cidr` / `macaddr`. v7.37.5 ζ-A promoted
17191 // each to a first-class type; the keywords are
17192 // bound below in the ζ-A block.
17193 // v7.12.0 — PG full-text search types. mailrs G-CRIT-3.
17194 // The actual `to_tsvector` / `@@` / `ts_rank` surface
17195 // arrives in v7.12.1+; the type itself loads here so
17196 // mailrs's `scripts/init-schema.sql` runs unmodified.
17197 "tsvector" => ColumnTypeName::TsVector,
17198 "tsquery" => ColumnTypeName::TsQuery,
17199 // v7.17.0 — PG `UUID`. Wire OID 2950. The drop-in PG
17200 // surface for Django / Rails / Hibernate's default
17201 // PK pattern.
17202 "uuid" => ColumnTypeName::Uuid,
17203 // v7.37.5 β-P2 — PG `INTERVAL` as a column type.
17204 // Storage = three-field {months, days, micros}, catalog
17205 // tag 34, FILE_VERSION 48+, wire OID 1186. Prior to this
17206 // line `INTERVAL` was parser-rejected at CREATE TABLE.
17207 "interval" => {
17208 // pg_dump emits field-qualified forms like `INTERVAL DAY TO
17209 // SECOND` and an optional `(p)` precision. SPG stores the full
17210 // {months,days,micros}; consume + ignore the qualifier/precision.
17211 while matches!(self.peek(), Token::To)
17212 || matches!(self.peek(), Token::Ident(s) if matches!(
17213 s.to_ascii_lowercase().as_str(),
17214 "year" | "month" | "day" | "hour" | "minute" | "second"
17215 ))
17216 {
17217 self.advance();
17218 }
17219 self.consume_optional_paren_size();
17220 ColumnTypeName::Interval
17221 }
17222 // v7.17.0 Phase 3.P0-32 — PG `TIME` (without time zone).
17223 // i64 microseconds since 00:00:00. Wire OID 1083.
17224 // pg_dump emits `TIME(6)` and `TIME(6) WITH TIME ZONE`.
17225 "time" => {
17226 // v7.39 (round 424) — MySQL TIME carries a semantic
17227 // fractional-seconds precision, bare meaning 0.
17228 let n = self.take_optional_paren_size();
17229 if self.mysql_dialect {
17230 mysql_fsp = Some(n.unwrap_or(0).min(6));
17231 }
17232 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with"))
17233 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("time"))
17234 && matches!(self.tokens.get(self.pos + 2), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("zone"))
17235 {
17236 self.advance();
17237 self.advance();
17238 self.advance();
17239 ColumnTypeName::TimeTz
17240 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("without"))
17241 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("time"))
17242 && matches!(self.tokens.get(self.pos + 2), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("zone"))
17243 {
17244 self.advance();
17245 self.advance();
17246 self.advance();
17247 ColumnTypeName::Time
17248 } else {
17249 ColumnTypeName::Time
17250 }
17251 }
17252 // v7.17.0 Phase 3.P0-33 — MySQL `YEAR`. u16 in
17253 // 1901..=2155 + zero-year sentinel 0. Wire = INT4.
17254 "year" => ColumnTypeName::Year,
17255 // v7.17.0 Phase 3.P0-34 — PG `TIMETZ` / `TIME WITH
17256 // TIME ZONE`. i64 us + i32 offset_secs. Wire OID 1266.
17257 "timetz" => ColumnTypeName::TimeTz,
17258 // v7.17.0 Phase 3.P0-35 — PG `MONEY` — i64 cents.
17259 // Wire OID 790.
17260 "money" => ColumnTypeName::Money,
17261 // v7.17.0 Phase 3.P0-38 — PG range types.
17262 "int4range" => ColumnTypeName::Range(RangeKindAst::Int4),
17263 "int8range" => ColumnTypeName::Range(RangeKindAst::Int8),
17264 "numrange" => ColumnTypeName::Range(RangeKindAst::Num),
17265 "tsrange" => ColumnTypeName::Range(RangeKindAst::Ts),
17266 "tstzrange" => ColumnTypeName::Range(RangeKindAst::TsTz),
17267 "daterange" => ColumnTypeName::Range(RangeKindAst::Date),
17268 // v7.37.5 δ — PG 14+ multirange keywords.
17269 "int4multirange" => ColumnTypeName::Multirange(RangeKindAst::Int4),
17270 "int8multirange" => ColumnTypeName::Multirange(RangeKindAst::Int8),
17271 "nummultirange" => ColumnTypeName::Multirange(RangeKindAst::Num),
17272 "tsmultirange" => ColumnTypeName::Multirange(RangeKindAst::Ts),
17273 "tstzmultirange" => ColumnTypeName::Multirange(RangeKindAst::TsTz),
17274 "datemultirange" => ColumnTypeName::Multirange(RangeKindAst::Date),
17275 // v7.37.5 ε — PG geometry scalar keywords.
17276 "point" => ColumnTypeName::Point,
17277 "lseg" => ColumnTypeName::Lseg,
17278 "path" => ColumnTypeName::Path,
17279 "box" => ColumnTypeName::PgBox,
17280 "polygon" => ColumnTypeName::Polygon,
17281 "line" => ColumnTypeName::Line,
17282 "circle" => ColumnTypeName::Circle,
17283 // v7.37.5 ζ-A — network / bit / xml / "char" keywords.
17284 "inet" => ColumnTypeName::Inet,
17285 "cidr" => ColumnTypeName::Cidr,
17286 "macaddr" => ColumnTypeName::Macaddr,
17287 "macaddr8" => ColumnTypeName::Macaddr8,
17288 // `bit`, `bit(N)`, `bit varying`, `bit varying(N)`. SPG carries the
17289 // width in the value, so the optional `(N)` typmod is accepted and
17290 // ignored (the column stores whatever width it's given).
17291 "bit" => {
17292 let varying = matches!(
17293 self.peek(),
17294 Token::Ident(k) if k.eq_ignore_ascii_case("varying")
17295 );
17296 if varying {
17297 self.advance();
17298 }
17299 // v7.39 (round 281) — the length used to be parsed and
17300 // dropped, so `bit(3)` accepted a five-bit string.
17301 let n = if matches!(self.peek(), Token::LParen) {
17302 self.parse_paren_size("BIT")?
17303 } else {
17304 0
17305 };
17306 if varying {
17307 ColumnTypeName::BitVarying(n)
17308 } else {
17309 ColumnTypeName::Bit(n)
17310 }
17311 }
17312 "varbit" => {
17313 let n = if matches!(self.peek(), Token::LParen) {
17314 self.parse_paren_size("VARBIT")?
17315 } else {
17316 0
17317 };
17318 ColumnTypeName::BitVarying(n)
17319 }
17320 "xml" => ColumnTypeName::Xml,
17321 // v7.17.0 Phase 3.P0-39 — PG hstore extension type.
17322 "hstore" => ColumnTypeName::Hstore,
17323 // v7.17.0 Phase 3.P0-36 — MySQL inline ENUM
17324 // `ENUM('a','b','c')`. Storage is TEXT; the value
17325 // list lands on `inline_enum_variants` for the
17326 // engine to validate INSERT cells against. Empty
17327 // value list is a parse error (matches MySQL).
17328 "enum" => {
17329 // Expect the opening `(`.
17330 if !matches!(self.peek(), Token::LParen) {
17331 return Err(self.err(alloc::format!(
17332 "expected '(' after ENUM, got {:?}",
17333 self.peek()
17334 )));
17335 }
17336 self.advance();
17337 let mut variants: Vec<String> = Vec::new();
17338 loop {
17339 match self.advance() {
17340 Token::String(s) => variants.push(s),
17341 other => {
17342 return Err(self.err(alloc::format!(
17343 "ENUM(...) expects string literal variants, got {other:?}"
17344 )));
17345 }
17346 }
17347 match self.peek() {
17348 Token::Comma => {
17349 self.advance();
17350 continue;
17351 }
17352 Token::RParen => {
17353 self.advance();
17354 break;
17355 }
17356 other => {
17357 return Err(self.err(alloc::format!(
17358 "expected ',' or ')' in ENUM(...), got {other:?}"
17359 )));
17360 }
17361 }
17362 }
17363 if variants.is_empty() {
17364 return Err(self.err("ENUM(...) must declare at least one variant".into()));
17365 }
17366 inline_enum_variants = Some(variants);
17367 // Storage is plain TEXT; the variant list lives on
17368 // the ColumnSchema side.
17369 ColumnTypeName::Text
17370 }
17371 // v7.17.0 Phase 3.P0-37 — MySQL inline SET
17372 // `SET('a','b','c')`. Same parse shape as ENUM;
17373 // semantics differ (subset rather than pick-one).
17374 "set" => {
17375 if !matches!(self.peek(), Token::LParen) {
17376 return Err(self.err(alloc::format!(
17377 "expected '(' after SET, got {:?}",
17378 self.peek()
17379 )));
17380 }
17381 self.advance();
17382 let mut variants: Vec<String> = Vec::new();
17383 loop {
17384 match self.advance() {
17385 Token::String(s) => variants.push(s),
17386 other => {
17387 return Err(self.err(alloc::format!(
17388 "SET(...) expects string literal variants, got {other:?}"
17389 )));
17390 }
17391 }
17392 match self.peek() {
17393 Token::Comma => {
17394 self.advance();
17395 continue;
17396 }
17397 Token::RParen => {
17398 self.advance();
17399 break;
17400 }
17401 other => {
17402 return Err(self.err(alloc::format!(
17403 "expected ',' or ')' in SET(...), got {other:?}"
17404 )));
17405 }
17406 }
17407 }
17408 if variants.is_empty() {
17409 return Err(self.err("SET(...) must declare at least one variant".into()));
17410 }
17411 inline_set_variants = Some(variants);
17412 ColumnTypeName::Text
17413 }
17414 _other => {
17415 // v7.17.0 Phase 1.4 — unknown ident → defer
17416 // resolution to the engine. Stored as Text in
17417 // ColumnTypeName + the original name carried as
17418 // `user_type_ref` so CREATE TABLE can look up
17419 // user-defined enum / domain types.
17420 user_type_ref = Some(ty_ident.clone());
17421 ColumnTypeName::Text
17422 }
17423 };
17424 // v7.17.0 Phase 4.4 — MySQL's `UNSIGNED` modifier sits
17425 // right after the type keyword. Pre-4.4 SPG consumed +
17426 // discarded the keyword, leaving a customer column
17427 // declared `id INT UNSIGNED NOT NULL` silently accepting
17428 // negative values — a Tier-A correctness drift where
17429 // application invariants (auto-increment-IDs never
17430 // negative) silently broke on cutover. Now: capture as
17431 // a column flag, persist on the schema, enforce at
17432 // INSERT / UPDATE time.
17433 let is_unsigned = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("unsigned"))
17434 {
17435 self.advance();
17436 true
17437 } else {
17438 false
17439 };
17440 // v7.14.0 — mysqldump emits `<type> CHARACTER SET <name>` and
17441 // `<type> COLLATE <name>` post-fixes on text columns. SPG
17442 // stores text as UTF-8 always so CHARACTER SET is still a
17443 // no-op. v7.17.0 Phase 2.5 — COLLATE no longer drops the
17444 // name: it gets classified into a `Collation` variant the
17445 // engine consults at WHERE-eval time. PG `default` /
17446 // `pg_catalog.default` / `C` / `POSIX` collations all
17447 // resolve to `Binary` (the prior behaviour); `_ci` /
17448 // `case_insensitive` / `nocase` shift to CaseInsensitive.
17449 // The schema-qualifier form (`pg_catalog.default`) lexes
17450 // as `Ident '.' Ident` — peek for the `.` and consume both
17451 // halves so it's treated as one collation name. PG's
17452 // `IDENT.IDENT` collation form (which can appear here) is
17453 // resolved by Collation::from_collation_name on the bare
17454 // identifier after the dot.
17455 let mut collation = Collation::Binary;
17456 // v7.39 (round 370, M4 P4a) — whether an explicit `COLLATE <name>`
17457 // clause was written. The engine needs this to tell an explicit
17458 // `COLLATE utf8mb4_bin` (byte-wise) apart from a column with no
17459 // clause at all: both resolve to `Collation::Binary`, but under the
17460 // MySQL dialect the latter takes the folding default collation.
17461 let mut collation_explicit = false;
17462 let mut collation_name: Option<alloc::string::String> = None;
17463 loop {
17464 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("character"))
17465 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("set"))
17466 {
17467 self.advance(); // CHARACTER
17468 self.advance(); // SET
17469 if matches!(
17470 self.peek(),
17471 Token::Ident(_) | Token::QuotedIdent(_) | Token::String(_)
17472 ) {
17473 self.advance();
17474 }
17475 continue;
17476 }
17477 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("collate")) {
17478 self.advance(); // COLLATE
17479 // Accept Ident / QuotedIdent / String AND the
17480 // keyword-tokenised `Default` (PG `pg_catalog.default`
17481 // and bare `DEFAULT` collation names — `default` is a
17482 // reserved word so the lexer hands back Token::Default
17483 // not Token::Ident).
17484 let read_collation_atom = |this: &mut Self| -> Option<alloc::string::String> {
17485 match this.peek().clone() {
17486 Token::Ident(s) | Token::QuotedIdent(s) | Token::String(s) => {
17487 this.advance();
17488 Some(s)
17489 }
17490 Token::Default => {
17491 this.advance();
17492 Some(alloc::string::String::from("default"))
17493 }
17494 _ => None,
17495 }
17496 };
17497 let raw = if let Some(head) = read_collation_atom(self) {
17498 // Schema-qualified PG form: `pg_catalog.default`.
17499 if matches!(self.peek(), Token::Dot) {
17500 self.advance();
17501 let tail = read_collation_atom(self).unwrap_or_default();
17502 alloc::format!("{head}.{tail}")
17503 } else {
17504 head
17505 }
17506 } else {
17507 alloc::string::String::new()
17508 };
17509 if !raw.is_empty() {
17510 collation_explicit = true;
17511 // v7.39 (round 676) — keep the name too. The enum below
17512 // folds C / POSIX / en_US / default into one value, and
17513 // `pg_attribute.attcollation` has to tell them apart.
17514 // The schema qualifier goes: PG's `pg_catalog.default`
17515 // and a bare `default` name the same collation.
17516 // v7.39 (round 679) — strip a SCHEMA qualifier, not an
17517 // encoding suffix. Round 676 used `rsplit('.')` for
17518 // both, and `COLLATE "en_US.utf8"` came out as `utf8`:
17519 // PG writes `pg_catalog.default` (qualifier) and
17520 // `en_US.utf8` (locale + encoding) with the same
17521 // separator. Only `pg_catalog.` is a qualifier, and it
17522 // is the only one PG's own dumps emit.
17523 let bare = raw.trim_matches(|c: char| c == '"' || c == '\'');
17524 let bare = bare.strip_prefix("pg_catalog.").unwrap_or(bare);
17525 collation_name = Some(alloc::string::String::from(bare));
17526 let parsed = Collation::from_collation_name(&raw);
17527 // Last COLLATE clause wins, but `Binary` from a
17528 // bare keyword like `default` should not
17529 // silently downgrade a stronger one set earlier
17530 // on the same column. v7.17 only ships one
17531 // non-Binary variant so a simple OR is enough.
17532 if parsed != Collation::Binary {
17533 collation = parsed;
17534 }
17535 }
17536 continue;
17537 }
17538 break;
17539 }
17540 // v7.10.10 — postfix `[]` widens the base type to its array
17541 // type. PG accepts `TYPE[]` after any base type and so does
17542 // SPG now (round-753 probe: INT[] / NUMERIC[] / TIMESTAMP[]
17543 // all through; the old "only TEXT[]" note was stale).
17544 if matches!(self.peek(), Token::LBracket) {
17545 self.advance();
17546 if !matches!(self.peek(), Token::RBracket) {
17547 return Err(self.err(alloc::format!(
17548 "TEXT[] takes no dimension; got {:?}",
17549 self.peek()
17550 )));
17551 }
17552 self.advance();
17553 // v7.11.13 — widened to INT[] and BIGINT[] in addition
17554 // to TEXT[]. Other base types (BOOL[], NUMERIC[], etc.)
17555 // still error here.
17556 ty = match ty {
17557 ColumnTypeName::Text => ColumnTypeName::TextArray,
17558 ColumnTypeName::Int => ColumnTypeName::IntArray,
17559 ColumnTypeName::BigInt => ColumnTypeName::BigIntArray,
17560 // v7.37.5 β-P4 — INTERVAL[] via the same postfix
17561 // `[]` grammar. Wire OID 1187.
17562 ColumnTypeName::Interval => ColumnTypeName::IntervalArray,
17563 // v7.37.5 γ — full PG array-of-scalar family.
17564 ColumnTypeName::Bool => ColumnTypeName::BoolArray,
17565 ColumnTypeName::SmallInt => ColumnTypeName::SmallIntArray,
17566 ColumnTypeName::Float => ColumnTypeName::FloatArray,
17567 // NUMERIC(p, s) loses its precision params at the
17568 // array level (matches PG: `NUMERIC[]` is untyped,
17569 // per-element precision flows through values).
17570 ColumnTypeName::Numeric(_, _) => ColumnTypeName::NumericArray,
17571 ColumnTypeName::Date => ColumnTypeName::DateArray,
17572 ColumnTypeName::Timestamp => ColumnTypeName::TimestampArray,
17573 ColumnTypeName::Timestamptz => ColumnTypeName::TimestamptzArray,
17574 ColumnTypeName::Uuid => ColumnTypeName::UuidArray,
17575 ColumnTypeName::Json => ColumnTypeName::JsonArray,
17576 ColumnTypeName::Jsonb => ColumnTypeName::JsonbArray,
17577 ColumnTypeName::Bytes => ColumnTypeName::BytesArray,
17578 // VARCHAR(n)[] / CHAR(n)[] drop the length cap at
17579 // the array level (matches PG semantics where the
17580 // element precision is per-row, not column-wide).
17581 ColumnTypeName::Varchar(_) => ColumnTypeName::VarcharArray,
17582 ColumnTypeName::Char(_) => ColumnTypeName::CharArray,
17583 // v7.37.5 ζ-A — MONEY[] (OID 791) ship-triage
17584 // follow-up.
17585 ColumnTypeName::Money => ColumnTypeName::MoneyArray,
17586 other => {
17587 return Err(self.err(alloc::format!("{other:?}[] not yet supported")));
17588 }
17589 };
17590 // v7.17.0 Phase 3.P0-40 — second `[]` widens 1D → 2D
17591 // for INT/TEXT/BIGINT. Anything else is an error.
17592 if matches!(self.peek(), Token::LBracket) {
17593 self.advance();
17594 if !matches!(self.peek(), Token::RBracket) {
17595 return Err(self.err(alloc::format!(
17596 "TYPE[][] second dimension takes no size; got {:?}",
17597 self.peek()
17598 )));
17599 }
17600 self.advance();
17601 ty = match ty {
17602 ColumnTypeName::IntArray => ColumnTypeName::IntArray2D,
17603 ColumnTypeName::BigIntArray => ColumnTypeName::BigIntArray2D,
17604 ColumnTypeName::TextArray => ColumnTypeName::TextArray2D,
17605 // v7.39 (read01 round 75) — bool[][].
17606 ColumnTypeName::BoolArray => ColumnTypeName::BoolArray2D,
17607 other => {
17608 return Err(self.err(alloc::format!(
17609 "v7.17 2D arrays support INT[][] / BIGINT[][] / \
17610 TEXT[][] only; got {other:?}"
17611 )));
17612 }
17613 };
17614 }
17615 }
17616 Ok((
17617 ty,
17618 implied_auto_increment,
17619 implied_not_null,
17620 user_type_ref,
17621 collation,
17622 collation_explicit,
17623 collation_name,
17624 is_unsigned,
17625 inline_enum_variants,
17626 inline_set_variants,
17627 mysql_int_width,
17628 mysql_fsp,
17629 mysql_declared_timestamp,
17630 mysql_float_md,
17631 ))
17632 }
17633
17634 fn parse_column_def(&mut self) -> Result<ColumnDef, ParseError> {
17635 // v7.20 — PG reserves the table-constraint keywords, so a
17636 // BARE `UNIQUE` / `PRIMARY` / … in column position is a
17637 // malformed constraint clause (e.g. `UNIQUE a` missing its
17638 // parens), not a column named "unique". Since v7.17's
17639 // unknown-type leniency (`user_type_ref`) such a clause
17640 // would otherwise parse as a column with a user-defined
17641 // type — silently accepting invalid DDL. Quoted
17642 // identifiers ("unique" / `unique`) remain valid names.
17643 if let Token::Ident(s) = self.peek()
17644 && [
17645 "unique",
17646 "primary",
17647 "foreign",
17648 "constraint",
17649 "check",
17650 "references",
17651 "exclude",
17652 ]
17653 .iter()
17654 .any(|kw| s.eq_ignore_ascii_case(kw))
17655 {
17656 return Err(self.err(alloc::format!(
17657 "unexpected reserved keyword '{s}' at start of column definition \
17658 (malformed table constraint?)"
17659 )));
17660 }
17661 let name_tok = self.pos;
17662 let name = self.expect_ident_like()?;
17663 // v7.39.3 — MySQL 9.7.2 reports a column by the SPELLING it was
17664 // declared with: `MyCol` stays `MyCol` in SHOW COLUMNS, in
17665 // information_schema, and in SHOW CREATE (measured). SPG folded
17666 // an unquoted name, so a table restored from a dump reported
17667 // names the application had never written.
17668 //
17669 // The written form comes back from the source span, which only
17670 // the MySQL dialect keeps. The span runs to the START of the
17671 // next token, so a comment or unusual spacing between them
17672 // arrives with it — hence the check that what came back is the
17673 // same identifier. It is not decoration: without it,
17674 // `CREATE TABLE t (MyCol /* c */ INT)` names the column
17675 // `MyCol /* c */`.
17676 let name = self
17677 .source_span(name_tok, name_tok)
17678 .map(|raw| raw.trim().trim_matches('`').trim_matches('"'))
17679 .filter(|raw| raw.eq_ignore_ascii_case(&name))
17680 .map_or(name, alloc::string::String::from);
17681 let (
17682 ty,
17683 implied_auto_increment,
17684 implied_not_null,
17685 user_type_ref,
17686 collation,
17687 collation_explicit,
17688 collation_name,
17689 is_unsigned,
17690 inline_enum_variants,
17691 inline_set_variants,
17692 mysql_int_width,
17693 mysql_fsp,
17694 mysql_declared_timestamp,
17695 mysql_float_md,
17696 ) = self.parse_type_with_implied_flags()?;
17697 // Column constraints: `DEFAULT <expr>`, `NOT NULL`, and the
17698 // MySQL-flavoured `AUTO_INCREMENT` may appear in any order;
17699 // each at most once.
17700 let mut default: Option<Expr> = None;
17701 let mut nullable = !implied_not_null;
17702 let mut nullability_seen = implied_not_null;
17703 let mut auto_increment = implied_auto_increment;
17704 let mut is_primary_key = false;
17705 let mut is_unique = false;
17706 let mut unique_nulls_not_distinct = false;
17707 let mut constraint_deferrable = false;
17708 let mut constraint_initially_deferred = false;
17709 let mut check: Option<Expr> = None;
17710 let mut on_update_runtime: Option<Expr> = None;
17711 let mut generated_stored_expr: Option<Box<Expr>> = None;
17712 let mut identity_always = false;
17713 loop {
17714 // v7.22 (mailrs round-13 gap 3) — PG 18 catalogs
17715 // not-null constraints by name and pg_dump emits them
17716 // inline: `id bigint CONSTRAINT contacts_id_not_null1
17717 // NOT NULL`. Accept and discard the name; whatever
17718 // constraint follows is parsed by the arms below.
17719 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint")) {
17720 // v7.39 (round 308, V29) — a name on an inline
17721 // REFERENCES belongs to the FOREIGN KEY, and the caller
17722 // (`parse_column_def_with_fk`) is what builds it, so
17723 // leave the whole clause for it. Dropping the name here
17724 // is what made `CONSTRAINT fk_a REFERENCES …` come back
17725 // as the synthesised `c_pid_fkey` — which then could
17726 // not be matched by `SET CONSTRAINTS fk_a`. Peek only:
17727 // `advance()` takes tokens by `mem::replace`, so there
17728 // is no rewinding once consumed.
17729 if matches!(
17730 self.tokens.get(self.pos + 2),
17731 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("references")
17732 ) {
17733 break;
17734 }
17735 self.advance();
17736 let _name = self.expect_ident_like()?;
17737 continue;
17738 }
17739 // v7.39 (round 379) — MySQL's SHORT generated-column form
17740 // omits `GENERATED ALWAYS`: `<col> <type> AS (<expr>)
17741 // [STORED | VIRTUAL]`. mysqldump emits the long form (handled
17742 // below), but hand-written schemas and app migrations use this.
17743 // STORED / VIRTUAL is optional (MySQL defaults to VIRTUAL);
17744 // SPG computes-and-stores either way, like the long form.
17745 if matches!(self.peek(), Token::As) {
17746 self.advance();
17747 if !matches!(self.peek(), Token::LParen) {
17748 return Err(self.err(alloc::format!(
17749 "expected '(' after AS in a generated column, got {:?}",
17750 self.peek()
17751 )));
17752 }
17753 self.advance();
17754 let expr = self.parse_expr(0)?;
17755 if !matches!(self.peek(), Token::RParen) {
17756 return Err(self.err(alloc::format!(
17757 "expected ')' after AS (<expr>), got {:?}",
17758 self.peek()
17759 )));
17760 }
17761 self.advance();
17762 if matches!(self.peek(), Token::Ident(s)
17763 if s.eq_ignore_ascii_case("stored") || s.eq_ignore_ascii_case("virtual"))
17764 {
17765 self.advance();
17766 }
17767 generated_stored_expr = Some(alloc::boxed::Box::new(expr));
17768 continue;
17769 }
17770 // v7.22 (round-13 T2) — inline `GENERATED { ALWAYS |
17771 // BY DEFAULT } AS IDENTITY [(seq options)]` (PG 10+;
17772 // the modern replacement for SERIAL in hand-written
17773 // schemas). Both flavours map onto the auto-increment
17774 // machinery — SPG's serial semantics ≈ BY DEFAULT;
17775 // ALWAYS's reject-explicit-values nuance is documented
17776 // leniency. Generated EXPRESSION columns
17777 // (`AS (expr) STORED`) are not supported: error loudly
17778 // instead of silently storing NULLs.
17779 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("generated")) {
17780 self.advance();
17781 let mut saw_generated_always = false;
17782 match self.peek().clone() {
17783 Token::Ident(s) if s.eq_ignore_ascii_case("always") => {
17784 self.advance();
17785 saw_generated_always = true;
17786 }
17787 Token::Ident(b) | Token::QuotedIdent(b) if b.eq_ignore_ascii_case("by") => {
17788 self.advance();
17789 if !matches!(self.peek(), Token::Default) {
17790 return Err(self.err(alloc::format!(
17791 "expected DEFAULT after GENERATED BY, got {:?}",
17792 self.peek()
17793 )));
17794 }
17795 self.advance();
17796 }
17797 other => {
17798 return Err(self.err(alloc::format!(
17799 "expected ALWAYS or BY DEFAULT after GENERATED, got {other:?}"
17800 )));
17801 }
17802 }
17803 if !matches!(self.peek(), Token::As) {
17804 return Err(self.err(alloc::format!(
17805 "expected AS after GENERATED ALWAYS/BY DEFAULT, got {:?}",
17806 self.peek()
17807 )));
17808 }
17809 self.advance();
17810 // v7.37.7(sentori Epic 3 P1)— `GENERATED ALWAYS AS
17811 // ( <expr> ) STORED` stored computed-column. The
17812 // expression is captured for the engine to recompute
17813 // on every INSERT / UPDATE. v7.37.7 accepts the
17814 // STORED keyword only; PG also has VIRTUAL, which
17815 // v7.37.7 carves out (sentori only uses STORED).
17816 if matches!(self.peek(), Token::LParen) {
17817 self.advance();
17818 let expr = self.parse_expr(0)?;
17819 if !matches!(self.peek(), Token::RParen) {
17820 return Err(self.err(alloc::format!(
17821 "expected ')' after GENERATED ALWAYS AS (<expr>), got {:?}",
17822 self.peek()
17823 )));
17824 }
17825 self.advance();
17826 let stored = match self.peek() {
17827 Token::Ident(s) | Token::QuotedIdent(s)
17828 if s.eq_ignore_ascii_case("stored") =>
17829 {
17830 self.advance();
17831 true
17832 }
17833 // v7.38 (read01 P4.14) — accept PG 18's VIRTUAL
17834 // generated columns. SPG computes them on write and
17835 // persists like STORED; the two are observably
17836 // identical for query results (the value, recompute
17837 // on base-column change, and NOT NULL enforcement all
17838 // match), so a PG 18 schema/dump using VIRTUAL loads
17839 // and behaves correctly. The compute-on-read storage
17840 // saving is an invisible internal difference.
17841 Token::Ident(s) | Token::QuotedIdent(s)
17842 if s.eq_ignore_ascii_case("virtual") =>
17843 {
17844 self.advance();
17845 false
17846 }
17847 other => {
17848 return Err(self.err(alloc::format!(
17849 "expected STORED or VIRTUAL after GENERATED ALWAYS AS (<expr>), \
17850 got {other:?}"
17851 )));
17852 }
17853 };
17854 let _ = stored; // STORED / VIRTUAL both compute-and-store.
17855 generated_stored_expr = Some(Box::new(expr));
17856 continue;
17857 }
17858 self.expect_keyword_ident("identity")?;
17859 // Optional `(START WITH 1 INCREMENT BY 1 …)` —
17860 // consume the balanced parens and discard (SPG's
17861 // auto-increment is max+1-scan based).
17862 if matches!(self.peek(), Token::LParen) {
17863 let mut depth = 0usize;
17864 loop {
17865 match self.advance() {
17866 Token::LParen => depth += 1,
17867 Token::RParen => {
17868 depth -= 1;
17869 if depth == 0 {
17870 break;
17871 }
17872 }
17873 Token::Eof => {
17874 return Err(self.err(
17875 "unterminated sequence-options parens after IDENTITY".into(),
17876 ));
17877 }
17878 _ => {}
17879 }
17880 }
17881 }
17882 auto_increment = true;
17883 // v7.38 (read01) — remember the ALWAYS flavour so the engine
17884 // can reject explicit non-DEFAULT INSERT values (unless
17885 // OVERRIDING SYSTEM VALUE) the way PG does.
17886 identity_always = saw_generated_always;
17887 // PG identity columns are implicitly NOT NULL.
17888 nullable = false;
17889 continue;
17890 }
17891 // v7.17.0 Phase 2.1 — MySQL `ON UPDATE
17892 // CURRENT_TIMESTAMP[(N)]`. Only CURRENT_TIMESTAMP
17893 // is accepted today. The "ON" token is an Ident
17894 // (not reserved) — peek before consuming.
17895 if matches!(self.peek(), Token::On)
17896 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("update"))
17897 {
17898 self.advance(); // ON
17899 self.advance(); // update
17900 // Accept CURRENT_TIMESTAMP / CURRENT_TIMESTAMP(N).
17901 let next = self.peek().clone();
17902 match next {
17903 Token::Ident(s) | Token::QuotedIdent(s)
17904 if s.eq_ignore_ascii_case("current_timestamp") =>
17905 {
17906 self.advance();
17907 // Optional `(N)` precision.
17908 if matches!(self.peek(), Token::LParen) {
17909 self.advance();
17910 if !matches!(self.peek(), Token::Integer(_)) {
17911 return Err(self.err(alloc::format!(
17912 "expected integer precision inside CURRENT_TIMESTAMP(…), got {:?}",
17913 self.peek()
17914 )));
17915 }
17916 self.advance();
17917 if !matches!(self.peek(), Token::RParen) {
17918 return Err(self.err(alloc::format!(
17919 "expected ')' after CURRENT_TIMESTAMP precision, got {:?}",
17920 self.peek()
17921 )));
17922 }
17923 self.advance();
17924 }
17925 on_update_runtime = Some(Expr::FunctionCall {
17926 name: "now".into(),
17927 args: Vec::new(),
17928 });
17929 continue;
17930 }
17931 other => {
17932 return Err(self.err(alloc::format!(
17933 "v7.17 only supports ON UPDATE CURRENT_TIMESTAMP, got {other:?}"
17934 )));
17935 }
17936 }
17937 }
17938 if matches!(self.peek(), Token::Default) {
17939 if default.is_some() {
17940 return Err(self.err("DEFAULT specified twice".into()));
17941 }
17942 self.advance();
17943 default = Some(self.parse_expr(0)?);
17944 continue;
17945 }
17946 // v7.39 (round 621) — `NOT DEFERRABLE` shares this arm's leading
17947 // token with NOT NULL and sits EARLIER in the loop than the
17948 // deferrability arm, so without the lookahead it was reported as
17949 // "NOT NULL specified twice" (or "expected NULL after NOT").
17950 if matches!(self.peek(), Token::Not)
17951 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("deferrable"))
17952 {
17953 // NOT DEFERRABLE — explicit immediate; nothing to carry.
17954 self.consume_optional_deferrable_clauses()?;
17955 continue;
17956 }
17957 if matches!(self.peek(), Token::Not) {
17958 if nullability_seen {
17959 return Err(self.err("NOT NULL specified twice".into()));
17960 }
17961 self.advance();
17962 if !matches!(self.peek(), Token::Null) {
17963 return Err(self.err(format!(
17964 "expected NULL after NOT in column def, got {:?}",
17965 self.peek()
17966 )));
17967 }
17968 self.advance();
17969 nullable = false;
17970 nullability_seen = true;
17971 continue;
17972 }
17973 // v7.14.0 — MySQL accepts a bare `NULL` as an explicit
17974 // "this column is nullable" marker (the default in
17975 // standard SQL anyway). mysqldump emits it routinely
17976 // (`col TYPE NULL DEFAULT NULL` for nullable
17977 // timestamps etc). Accept + no-op.
17978 if matches!(self.peek(), Token::Null) {
17979 if nullability_seen && !nullable {
17980 // v7.39 (round 761, F31 tranche 2 #31) — PG's
17981 // sentence, PG18-measured (the table name is the
17982 // caller's; the column half is exact).
17983 return Err(self.err(alloc::format!(
17984 "conflicting NULL/NOT NULL declarations for column \"{name}\""
17985 )));
17986 }
17987 self.advance();
17988 nullable = true;
17989 nullability_seen = true;
17990 continue;
17991 }
17992 // `AUTO_INCREMENT` or its abbreviated form `AUTOINCREMENT`
17993 // arrives as a bare Ident. Match either, case-insensitive.
17994 if let Token::Ident(s) = self.peek()
17995 && (s.eq_ignore_ascii_case("auto_increment")
17996 || s.eq_ignore_ascii_case("autoincrement"))
17997 {
17998 if auto_increment {
17999 return Err(self.err("AUTO_INCREMENT specified twice".into()));
18000 }
18001 self.advance();
18002 auto_increment = true;
18003 continue;
18004 }
18005 // v7.9.13 — inline `PRIMARY KEY` column constraint
18006 // (mailrs F1). Implies `NOT NULL`. The engine creates
18007 // a BTree index for the PK column at CREATE TABLE time
18008 // so FK parent-side index lookups resolve.
18009 // v7.39 (round 621) — `[NOT] DEFERRABLE [INITIALLY {DEFERRED |
18010 // IMMEDIATE}]` after an inline PK / UNIQUE / REFERENCES. Every
18011 // spelling was a parse error, so a pg_dump carrying one stopped
18012 // mid-restore. The clauses are consumed by the same helper the FK
18013 // path has used since round 288 and recorded nowhere: SPG enforces
18014 // the constraint IMMEDIATELY either way, which fails earlier than
18015 // PG inside a transaction that violates-then-repairs — a refusal,
18016 // not a wrong answer. True deferral is the open remainder of F08.
18017 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("deferrable") || s.eq_ignore_ascii_case("initially"))
18018 || (matches!(self.peek(), Token::Not)
18019 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("deferrable")))
18020 {
18021 // v7.39 (round 711) — CARRIED now (the storing half of
18022 // F08); round 621 only consumed.
18023 let (d, idef) = self.consume_deferrable_clauses_timed()?;
18024 constraint_deferrable |= d;
18025 constraint_initially_deferred |= idef;
18026 continue;
18027 }
18028 if let Token::Ident(s) = self.peek()
18029 && s.eq_ignore_ascii_case("primary")
18030 {
18031 if is_primary_key {
18032 return Err(self.err("PRIMARY KEY specified twice".into()));
18033 }
18034 // Peek-ahead for the required `KEY` token.
18035 let next = self.tokens.get(self.pos + 1);
18036 let next_is_key = matches!(
18037 next,
18038 Some(Token::Ident(k)) if k.eq_ignore_ascii_case("key")
18039 );
18040 if !next_is_key {
18041 return Err(self.err(format!(
18042 "expected KEY after PRIMARY in column def, got {:?}",
18043 next
18044 )));
18045 }
18046 self.advance(); // PRIMARY
18047 self.advance(); // KEY
18048 is_primary_key = true;
18049 if nullability_seen && nullable {
18050 return Err(self.err(
18051 "column declared NULL but inline PRIMARY KEY implies NOT NULL".into(),
18052 ));
18053 }
18054 nullable = false;
18055 nullability_seen = true;
18056 continue;
18057 }
18058 // v7.13.0 — inline `UNIQUE` column constraint
18059 // (mailrs round-5 G2). Fold into a single-column
18060 // table-level UNIQUE at CREATE TABLE post-process time.
18061 if let Token::Ident(s) = self.peek()
18062 && s.eq_ignore_ascii_case("unique")
18063 {
18064 if is_unique {
18065 return Err(self.err("UNIQUE specified twice".into()));
18066 }
18067 self.advance();
18068 is_unique = true;
18069 // v7.38 (read01 P4.19) — optional `NULLS [NOT] DISTINCT`
18070 // (PG 15+); default is NULLS DISTINCT per the SQL standard.
18071 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("nulls")) {
18072 let n1 = self.tokens.get(self.pos + 1);
18073 let n2 = self.tokens.get(self.pos + 2);
18074 if matches!(n1, Some(Token::Not)) && matches!(n2, Some(Token::Distinct)) {
18075 self.advance(); // NULLS
18076 self.advance(); // NOT
18077 self.advance(); // DISTINCT
18078 unique_nulls_not_distinct = true;
18079 } else if matches!(n1, Some(Token::Distinct)) {
18080 self.advance(); // NULLS
18081 self.advance(); // DISTINCT
18082 }
18083 }
18084 continue;
18085 }
18086 // v7.13.0 — inline `CHECK (<expr>)` column constraint
18087 // (mailrs round-5 G3). PG semantics: column-level
18088 // CHECK is equivalent to a table-level CHECK. Multiple
18089 // inline CHECKs on the same column AND together.
18090 if let Token::Ident(s) = self.peek()
18091 && s.eq_ignore_ascii_case("check")
18092 {
18093 self.advance();
18094 if !matches!(self.peek(), Token::LParen) {
18095 return Err(self.err(alloc::format!(
18096 "expected '(' after CHECK in column def, got {:?}",
18097 self.peek()
18098 )));
18099 }
18100 self.advance();
18101 let pred = self.parse_expr(0)?;
18102 if !matches!(self.peek(), Token::RParen) {
18103 return Err(self.err(alloc::format!(
18104 "expected ')' to close CHECK predicate, got {:?}",
18105 self.peek()
18106 )));
18107 }
18108 self.advance();
18109 check = Some(match check.take() {
18110 Some(prev) => Expr::Binary {
18111 op: BinOp::And,
18112 lhs: Box::new(prev),
18113 rhs: Box::new(pred),
18114 },
18115 None => pred,
18116 });
18117 continue;
18118 }
18119 break;
18120 }
18121 Ok(ColumnDef {
18122 name,
18123 ty,
18124 nullable,
18125 default,
18126 auto_increment,
18127 is_primary_key,
18128 is_unique,
18129 unique_nulls_not_distinct,
18130 constraint_deferrable,
18131 constraint_initially_deferred,
18132 check,
18133 user_type_ref,
18134 on_update_runtime,
18135 collation,
18136 collation_explicit,
18137 collation_name,
18138 is_unsigned,
18139 inline_enum_variants,
18140 inline_set_variants,
18141 generated_stored_expr,
18142 identity_always,
18143 mysql_int_width,
18144 mysql_fsp,
18145 mysql_declared_timestamp,
18146 mysql_float_md,
18147 })
18148 }
18149
18150 /// `NUMERIC` may appear without parameters, with one (precision
18151 /// only, scale=0), or with both. Returns `(precision, scale)` with
18152 /// 0 = unspecified for the bare form.
18153 fn parse_optional_numeric_params(&mut self) -> Result<(u16, i16), ParseError> {
18154 if !matches!(self.peek(), Token::LParen) {
18155 // Bare `NUMERIC` — PG treats this as "unlimited precision";
18156 // we surface it as precision=0 to mean "unconstrained" so
18157 // the engine doesn't need a separate variant.
18158 return Ok((0, 0));
18159 }
18160 self.advance();
18161 // v7.39 (round 272) — PG's declared precision runs to 1000, and
18162 // it words the out-of-range case with the value it saw. SPG
18163 // capped at 38 (i128's width), so a `numeric(50,10)` column PG
18164 // accepts failed to parse at all; values wider than i128 are
18165 // carried by the arbitrary-precision form.
18166 let precision = match self.advance() {
18167 Token::Integer(n) if (1..=1000).contains(&n) => {
18168 u16::try_from(n).expect("range-checked")
18169 }
18170 Token::Integer(n) => {
18171 return Err(ParseError {
18172 message: format!("NUMERIC precision {n} must be between 1 and 1000"),
18173 token_pos: self.consumed_pos(),
18174 });
18175 }
18176 other => {
18177 return Err(ParseError {
18178 message: format!(
18179 "NUMERIC precision must be an integer in 1..=1000, got {other:?}"
18180 ),
18181 token_pos: self.consumed_pos(),
18182 });
18183 }
18184 };
18185 // v7.39 (round 273) — PG's declared scale runs -1000..=1000 and is
18186 // NOT bounded by the precision (`numeric(10,11)` is legal; a value
18187 // then overflows). A negative scale rounds to tens / hundreds / …
18188 let scale = if matches!(self.peek(), Token::Comma) {
18189 self.advance();
18190 let neg = if matches!(self.peek(), Token::Minus) {
18191 self.advance();
18192 true
18193 } else {
18194 false
18195 };
18196 match self.advance() {
18197 Token::Integer(n) => {
18198 let signed = if neg { -n } else { n };
18199 if !(-1000..=1000).contains(&signed) {
18200 return Err(ParseError {
18201 message: format!(
18202 "NUMERIC scale {signed} must be between -1000 and 1000"
18203 ),
18204 token_pos: self.consumed_pos(),
18205 });
18206 }
18207 i16::try_from(signed).expect("range-checked")
18208 }
18209 other => {
18210 return Err(ParseError {
18211 message: format!("NUMERIC scale must be an integer, got {other:?}"),
18212 token_pos: self.consumed_pos(),
18213 });
18214 }
18215 }
18216 } else {
18217 0
18218 };
18219 if !matches!(self.peek(), Token::RParen) {
18220 return Err(self.err(format!(
18221 "expected ')' to close NUMERIC params, got {:?}",
18222 self.peek()
18223 )));
18224 }
18225 self.advance();
18226 Ok((precision, scale))
18227 }
18228
18229 /// Parse `(N)` where `N` is a positive integer literal — used by the
18230 /// `VARCHAR`/`CHAR`/`VECTOR` column types. `label` is the type name
18231 /// for the error message.
18232 /// v6.0.1: parse the optional `USING <encoding>` clause that
18233 /// follows `VECTOR(N)` in a column definition. Missing clause
18234 /// → `VecEncoding::F32` (pre-v6 default). Unknown encoding
18235 /// ident → `ParseError` listing the encodings recognised today.
18236 fn parse_optional_vector_encoding(&mut self) -> Result<VecEncoding, ParseError> {
18237 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
18238 return Ok(VecEncoding::F32);
18239 }
18240 // v7.13.2 — mailrs round-6 S6: `USING` after a vector type
18241 // overlaps with `ALTER COLUMN TYPE … USING <expr>`. Only
18242 // consume the token when the very next token is a known
18243 // vector-encoding keyword (SQ8 / HALF). Otherwise leave
18244 // `USING` for the caller — it's the rewrite-expression form.
18245 let n1 = self.tokens.get(self.pos + 1);
18246 let next_is_encoding = matches!(
18247 n1,
18248 Some(Token::Ident(s))
18249 if s.eq_ignore_ascii_case("sq8") || s.eq_ignore_ascii_case("half")
18250 );
18251 if !next_is_encoding {
18252 return Ok(VecEncoding::F32);
18253 }
18254 self.advance();
18255 let enc_ident = match self.advance() {
18256 Token::Ident(s) => s,
18257 other => {
18258 return Err(self.err(format!(
18259 "expected vector encoding after USING, got {other:?}"
18260 )));
18261 }
18262 };
18263 match enc_ident.to_ascii_lowercase().as_str() {
18264 "sq8" => Ok(VecEncoding::Sq8),
18265 // v6.0.3: `HALF` (pgvector convention) selects IEEE-754
18266 // binary16 per-element storage.
18267 "half" => Ok(VecEncoding::F16),
18268 other => Err(self.err(format!(
18269 "unknown vector encoding {other:?}; supported: SQ8, HALF"
18270 ))),
18271 }
18272 }
18273
18274 /// v7.17.0 Phase 4.3 — peek at the MySQL display-width
18275 /// without consuming it. Returns `Some(N)` when the next
18276 /// tokens are `( <int> )`; None otherwise. Used by the
18277 /// TINYINT classifier to decide whether to map to Bool or
18278 /// SmallInt.
18279 fn peek_optional_paren_size_value(&self) -> Option<i64> {
18280 if !matches!(self.peek(), Token::LParen) {
18281 return None;
18282 }
18283 let next = self.tokens.get(self.pos + 1)?;
18284 let n = match next {
18285 Token::Integer(n) => *n,
18286 _ => return None,
18287 };
18288 if !matches!(self.tokens.get(self.pos + 2), Some(Token::RParen)) {
18289 return None;
18290 }
18291 Some(n)
18292 }
18293
18294 /// v7.14.0 — consume an optional MySQL display-width
18295 /// parenthesised number after an integer type, returning
18296 /// nothing. `TINYINT(1)` etc.
18297 /// v7.39 (round 360) — does the parenthesised group ahead contain a
18298 /// comma, i.e. is it MySQL's `(m,d)` rather than PG's `(p)`?
18299 fn peek_paren_has_comma(&self) -> bool {
18300 let mut i = self.pos + 1;
18301 let mut depth = 1usize;
18302 while depth > 0 {
18303 match self.tokens.get(i) {
18304 Some(Token::LParen) => depth += 1,
18305 Some(Token::RParen) => depth -= 1,
18306 Some(Token::Comma) if depth == 1 => return true,
18307 None | Some(Token::Eof) => return false,
18308 _ => {}
18309 }
18310 i += 1;
18311 }
18312 false
18313 }
18314
18315 /// v7.39 (round 424) — the same optional `(N)` modifier, but RETURNING
18316 /// the number. Temporal columns need it: MySQL's `DATETIME(3)` declares a
18317 /// fractional-seconds precision that drives write truncation and render
18318 /// padding, where `consume_optional_paren_size` throws it away.
18319 /// `Some(0)` for an explicit `(0)`, `None` when no modifier is written.
18320 fn take_optional_paren_size(&mut self) -> Option<u8> {
18321 let Some(Token::Integer(n)) = self
18322 .tokens
18323 .get(self.pos + 1)
18324 .filter(|_| matches!(self.peek(), Token::LParen))
18325 .cloned()
18326 else {
18327 self.consume_optional_paren_size();
18328 return None;
18329 };
18330 if !matches!(self.tokens.get(self.pos + 2), Some(Token::RParen)) {
18331 self.consume_optional_paren_size();
18332 return None;
18333 }
18334 self.consume_optional_paren_size();
18335 u8::try_from(n).ok()
18336 }
18337
18338 fn consume_optional_paren_size(&mut self) {
18339 if !matches!(self.peek(), Token::LParen) {
18340 return;
18341 }
18342 self.advance();
18343 // Skip until matching RParen (allow nested or any tokens).
18344 let mut depth = 1usize;
18345 while depth > 0 {
18346 match self.peek() {
18347 Token::LParen => depth += 1,
18348 Token::RParen => depth -= 1,
18349 Token::Eof => return,
18350 _ => {}
18351 }
18352 self.advance();
18353 }
18354 }
18355
18356 fn parse_paren_size(&mut self, label: &str) -> Result<u32, ParseError> {
18357 if !matches!(self.peek(), Token::LParen) {
18358 return Err(self.err(format!("{label} type requires (N), got {:?}", self.peek())));
18359 }
18360 self.advance();
18361 let n = match self.advance() {
18362 Token::Integer(n) if n > 0 => u32::try_from(n).map_err(|_| ParseError {
18363 message: format!("{label} size too large: {n}"),
18364 token_pos: self.consumed_pos(),
18365 })?,
18366 other => {
18367 return Err(ParseError {
18368 message: format!("expected positive integer {label} size, got {other:?}"),
18369 token_pos: self.consumed_pos(),
18370 });
18371 }
18372 };
18373 if !matches!(self.peek(), Token::RParen) {
18374 return Err(self.err(format!(
18375 "expected ')' after {label} size, got {:?}",
18376 self.peek()
18377 )));
18378 }
18379 self.advance();
18380 Ok(n)
18381 }
18382
18383 /// v7.39 (round 406) — the `ON CONFLICT DO NOTHING` clause that MySQL's
18384 /// `INSERT IGNORE` lowers to: a bare target (arbitrate on every unique
18385 /// key, like MySQL) whose action skips conflicting rows.
18386 /// v7.39 (round 419) — resolve the conflict clause for ANY of the four
18387 /// INSERT source forms (VALUES / SELECT / parenthesized source / WITH).
18388 /// Before this the MySQL upsert lowerings (`ON DUPLICATE KEY UPDATE`,
18389 /// `REPLACE INTO`) were wired into the VALUES branch ONLY, so the very
18390 /// common bulk-upsert spellings —
18391 /// INSERT INTO t SELECT … ON DUPLICATE KEY UPDATE c = VALUES(c)
18392 /// REPLACE INTO t SELECT …
18393 /// — were a parse error / a duplicate-key failure respectively.
18394 ///
18395 /// Precedence: an explicitly written clause beats a statement-level flag.
18396 /// `ON DUPLICATE KEY UPDATE` first, then PG's own `ON CONFLICT`, then the
18397 /// implicit `REPLACE` and `IGNORE` lowerings.
18398 fn parse_insert_conflict_clause(
18399 &mut self,
18400 replace: bool,
18401 ignore: bool,
18402 ) -> Result<Option<crate::ast::OnConflictClause>, ParseError> {
18403 if let Some(c) = self.parse_optional_on_duplicate_key()? {
18404 return Ok(Some(c));
18405 }
18406 if let Some(c) = self.parse_optional_on_conflict()? {
18407 return Ok(Some(c));
18408 }
18409 if replace {
18410 // REPLACE INTO = delete-then-insert, which PG spells as
18411 // `ON CONFLICT DO UPDATE SET` over every column; the engine
18412 // reads an empty assignment list as "take the incoming row".
18413 return Ok(Some(crate::ast::OnConflictClause {
18414 target_columns: Vec::new(),
18415 index_where: None,
18416 constraint_name: None,
18417 mysql_lowered: true,
18418 action: crate::ast::OnConflictAction::Update {
18419 assignments: Vec::new(),
18420 where_: None,
18421 },
18422 }));
18423 }
18424 if ignore {
18425 return Ok(Some(Self::insert_ignore_clause()));
18426 }
18427 Ok(None)
18428 }
18429
18430 /// v7.39 (round 419, extracted from the VALUES branch) — MySQL's
18431 /// `ON DUPLICATE KEY UPDATE col = expr [, …]`. Bare target (MySQL
18432 /// watches every unique key, which `mysql_lowered` records); `VALUES(col)`
18433 /// in an assignment is MySQL's spelling of `EXCLUDED.col`.
18434 fn parse_optional_on_duplicate_key(
18435 &mut self,
18436 ) -> Result<Option<crate::ast::OnConflictClause>, ParseError> {
18437 if !(matches!(self.peek(), Token::On)
18438 && matches!(self.tokens.get(self.pos + 1),
18439 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("duplicate")))
18440 {
18441 return Ok(None);
18442 }
18443 self.advance(); // ON
18444 self.advance(); // DUPLICATE
18445 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("key")) {
18446 return Err(self.err(format!(
18447 "expected KEY after ON DUPLICATE, got {:?}",
18448 self.peek()
18449 )));
18450 }
18451 self.advance();
18452 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("update")) {
18453 return Err(self.err(format!(
18454 "expected UPDATE after ON DUPLICATE KEY, got {:?}",
18455 self.peek()
18456 )));
18457 }
18458 self.advance();
18459 let mut assignments: Vec<(String, Expr)> = Vec::new();
18460 loop {
18461 let col = self.expect_ident_like()?;
18462 if !matches!(self.peek(), Token::Eq) {
18463 return Err(self.err(format!(
18464 "expected '=' in ON DUPLICATE KEY UPDATE, got {:?}",
18465 self.peek()
18466 )));
18467 }
18468 self.advance();
18469 let mut expr = self.parse_expr(0)?;
18470 Self::rewrite_mysql_values_refs(&mut expr);
18471 assignments.push((col, expr));
18472 if matches!(self.peek(), Token::Comma) {
18473 self.advance();
18474 continue;
18475 }
18476 break;
18477 }
18478 Ok(Some(crate::ast::OnConflictClause {
18479 target_columns: Vec::new(),
18480 index_where: None,
18481 constraint_name: None,
18482 mysql_lowered: true,
18483 action: crate::ast::OnConflictAction::Update {
18484 assignments,
18485 where_: None,
18486 },
18487 }))
18488 }
18489
18490 fn insert_ignore_clause() -> crate::ast::OnConflictClause {
18491 crate::ast::OnConflictClause {
18492 target_columns: Vec::new(),
18493 index_where: None,
18494 constraint_name: None,
18495 mysql_lowered: true,
18496 action: crate::ast::OnConflictAction::Nothing,
18497 }
18498 }
18499
18500 fn parse_insert_stmt(&mut self, replace: bool) -> Result<Statement, ParseError> {
18501 debug_assert!(
18502 matches!(self.peek(), Token::Insert)
18503 || (replace
18504 && matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("replace")))
18505 );
18506 self.advance();
18507 // v7.39 (round 406) — MySQL `INSERT IGNORE INTO t …` skips a row that
18508 // would raise a duplicate-key error instead of failing the statement,
18509 // i.e. `ON CONFLICT DO NOTHING` over every unique key. IGNORE is a
18510 // plain ident to the lexer; only the MySQL dialect accepts it here.
18511 let ignore = self.mysql_dialect
18512 && matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("ignore"));
18513 if ignore {
18514 self.advance();
18515 }
18516 if !matches!(self.peek(), Token::Into) {
18517 return Err(self.err(format!("expected INTO after INSERT, got {:?}", self.peek())));
18518 }
18519 self.advance();
18520 let table = self.expect_ident_like()?;
18521 // v7.39 (round 240) — `INSERT INTO t AS alias`: PG's insert_target
18522 // grammar requires the AS keyword here (a bare identifier would be
18523 // ambiguous with a column list). The alias is what the ON CONFLICT
18524 // DO UPDATE expressions refer to the target row by.
18525 let alias = if matches!(self.peek(), Token::As) {
18526 self.advance();
18527 Some(self.expect_ident_like()?)
18528 } else {
18529 None
18530 };
18531 // v7.39 (round 428) — MySQL's SET-form INSERT:
18532 // INSERT INTO t SET a = 1, b = 'x'
18533 // It is exactly `INSERT INTO t (a, b) VALUES (1, 'x')` — omitted
18534 // columns take their DEFAULT, `SET a = DEFAULT` is legal, and it
18535 // composes with IGNORE / ON DUPLICATE KEY UPDATE / REPLACE (all
18536 // measured). So it lowers to the column list + one VALUES row and
18537 // rejoins the ordinary path, which already handles every one of
18538 // those. PG has no such spelling, hence the dialect gate.
18539 if self.mysql_dialect
18540 && matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("set"))
18541 {
18542 self.advance(); // SET
18543 let mut names = Vec::new();
18544 let mut values = Vec::new();
18545 loop {
18546 names.push(self.expect_ident_like()?);
18547 if !matches!(self.peek(), Token::Eq) {
18548 return Err(self.err(alloc::format!(
18549 "expected '=' in INSERT … SET, got {:?}",
18550 self.peek()
18551 )));
18552 }
18553 self.advance();
18554 // `SET a = DEFAULT` rides the same `__column_default` marker
18555 // the VALUES-row and UPDATE-SET paths use; the INSERT
18556 // executor resolves it against the target column.
18557 if matches!(self.peek(), Token::Default) {
18558 self.advance();
18559 values.push(Expr::FunctionCall {
18560 name: "__column_default".to_string(),
18561 args: Vec::new(),
18562 });
18563 } else {
18564 values.push(self.parse_expr(0)?);
18565 }
18566 if matches!(self.peek(), Token::Comma) {
18567 self.advance();
18568 continue;
18569 }
18570 break;
18571 }
18572 let on_conflict = self.parse_insert_conflict_clause(replace, ignore)?;
18573 let returning = self.parse_optional_returning()?;
18574 return Ok(Statement::Insert(InsertStatement {
18575 ctes: Vec::new(),
18576 table,
18577 alias,
18578 columns: Some(names),
18579 rows: alloc::vec![values],
18580 select_source: None,
18581 // MySQL's SET form has no `OVERRIDING …` clause (that is
18582 // PG's identity-column spelling).
18583 overriding: Overriding::None,
18584 mysql_ignore: ignore,
18585 on_conflict,
18586 returning,
18587 }));
18588 }
18589 // Optional column list — `INSERT INTO t (a, b) VALUES ...`.
18590 // v7.39 (round 151) — a SELECT or WITH right after the paren is
18591 // a parenthesized query source instead (PG select_with_parens:
18592 // `INSERT INTO t (SELECT …)` / `INSERT INTO t (WITH … SELECT …)`);
18593 // both keywords are reserved in PG, so no column list can start
18594 // with them.
18595 let columns = if matches!(self.peek(), Token::LParen) {
18596 self.advance();
18597 if matches!(self.peek(), Token::Select) || self.peek_is_with_kw() {
18598 let select_stmt = if self.peek_is_with_kw() {
18599 self.advance();
18600 self.parse_nested_with_select()?
18601 } else {
18602 match self.parse_select_stmt()? {
18603 Statement::Select(s) => s,
18604 other => {
18605 return Err(self.err(alloc::format!(
18606 "expected SELECT in parenthesized INSERT source, got {other:?}"
18607 )));
18608 }
18609 }
18610 };
18611 if !matches!(self.peek(), Token::RParen) {
18612 return Err(self.err(format!(
18613 "expected ')' after parenthesized INSERT source, got {:?}",
18614 self.peek()
18615 )));
18616 }
18617 self.advance();
18618 let on_conflict = self.parse_insert_conflict_clause(replace, ignore)?;
18619 let returning = self.parse_optional_returning()?;
18620 return Ok(Statement::Insert(InsertStatement {
18621 ctes: Vec::new(),
18622 table,
18623 alias: alias.clone(),
18624 columns: None,
18625 rows: Vec::new(),
18626 select_source: Some(Box::new(select_stmt)),
18627 on_conflict,
18628 returning,
18629 overriding: Overriding::None,
18630 mysql_ignore: ignore,
18631 }));
18632 }
18633 let mut names = Vec::new();
18634 loop {
18635 names.push(self.expect_ident_like()?);
18636 match self.peek() {
18637 Token::Comma => {
18638 self.advance();
18639 }
18640 Token::RParen => {
18641 self.advance();
18642 break;
18643 }
18644 other => {
18645 return Err(self.err(format!(
18646 "expected ',' or ')' in INSERT column list, got {other:?}"
18647 )));
18648 }
18649 }
18650 }
18651 Some(names)
18652 } else {
18653 None
18654 };
18655 // PG 10+ `OVERRIDING {SYSTEM | USER} VALUE` — pg_dump emits
18656 // OVERRIDING SYSTEM VALUE for its identity columns. The clause
18657 // is captured on the statement so the engine can apply PG's
18658 // GENERATED ALWAYS / BY DEFAULT interaction (v7.38, read01).
18659 let overriding = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("overriding"))
18660 {
18661 self.advance();
18662 let which = self.expect_ident_like()?;
18663 let ov = if which.eq_ignore_ascii_case("system") {
18664 Overriding::System
18665 } else if which.eq_ignore_ascii_case("user") {
18666 Overriding::User
18667 } else {
18668 return Err(self.err(format!(
18669 "expected SYSTEM or USER after OVERRIDING, got {which:?}"
18670 )));
18671 };
18672 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("value")) {
18673 return Err(self.err(format!(
18674 "expected VALUE after OVERRIDING {}, got {:?}",
18675 which.to_ascii_uppercase(),
18676 self.peek()
18677 )));
18678 }
18679 self.advance();
18680 ov
18681 } else {
18682 Overriding::None
18683 };
18684 // `INSERT INTO t DEFAULT VALUES` — a single row made
18685 // entirely of column defaults. Lower to the permuted
18686 // column-list path with an empty list: every schema column
18687 // is unmapped, so the engine fills each from its default
18688 // (serials advance, plain defaults evaluate, the rest NULL).
18689 if matches!(self.peek(), Token::Default) {
18690 self.advance();
18691 if !matches!(self.peek(), Token::Values) {
18692 return Err(self.err(format!(
18693 "expected VALUES after DEFAULT in INSERT, got {:?}",
18694 self.peek()
18695 )));
18696 }
18697 self.advance();
18698 if columns.is_some() {
18699 return Err(self.err("DEFAULT VALUES cannot follow an INSERT column list".into()));
18700 }
18701 let on_conflict = self.parse_insert_conflict_clause(replace, ignore)?;
18702 let returning = self.parse_optional_returning()?;
18703 return Ok(Statement::Insert(InsertStatement {
18704 ctes: Vec::new(),
18705 table,
18706 alias: alias.clone(),
18707 columns: Some(Vec::new()),
18708 rows: alloc::vec![Vec::new()],
18709 select_source: None,
18710 on_conflict,
18711 returning,
18712 overriding,
18713 mysql_ignore: ignore,
18714 }));
18715 }
18716 // v7.13.0 — `INSERT INTO t [(cols)] SELECT …` (mailrs
18717 // round-5 G4). Dispatch on VALUES vs SELECT. v7.39 (round 151)
18718 // — a WITH-headed source query (`INSERT INTO t WITH c AS (…)
18719 // SELECT …`) heads the SOURCE select, as in PG (the statement's
18720 // own WITH comes before INSERT).
18721 if matches!(self.peek(), Token::Select) || self.peek_is_with_kw() {
18722 let select_stmt = if self.peek_is_with_kw() {
18723 self.advance();
18724 self.parse_nested_with_select()?
18725 } else {
18726 match self.parse_select_stmt()? {
18727 Statement::Select(s) => s,
18728 other => {
18729 return Err(self.err(alloc::format!(
18730 "expected SELECT after INSERT INTO ... target, got {other:?}"
18731 )));
18732 }
18733 }
18734 };
18735 let on_conflict = self.parse_insert_conflict_clause(replace, ignore)?;
18736 let returning = self.parse_optional_returning()?;
18737 return Ok(Statement::Insert(InsertStatement {
18738 ctes: Vec::new(),
18739 table,
18740 alias: alias.clone(),
18741 columns,
18742 rows: Vec::new(),
18743 select_source: Some(Box::new(select_stmt)),
18744 on_conflict,
18745 returning,
18746 overriding,
18747 mysql_ignore: ignore,
18748 }));
18749 }
18750 if !matches!(self.peek(), Token::Values) {
18751 return Err(self.err(format!(
18752 "expected VALUES or SELECT after table name, got {:?}",
18753 self.peek()
18754 )));
18755 }
18756 self.advance();
18757 if !matches!(self.peek(), Token::LParen) {
18758 return Err(self.err(format!("expected '(' after VALUES, got {:?}", self.peek())));
18759 }
18760 let mut rows = Vec::new();
18761 loop {
18762 // Each iteration consumes one `(expr, expr, …)` tuple.
18763 if !matches!(self.peek(), Token::LParen) {
18764 return Err(self.err(format!(
18765 "expected '(' for next VALUES tuple, got {:?}",
18766 self.peek()
18767 )));
18768 }
18769 self.advance();
18770 let mut tuple = Vec::new();
18771 loop {
18772 // v7.38 (read01) — `INSERT INTO t VALUES (…, DEFAULT, …)` uses
18773 // the column's declared default for that slot. Rides out as the
18774 // same `__column_default` marker call the UPDATE `SET c = DEFAULT`
18775 // path uses; the INSERT executor resolves it per target column.
18776 if matches!(self.peek(), Token::Default) {
18777 self.advance();
18778 tuple.push(Expr::FunctionCall {
18779 name: "__column_default".to_string(),
18780 args: Vec::new(),
18781 });
18782 } else {
18783 tuple.push(self.parse_expr(0)?);
18784 }
18785 match self.peek() {
18786 Token::Comma => {
18787 self.advance();
18788 }
18789 Token::RParen => {
18790 self.advance();
18791 break;
18792 }
18793 other => {
18794 return Err(self.err(format!(
18795 "expected ',' or ')' in VALUES tuple, got {other:?}"
18796 )));
18797 }
18798 }
18799 }
18800 if tuple.is_empty() {
18801 return Err(self.err("INSERT VALUES tuple requires at least one value".into()));
18802 }
18803 rows.push(tuple);
18804 // Continue with comma-separated tuples.
18805 if matches!(self.peek(), Token::Comma) {
18806 self.advance();
18807 } else {
18808 break;
18809 }
18810 }
18811 // MySQL `ON DUPLICATE KEY UPDATE col = expr [, …]` — lowers
18812 // to ON CONFLICT DO UPDATE with an empty conflict target
18813 // (the engine picks the table's first unique index, which
18814 // matches MySQL's any-unique-key behaviour for the common
18815 // single-key case). `VALUES(col)` in the assignments is
18816 // MySQL's spelling of EXCLUDED.col.
18817 let on_conflict = self.parse_insert_conflict_clause(replace, ignore)?;
18818 let returning = self.parse_optional_returning()?;
18819 Ok(Statement::Insert(InsertStatement {
18820 ctes: Vec::new(),
18821 table,
18822 alias,
18823 columns,
18824 rows,
18825 select_source: None,
18826 on_conflict,
18827 returning,
18828 overriding,
18829 mysql_ignore: ignore,
18830 }))
18831 }
18832
18833 /// MySQL's `VALUES(col)` inside ON DUPLICATE KEY UPDATE reads
18834 /// the incoming row's value — exactly PG's EXCLUDED.col.
18835 fn rewrite_mysql_values_refs(e: &mut Expr) {
18836 match e {
18837 Expr::FunctionCall { name, args }
18838 if name.eq_ignore_ascii_case("values")
18839 && args.len() == 1
18840 && matches!(&args[0], Expr::Column(c) if c.qualifier.is_none()) =>
18841 {
18842 let Expr::Column(c) = &args[0] else {
18843 unreachable!("guarded above");
18844 };
18845 *e = Expr::Column(crate::ast::ColumnName {
18846 qualifier: Some("EXCLUDED".to_string()),
18847 name: c.name.clone(),
18848 });
18849 }
18850 Expr::FunctionCall { args, .. } => {
18851 for a in args {
18852 Self::rewrite_mysql_values_refs(a);
18853 }
18854 }
18855 Expr::Binary { lhs, rhs, .. } => {
18856 Self::rewrite_mysql_values_refs(lhs);
18857 Self::rewrite_mysql_values_refs(rhs);
18858 }
18859 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } => {
18860 Self::rewrite_mysql_values_refs(expr);
18861 }
18862 Expr::Case {
18863 operand,
18864 branches,
18865 else_branch,
18866 } => {
18867 if let Some(op) = operand {
18868 Self::rewrite_mysql_values_refs(op);
18869 }
18870 for (w, t) in branches {
18871 Self::rewrite_mysql_values_refs(w);
18872 Self::rewrite_mysql_values_refs(t);
18873 }
18874 if let Some(el) = else_branch {
18875 Self::rewrite_mysql_values_refs(el);
18876 }
18877 }
18878 _ => {}
18879 }
18880 }
18881
18882 /// v7.9.7 — parse the optional `ON CONFLICT (cols) DO …`
18883 /// clause sitting between the INSERT body and the trailing
18884 /// RETURNING. All keywords come in as bare idents; `ON` is
18885 /// a reserved Token though.
18886 fn parse_optional_on_conflict(
18887 &mut self,
18888 ) -> Result<Option<crate::ast::OnConflictClause>, ParseError> {
18889 if !matches!(self.peek(), Token::On) {
18890 return Ok(None);
18891 }
18892 // Peek further: we want exactly "ON CONFLICT ...". If the
18893 // next ident isn't "conflict", let some other parser handle.
18894 let next_is_conflict = matches!(
18895 self.tokens.get(self.pos + 1),
18896 Some(Token::Ident(s) | Token::QuotedIdent(s)) if s.eq_ignore_ascii_case("conflict")
18897 );
18898 if !next_is_conflict {
18899 return Ok(None);
18900 }
18901 self.advance(); // ON
18902 self.advance(); // CONFLICT
18903 // v7.37.17 (17.6 siblings) — `ON CONSTRAINT <name>` names
18904 // the constraint instead of listing columns (the pg_dump
18905 // form); the engine resolves it.
18906 let mut constraint_name: Option<String> = None;
18907 if matches!(self.peek(), Token::On) {
18908 self.advance(); // ON
18909 match self.advance() {
18910 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("constraint") => {
18911 }
18912 other => {
18913 return Err(self.err(alloc::format!(
18914 "expected CONSTRAINT after ON CONFLICT ON, got {other:?}"
18915 )));
18916 }
18917 }
18918 constraint_name = Some(self.expect_ident_like()?);
18919 }
18920 // Optional `(col [, col]*)` target list.
18921 let mut target_columns: Vec<String> = Vec::new();
18922 if matches!(self.peek(), Token::LParen) {
18923 self.advance();
18924 loop {
18925 target_columns.push(self.expect_ident_like()?);
18926 match self.peek() {
18927 Token::Comma => {
18928 self.advance();
18929 }
18930 Token::RParen => {
18931 self.advance();
18932 break;
18933 }
18934 other => {
18935 return Err(self.err(alloc::format!(
18936 "expected ',' or ')' in ON CONFLICT target list, got {other:?}"
18937 )));
18938 }
18939 }
18940 }
18941 }
18942 // v7.39 (round 240) — optional index predicate after the target
18943 // list: `ON CONFLICT (col) WHERE pred DO …`. PG uses it to infer a
18944 // PARTIAL unique index; SPG's arbiters are full indexes, which
18945 // satisfy any predicate, so it is parsed and carried but not
18946 // consulted (recorded residual: partial-unique-index arbiters).
18947 let index_where = if !target_columns.is_empty() && matches!(self.peek(), Token::Where) {
18948 self.advance();
18949 Some(self.parse_expr(0)?)
18950 } else {
18951 None
18952 };
18953 // Required `DO`.
18954 match self.advance() {
18955 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("do") => {}
18956 other => {
18957 return Err(self.err(alloc::format!(
18958 "expected DO after ON CONFLICT [(…)], got {other:?}"
18959 )));
18960 }
18961 }
18962 // Action: NOTHING | UPDATE SET …
18963 let action = match self.advance() {
18964 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("nothing") => {
18965 crate::ast::OnConflictAction::Nothing
18966 }
18967 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("update") => {
18968 self.parse_on_conflict_update_action()?
18969 }
18970 other => {
18971 return Err(self.err(alloc::format!(
18972 "expected NOTHING or UPDATE after ON CONFLICT DO, got {other:?}"
18973 )));
18974 }
18975 };
18976 Ok(Some(crate::ast::OnConflictClause {
18977 target_columns,
18978 index_where,
18979 constraint_name,
18980 mysql_lowered: false,
18981 action,
18982 }))
18983 }
18984
18985 /// v7.9.7 — tail of `ON CONFLICT … DO UPDATE`: parse
18986 /// `SET col = expr [, …] [WHERE cond]`. Caller already
18987 /// consumed `UPDATE`.
18988 fn parse_on_conflict_update_action(
18989 &mut self,
18990 ) -> Result<crate::ast::OnConflictAction, ParseError> {
18991 // `SET`
18992 match self.advance() {
18993 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("set") => {}
18994 other => {
18995 return Err(self.err(alloc::format!(
18996 "expected SET after ON CONFLICT DO UPDATE, got {other:?}"
18997 )));
18998 }
18999 }
19000 let mut assignments: Vec<(String, Expr)> = Vec::new();
19001 loop {
19002 let col = self.expect_ident_like()?;
19003 if !matches!(self.peek(), Token::Eq) {
19004 return Err(self.err(alloc::format!(
19005 "expected `=` after column in ON CONFLICT DO UPDATE SET, got {:?}",
19006 self.peek()
19007 )));
19008 }
19009 self.advance();
19010 let value = self.parse_expr(0)?;
19011 assignments.push((col, value));
19012 if matches!(self.peek(), Token::Comma) {
19013 self.advance();
19014 continue;
19015 }
19016 break;
19017 }
19018 let where_ = if matches!(self.peek(), Token::Where) {
19019 self.advance();
19020 Some(self.parse_expr(0)?)
19021 } else {
19022 None
19023 };
19024 Ok(crate::ast::OnConflictAction::Update {
19025 assignments,
19026 where_,
19027 })
19028 }
19029
19030 fn parse_select_list(&mut self) -> Result<Vec<SelectItem>, ParseError> {
19031 let mut items = Vec::new();
19032 // v7.39 (round 341, V66) — PG's target list may be EMPTY
19033 // (`opt_target_list: target_list | /*EMPTY*/`): `SELECT FROM t`
19034 // answers one zero-column row per row of t, and a bare `SELECT`
19035 // answers a single zero-column row. SPG required at least one
19036 // item, so both were syntax errors. Recognised by the token that
19037 // follows — nothing that can start an expression appears here.
19038 if self.select_list_is_empty_here() {
19039 return Ok(items);
19040 }
19041 loop {
19042 items.push(self.parse_select_item()?);
19043 if matches!(self.peek(), Token::Comma) {
19044 self.advance();
19045 } else {
19046 break;
19047 }
19048 }
19049 Ok(items)
19050 }
19051
19052 /// Is the target list empty at this point — i.e. does the next token
19053 /// end the SELECT's item list rather than start an item?
19054 fn select_list_is_empty_here(&self) -> bool {
19055 match self.peek() {
19056 Token::From
19057 | Token::Where
19058 | Token::Group
19059 | Token::Having
19060 | Token::Order
19061 | Token::Limit
19062 | Token::Offset
19063 | Token::Semicolon
19064 | Token::RParen
19065 | Token::Union
19066 | Token::Except
19067 | Token::Eof => true,
19068 // `FETCH FIRST … ROWS ONLY` and `WINDOW w AS …` are spelled
19069 // with unreserved keywords, so they arrive as plain idents.
19070 Token::Ident(s) => {
19071 s.eq_ignore_ascii_case("fetch")
19072 || s.eq_ignore_ascii_case("window")
19073 || s.eq_ignore_ascii_case("intersect")
19074 }
19075 _ => false,
19076 }
19077 }
19078
19079 fn parse_select_item(&mut self) -> Result<SelectItem, ParseError> {
19080 if matches!(self.peek(), Token::Star) {
19081 self.advance();
19082 return Ok(SelectItem::Wildcard);
19083 }
19084 // v7.39 (read01 round 128) — qualified wildcard `qualifier.*`. Intercept
19085 // BEFORE `parse_expr`, which would treat `q.` as a qualified column and
19086 // choke on the `*` ("expected identifier, got Star"). The lookahead is
19087 // `<ident> . *` with nothing binding tighter.
19088 if let Token::Ident(q) | Token::QuotedIdent(q) = self.peek().clone() {
19089 if matches!(self.tokens.get(self.pos + 1), Some(Token::Dot))
19090 && matches!(self.tokens.get(self.pos + 2), Some(Token::Star))
19091 {
19092 self.advance(); // qualifier
19093 self.advance(); // .
19094 self.advance(); // *
19095 return Ok(SelectItem::QualifiedWildcard(q));
19096 }
19097 }
19098 let start_tok = self.pos;
19099 let expr = self.parse_expr(0)?;
19100 let end_tok = self.consumed_pos();
19101 // v7.39 (read01 round 69) — `(f(args)).*`: expand the RECORD a
19102 // multi-column function returns into columns. Marked here and lowered in
19103 // `parse_bare_select`, where the FROM clause is in hand.
19104 if matches!(self.peek(), Token::Dot)
19105 && matches!(self.tokens.get(self.pos + 1), Some(Token::Star))
19106 {
19107 self.advance(); // .
19108 self.advance(); // *
19109 return Ok(SelectItem::Expr {
19110 expr: Expr::FunctionCall {
19111 name: "__record_expand".to_string(),
19112 args: alloc::vec![expr],
19113 },
19114 alias: None,
19115 });
19116 }
19117 // v7.39.2 — MySQL lets a STRING name a projection item, with or
19118 // without `AS`: `SELECT 1 'x'`, `SELECT COUNT(*) 'total'`,
19119 // `SELECT 1 'a b'` (which is why one quotes it). SPG answered
19120 // `syntax error at or near "'x'"` to all of them.
19121 //
19122 // Only here, not in `parse_optional_alias`: that one also names
19123 // TABLES, and MySQL 9.7.2 refuses a string there — `FROM t 'ta'`
19124 // and `FROM t AS 'ta'` are both syntax errors, measured. And only
19125 // after the lexer's own rule has joined adjacent literals, or
19126 // `SELECT 'a' 'b'` would read as a literal aliased `b` where
19127 // MySQL answers the concatenation `ab`.
19128 if self.mysql_dialect {
19129 let at_as = matches!(self.peek(), Token::As)
19130 && matches!(self.tokens.get(self.pos + 1), Some(Token::String(_)));
19131 if at_as {
19132 self.advance();
19133 }
19134 if let Token::String(name) = self.peek().clone() {
19135 self.advance();
19136 return Ok(SelectItem::Expr {
19137 expr,
19138 alias: Some(name),
19139 });
19140 }
19141 }
19142 let alias = match self.parse_optional_alias()? {
19143 Some(a) => Some(a),
19144 None => self.mysql_item_label(&expr, start_tok, end_tok),
19145 };
19146 Ok(SelectItem::Expr { expr, alias })
19147 }
19148
19149 /// v7.39 (round 506) — the name MariaDB 11 gives a projection item that
19150 /// carries no `AS`, filled in here so every downstream path reports it
19151 /// without knowing the rule. `None` leaves the item un-aliased, which is
19152 /// what a PG session always gets.
19153 ///
19154 /// Measured against MariaDB 11, three rules and no more:
19155 ///
19156 /// | item | label | why |
19157 /// |------------------|------------|------------------------------|
19158 /// | `lbl.a` | `a` | a column reports its name |
19159 /// | `'it''s'` | `it's` | a string reports its VALUE |
19160 /// | `a + b` | `a + b` | anything else, source text |
19161 ///
19162 /// The third is why this lives in the parser at all: the label is the
19163 /// text the client WROTE, down to the spacing, so it cannot be printed
19164 /// back out of the parsed shape. `COUNT( * )` names itself `COUNT( * )`.
19165 ///
19166 /// Comments survive, and that is right: through a `mariadb` CLI both
19167 /// servers answer `a + b` for `SELECT a /* c */ + b`, but that is the
19168 /// CLIENT stripping the comment before it sends. Asked over the raw
19169 /// protocol, MariaDB answers `a /* c */ + b` — byte for byte what this
19170 /// produces.
19171 fn mysql_item_label(&self, expr: &Expr, start_tok: usize, end_tok: usize) -> Option<String> {
19172 if !self.mysql_dialect {
19173 return None;
19174 }
19175 match expr {
19176 // A column already reports its own name downstream; naming it
19177 // again here would only re-state the qualifier the label drops.
19178 Expr::Column(_) => None,
19179 // v7.39.3 — `SELECT 'a' 'b'` is ONE literal whose value is
19180 // `ab`, and MySQL 9.7.2 names the column `a`: the label is
19181 // the first segment as written, not the joined value
19182 // (measured). The lexer logs where it joined them.
19183 Expr::Literal(Literal::String(v)) => Some(
19184 self.merged_first_len(start_tok)
19185 .and_then(|n| v.get(..n))
19186 .map_or_else(|| v.clone(), String::from),
19187 ),
19188 _ => self.source_span(start_tok, end_tok).map(str::to_string),
19189 }
19190 }
19191
19192 /// v7.37.17 (17.6 siblings) — parse `(row), (row), …` after a
19193 /// consumed VALUES keyword. Each row lowers to a constant SELECT
19194 /// with PG's default column1..columnN names; subsequent rows
19195 /// chain as UNION ALL peers. Shared by the FROM-position
19196 /// `( VALUES … )` arm and the top-level bare VALUES statement.
19197 fn parse_values_rows_body(&mut self) -> Result<SelectStatement, ParseError> {
19198 let mut row_selects: Vec<SelectStatement> = Vec::new();
19199 loop {
19200 if !matches!(self.peek(), Token::LParen) {
19201 return Err(self.err(alloc::format!(
19202 "expected '(' to start a VALUES row, got {:?}",
19203 self.peek()
19204 )));
19205 }
19206 self.advance(); // (
19207 let mut items: Vec<SelectItem> = Vec::new();
19208 loop {
19209 let expr = self.parse_expr(0)?;
19210 items.push(SelectItem::Expr {
19211 expr,
19212 alias: Some(alloc::format!("column{}", items.len() + 1)),
19213 });
19214 match self.peek() {
19215 Token::Comma => {
19216 self.advance();
19217 }
19218 Token::RParen => break,
19219 other => {
19220 return Err(self.err(alloc::format!(
19221 "expected ',' or ')' in VALUES row, got {other:?}"
19222 )));
19223 }
19224 }
19225 }
19226 self.advance(); // )
19227 row_selects.push(SelectStatement {
19228 locking: None,
19229 ctes: Vec::new(),
19230 distinct: false,
19231 distinct_on: Vec::new(),
19232 items,
19233 from: None,
19234 where_: None,
19235 group_by: None,
19236 group_by_all: false,
19237 having: None,
19238 unions: Vec::new(),
19239 order_by: Vec::new(),
19240 limit: None,
19241 offset: None,
19242 limit_with_ties: false,
19243 window_check_exprs: Vec::new(),
19244 });
19245 if matches!(self.peek(), Token::Comma) {
19246 self.advance();
19247 continue;
19248 }
19249 break;
19250 }
19251 let mut head = row_selects.remove(0);
19252 head.unions = row_selects
19253 .into_iter()
19254 .map(|s| (UnionKind::All, s))
19255 .collect();
19256 Ok(head)
19257 }
19258
19259 fn parse_table_ref(&mut self) -> Result<TableRef, ParseError> {
19260 // v7.39 (round 621) — `FROM ONLY <table>` excludes a table's
19261 // children. It was read as a table NAMED `only`, so the query
19262 // failed on `relation "only" does not exist`.
19263 //
19264 // v7.39 (round 644) — and it is no longer a no-op. Round 621
19265 // absorbed the keyword, reasoning that SPG's children are
19266 // separate relations a plain scan does not descend into, so ONLY
19267 // already described the scan. That stopped being true when a
19268 // partition parent started unioning its children: measured,
19269 // `SELECT count(*) FROM ONLY <partitioned parent>` answered 2
19270 // where PG answers 0. The flag is carried now.
19271 let mut only = false;
19272 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("only"))
19273 && matches!(
19274 self.tokens.get(self.pos + 1),
19275 Some(Token::Ident(_) | Token::QuotedIdent(_))
19276 )
19277 {
19278 only = true;
19279 self.advance();
19280 }
19281 // `LATERAL generate_series(...)` / `LATERAL unnest(...)` —
19282 // for these SRFs the keyword is noise at parse time: the
19283 // join executor already substitutes outer-column references
19284 // into unnest_expr / generate_series_args per outer row
19285 // (v7.37.43-T4.5 substitute_outer_in_table_ref), and PG
19286 // licences the correlation even without the keyword. Absorb
19287 // it and fall through to the SRF arms below.
19288 // v7.39 (read01 round 69) — `LATERAL <fn>(args)` for ANY function, not
19289 // just the four builtin SRFs: a user set-returning function on a JOIN's
19290 // right side is the whole point of LATERAL. The keyword stays noise at
19291 // parse time — the join executor substitutes the outer row into the
19292 // call's arguments per outer row.
19293 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("lateral"))
19294 && matches!(
19295 self.tokens.get(self.pos + 1),
19296 // The json_each family has its OWN `LATERAL …` arm below, which
19297 // needs to see the keyword — absorbing it here would send those
19298 // calls down the generic table-function channel instead.
19299 Some(Token::Ident(s) | Token::QuotedIdent(s)) if !is_json_each_name(s)
19300 )
19301 && matches!(self.tokens.get(self.pos + 2), Some(Token::LParen))
19302 {
19303 self.advance(); // LATERAL
19304 }
19305 // v7.37.43-T4.5 — `LATERAL jsonb_each_text(<expr>)` —
19306 // set-returning function whose argument may reference a
19307 // preceding FROM item. We rewrite this to
19308 // `LATERAL (SELECT key, value FROM jsonb_each_text(<expr>)
19309 // AS __srf__) AS <alias>` so the existing LATERAL subquery
19310 // executor handles per-outer-row evaluation and the
19311 // SRF-primary jsonb_each_text path handles the inner
19312 // materialisation. Sentori 0067 backfill is the dogfood
19313 // shape.
19314 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("lateral"))
19315 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s) | Token::QuotedIdent(s)) if is_json_each_name(s))
19316 && matches!(self.tokens.get(self.pos + 2), Some(Token::LParen))
19317 {
19318 self.advance(); // LATERAL
19319 let each_fn = match self.peek() {
19320 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
19321 _ => unreachable!(),
19322 };
19323 self.advance(); // jsonb_each[_text] / json_each[_text]
19324 self.advance(); // (
19325 let arg = self.parse_expr(0)?;
19326 if !matches!(self.peek(), Token::RParen) {
19327 return Err(self.err(alloc::format!(
19328 "expected ')' after LATERAL {each_fn}() argument, got {:?}",
19329 self.peek()
19330 )));
19331 }
19332 self.advance();
19333 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
19334 let alias = alias_ident.clone().unwrap_or_else(|| each_fn.clone());
19335 // Synthesise: SELECT __srf__.key AS <key_alias>, __srf__.value AS <value_alias>
19336 // FROM jsonb_each_text(<arg>) AS __srf__
19337 // PG's `AS kv(key, value)` column-alias list maps
19338 // positions to names; default to (key, value) when
19339 // omitted (matching the SRF's natural column names).
19340 let srf_alias = "__srf__".to_string();
19341 let key_alias = column_aliases
19342 .first()
19343 .cloned()
19344 .unwrap_or_else(|| "key".to_string());
19345 let value_alias = column_aliases
19346 .get(1)
19347 .cloned()
19348 .unwrap_or_else(|| "value".to_string());
19349 let inner_select = crate::ast::SelectStatement {
19350 locking: None,
19351 ctes: Vec::new(),
19352 distinct: false,
19353 distinct_on: Vec::new(),
19354 items: alloc::vec![
19355 crate::ast::SelectItem::Expr {
19356 expr: crate::ast::Expr::Column(crate::ast::ColumnName {
19357 qualifier: Some(srf_alias.clone()),
19358 name: "key".to_string(),
19359 }),
19360 alias: Some(key_alias),
19361 },
19362 crate::ast::SelectItem::Expr {
19363 expr: crate::ast::Expr::Column(crate::ast::ColumnName {
19364 qualifier: Some(srf_alias.clone()),
19365 name: "value".to_string(),
19366 }),
19367 alias: Some(value_alias),
19368 },
19369 ],
19370 from: Some(crate::ast::FromClause {
19371 primary: TableRef {
19372 name: srf_alias.clone(),
19373 alias: Some(srf_alias.clone()),
19374 only: false,
19375 as_of_segment: None,
19376 unnest_expr: None,
19377 unnest_column_aliases: Vec::new(),
19378 with_ordinality: false,
19379 generate_series_args: None,
19380 lateral_subquery: None,
19381 jsonb_each_text_arg: Some((each_fn, Box::new(arg))),
19382 table_fn_call: None,
19383 rows_from: None,
19384 json_table: None,
19385 scalar_fn_item: false,
19386 },
19387 joins: Vec::new(),
19388 }),
19389 where_: None,
19390 group_by: None,
19391 group_by_all: false,
19392 having: None,
19393 unions: Vec::new(),
19394 order_by: Vec::new(),
19395 limit: None,
19396 offset: None,
19397 limit_with_ties: false,
19398 window_check_exprs: Vec::new(),
19399 };
19400 return Ok(TableRef {
19401 name: alias.clone(),
19402 alias: Some(alias),
19403 only: false,
19404 as_of_segment: None,
19405 unnest_expr: None,
19406 unnest_column_aliases: Vec::new(),
19407 with_ordinality: false,
19408 generate_series_args: None,
19409 lateral_subquery: Some(Box::new(inner_select)),
19410 jsonb_each_text_arg: None,
19411 table_fn_call: None,
19412 rows_from: None,
19413 json_table: None,
19414 scalar_fn_item: false,
19415 });
19416 }
19417 // v7.37.43-T4.5 — bare `CROSS JOIN jsonb_each_text(t.col)`
19418 // without an explicit `LATERAL` keyword is the same shape
19419 // PG accepts (SRF naturally licences lateral correlation).
19420 // We mirror the LATERAL rewrite when the argument syntactic-
19421 // ally references an outer column (Column { qualifier:
19422 // Some(_), … }). For simplicity we apply the rewrite
19423 // whenever the SRF directly follows JOIN/CROSS JOIN/comma
19424 // in the FROM-list — caller-side join parsing positions
19425 // this peek correctly.
19426 // (Implementation note: detection lives below; the LATERAL
19427 // branch above already covers the explicit form; the bare
19428 // form falls through to the plain SRF arm and the engine
19429 // treats it as a constant-arg SRF if no outer reference is
19430 // present.)
19431 // v7.17.0 Phase 3.P0-41 — `LATERAL ( SELECT … )` derived
19432 // table. Detect at the head so it claims precedence over
19433 // every other table-ref shape (unnest / generate_series /
19434 // bare ident); the lateral subquery itself follows the
19435 // regular SELECT grammar.
19436 // v7.37.17 (17.6 siblings) — `FROM ( VALUES (…), (…) ) [AS]
19437 // t(cols)`. Each row lowers to a constant SELECT with PG's
19438 // default column1..columnN names; subsequent rows chain as
19439 // UNION ALL peers. The result rides the derived-table
19440 // lateral_subquery channel — zero executor work.
19441 if matches!(self.peek(), Token::LParen)
19442 && matches!(self.tokens.get(self.pos + 1), Some(Token::Values))
19443 {
19444 self.advance(); // (
19445 self.advance(); // VALUES
19446 let head = self.parse_values_rows_body()?;
19447 if !matches!(self.peek(), Token::RParen) {
19448 return Err(self.err(alloc::format!(
19449 "expected ')' after VALUES list, got {:?}",
19450 self.peek()
19451 )));
19452 }
19453 self.advance();
19454 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
19455 let name = alias_ident.clone().unwrap_or_else(|| "values".to_string());
19456 return Ok(TableRef {
19457 name,
19458 alias: alias_ident,
19459 only: false,
19460 as_of_segment: None,
19461 unnest_expr: None,
19462 unnest_column_aliases: column_aliases,
19463 with_ordinality: false,
19464 generate_series_args: None,
19465 lateral_subquery: Some(Box::new(head)),
19466 jsonb_each_text_arg: None,
19467 table_fn_call: None,
19468 rows_from: None,
19469 json_table: None,
19470 scalar_fn_item: false,
19471 });
19472 }
19473 // v7.37.17 (17.6 siblings) — plain derived table:
19474 // `FROM ( SELECT … ) [AS] alias`. Rides the same
19475 // lateral_subquery channel the explicit LATERAL form uses —
19476 // an uncorrelated inner SELECT executes identically. The
19477 // inner parse carries UNION tails (they live on
19478 // SelectStatement.unions).
19479 // v7.37 D.20 — the derived-table inner may itself be a
19480 // parenthesized set-operation group (`FROM ((SELECT…) UNION
19481 // (SELECT…)) s`) or a CTE (`FROM (WITH … SELECT …) z`), not just a
19482 // bare `(SELECT …)`. parse_one_statement already routes a leading
19483 // `(` set-op group (its LParen arm) and a leading WITH
19484 // (parse_with_cte_then_select), so widen the second-token gate to
19485 // Select | LParen | WITH.
19486 // v7.39 (round 869) — `Table` joins that gate. `TABLE t` is
19487 // PG's spelling of `SELECT * FROM t` and is accepted wherever a
19488 // SELECT is, so `FROM (TABLE t) x` has to parse. The desugaring
19489 // has existed since the shorthand landed and `parse_bare_select`
19490 // already routes it ("valid anywhere a SELECT head is"); what was
19491 // missing is this second-token gate, and the CTE body's dispatch
19492 // below. Round 868 found both by putting the shorthand in a
19493 // subquery — the top-level forms had been the only ones tested.
19494 if matches!(self.peek(), Token::LParen)
19495 && (matches!(
19496 self.tokens.get(self.pos + 1),
19497 Some(Token::Select | Token::LParen | Token::Table)
19498 ) || matches!(self.tokens.get(self.pos + 1),
19499 Some(Token::Ident(s) | Token::QuotedIdent(s)) if s.eq_ignore_ascii_case("with")))
19500 {
19501 self.advance(); // (
19502 let inner = match self.parse_one_statement()? {
19503 Statement::Select(s) => s,
19504 other => {
19505 return Err(self.err(alloc::format!(
19506 "expected SELECT inside derived table ( … ), got {other:?}"
19507 )));
19508 }
19509 };
19510 if !matches!(self.peek(), Token::RParen) {
19511 return Err(self.err(alloc::format!(
19512 "expected ')' after derived-table subquery, got {:?}",
19513 self.peek()
19514 )));
19515 }
19516 self.advance();
19517 // `AS t(a, b)` column-alias list rides the
19518 // unnest_column_aliases field (same positional-rename
19519 // contract the unnest SRFs use).
19520 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
19521 let name = alias_ident
19522 .clone()
19523 .unwrap_or_else(|| "subquery".to_string());
19524 return Ok(TableRef {
19525 name,
19526 alias: alias_ident,
19527 only: false,
19528 as_of_segment: None,
19529 unnest_expr: None,
19530 unnest_column_aliases: column_aliases,
19531 with_ordinality: false,
19532 generate_series_args: None,
19533 lateral_subquery: Some(Box::new(inner)),
19534 jsonb_each_text_arg: None,
19535 table_fn_call: None,
19536 rows_from: None,
19537 json_table: None,
19538 scalar_fn_item: false,
19539 });
19540 }
19541 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("lateral"))
19542 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
19543 {
19544 self.advance(); // LATERAL
19545 self.advance(); // (
19546 // Parse the inner SELECT.
19547 let inner = match self.parse_one_statement()? {
19548 Statement::Select(s) => s,
19549 other => {
19550 return Err(self.err(alloc::format!(
19551 "expected SELECT inside LATERAL ( … ), got {other:?}"
19552 )));
19553 }
19554 };
19555 if !matches!(self.peek(), Token::RParen) {
19556 return Err(self.err(alloc::format!(
19557 "expected ')' after LATERAL subquery, got {:?}",
19558 self.peek()
19559 )));
19560 }
19561 self.advance();
19562 // v7.37 D.28 — `LATERAL (…) AS t(cols)` column-alias list (also how a
19563 // `(VALUES …) t(g)` derived table round-trips through view-body
19564 // Display, which renders on the lateral_subquery channel).
19565 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
19566 let name = alias_ident.clone().unwrap_or_else(|| "lateral".to_string());
19567 return Ok(TableRef {
19568 name,
19569 alias: alias_ident,
19570 only: false,
19571 as_of_segment: None,
19572 unnest_expr: None,
19573 unnest_column_aliases: column_aliases,
19574 with_ordinality: false,
19575 generate_series_args: None,
19576 lateral_subquery: Some(Box::new(inner)),
19577 jsonb_each_text_arg: None,
19578 table_fn_call: None,
19579 rows_from: None,
19580 json_table: None,
19581 scalar_fn_item: false,
19582 });
19583 }
19584 // v7.37.43-T4.5 — `jsonb_each_text(<expr>)` set-returning
19585 // function as a FROM item. Emits one row per (key, value)
19586 // pair in the JSONB object argument as TEXT columns. May
19587 // be wrapped in CROSS JOIN LATERAL when the argument
19588 // references a preceding FROM item (sentori migration
19589 // 0067 backfill shape: `CROSS JOIN LATERAL
19590 // jsonb_each_text(t.json_col) AS kv(key, value)`).
19591 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if is_json_each_name(s))
19592 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
19593 {
19594 let each_fn = match self.peek() {
19595 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
19596 _ => unreachable!(),
19597 };
19598 self.advance(); // jsonb_each[_text] / json_each[_text]
19599 self.advance(); // (
19600 let arg = self.parse_expr(0)?;
19601 if !matches!(self.peek(), Token::RParen) {
19602 return Err(self.err(alloc::format!(
19603 "expected ')' after {each_fn}() argument, got {:?}",
19604 self.peek()
19605 )));
19606 }
19607 self.advance();
19608 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
19609 let name = alias_ident.clone().unwrap_or_else(|| each_fn.clone());
19610 return Ok(TableRef {
19611 name,
19612 alias: alias_ident,
19613 only: false,
19614 as_of_segment: None,
19615 unnest_expr: None,
19616 // `AS t(k, v)` renames key/value positionally, same as the
19617 // LATERAL-position form already does.
19618 unnest_column_aliases: column_aliases,
19619 with_ordinality: false,
19620 generate_series_args: None,
19621 lateral_subquery: None,
19622 jsonb_each_text_arg: Some((each_fn, Box::new(arg))),
19623 table_fn_call: None,
19624 rows_from: None,
19625 json_table: None,
19626 scalar_fn_item: false,
19627 });
19628 }
19629 // `jsonb_to_recordset(J) AS t(a int, b text)` / `jsonb_to_record`
19630 // (+ json_ variants) — record-returning JSON functions with a
19631 // column-definition list. Desugar to a derived table that
19632 // projects each declared column from the JSON via `->>` + a cast,
19633 // over `jsonb_array_elements(J)` for the *set (per-element) form.
19634 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if is_json_to_record_name(s))
19635 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
19636 {
19637 return self.parse_json_to_record_from();
19638 }
19639 // v7.38 (T15) — `regexp_matches(s, pat[, flags])` as a FROM item. Each
19640 // row is a text[] of capture groups, so it cannot desugar to unnest
19641 // (that would flatten the array). Wrap it as a derived table
19642 // `(SELECT regexp_matches(args)) AS <alias>(<col>)` — the SELECT-list
19643 // SRF path already emits one text[] row per match. PG names the column
19644 // `regexp_matches`; an `AS a(col)` alias overrides it.
19645 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
19646 if s.eq_ignore_ascii_case("regexp_matches"))
19647 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
19648 {
19649 self.advance(); // fn name
19650 self.advance(); // (
19651 let mut fn_args: Vec<Expr> = Vec::new();
19652 loop {
19653 fn_args.push(self.parse_expr(0)?);
19654 if matches!(self.peek(), Token::Comma) {
19655 self.advance();
19656 continue;
19657 }
19658 break;
19659 }
19660 if !matches!(self.peek(), Token::RParen) {
19661 return Err(self.err(alloc::format!(
19662 "expected ')' after regexp_matches() arguments, got {:?}",
19663 self.peek()
19664 )));
19665 }
19666 self.advance();
19667 // v7.39 (read01 round 78) — WITH ORDINALITY sits BEFORE the alias
19668 // (`f(x) WITH ORDINALITY AS t(v, o)`), and this arm never looked for
19669 // it, so it died on the `with` token while every other table function
19670 // accepted it.
19671 let with_ordinality = self.absorb_with_ordinality();
19672 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
19673 let table_alias = alias_ident
19674 .clone()
19675 .unwrap_or_else(|| "regexp_matches".to_string());
19676 // PG names a single-column function's output column after the ALIAS
19677 // when one is given (`FROM regexp_matches(…) AS m` → column `m`), so
19678 // `m` reads as that column and not as a whole-row composite. Naming
19679 // it after the function regardless made `SELECT m[1] FROM … AS m`
19680 // subscript a record.
19681 let col_name = column_aliases
19682 .first()
19683 .cloned()
19684 .or_else(|| alias_ident.clone())
19685 .unwrap_or_else(|| "regexp_matches".to_string());
19686 let inner = crate::ast::SelectStatement {
19687 locking: None,
19688 ctes: Vec::new(),
19689 distinct: false,
19690 distinct_on: Vec::new(),
19691 items: alloc::vec![SelectItem::Expr {
19692 expr: Expr::FunctionCall {
19693 name: "regexp_matches".to_string(),
19694 args: fn_args,
19695 },
19696 alias: Some(col_name),
19697 }],
19698 from: None,
19699 where_: None,
19700 group_by: None,
19701 group_by_all: false,
19702 having: None,
19703 unions: Vec::new(),
19704 order_by: Vec::new(),
19705 limit: None,
19706 offset: None,
19707 limit_with_ties: false,
19708 window_check_exprs: Vec::new(),
19709 };
19710 return Ok(TableRef {
19711 name: table_alias.clone(),
19712 alias: Some(table_alias),
19713 only: false,
19714 as_of_segment: None,
19715 unnest_expr: None,
19716 unnest_column_aliases: column_aliases,
19717 with_ordinality,
19718 generate_series_args: None,
19719 lateral_subquery: Some(Box::new(inner)),
19720 jsonb_each_text_arg: None,
19721 table_fn_call: None,
19722 rows_from: None,
19723 json_table: None,
19724 // regexp_matches returns text[], a base type: `SELECT m FROM
19725 // regexp_matches(…) AS m` is the array, not a composite wrapping it.
19726 scalar_fn_item: true,
19727 });
19728 }
19729 // v7.37.17 (17.6 siblings) — `jsonb_array_elements[_text](<expr>)`
19730 // / json_ variants as a FROM item. Rewritten into
19731 // `unnest(<same fn>(<expr>))`: the scalar form returns the
19732 // elements as a TEXT array, and the existing unnest SRF path
19733 // materialises one row per element. PG's natural column name
19734 // is `value`; an `AS a(col)` column-alias list overrides it.
19735 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
19736 if s.eq_ignore_ascii_case("jsonb_array_elements")
19737 || s.eq_ignore_ascii_case("json_array_elements")
19738 || s.eq_ignore_ascii_case("jsonb_array_elements_text")
19739 || s.eq_ignore_ascii_case("json_array_elements_text")
19740 || s.eq_ignore_ascii_case("jsonb_object_keys")
19741 || s.eq_ignore_ascii_case("json_object_keys")
19742 || s.eq_ignore_ascii_case("jsonb_path_query")
19743 || s.eq_ignore_ascii_case("json_path_query")
19744 || s.eq_ignore_ascii_case("generate_subscripts")
19745 || s.eq_ignore_ascii_case("string_to_table")
19746 || s.eq_ignore_ascii_case("regexp_split_to_table"))
19747 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
19748 {
19749 let fn_name = match self.peek() {
19750 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
19751 _ => unreachable!(),
19752 };
19753 self.advance(); // fn name
19754 self.advance(); // (
19755 let mut fn_args: Vec<Expr> = Vec::new();
19756 loop {
19757 fn_args.push(self.parse_expr(0)?);
19758 if matches!(self.peek(), Token::Comma) {
19759 self.advance();
19760 continue;
19761 }
19762 break;
19763 }
19764 if !matches!(self.peek(), Token::RParen) {
19765 return Err(self.err(alloc::format!(
19766 "expected ')' after {fn_name}() arguments, got {:?}",
19767 self.peek()
19768 )));
19769 }
19770 self.advance();
19771 let with_ordinality = self.absorb_with_ordinality();
19772 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
19773 let name = alias_ident.clone().unwrap_or_else(|| fn_name.clone());
19774 // PG's natural column name: the array-elements SRFs
19775 // declare an OUT parameter `value`; jsonb_object_keys
19776 // and generate_subscripts have none, so the column is
19777 // named after the function. A bare table alias on a
19778 // single-column SRF renames the column too (PG: `FROM
19779 // generate_subscripts(a, 1) AS s` projects column s) —
19780 // except for the OUT-parameter SRFs, whose column stays
19781 // `value` under a bare alias.
19782 let natural_col = if fn_name.ends_with("_array_elements")
19783 || fn_name.ends_with("_array_elements_text")
19784 {
19785 "value".to_string()
19786 } else {
19787 alias_ident.clone().unwrap_or_else(|| fn_name.clone())
19788 };
19789 let mut srf_cols = alloc::vec![column_aliases.first().cloned().unwrap_or(natural_col)];
19790 // Keep any further entries — the second names the
19791 // ordinality column under WITH ORDINALITY.
19792 srf_cols.extend(column_aliases.into_iter().skip(1));
19793 // The *_to_table SRFs are row-streams over the existing
19794 // *_to_array scalars — map the call target; the display
19795 // name (alias / column defaults) keeps the SRF spelling.
19796 let call_name = match fn_name.as_str() {
19797 "string_to_table" => "string_to_array".to_string(),
19798 "regexp_split_to_table" => "regexp_split_to_array".to_string(),
19799 _ => fn_name,
19800 };
19801 // v7.38 (read01, T-srf/T-lateral) — an SRF argument that references a
19802 // preceding FROM item (bare or qualified column) is correlated;
19803 // route it through the per-outer-row lateral channel.
19804 let expr = crate::ast::Expr::FunctionCall {
19805 name: call_name,
19806 args: fn_args,
19807 };
19808 let correlated = Self::expr_has_any_column(&expr);
19809 let tref = TableRef {
19810 name,
19811 alias: alias_ident,
19812 only: false,
19813 as_of_segment: None,
19814 unnest_expr: Some(Box::new(expr)),
19815 unnest_column_aliases: srf_cols,
19816 with_ordinality,
19817 generate_series_args: None,
19818 lateral_subquery: None,
19819 jsonb_each_text_arg: None,
19820 table_fn_call: None,
19821 rows_from: None,
19822 json_table: None,
19823 // Each of these returns a BASE type (jsonb / text / int), so the item's
19824 // row type is that scalar: `SELECT j FROM jsonb_array_elements('[1]') j`
19825 // is `1`, not `(1)`. WITH ORDINALITY makes it a real two-column item.
19826 scalar_fn_item: !with_ordinality,
19827 };
19828 return Ok(if correlated {
19829 Self::wrap_correlated_srf(tref)
19830 } else {
19831 tref
19832 });
19833 }
19834 // `ROWS FROM ( srf(args) [, srf(args)]* )` — SQL-standard
19835 // explicit parallel-zip syntax. Each entry lowers to its
19836 // array-returning scalar form (unnest(x) → x itself; the
19837 // FROM-SRF rewrite family → their scalar array calls) and
19838 // the list rides the multi-arg unnest zip channel:
19839 // NULL-padded to the longest, WITH ORDINALITY appends the
19840 // counter. generate_series has no scalar array form and
19841 // errors honestly.
19842 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("rows"))
19843 && matches!(self.tokens.get(self.pos + 1), Some(Token::From))
19844 && matches!(self.tokens.get(self.pos + 2), Some(Token::LParen))
19845 {
19846 self.advance(); // ROWS
19847 self.advance(); // FROM
19848 self.advance(); // (
19849 let mut entries: Vec<Expr> = Vec::new();
19850 // v7.39 (read01 round 74) — the generic channel, filled in parallel.
19851 // Used only when some entry has no array form.
19852 let mut generic: Vec<(String, Vec<Expr>)> = Vec::new();
19853 loop {
19854 let fn_name = self.expect_ident_like()?.to_ascii_lowercase();
19855 if !matches!(self.peek(), Token::LParen) {
19856 return Err(self.err(alloc::format!(
19857 "expected '(' after {fn_name} in ROWS FROM, got {:?}",
19858 self.peek()
19859 )));
19860 }
19861 self.advance();
19862 let mut fn_args: Vec<Expr> = Vec::new();
19863 if !matches!(self.peek(), Token::RParen) {
19864 loop {
19865 fn_args.push(self.parse_expr(0)?);
19866 if matches!(self.peek(), Token::Comma) {
19867 self.advance();
19868 continue;
19869 }
19870 break;
19871 }
19872 }
19873 if !matches!(self.peek(), Token::RParen) {
19874 return Err(self.err(alloc::format!(
19875 "expected ')' after {fn_name}() arguments in ROWS FROM, got {:?}",
19876 self.peek()
19877 )));
19878 }
19879 self.advance();
19880 let entry = match fn_name.as_str() {
19881 "unnest" => {
19882 if fn_args.len() != 1 {
19883 return Err(
19884 self.err("unnest inside ROWS FROM takes exactly one array".into())
19885 );
19886 }
19887 fn_args.pop().expect("len checked")
19888 }
19889 "jsonb_array_elements"
19890 | "json_array_elements"
19891 | "jsonb_array_elements_text"
19892 | "json_array_elements_text"
19893 | "jsonb_object_keys"
19894 | "json_object_keys"
19895 | "generate_subscripts" => crate::ast::Expr::FunctionCall {
19896 name: fn_name,
19897 args: fn_args,
19898 },
19899 "string_to_table" => crate::ast::Expr::FunctionCall {
19900 name: "string_to_array".to_string(),
19901 args: fn_args,
19902 },
19903 "regexp_split_to_table" => crate::ast::Expr::FunctionCall {
19904 name: "regexp_split_to_array".to_string(),
19905 args: fn_args,
19906 },
19907 // v7.39 (read01 round 74) — an SRF with no array form
19908 // (`generate_series`, a user `RETURNS SETOF` function) has no
19909 // scalar expression to zip, so the WHOLE list switches to the
19910 // rows_from channel, which runs each function and zips the
19911 // rows themselves. The all-array case keeps the old lowering:
19912 // it is well-trodden and this must not disturb it.
19913 _ => {
19914 generic.push((fn_name, fn_args));
19915 if matches!(self.peek(), Token::Comma) {
19916 self.advance();
19917 continue;
19918 }
19919 break;
19920 }
19921 };
19922 generic.push((
19923 // The array-able entries carry their lowered expr along, so a
19924 // MIXED list still works: the engine sees the scalar array
19925 // form and unnests it.
19926 "__array".to_string(),
19927 alloc::vec![entry.clone()],
19928 ));
19929 entries.push(entry);
19930 if matches!(self.peek(), Token::Comma) {
19931 self.advance();
19932 continue;
19933 }
19934 break;
19935 }
19936 if !matches!(self.peek(), Token::RParen) {
19937 return Err(self.err(alloc::format!(
19938 "expected ')' to close ROWS FROM, got {:?}",
19939 self.peek()
19940 )));
19941 }
19942 self.advance();
19943 let with_ordinality = self.absorb_with_ordinality();
19944 let (alias_ident, unnest_column_aliases) = self.parse_optional_alias_with_columns()?;
19945 let name = alias_ident.clone().unwrap_or_else(|| "rows".to_string());
19946 // v7.39 (read01 round 74) — some entry had no array form, so the whole
19947 // list rides the generic channel.
19948 if generic.iter().any(|(n, _)| n != "__array") {
19949 let correlated = generic
19950 .iter()
19951 .any(|(_, a)| a.iter().any(Self::expr_has_any_column));
19952 let tref = TableRef {
19953 name,
19954 alias: alias_ident,
19955 only: false,
19956 as_of_segment: None,
19957 unnest_expr: None,
19958 unnest_column_aliases,
19959 with_ordinality,
19960 generate_series_args: None,
19961 lateral_subquery: None,
19962 jsonb_each_text_arg: None,
19963 table_fn_call: None,
19964 rows_from: Some(generic),
19965 json_table: None,
19966 scalar_fn_item: false,
19967 };
19968 return Ok(if correlated {
19969 Self::wrap_correlated_srf(tref)
19970 } else {
19971 tref
19972 });
19973 }
19974 let correlated = entries.iter().any(Self::expr_has_any_column);
19975 let expr = if entries.len() == 1 {
19976 entries.pop().expect("len checked")
19977 } else {
19978 crate::ast::Expr::FunctionCall {
19979 name: "__unnest_zip".to_string(),
19980 args: entries,
19981 }
19982 };
19983 let tref = TableRef {
19984 name,
19985 alias: alias_ident,
19986 only: false,
19987 as_of_segment: None,
19988 unnest_expr: Some(Box::new(expr)),
19989 unnest_column_aliases,
19990 with_ordinality,
19991 generate_series_args: None,
19992 lateral_subquery: None,
19993 jsonb_each_text_arg: None,
19994 table_fn_call: None,
19995 rows_from: None,
19996 json_table: None,
19997 scalar_fn_item: false,
19998 };
19999 return Ok(if correlated {
20000 Self::wrap_correlated_srf(tref)
20001 } else {
20002 tref
20003 });
20004 }
20005 // v7.11.7 — `FROM unnest(<expr>) [AS] <alias>` set-returning
20006 // source. Detect at the head before the bare-ident fallback;
20007 // unnest is not a reserved token.
20008 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("unnest"))
20009 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
20010 {
20011 self.advance(); // unnest
20012 self.advance(); // (
20013 let mut srf_args = alloc::vec![self.parse_expr(0)?];
20014 while matches!(self.peek(), Token::Comma) {
20015 self.advance();
20016 srf_args.push(self.parse_expr(0)?);
20017 }
20018 if !matches!(self.peek(), Token::RParen) {
20019 return Err(self.err(alloc::format!(
20020 "expected ')' after unnest() argument, got {:?}",
20021 self.peek()
20022 )));
20023 }
20024 self.advance();
20025 // Multi-arg unnest(a, b, …) zips the arrays in
20026 // parallel, NULL-padding to the longest (PG's ROWS
20027 // FROM shorthand). Lower onto the unnest channel as an
20028 // internal marker call the executors unpack.
20029 let expr = if srf_args.len() == 1 {
20030 srf_args.pop().expect("len checked")
20031 } else {
20032 crate::ast::Expr::FunctionCall {
20033 name: "__unnest_zip".to_string(),
20034 args: srf_args,
20035 }
20036 };
20037 let with_ordinality = self.absorb_with_ordinality();
20038 let (alias_ident, unnest_column_aliases) = self.parse_optional_alias_with_columns()?;
20039 let name = alias_ident.clone().unwrap_or_else(|| "unnest".to_string());
20040 let correlated = Self::expr_has_any_column(&expr);
20041 let tref = TableRef {
20042 name,
20043 alias: alias_ident,
20044 only: false,
20045 as_of_segment: None,
20046 unnest_expr: Some(Box::new(expr)),
20047 unnest_column_aliases,
20048 with_ordinality,
20049 generate_series_args: None,
20050 lateral_subquery: None,
20051 jsonb_each_text_arg: None,
20052 table_fn_call: None,
20053 rows_from: None,
20054 json_table: None,
20055 scalar_fn_item: false,
20056 };
20057 return Ok(if correlated {
20058 Self::wrap_correlated_srf(tref)
20059 } else {
20060 tref
20061 });
20062 }
20063 // v7.39 (round 205, JSON_TABLE epic) — `JSON_TABLE(doc, '$path'
20064 // COLUMNS (...))` has bespoke syntax (a COLUMNS clause the
20065 // generic table-fn arg parser can't read), so it is intercepted
20066 // here BEFORE the generic dispatch. The doc expr may reference
20067 // outer columns (implicit LATERAL) — same correlated-wrap rule.
20068 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
20069 if s.eq_ignore_ascii_case("json_table"))
20070 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
20071 {
20072 let tref = self.parse_json_table_ref()?;
20073 let correlated = tref
20074 .json_table
20075 .as_deref()
20076 .is_some_and(|jt| Self::expr_has_any_column(&jt.doc));
20077 return Ok(if correlated {
20078 Self::wrap_correlated_srf(tref)
20079 } else {
20080 tref
20081 });
20082 }
20083 // v7.39 (read01 partitionfuncs.c) — generic FROM-position table
20084 // functions dispatched by name (`pg_partition_tree('t')`,
20085 // `pg_partition_ancestors('t')`). Same head-detection shape as
20086 // unnest; the engine executor owns the row shape per function.
20087 // v7.39 (read01 round 65) — and a USER function in FROM position
20088 // (`FROM rows_of(2)`). The SRFs with their own FROM pipeline
20089 // (generate_series / unnest / the json_each family) keep it — their arms
20090 // sit further down, so they are excluded here by name rather than by
20091 // ordering. Anything else that is an ident followed by `(` is a table
20092 // function; the engine executor decides whether it is a builtin, a
20093 // set-returning user function, or an error.
20094 // 7.38.1 S5.1 — pg_dump spells its table functions
20095 // schema-qualified (`pg_catalog.pg_options_to_table(...)`);
20096 // strip the pg_catalog prefix here so the same head-detection
20097 // fires. Only pg_catalog: a user schema's `s.f(x)` keeps its
20098 // meaning.
20099 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("pg_catalog"))
20100 && matches!(self.tokens.get(self.pos + 1), Some(Token::Dot))
20101 && matches!(
20102 self.tokens.get(self.pos + 2),
20103 Some(Token::Ident(_) | Token::QuotedIdent(_))
20104 )
20105 && matches!(self.tokens.get(self.pos + 3), Some(Token::LParen))
20106 {
20107 self.advance(); // pg_catalog
20108 self.advance(); // .
20109 }
20110 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
20111 if !s.eq_ignore_ascii_case("generate_series")
20112 && !s.eq_ignore_ascii_case("unnest")
20113 && !is_json_each_name(s))
20114 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
20115 {
20116 // Body out-of-line — this parse sits on the FROM/subquery
20117 // recursion chain (debug frame-cliff discipline).
20118 // v7.39 (read01 round 69) — a call whose arguments reference an outer
20119 // column (`t JOIN LATERAL dbl(t.id)`) is CORRELATED: it runs once per
20120 // outer row, so it rides the lateral channel. Same rule the unnest
20121 // arm uses.
20122 let tref = self.parse_table_fn_ref()?;
20123 let correlated = tref
20124 .table_fn_call
20125 .as_deref()
20126 .is_some_and(|(_, args)| args.iter().any(Self::expr_has_any_column));
20127 return Ok(if correlated {
20128 Self::wrap_correlated_srf(tref)
20129 } else {
20130 tref
20131 });
20132 }
20133 // v7.17.0 Phase 3.10 — `FROM generate_series(start, stop
20134 // [, step])` set-returning source. Same shape as unnest:
20135 // detect at the head, parse the comma-separated arg list,
20136 // dispatch downstream through the engine's set-returning
20137 // path. Supports integer triplets (mailrs's `WITH row_no AS
20138 // (SELECT * FROM generate_series(1, N))` pattern) and
20139 // TIMESTAMP + INTERVAL triplets (the Tier-A audit's
20140 // date-range iteration pattern, which pre-3.10 had no
20141 // direct equivalent in SPG).
20142 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("generate_series"))
20143 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
20144 {
20145 self.advance(); // generate_series
20146 self.advance(); // (
20147 let mut args: Vec<Expr> = Vec::new();
20148 loop {
20149 args.push(self.parse_expr(0)?);
20150 if matches!(self.peek(), Token::Comma) {
20151 self.advance();
20152 continue;
20153 }
20154 break;
20155 }
20156 if !matches!(self.peek(), Token::RParen) {
20157 return Err(self.err(alloc::format!(
20158 "expected ')' after generate_series() arguments, got {:?}",
20159 self.peek()
20160 )));
20161 }
20162 self.advance();
20163 if args.len() < 2 || args.len() > 3 {
20164 return Err(self.err(alloc::format!(
20165 "generate_series() expects 2 or 3 arguments (start, stop [, step]); got {}",
20166 args.len()
20167 )));
20168 }
20169 let with_ordinality = self.absorb_with_ordinality();
20170 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
20171 let name = alias_ident
20172 .clone()
20173 .unwrap_or_else(|| "generate_series".to_string());
20174 let correlated = args.iter().any(Self::expr_has_any_column);
20175 let tref = TableRef {
20176 name,
20177 alias: alias_ident,
20178 only: false,
20179 as_of_segment: None,
20180 unnest_expr: None,
20181 unnest_column_aliases: column_aliases,
20182 with_ordinality,
20183 generate_series_args: Some(args),
20184 lateral_subquery: None,
20185 jsonb_each_text_arg: None,
20186 table_fn_call: None,
20187 rows_from: None,
20188 json_table: None,
20189 scalar_fn_item: false,
20190 };
20191 return Ok(if correlated {
20192 Self::wrap_correlated_srf(tref)
20193 } else {
20194 tref
20195 });
20196 }
20197 // v7.16.2 — preserve information_schema / pg_catalog
20198 // qualifiers (mailrs round-10 A.3). The generic
20199 // `expect_ident_like` strip silently drops the schema;
20200 // we want the engine to recognise these PG meta tables
20201 // and synthesise rows from the live catalog. Produce a
20202 // synthetic name (`__spg_info_columns` etc.) so the
20203 // engine's SELECT-side router can dispatch without
20204 // clashing with any user-defined `columns` table.
20205 let (name, meta_original) = if let Some((synth, orig)) = self.try_peek_meta_qualified() {
20206 (synth, Some(orig))
20207 } else if let Some((synth, orig)) = self.try_peek_meta_bare() {
20208 (synth, Some(orig))
20209 } else {
20210 (self.expect_ident_like()?, None)
20211 };
20212 // v6.10.2 — optional `AS OF SEGMENT '<id>'` cold-tier
20213 // time-travel clause. Parse BEFORE the alias so the
20214 // alias can still ride at the tail (`tbl AS OF SEGMENT
20215 // '5' alias`). `AS` is a reserved keyword token, while
20216 // `OF` and `SEGMENT` are bare idents.
20217 let as_of_segment = if matches!(self.peek(), Token::As)
20218 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s) | Token::QuotedIdent(s)) if s.eq_ignore_ascii_case("of"))
20219 {
20220 self.advance(); // AS
20221 self.advance(); // OF
20222 let kw = match self.peek().clone() {
20223 Token::Ident(s) | Token::QuotedIdent(s) => s,
20224 other => {
20225 return Err(self.err(format!("expected SEGMENT after AS OF, got {other:?}")));
20226 }
20227 };
20228 if !kw.eq_ignore_ascii_case("segment") {
20229 return Err(self.err(format!(
20230 "expected SEGMENT after AS OF, got {kw:?}; v6.10.2 supports SEGMENT only"
20231 )));
20232 }
20233 self.advance();
20234 // Segment id literal — accept either a string or
20235 // integer for operator ergonomics.
20236 let id = match self.advance() {
20237 Token::String(s) => s
20238 .parse::<u32>()
20239 .map_err(|e| self.err(format!("AS OF SEGMENT id parse: {e}")))?,
20240 Token::Integer(n) => u32::try_from(n)
20241 .map_err(|e| self.err(format!("AS OF SEGMENT id parse: {e}")))?,
20242 other => {
20243 return Err(self.err(format!(
20244 "expected segment id literal after AS OF SEGMENT, got {other:?}"
20245 )));
20246 }
20247 };
20248 Some(id)
20249 } else {
20250 None
20251 };
20252 // TABLESAMPLE is not a reserved token — keep the bare-ident
20253 // alias rule from swallowing it (`FROM t TABLESAMPLE …`).
20254 let alias = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("tablesample"))
20255 {
20256 None
20257 } else {
20258 self.parse_optional_alias()?
20259 };
20260 // r1052 — a catalog name rewritten to its synthetic form keeps
20261 // the WRITTEN name as the relation's alias, so `pg_cast.oid`
20262 // still binds after `pg_cast` became `__spg_pg_cast`. PG
20263 // semantics: the visible name of `pg_catalog.pg_cast` IS
20264 // `pg_cast`. Without this, every table-name-qualified column
20265 // on a synthesised catalog answered "missing FROM-clause
20266 // entry" — which is the wall pg_dump hit on its first
20267 // pg_proc/pg_cast query.
20268 let alias = match (&alias, &meta_original) {
20269 (None, Some(orig)) if *orig != name => Some(orig.clone()),
20270 _ => alias,
20271 };
20272 // `TABLESAMPLE BERNOULLI(p) | SYSTEM(p)` follows the alias
20273 // (PG grammar). BERNOULLI lowers to a per-row
20274 // `random() < p/100` conjunct on the enclosing SELECT's
20275 // WHERE — exact row-level Bernoulli semantics. SYSTEM
20276 // shares the lowering: SPG has no page structure to
20277 // sample, and the row-level form returns the same expected
20278 // fraction. REPEATABLE(seed) promises a deterministic
20279 // sample SPG cannot honour yet — honest error rather than
20280 // a silently ignored seed.
20281 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("tablesample")) {
20282 self.advance();
20283 let method = self.expect_ident_like()?;
20284 if !method.eq_ignore_ascii_case("bernoulli") && !method.eq_ignore_ascii_case("system") {
20285 return Err(self.err(alloc::format!(
20286 "TABLESAMPLE method {method:?} not supported; use BERNOULLI or SYSTEM"
20287 )));
20288 }
20289 if !matches!(self.peek(), Token::LParen) {
20290 return Err(self.err(alloc::format!(
20291 "expected '(' after TABLESAMPLE {}, got {:?}",
20292 method.to_ascii_uppercase(),
20293 self.peek()
20294 )));
20295 }
20296 self.advance();
20297 let percent = self.parse_expr(0)?;
20298 if !matches!(self.peek(), Token::RParen) {
20299 return Err(self.err(alloc::format!(
20300 "expected ')' after TABLESAMPLE percentage, got {:?}",
20301 self.peek()
20302 )));
20303 }
20304 self.advance();
20305 // REPEATABLE(seed) → a deterministic per-row draw seeded by
20306 // `seed`, so the sample is stable across repeats and rescans.
20307 // Non-REPEATABLE keeps the non-deterministic `random()` draw.
20308 let mut sample_seed: Option<Expr> = None;
20309 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("repeatable")) {
20310 self.advance();
20311 if !matches!(self.peek(), Token::LParen) {
20312 return Err(self.err(alloc::format!(
20313 "expected '(' after REPEATABLE, got {:?}",
20314 self.peek()
20315 )));
20316 }
20317 self.advance();
20318 let seed = self.parse_expr(0)?;
20319 if !matches!(self.peek(), Token::RParen) {
20320 return Err(self.err(alloc::format!(
20321 "expected ')' after REPEATABLE seed, got {:?}",
20322 self.peek()
20323 )));
20324 }
20325 self.advance();
20326 sample_seed = Some(seed);
20327 }
20328 let draw = match sample_seed {
20329 Some(seed) => Expr::FunctionCall {
20330 name: "__tsm_fract".to_string(),
20331 args: alloc::vec![seed],
20332 },
20333 None => Expr::FunctionCall {
20334 name: "random".to_string(),
20335 args: Vec::new(),
20336 },
20337 };
20338 self.pending_sample_preds.push(Expr::Binary {
20339 lhs: Box::new(draw),
20340 op: crate::ast::BinOp::Lt,
20341 rhs: Box::new(Expr::Binary {
20342 lhs: Box::new(percent),
20343 op: crate::ast::BinOp::Div,
20344 rhs: Box::new(Expr::Literal(crate::ast::Literal::Float(100.0))),
20345 }),
20346 });
20347 }
20348 Ok(TableRef {
20349 name,
20350 alias,
20351 only,
20352 as_of_segment,
20353 unnest_expr: None,
20354 unnest_column_aliases: Vec::new(),
20355 with_ordinality: false,
20356 generate_series_args: None,
20357 lateral_subquery: None,
20358 jsonb_each_text_arg: None,
20359 table_fn_call: None,
20360 rows_from: None,
20361 json_table: None,
20362 scalar_fn_item: false,
20363 })
20364 }
20365
20366 /// v7.13.2 — mailrs round-6 S5. Like `parse_optional_alias`
20367 /// but also accepts `AS alias(col [, col, …])` — the
20368 /// PG-standard table-function column-list form. The column
20369 /// list is only honoured when paired with `UNNEST(...)` in
20370 /// the parent; other call sites currently discard it.
20371 /// True when the expression tree contains a qualified column
20372 /// reference (`t.col`) — the syntactic marker that an SRF
20373 /// argument correlates with a preceding FROM item.
20374 fn expr_has_qualified_column(e: &Expr) -> bool {
20375 match e {
20376 Expr::Column(c) => c.qualifier.is_some(),
20377 Expr::Binary { lhs, rhs, .. } => {
20378 Self::expr_has_qualified_column(lhs) || Self::expr_has_qualified_column(rhs)
20379 }
20380 Expr::Unary { expr, .. } => Self::expr_has_qualified_column(expr),
20381 Expr::Cast { expr, .. } => Self::expr_has_qualified_column(expr),
20382 Expr::FunctionCall { args, .. } => args.iter().any(Self::expr_has_qualified_column),
20383 Expr::Case {
20384 operand,
20385 branches,
20386 else_branch,
20387 } => {
20388 operand
20389 .as_deref()
20390 .is_some_and(Self::expr_has_qualified_column)
20391 || branches.iter().any(|(w, t)| {
20392 Self::expr_has_qualified_column(w) || Self::expr_has_qualified_column(t)
20393 })
20394 || else_branch
20395 .as_deref()
20396 .is_some_and(Self::expr_has_qualified_column)
20397 }
20398 _ => false,
20399 }
20400 }
20401
20402 /// v7.38 (read01, T-lateral) — like `expr_has_qualified_column` but also
20403 /// counts a bare (unqualified) column. A set-returning function has no
20404 /// input columns of its own, so ANY column in its arguments is an outer
20405 /// (correlated) reference — `generate_series(1, n)` correlates on `n`.
20406 fn expr_has_any_column(e: &Expr) -> bool {
20407 match e {
20408 Expr::Column(_) => true,
20409 Expr::Binary { lhs, rhs, .. } => {
20410 Self::expr_has_any_column(lhs) || Self::expr_has_any_column(rhs)
20411 }
20412 Expr::Unary { expr, .. } => Self::expr_has_any_column(expr),
20413 Expr::Cast { expr, .. } => Self::expr_has_any_column(expr),
20414 Expr::FunctionCall { args, .. } => args.iter().any(Self::expr_has_any_column),
20415 // v7.39 (round 759, F31-B8b) — a column INSIDE an array
20416 // constructor or subscript fell to the `_ => false` arm, so
20417 // `unnest(ARRAY[x, x + 1])` never wrapped into the lateral
20418 // channel and the eager peer eval answered `column "x" does
20419 // not exist` (the substitution walker already recurses both
20420 // shapes; only this detector was blind to them).
20421 Expr::Array(items) => items.iter().any(Self::expr_has_any_column),
20422 Expr::ArraySubscript { target, index } => {
20423 Self::expr_has_any_column(target) || Self::expr_has_any_column(index)
20424 }
20425 Expr::Case {
20426 operand,
20427 branches,
20428 else_branch,
20429 } => {
20430 operand.as_deref().is_some_and(Self::expr_has_any_column)
20431 || branches
20432 .iter()
20433 .any(|(w, t)| Self::expr_has_any_column(w) || Self::expr_has_any_column(t))
20434 || else_branch
20435 .as_deref()
20436 .is_some_and(Self::expr_has_any_column)
20437 }
20438 _ => false,
20439 }
20440 }
20441
20442 /// Wrap a correlated SRF table ref (`unnest(t.col)` /
20443 /// `generate_series(1, t.n)`) into the lateral_subquery
20444 /// channel: `SELECT * FROM <srf>` executes per outer row with
20445 /// outer references substituted (v7.37.43-T4.5 machinery).
20446 /// Uncorrelated SRFs stay on their plain channels.
20447 fn wrap_correlated_srf(srf: TableRef) -> TableRef {
20448 let name = srf.name.clone();
20449 let alias = srf.alias.clone();
20450 let inner = crate::ast::SelectStatement {
20451 locking: None,
20452 ctes: Vec::new(),
20453 distinct: false,
20454 distinct_on: Vec::new(),
20455 items: alloc::vec![crate::ast::SelectItem::Wildcard],
20456 from: Some(crate::ast::FromClause {
20457 primary: srf,
20458 joins: Vec::new(),
20459 }),
20460 where_: None,
20461 group_by: None,
20462 group_by_all: false,
20463 having: None,
20464 unions: Vec::new(),
20465 order_by: Vec::new(),
20466 limit: None,
20467 offset: None,
20468 limit_with_ties: false,
20469 window_check_exprs: Vec::new(),
20470 };
20471 TableRef {
20472 name,
20473 alias,
20474 only: false,
20475 as_of_segment: None,
20476 unnest_expr: None,
20477 unnest_column_aliases: Vec::new(),
20478 with_ordinality: false,
20479 generate_series_args: None,
20480 lateral_subquery: Some(Box::new(inner)),
20481 jsonb_each_text_arg: None,
20482 table_fn_call: None,
20483 rows_from: None,
20484 json_table: None,
20485 scalar_fn_item: false,
20486 }
20487 }
20488
20489 /// True when the expression tree contains an unresolved
20490 /// `OVER w` marker (see parse_over_clause).
20491 fn expr_has_named_window(e: &Expr) -> bool {
20492 match e {
20493 Expr::WindowFunction { partition_by, .. } => matches!(
20494 partition_by.as_slice(),
20495 [Expr::Column(c)] if matches!(
20496 c.qualifier.as_deref(),
20497 Some("__named_window__") | Some("__named_window_ref__")
20498 )
20499 ),
20500 Expr::Binary { lhs, rhs, .. } => {
20501 Self::expr_has_named_window(lhs) || Self::expr_has_named_window(rhs)
20502 }
20503 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } => Self::expr_has_named_window(expr),
20504 Expr::FunctionCall { args, .. } => args.iter().any(Self::expr_has_named_window),
20505 Expr::Case {
20506 operand,
20507 branches,
20508 else_branch,
20509 } => {
20510 operand.as_deref().is_some_and(Self::expr_has_named_window)
20511 || branches.iter().any(|(w, t)| {
20512 Self::expr_has_named_window(w) || Self::expr_has_named_window(t)
20513 })
20514 || else_branch
20515 .as_deref()
20516 .is_some_and(Self::expr_has_named_window)
20517 }
20518 _ => false,
20519 }
20520 }
20521
20522 /// v7.39 (round 705) — the NAMES the expression references through the
20523 /// `OVER w` markers, so `parse_bare_select` can tell which WINDOW
20524 /// definitions nothing referenced. Traversal mirrors
20525 /// `expr_has_named_window` above.
20526 fn collect_named_window_refs(e: &Expr, into: &mut Vec<String>) {
20527 match e {
20528 Expr::WindowFunction { partition_by, .. } => {
20529 if let [Expr::Column(c)] = partition_by.as_slice()
20530 && matches!(
20531 c.qualifier.as_deref(),
20532 Some("__named_window__") | Some("__named_window_ref__")
20533 )
20534 {
20535 into.push(c.name.clone());
20536 }
20537 }
20538 Expr::Binary { lhs, rhs, .. } => {
20539 Self::collect_named_window_refs(lhs, into);
20540 Self::collect_named_window_refs(rhs, into);
20541 }
20542 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } => {
20543 Self::collect_named_window_refs(expr, into);
20544 }
20545 Expr::FunctionCall { args, .. } => {
20546 for a in args {
20547 Self::collect_named_window_refs(a, into);
20548 }
20549 }
20550 Expr::Case {
20551 operand,
20552 branches,
20553 else_branch,
20554 } => {
20555 if let Some(o) = operand.as_deref() {
20556 Self::collect_named_window_refs(o, into);
20557 }
20558 for (w, t) in branches {
20559 Self::collect_named_window_refs(w, into);
20560 Self::collect_named_window_refs(t, into);
20561 }
20562 if let Some(eb) = else_branch.as_deref() {
20563 Self::collect_named_window_refs(eb, into);
20564 }
20565 }
20566 _ => {}
20567 }
20568 }
20569
20570 /// Inline named-window definitions into the `OVER w` markers.
20571 /// An unknown name errors (PG: window "w" does not exist).
20572 #[allow(clippy::type_complexity)]
20573 fn substitute_named_windows(
20574 e: &mut Expr,
20575 defs: &[(
20576 String,
20577 (
20578 Vec<Expr>,
20579 Vec<(Expr, bool, Option<bool>)>,
20580 Option<WindowFrame>,
20581 ),
20582 )],
20583 ) -> Result<(), String> {
20584 match e {
20585 Expr::WindowFunction {
20586 partition_by,
20587 order_by,
20588 frame,
20589 ..
20590 } => {
20591 // `is_copy` distinguishes `OVER (w1 …)` (a refinable copy)
20592 // from the bare `OVER w1` (a plain reference).
20593 let named = match partition_by.as_slice() {
20594 [Expr::Column(c)] => match c.qualifier.as_deref() {
20595 Some("__named_window__") => Some((c.name.clone(), false)),
20596 Some("__named_window_ref__") => Some((c.name.clone(), true)),
20597 _ => None,
20598 },
20599 _ => None,
20600 };
20601 if let Some((wname, is_copy)) = named {
20602 let Some((_, def)) = defs.iter().find(|(n, _)| n.eq_ignore_ascii_case(&wname))
20603 else {
20604 return Err(alloc::format!("window {wname:?} does not exist"));
20605 };
20606 if !is_copy {
20607 *partition_by = def.0.clone();
20608 *order_by = def.1.clone();
20609 *frame = def.2.clone();
20610 return Ok(());
20611 }
20612 // v7.39 (round 229) — PG's copy rules, probed against
20613 // 18.4: a copy inherits the partitioning, may supply an
20614 // ordering only when the base has none, and may not copy
20615 // a base that already carries a frame (its own frame
20616 // would be ambiguous with the inherited one).
20617 if !def.1.is_empty() && !order_by.is_empty() {
20618 return Err(alloc::format!(
20619 "cannot override ORDER BY clause of window \"{wname}\""
20620 ));
20621 }
20622 if def.2.is_some() {
20623 return Err(alloc::format!(
20624 "cannot copy window \"{wname}\" because it has a frame clause"
20625 ));
20626 }
20627 *partition_by = def.0.clone();
20628 if order_by.is_empty() {
20629 *order_by = def.1.clone();
20630 }
20631 }
20632 Ok(())
20633 }
20634 Expr::Binary { lhs, rhs, .. } => {
20635 Self::substitute_named_windows(lhs, defs)?;
20636 Self::substitute_named_windows(rhs, defs)
20637 }
20638 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } => {
20639 Self::substitute_named_windows(expr, defs)
20640 }
20641 Expr::FunctionCall { args, .. } => {
20642 for a in args {
20643 Self::substitute_named_windows(a, defs)?;
20644 }
20645 Ok(())
20646 }
20647 Expr::Case {
20648 operand,
20649 branches,
20650 else_branch,
20651 } => {
20652 if let Some(op) = operand {
20653 Self::substitute_named_windows(op, defs)?;
20654 }
20655 for (w, t) in branches {
20656 Self::substitute_named_windows(w, defs)?;
20657 Self::substitute_named_windows(t, defs)?;
20658 }
20659 if let Some(el) = else_branch {
20660 Self::substitute_named_windows(el, defs)?;
20661 }
20662 Ok(())
20663 }
20664 _ => Ok(()),
20665 }
20666 }
20667
20668 /// SQL-standard `TABLE name` shorthand — builds the equivalent
20669 /// `SELECT * FROM name` head. Callers own set-op chain / tail
20670 /// composition.
20671 fn parse_table_shorthand(&mut self) -> Result<SelectStatement, ParseError> {
20672 debug_assert!(matches!(self.peek(), Token::Table));
20673 self.advance(); // TABLE
20674 let tname = self.expect_ident_like()?;
20675 Ok(SelectStatement {
20676 locking: None,
20677 ctes: Vec::new(),
20678 distinct: false,
20679 distinct_on: Vec::new(),
20680 items: alloc::vec![SelectItem::Wildcard],
20681 from: Some(FromClause {
20682 primary: TableRef {
20683 name: tname,
20684 alias: None,
20685 only: false,
20686 as_of_segment: None,
20687 unnest_expr: None,
20688 unnest_column_aliases: Vec::new(),
20689 with_ordinality: false,
20690 generate_series_args: None,
20691 lateral_subquery: None,
20692 jsonb_each_text_arg: None,
20693 table_fn_call: None,
20694 rows_from: None,
20695 json_table: None,
20696 scalar_fn_item: false,
20697 },
20698 joins: Vec::new(),
20699 }),
20700 where_: None,
20701 group_by: None,
20702 group_by_all: false,
20703 having: None,
20704 unions: Vec::new(),
20705 order_by: Vec::new(),
20706 limit: None,
20707 offset: None,
20708 limit_with_ties: false,
20709 window_check_exprs: Vec::new(),
20710 })
20711 }
20712
20713 /// `jsonb_to_recordset(J) AS t(c1 t1, c2 t2, …)` (and record / json_
20714 /// variants) → a derived table that reads each declared column out of
20715 /// the JSON with `(row ->> 'ci')::ti`. The *set form iterates
20716 /// `jsonb_array_elements(J)` (one row per element, column `value`);
20717 /// the scalar *record form projects a single row straight off `J`.
20718 /// Rides the existing lateral-subquery channel, so no new executor or
20719 /// AST is needed.
20720 fn parse_json_to_record_from(&mut self) -> Result<TableRef, ParseError> {
20721 use crate::ast::{
20722 BinOp, ColumnName, Expr, FromClause, Literal, SelectItem, SelectStatement,
20723 };
20724 let fn_name = match self.peek() {
20725 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
20726 _ => unreachable!("caller guarded is_json_to_record_name"),
20727 };
20728 self.advance(); // fn name
20729 self.advance(); // (
20730 let mut arg = self.parse_expr(0)?;
20731 // populate_record(base, json): the base only carries the record
20732 // type here — the JSON argument is the second expression.
20733 let mut base: Option<Expr> = None;
20734 if matches!(self.peek(), Token::Comma) {
20735 self.advance();
20736 base = Some(arg);
20737 arg = self.parse_expr(0)?;
20738 }
20739 if !matches!(self.peek(), Token::RParen) {
20740 return Err(self.err(alloc::format!(
20741 "expected ')' after {fn_name}() argument, got {:?}",
20742 self.peek()
20743 )));
20744 }
20745 self.advance(); // )
20746 let is_set = fn_name.ends_with("recordset");
20747 // `[AS] alias ( col type [, …] )` column-definition list.
20748 if matches!(self.peek(), Token::As) {
20749 self.advance();
20750 }
20751 let alias_opt = match self.peek() {
20752 Token::Ident(s) | Token::QuotedIdent(s) => {
20753 let a = s.clone();
20754 self.advance();
20755 Some(a)
20756 }
20757 _ => None,
20758 };
20759 // v7.39 (read01 round 76) — the populate family's canonical PG
20760 // spelling carries no column list at all: the row shape comes from
20761 // the BASE argument's declared type (`jsonb_populate_record(NULL::t,
20762 // j)`). The parser has no catalog, so hand the two arguments to the
20763 // engine's table-function channel, which does. Only `*_to_record*`
20764 // (whose base is bare `record`) genuinely requires the list.
20765 if !matches!(self.peek(), Token::LParen) {
20766 if let Some(base_expr) = base {
20767 let alias = alias_opt.unwrap_or_else(|| fn_name.clone());
20768 return Ok(TableRef {
20769 name: alias.clone(),
20770 alias: Some(alias),
20771 only: false,
20772 as_of_segment: None,
20773 unnest_expr: None,
20774 unnest_column_aliases: Vec::new(),
20775 with_ordinality: false,
20776 generate_series_args: None,
20777 lateral_subquery: None,
20778 jsonb_each_text_arg: None,
20779 table_fn_call: Some(Box::new((fn_name, alloc::vec![base_expr, arg]))),
20780 rows_from: None,
20781 json_table: None,
20782 scalar_fn_item: false,
20783 });
20784 }
20785 return Err(self.err(alloc::format!(
20786 "expected '(' to start the {fn_name} column-definition list, got {:?}",
20787 self.peek()
20788 )));
20789 }
20790 let Some(alias) = alias_opt else {
20791 return Err(self.err(alloc::format!(
20792 "{fn_name}(...) needs a column-definition list, e.g. AS t(a int, b text)"
20793 )));
20794 };
20795 self.advance(); // (
20796 let mut coldefs: Vec<(String, crate::ast::CastTarget)> = Vec::new();
20797 loop {
20798 let col = self.expect_ident_like()?;
20799 let ty = self.parse_cast_target()?;
20800 coldefs.push((col, ty));
20801 if matches!(self.peek(), Token::Comma) {
20802 self.advance();
20803 continue;
20804 }
20805 if matches!(self.peek(), Token::RParen) {
20806 self.advance();
20807 break;
20808 }
20809 return Err(self.err(alloc::format!(
20810 "expected ',' or ')' in {fn_name} column list, got {:?}",
20811 self.peek()
20812 )));
20813 }
20814 if coldefs.is_empty() {
20815 return Err(self.err(alloc::format!(
20816 "{fn_name} column-definition list must declare at least one column"
20817 )));
20818 }
20819 // Per column: (base ->> 'col')::type AS col. The base is the
20820 // per-element `value` column for the *set form, or the argument
20821 // itself for the scalar record form.
20822 let items: Vec<SelectItem> = coldefs
20823 .into_iter()
20824 .map(|(col, ty)| {
20825 let base = if is_set {
20826 Expr::Column(ColumnName {
20827 qualifier: None,
20828 name: "value".to_string(),
20829 })
20830 } else {
20831 arg.clone()
20832 };
20833 SelectItem::Expr {
20834 expr: Expr::Cast {
20835 expr: Box::new(Expr::Binary {
20836 lhs: Box::new(base),
20837 op: BinOp::JsonGetText,
20838 rhs: Box::new(Expr::Literal(Literal::String(col.clone()))),
20839 }),
20840 target: ty,
20841 },
20842 alias: Some(col),
20843 }
20844 })
20845 .collect();
20846 let from = if is_set {
20847 let elem_fn = if fn_name.starts_with("jsonb") {
20848 "jsonb_array_elements"
20849 } else {
20850 "json_array_elements"
20851 };
20852 Some(FromClause {
20853 primary: TableRef {
20854 name: "value".to_string(),
20855 alias: None,
20856 only: false,
20857 as_of_segment: None,
20858 unnest_expr: Some(Box::new(Expr::FunctionCall {
20859 name: elem_fn.to_string(),
20860 args: alloc::vec![arg],
20861 })),
20862 unnest_column_aliases: alloc::vec!["value".to_string()],
20863 with_ordinality: false,
20864 generate_series_args: None,
20865 lateral_subquery: None,
20866 jsonb_each_text_arg: None,
20867 table_fn_call: None,
20868 rows_from: None,
20869 json_table: None,
20870 scalar_fn_item: false,
20871 },
20872 joins: Vec::new(),
20873 })
20874 } else {
20875 None
20876 };
20877 let inner = SelectStatement {
20878 locking: None,
20879 ctes: Vec::new(),
20880 distinct: false,
20881 distinct_on: Vec::new(),
20882 items,
20883 from,
20884 where_: None,
20885 group_by: None,
20886 group_by_all: false,
20887 having: None,
20888 unions: Vec::new(),
20889 order_by: Vec::new(),
20890 limit: None,
20891 offset: None,
20892 limit_with_ties: false,
20893 window_check_exprs: Vec::new(),
20894 };
20895 Ok(TableRef {
20896 name: alias.clone(),
20897 alias: Some(alias),
20898 only: false,
20899 as_of_segment: None,
20900 unnest_expr: None,
20901 unnest_column_aliases: Vec::new(),
20902 with_ordinality: false,
20903 generate_series_args: None,
20904 lateral_subquery: Some(Box::new(inner)),
20905 jsonb_each_text_arg: None,
20906 table_fn_call: None,
20907 rows_from: None,
20908 json_table: None,
20909 scalar_fn_item: false,
20910 })
20911 }
20912
20913 /// Absorb `WITH ORDINALITY` after an SRF call in FROM position.
20914 /// Returns true when the clause was present. `WITH` alone (a
20915 /// CTE can never start here) is not enough — the ORDINALITY
20916 /// ident must follow, so a stray WITH still errors downstream.
20917 fn absorb_with_ordinality(&mut self) -> bool {
20918 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with"))
20919 && matches!(self.tokens.get(self.pos + 1),
20920 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("ordinality"))
20921 {
20922 self.advance();
20923 self.advance();
20924 true
20925 } else {
20926 false
20927 }
20928 }
20929
20930 /// v7.39 (read01 partitionfuncs.c) — parse a FROM-position table
20931 /// function reference (`pg_partition_tree('t') [AS a(c, …)]`).
20932 /// Out-of-line: the caller sits on the FROM recursion chain.
20933 #[inline(never)]
20934 fn parse_table_fn_ref(&mut self) -> Result<TableRef, ParseError> {
20935 let fn_name = match self.advance() {
20936 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
20937 _ => unreachable!("caller peeked an ident"),
20938 };
20939 self.advance(); // (
20940 let mut args: Vec<Expr> = Vec::new();
20941 if !matches!(self.peek(), Token::RParen) {
20942 loop {
20943 args.push(self.parse_expr(0)?);
20944 if matches!(self.peek(), Token::Comma) {
20945 self.advance();
20946 continue;
20947 }
20948 break;
20949 }
20950 }
20951 if !matches!(self.peek(), Token::RParen) {
20952 return Err(self.err(alloc::format!(
20953 "expected ')' after {fn_name}() arguments, got {:?}",
20954 self.peek()
20955 )));
20956 }
20957 self.advance();
20958 // v7.39 (read01 round 68) — `f(args) WITH ORDINALITY AS a(x, n)`: the
20959 // counter column rides after the function's own, and the alias list
20960 // names it.
20961 let with_ordinality = self.absorb_with_ordinality();
20962 let (alias_ident, unnest_column_aliases) = self.parse_optional_alias_with_columns()?;
20963 let name = alias_ident.clone().unwrap_or_else(|| fn_name.clone());
20964 Ok(TableRef {
20965 name,
20966 alias: alias_ident,
20967 only: false,
20968 as_of_segment: None,
20969 unnest_expr: None,
20970 unnest_column_aliases,
20971 with_ordinality,
20972 generate_series_args: None,
20973 lateral_subquery: None,
20974 jsonb_each_text_arg: None,
20975 table_fn_call: Some(Box::new((fn_name, args))),
20976 rows_from: None,
20977 json_table: None,
20978 scalar_fn_item: false,
20979 })
20980 }
20981
20982 /// v7.39 (round 205, JSON_TABLE) — parse
20983 /// `JSON_TABLE(<doc>, '<row_path>' [PASSING …] COLUMNS (<coldefs>))
20984 /// [AS <alias>]`. The COLUMNS list is a recursive tree (NESTED
20985 /// PATH nests another COLUMNS). Out-of-line (FROM recursion chain).
20986 #[inline(never)]
20987 fn parse_json_table_ref(&mut self) -> Result<TableRef, ParseError> {
20988 self.advance(); // json_table
20989 self.advance(); // (
20990 let doc = Box::new(self.parse_expr(0)?);
20991 self.expect_comma_json_table()?;
20992 let row_path = self.parse_json_string_literal("JSON_TABLE row path")?;
20993 // Optional `PASSING <expr> AS <name> [, …]`.
20994 let mut passing: Vec<(String, Expr)> = Vec::new();
20995 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("passing")) {
20996 self.advance();
20997 loop {
20998 let e = self.parse_expr(0)?;
20999 if !matches!(self.peek(), Token::As) {
21000 return Err(self.err("expected AS after JSON_TABLE PASSING value".into()));
21001 }
21002 self.advance();
21003 let vname = match self.advance() {
21004 Token::Ident(s) | Token::QuotedIdent(s) => s,
21005 other => {
21006 return Err(self.err(alloc::format!(
21007 "expected PASSING variable name, got {other:?}"
21008 )));
21009 }
21010 };
21011 passing.push((vname, e));
21012 if matches!(self.peek(), Token::Comma) {
21013 self.advance();
21014 continue;
21015 }
21016 break;
21017 }
21018 }
21019 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("columns")) {
21020 return Err(self.err("expected COLUMNS in JSON_TABLE".into()));
21021 }
21022 self.advance();
21023 let columns = self.parse_json_table_columns()?;
21024 if !matches!(self.peek(), Token::RParen) {
21025 return Err(self.err(alloc::format!(
21026 "expected ')' to close JSON_TABLE, got {:?}",
21027 self.peek()
21028 )));
21029 }
21030 self.advance();
21031 let alias_ident = self.parse_optional_alias()?;
21032 let name = alias_ident
21033 .clone()
21034 .unwrap_or_else(|| String::from("json_table"));
21035 Ok(TableRef {
21036 name,
21037 alias: alias_ident,
21038 only: false,
21039 as_of_segment: None,
21040 unnest_expr: None,
21041 unnest_column_aliases: Vec::new(),
21042 with_ordinality: false,
21043 generate_series_args: None,
21044 lateral_subquery: None,
21045 jsonb_each_text_arg: None,
21046 table_fn_call: None,
21047 rows_from: None,
21048 json_table: Some(Box::new(crate::ast::JsonTable {
21049 doc,
21050 row_path,
21051 columns,
21052 passing,
21053 })),
21054 scalar_fn_item: false,
21055 })
21056 }
21057
21058 fn expect_comma_json_table(&mut self) -> Result<(), ParseError> {
21059 if !matches!(self.peek(), Token::Comma) {
21060 return Err(self.err(alloc::format!(
21061 "expected ',' after JSON_TABLE document, got {:?}",
21062 self.peek()
21063 )));
21064 }
21065 self.advance();
21066 Ok(())
21067 }
21068
21069 fn parse_json_string_literal(&mut self, what: &str) -> Result<String, ParseError> {
21070 match self.advance() {
21071 Token::String(s) => Ok(s),
21072 other => Err(self.err(alloc::format!(
21073 "expected {what} string literal, got {other:?}"
21074 ))),
21075 }
21076 }
21077
21078 /// v7.39 (round 205) — `( <coldef> [, <coldef>]* )`.
21079 #[inline(never)]
21080 fn parse_json_table_columns(
21081 &mut self,
21082 ) -> Result<alloc::vec::Vec<crate::ast::JsonTableColumn>, ParseError> {
21083 if !matches!(self.peek(), Token::LParen) {
21084 return Err(self.err("expected '(' after COLUMNS".into()));
21085 }
21086 self.advance();
21087 let mut cols = Vec::new();
21088 loop {
21089 cols.push(self.parse_json_table_one_column()?);
21090 if matches!(self.peek(), Token::Comma) {
21091 self.advance();
21092 continue;
21093 }
21094 break;
21095 }
21096 if !matches!(self.peek(), Token::RParen) {
21097 return Err(self.err(alloc::format!(
21098 "expected ')' after JSON_TABLE COLUMNS, got {:?}",
21099 self.peek()
21100 )));
21101 }
21102 self.advance();
21103 Ok(cols)
21104 }
21105
21106 fn parse_json_table_one_column(&mut self) -> Result<crate::ast::JsonTableColumn, ParseError> {
21107 use crate::ast::{JsonTableColumn, JsonTableOnBehavior};
21108 // NESTED [PATH] '<p>' COLUMNS (...)
21109 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("nested")) {
21110 self.advance();
21111 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("path")) {
21112 self.advance();
21113 }
21114 let path = self.parse_json_string_literal("NESTED PATH")?;
21115 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("columns")) {
21116 return Err(self.err("expected COLUMNS after NESTED PATH".into()));
21117 }
21118 self.advance();
21119 let columns = self.parse_json_table_columns()?;
21120 return Ok(JsonTableColumn::Nested { path, columns });
21121 }
21122 // <name> ...
21123 let name = match self.advance() {
21124 Token::Ident(s) | Token::QuotedIdent(s) => s,
21125 other => {
21126 return Err(self.err(alloc::format!("expected column name, got {other:?}")));
21127 }
21128 };
21129 // <name> FOR ORDINALITY
21130 if matches!(self.peek(), Token::For) {
21131 self.advance();
21132 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("ordinality")) {
21133 return Err(self.err("expected ORDINALITY after FOR".into()));
21134 }
21135 self.advance();
21136 return Ok(JsonTableColumn::Ordinality { name });
21137 }
21138 // <name> <type> [FORMAT JSON] {PATH '<p>' | EXISTS [PATH '<p>']} [WITH WRAPPER] [ON …]
21139 let ty = self.parse_column_type_name()?;
21140 let mut format_json = false;
21141 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("format")) {
21142 self.advance();
21143 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("json")) {
21144 return Err(self.err("expected JSON after FORMAT".into()));
21145 }
21146 self.advance();
21147 format_json = true;
21148 }
21149 let mut exists = false;
21150 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("exists")) {
21151 self.advance();
21152 exists = true;
21153 }
21154 let mut path = alloc::format!("$.{name}");
21155 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("path")) {
21156 self.advance();
21157 path = self.parse_json_string_literal("column PATH")?;
21158 }
21159 if !exists && matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("format")) {
21160 // `FORMAT JSON` after PATH (alternate placement).
21161 self.advance();
21162 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("json")) {
21163 self.advance();
21164 }
21165 format_json = true;
21166 }
21167 let mut wrapper = false;
21168 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with")) {
21169 self.advance();
21170 // optional CONDITIONAL/UNCONDITIONAL
21171 if matches!(self.peek(), Token::Ident(s)
21172 if s.eq_ignore_ascii_case("unconditional")
21173 || s.eq_ignore_ascii_case("conditional"))
21174 {
21175 self.advance();
21176 }
21177 if !matches!(self.peek(), Token::Ident(s)
21178 if s.eq_ignore_ascii_case("wrapper") || s.eq_ignore_ascii_case("array"))
21179 {
21180 return Err(self.err("expected WRAPPER after WITH".into()));
21181 }
21182 self.advance();
21183 // optional `ARRAY` after `WRAPPER`, or `WRAPPER` after `ARRAY`
21184 if matches!(self.peek(), Token::Ident(s)
21185 if s.eq_ignore_ascii_case("wrapper") || s.eq_ignore_ascii_case("array"))
21186 {
21187 self.advance();
21188 }
21189 wrapper = true;
21190 }
21191 // ON EMPTY / ON ERROR clauses (two, in any order).
21192 let mut on_empty = JsonTableOnBehavior::Null;
21193 let mut on_error = JsonTableOnBehavior::Null;
21194 for _ in 0..2 {
21195 let behavior = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("error"))
21196 {
21197 self.advance();
21198 Some(JsonTableOnBehavior::Error)
21199 } else if matches!(self.peek(), Token::Null) {
21200 self.advance();
21201 Some(JsonTableOnBehavior::Null)
21202 } else if matches!(self.peek(), Token::Default) {
21203 self.advance();
21204 Some(JsonTableOnBehavior::Default(Box::new(self.parse_expr(0)?)))
21205 } else {
21206 None
21207 };
21208 let Some(behavior) = behavior else { break };
21209 // `ON {EMPTY|ERROR}`
21210 if !matches!(self.peek(), Token::On) {
21211 return Err(self.err("expected ON after JSON_TABLE column behavior".into()));
21212 }
21213 self.advance();
21214 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("empty")) {
21215 self.advance();
21216 on_empty = behavior;
21217 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("error")) {
21218 self.advance();
21219 on_error = behavior;
21220 } else {
21221 return Err(self.err("expected EMPTY or ERROR after ON".into()));
21222 }
21223 }
21224 Ok(JsonTableColumn::Regular {
21225 name,
21226 ty,
21227 path,
21228 exists,
21229 format_json,
21230 wrapper,
21231 on_empty,
21232 on_error,
21233 })
21234 }
21235
21236 fn parse_optional_alias_with_columns(
21237 &mut self,
21238 ) -> Result<(Option<String>, Vec<String>), ParseError> {
21239 let alias = self.parse_optional_alias()?;
21240 if alias.is_none() {
21241 return Ok((None, Vec::new()));
21242 }
21243 let mut cols: Vec<String> = Vec::new();
21244 if matches!(self.peek(), Token::LParen) {
21245 self.advance();
21246 while let Token::Ident(s) | Token::QuotedIdent(s) = self.peek().clone() {
21247 self.advance();
21248 cols.push(s);
21249 if matches!(self.peek(), Token::Comma) {
21250 self.advance();
21251 continue;
21252 }
21253 break;
21254 }
21255 if matches!(self.peek(), Token::RParen) {
21256 self.advance();
21257 }
21258 }
21259 Ok((alias, cols))
21260 }
21261
21262 /// v7.37.16 — parse a `left(str, n)` / `right(str, n)` function call
21263 /// whose keyword token was already consumed and whose `(` is the
21264 /// current token. Factored out of `parse_atom` (and marked
21265 /// `#[inline(never)]`) so its `Vec`/loop locals stay OFF the giant
21266 /// recursive `parse_atom` frame — inlining them there enlarges the
21267 /// per-nesting-level stack cost that `MAX_NEST_DEPTH` is tuned
21268 /// against, risking an overflow before the budget triggers.
21269 #[inline(never)]
21270 fn parse_lr_string_function_call(&mut self, name: &str) -> Result<Expr, ParseError> {
21271 self.advance(); // (
21272 let mut args = Vec::new();
21273 if !matches!(self.peek(), Token::RParen) {
21274 loop {
21275 args.push(self.parse_expr(0)?);
21276 match self.peek() {
21277 Token::Comma => {
21278 self.advance();
21279 }
21280 Token::RParen => break,
21281 other => {
21282 return Err(self.err(alloc::format!(
21283 "expected ',' or ')' in {name}() args, got {other:?}"
21284 )));
21285 }
21286 }
21287 }
21288 }
21289 self.advance(); // )
21290 Ok(Expr::FunctionCall {
21291 name: name.into(),
21292 args,
21293 })
21294 }
21295
21296 /// FROM-clause: a primary table reference plus zero-or-more joined
21297 /// peers expressed via either `, <table>` (cross-product, no ON) or
21298 /// `[INNER|LEFT|RIGHT [OUTER]|FULL [OUTER]|CROSS] JOIN <table> [ON expr]`.
21299 /// v1.10 keeps the join list flat (left-associative nested-loop
21300 /// semantics).
21301 fn parse_from_clause(&mut self) -> Result<FromClause, ParseError> {
21302 let primary = self.parse_table_ref()?;
21303 let primary_qual = primary
21304 .alias
21305 .clone()
21306 .unwrap_or_else(|| primary.name.clone());
21307 let joins = self.parse_from_joins(&primary_qual)?;
21308 Ok(FromClause { primary, joins })
21309 }
21310
21311 /// v7.39 (round 420) — the join tail of a FROM clause, factored out of
21312 /// [`Self::parse_from_clause`] so MySQL's multi-table UPDATE can read the
21313 /// SAME grammar after its target table has already been consumed.
21314 /// (`advance()` destroys the tokens it returns — `mem::replace(.., Eof)`
21315 /// — so re-parsing by rewinding `self.pos` is not possible; the tail must
21316 /// be parsed forward, once.)
21317 /// `left_primary_qual` is the qualifier (alias, else name) of whatever
21318 /// sits to the LEFT of the first join — the FROM primary, or the UPDATE
21319 /// target in the MySQL multi-table form. It only feeds the `USING (…)`
21320 /// desugaring, which needs a name for the left side of each equality.
21321 fn parse_from_joins(&mut self, left_primary_qual: &str) -> Result<Vec<FromJoin>, ParseError> {
21322 let mut joins = Vec::new();
21323 loop {
21324 // `, <table>` — cross-product with no ON.
21325 if matches!(self.peek(), Token::Comma) {
21326 self.advance();
21327 let table = self.parse_table_ref()?;
21328 joins.push(FromJoin {
21329 kind: JoinKind::Cross,
21330 table,
21331 on: None,
21332 using_cols: None,
21333 natural: false,
21334 });
21335 continue;
21336 }
21337 // v7.37.16 — optional leading `NATURAL` before the join
21338 // kind: `NATURAL JOIN`, `NATURAL LEFT JOIN`, etc. NATURAL is
21339 // not a lexer keyword (it arrives as a bare Ident), so match
21340 // it case-insensitively here. When present, no ON/USING
21341 // clause is allowed — the common columns are resolved at
21342 // execution time.
21343 let natural = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("natural"));
21344 if natural {
21345 self.advance();
21346 }
21347 // Explicit JOIN syntax. Accept INNER JOIN, LEFT [OUTER] JOIN,
21348 // CROSS JOIN, and bare JOIN (defaults to INNER).
21349 let kind =
21350 match self.peek() {
21351 Token::Inner => {
21352 self.advance();
21353 if !matches!(self.peek(), Token::Join) {
21354 return Err(self
21355 .err(format!("expected JOIN after INNER, got {:?}", self.peek())));
21356 }
21357 self.advance();
21358 JoinKind::Inner
21359 }
21360 Token::Left => {
21361 self.advance();
21362 if matches!(self.peek(), Token::Outer) {
21363 self.advance();
21364 }
21365 if !matches!(self.peek(), Token::Join) {
21366 return Err(self.err(format!(
21367 "expected JOIN after LEFT [OUTER], got {:?}",
21368 self.peek()
21369 )));
21370 }
21371 self.advance();
21372 JoinKind::Left
21373 }
21374 Token::Cross => {
21375 self.advance();
21376 if !matches!(self.peek(), Token::Join) {
21377 return Err(self
21378 .err(format!("expected JOIN after CROSS, got {:?}", self.peek())));
21379 }
21380 self.advance();
21381 JoinKind::Cross
21382 }
21383 // v7.37.16 — RIGHT [OUTER] JOIN. OUTER is optional noise.
21384 Token::Right => {
21385 self.advance();
21386 if matches!(self.peek(), Token::Outer) {
21387 self.advance();
21388 }
21389 if !matches!(self.peek(), Token::Join) {
21390 return Err(self.err(format!(
21391 "expected JOIN after RIGHT [OUTER], got {:?}",
21392 self.peek()
21393 )));
21394 }
21395 self.advance();
21396 JoinKind::Right
21397 }
21398 // v7.37.16 — FULL [OUTER] JOIN. OUTER is optional noise.
21399 Token::Full => {
21400 self.advance();
21401 if matches!(self.peek(), Token::Outer) {
21402 self.advance();
21403 }
21404 if !matches!(self.peek(), Token::Join) {
21405 return Err(self.err(format!(
21406 "expected JOIN after FULL [OUTER], got {:?}",
21407 self.peek()
21408 )));
21409 }
21410 self.advance();
21411 JoinKind::FullOuter
21412 }
21413 Token::Join => {
21414 self.advance();
21415 JoinKind::Inner
21416 }
21417 _ => break,
21418 };
21419 let table = self.parse_table_ref()?;
21420 // v7.37.7 C.1 — USING (col_list) sugar. Desugars to
21421 // `prev_table.col1 = table.col1 AND prev_table.col2 = table.col2 …`
21422 // where prev_table is the most-recent left-side table
21423 // (the previous join's table if any, else the FROM primary).
21424 // PG semantics around column merging are richer (USING'd
21425 // cols become deduplicated single output columns); for
21426 // sugar purposes the predicate-only form covers the
21427 // baseline corpus shape and chained `… JOIN x USING (k)
21428 // JOIN y USING (k)` calls.
21429 // v7.37.16 — NATURAL joins carry no ON/USING clause; the
21430 // common columns resolve at execution time.
21431 if natural {
21432 joins.push(FromJoin {
21433 kind,
21434 table,
21435 on: None,
21436 using_cols: None,
21437 natural: true,
21438 });
21439 continue;
21440 }
21441 let using_match = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("using"));
21442 // v7.37.16 — capture the USING column list (in addition to
21443 // the ON desugar below) so the executor can perform PG's
21444 // column-merge on the output side.
21445 let mut using_cols: Option<Vec<String>> = None;
21446 let on = if matches!(self.peek(), Token::On) {
21447 self.advance();
21448 Some(self.parse_expr(0)?)
21449 } else if using_match {
21450 self.advance();
21451 if !matches!(self.peek(), Token::LParen) {
21452 return Err(
21453 self.err(format!("expected '(' after USING, got {:?}", self.peek()))
21454 );
21455 }
21456 self.advance();
21457 let mut cols: Vec<String> = Vec::new();
21458 loop {
21459 match self.peek().clone() {
21460 Token::Ident(s) | Token::QuotedIdent(s) => {
21461 self.advance();
21462 cols.push(s);
21463 }
21464 other => {
21465 return Err(self.err(format!(
21466 "expected column name inside USING (…), got {other:?}"
21467 )));
21468 }
21469 }
21470 match self.peek() {
21471 Token::Comma => {
21472 self.advance();
21473 continue;
21474 }
21475 Token::RParen => {
21476 self.advance();
21477 break;
21478 }
21479 other => {
21480 return Err(self.err(format!(
21481 "expected ',' or ')' inside USING (…), got {other:?}"
21482 )));
21483 }
21484 }
21485 }
21486 if cols.is_empty() {
21487 return Err(self.err("USING (…) requires at least one column".to_string()));
21488 }
21489 using_cols = Some(cols.clone());
21490 // Pick the left-side alias: prev join's table if any,
21491 // else FROM primary. Use alias when present, else
21492 // table name (PG-equivalent qualifier).
21493 let left_qual: String = joins
21494 .last()
21495 .map(|j| {
21496 j.table
21497 .alias
21498 .clone()
21499 .unwrap_or_else(|| j.table.name.clone())
21500 })
21501 .unwrap_or_else(|| alloc::string::String::from(left_primary_qual));
21502 let right_qual = table.alias.clone().unwrap_or_else(|| table.name.clone());
21503 let mut iter = cols.into_iter().map(|c| Expr::Binary {
21504 lhs: alloc::boxed::Box::new(Expr::Column(crate::ast::ColumnName {
21505 qualifier: Some(left_qual.clone()),
21506 name: c.clone(),
21507 })),
21508 op: crate::ast::BinOp::Eq,
21509 rhs: alloc::boxed::Box::new(Expr::Column(crate::ast::ColumnName {
21510 qualifier: Some(right_qual.clone()),
21511 name: c,
21512 })),
21513 });
21514 let first = iter.next().expect("at least one col");
21515 Some(iter.fold(first, |acc, pred| Expr::Binary {
21516 lhs: alloc::boxed::Box::new(acc),
21517 op: crate::ast::BinOp::And,
21518 rhs: alloc::boxed::Box::new(pred),
21519 }))
21520 } else if kind == JoinKind::Cross {
21521 None
21522 } else {
21523 return Err(self.err(format!(
21524 "expected ON or USING after {:?} JOIN, got {:?}",
21525 kind,
21526 self.peek()
21527 )));
21528 };
21529 joins.push(FromJoin {
21530 kind,
21531 table,
21532 on,
21533 using_cols,
21534 natural: false,
21535 });
21536 }
21537 Ok(joins)
21538 }
21539
21540 /// Optional alias after an expression or table:
21541 /// `AS <ident>` is unambiguous; a bare `<ident>` directly after is also
21542 /// accepted (PG-style implicit alias). Returns `None` if the next token
21543 /// is not alias-shaped (e.g. comma, FROM, WHERE, semicolon, EOF, operator).
21544 fn parse_optional_alias(&mut self) -> Result<Option<String>, ParseError> {
21545 if matches!(self.peek(), Token::As) {
21546 self.advance();
21547 // v7.39 (round 340, V56) — after AS the next token MUST be an
21548 // identifier. This used to return None and "let the caller
21549 // surface the error on the next expectation", but when AS is
21550 // the LAST token there is no next expectation: `SELECT 1 AS`
21551 // parsed clean and silently dropped the alias. PG rejects it.
21552 if let Token::Ident(_) | Token::QuotedIdent(_) = self.peek() {
21553 return self.expect_ident_like().map(Some);
21554 }
21555 return Err(self.err(alloc::format!(
21556 "expected an alias after AS, got {:?}",
21557 self.peek()
21558 )));
21559 }
21560 // v7.17.0 Phase 1.3 — implicit alias (no `AS`). PG's
21561 // grammar reserves a long list of follow-keywords from the
21562 // alias slot. SPG's bareword approximation: skip a small
21563 // set of idents that would otherwise be swallowed as the
21564 // table alias and break trailing clauses like CREATE
21565 // MATERIALIZED VIEW … WITH [NO] DATA or future ON
21566 // CONFLICT WHERE shapes.
21567 if let Token::Ident(s) | Token::QuotedIdent(s) = self.peek() {
21568 if is_alias_stopword(s) {
21569 return Ok(None);
21570 }
21571 return Ok(self.expect_ident_like().ok());
21572 }
21573 Ok(None)
21574 }
21575
21576 /// Pratt loop. `min_prec` is the minimum binary-op precedence we'll accept.
21577 fn parse_expr(&mut self, min_prec: u8) -> Result<Expr, ParseError> {
21578 // v7.30.2 (mailrs round-25 ask 2) — nesting budget: a parse
21579 // error beats a stack overflow (an overflow aborts the
21580 // embedding host process).
21581 self.enter_nested()?;
21582 let r = self.parse_expr_inner(min_prec);
21583 self.nest_depth -= 1;
21584 r
21585 }
21586
21587 /// `OPERATOR([schema.]<op>)` — PG's explicit-operator spelling.
21588 /// When the upcoming tokens form one, return the underlying
21589 /// operator token and the position just past the closing paren
21590 /// so the binary loop can dispatch on the plain operator.
21591 fn peek_explicit_operator(&self) -> Option<(usize, Token)> {
21592 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("operator")) {
21593 return None;
21594 }
21595 if !matches!(self.tokens.get(self.pos + 1), Some(Token::LParen)) {
21596 return None;
21597 }
21598 let mut i = self.pos + 2;
21599 // Optional schema qualifier (pg_catalog.<op> etc.).
21600 if matches!(self.tokens.get(i), Some(Token::Ident(_)))
21601 && matches!(self.tokens.get(i + 1), Some(Token::Dot))
21602 {
21603 i += 2;
21604 }
21605 let op_tok = self.tokens.get(i)?.clone();
21606 if !matches!(self.tokens.get(i + 1), Some(Token::RParen)) {
21607 return None;
21608 }
21609 Some((i + 2, op_tok))
21610 }
21611
21612 /// PG operator symbols that lower onto function calls in
21613 /// binary position: `~` / `~*` / `!~` / `!~*` (regex match
21614 /// family → regexp_like, comparison rung), `^@` (starts_with,
21615 /// comparison rung), `^` (power, tighter than `*`), `#`
21616 /// (integer XOR via `(a|b) - (a&b)` — the AND bits are a
21617 /// subset of the OR bits so the subtraction never borrows).
21618 fn try_symbol_operator(
21619 &mut self,
21620 lhs: &Expr,
21621 min_prec: u8,
21622 ) -> Result<Option<Expr>, ParseError> {
21623 enum Sym {
21624 Regex { ci: bool, negated: bool },
21625 Like { ci: bool, negated: bool },
21626 StartsWith,
21627 Power,
21628 Xor,
21629 RangeAdjacent,
21630 }
21631 // v7.39 (IS-precedence knife) — the low-precedence postfix
21632 // predicates ride this existing leaf call (zero new frame slots
21633 // on the nesting chain).
21634 if let Some(e) = self.parse_postfix_predicate(lhs, min_prec)? {
21635 return Ok(Some(e));
21636 }
21637 let (sym, prec): (Sym, u8) = match self.peek() {
21638 Token::Tilde => (
21639 Sym::Regex {
21640 ci: false,
21641 negated: false,
21642 },
21643 5,
21644 ),
21645 Token::TildeStar => (
21646 Sym::Regex {
21647 ci: true,
21648 negated: false,
21649 },
21650 5,
21651 ),
21652 Token::NotTilde => (
21653 Sym::Regex {
21654 ci: false,
21655 negated: true,
21656 },
21657 5,
21658 ),
21659 Token::NotTildeStar => (
21660 Sym::Regex {
21661 ci: true,
21662 negated: true,
21663 },
21664 5,
21665 ),
21666 // v7.37 D.25 — PG operator spellings of LIKE/ILIKE.
21667 Token::DoubleTilde => (
21668 Sym::Like {
21669 ci: false,
21670 negated: false,
21671 },
21672 5,
21673 ),
21674 Token::DoubleTildeStar => (
21675 Sym::Like {
21676 ci: true,
21677 negated: false,
21678 },
21679 5,
21680 ),
21681 Token::NotDoubleTilde => (
21682 Sym::Like {
21683 ci: false,
21684 negated: true,
21685 },
21686 5,
21687 ),
21688 Token::NotDoubleTildeStar => (
21689 Sym::Like {
21690 ci: true,
21691 negated: true,
21692 },
21693 5,
21694 ),
21695 Token::CaretAt => (Sym::StartsWith, 5),
21696 // PG `^` is exponentiation; MySQL `^` is bitwise XOR (and binds
21697 // tighter than `* / & |`, which the prec-9 rung preserves —
21698 // v7.39 round 407: +1 from the pre-XOR ladder's rung 8).
21699 Token::Caret if self.mysql_dialect => (Sym::Xor, 9),
21700 Token::Caret => (Sym::Power, 9),
21701 // v7.39 (round 760, F31-B1) — `#` is a generic operator too:
21702 // PG answers `5 # 3 + 1` as `5 # 4` = 1 (additive first).
21703 Token::Hash => (Sym::Xor, 6),
21704 Token::Adjacent => (Sym::RangeAdjacent, 5),
21705 _ => return Ok(None),
21706 };
21707 if prec < min_prec {
21708 return Ok(None);
21709 }
21710 self.advance();
21711 let rhs = self.parse_expr(prec + 1)?;
21712 let out = match sym {
21713 Sym::Regex { ci, negated } => {
21714 let mut args = alloc::vec![lhs.clone(), rhs];
21715 if ci {
21716 args.push(Expr::Literal(Literal::String(String::from("i"))));
21717 }
21718 maybe_not(
21719 Expr::FunctionCall {
21720 name: String::from("regexp_like"),
21721 args,
21722 },
21723 negated,
21724 )
21725 }
21726 Sym::Like { ci, negated } => Expr::Like {
21727 expr: alloc::boxed::Box::new(lhs.clone()),
21728 pattern: alloc::boxed::Box::new(rhs),
21729 negated,
21730 case_insensitive: ci,
21731 },
21732 Sym::StartsWith => Expr::FunctionCall {
21733 name: String::from("starts_with"),
21734 args: alloc::vec![lhs.clone(), rhs],
21735 },
21736 Sym::Power => Expr::FunctionCall {
21737 name: String::from("power"),
21738 args: alloc::vec![lhs.clone(), rhs],
21739 },
21740 // `#` bitwise XOR — a real operator now (was desugared to
21741 // `(a|b)-(a&b)`, algebraically identical for integers but
21742 // undefined for bit strings; the direct op handles both).
21743 Sym::Xor => Expr::Binary {
21744 lhs: Box::new(lhs.clone()),
21745 op: BinOp::BitXor,
21746 rhs: Box::new(rhs),
21747 },
21748 // range `-|-` "is adjacent to" — lowered to a catalog function.
21749 Sym::RangeAdjacent => Expr::FunctionCall {
21750 name: String::from("range_adjacent"),
21751 args: alloc::vec![lhs.clone(), rhs],
21752 },
21753 };
21754 Ok(Some(out))
21755 }
21756
21757 /// v7.39 (IS-precedence knife) — the LOW-precedence postfix
21758 /// predicates, moved out of the tight postfix-cast loop: PG binds
21759 /// `IS [NOT] NULL/TRUE/FALSE/UNKNOWN/DISTINCT FROM/JSON/NORMALIZED`
21760 /// looser than EVERY binary operator (only NOT/AND/OR are looser),
21761 /// and BETWEEN/IN/LIKE/ILIKE/SIMILAR at the comparison rung — so
21762 /// `1 + 1 IS NULL` is `(1+1) IS NULL`, not `1 + (1 IS NULL)`.
21763 /// Returns Ok(consumed expr) when a predicate fired, Err(expr back)
21764 /// when nothing at this position belongs to the family. Out-of-line
21765 /// (`inline(never)`): the caller sits on the per-nesting-level frame
21766 /// chain that MAX_NEST_DEPTH is tuned against.
21767 #[inline(never)]
21768 fn parse_postfix_predicate(
21769 &mut self,
21770 lhs: &Expr,
21771 min_prec: u8,
21772 ) -> Result<Option<Expr>, ParseError> {
21773 // Reached through try_symbol_operator (an existing leaf call of
21774 // the binary loop) so NO new stack slots land on the per-nesting
21775 // frame chain; the lhs clones only when a predicate actually
21776 // consumes it.
21777 match self.peek() {
21778 // v7.39 (round 407) — IS is rung 4, the BETWEEN/IN/LIKE
21779 // comparison family rung 5 (each +1 from the pre-XOR ladder).
21780 Token::Is if min_prec <= 4 => {}
21781 Token::Between | Token::In | Token::Like if min_prec <= 5 => {}
21782 Token::Not
21783 if min_prec <= 5
21784 && matches!(
21785 self.tokens.get(self.pos + 1),
21786 Some(Token::Between | Token::In | Token::Like)
21787 ) => {}
21788 Token::Not | Token::Ident(_)
21789 if min_prec <= 5
21790 && (matches!(self.peek(), Token::Ident(s)
21791 if s.eq_ignore_ascii_case("ilike")
21792 || (self.mysql_dialect
21793 && (s.eq_ignore_ascii_case("regexp")
21794 || s.eq_ignore_ascii_case("rlike")))
21795 || (s.eq_ignore_ascii_case("similar")
21796 && matches!(self.tokens.get(self.pos + 1), Some(Token::To))))
21797 || (matches!(self.peek(), Token::Not)
21798 && matches!(self.tokens.get(self.pos + 1),
21799 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("ilike")
21800 || (self.mysql_dialect
21801 && (s.eq_ignore_ascii_case("regexp")
21802 || s.eq_ignore_ascii_case("rlike")))
21803 || s.eq_ignore_ascii_case("similar")))) => {}
21804 _ => return Ok(None),
21805 }
21806 let mut expr = lhs.clone();
21807 // IS family: rung 4 (NOT's operand parses at 4, so `NOT x IS NULL`
21808 // still groups as NOT (x IS NULL); OR/XOR/AND at 1-3 stay outside).
21809 if min_prec <= 4 {
21810 if matches!(self.peek(), Token::Is) {
21811 self.advance();
21812 let negated = if matches!(self.peek(), Token::Not) {
21813 self.advance();
21814 true
21815 } else {
21816 false
21817 };
21818 // v7.9.27b — `IS [NOT] DISTINCT FROM <rhs>`.
21819 // mailrs pg_dump.
21820 if matches!(self.peek(), Token::Distinct) {
21821 self.advance();
21822 if !matches!(self.peek(), Token::From) {
21823 return Err(self.err(format!(
21824 "expected FROM after IS{} DISTINCT, got {:?}",
21825 if negated { " NOT" } else { "" },
21826 self.peek()
21827 )));
21828 }
21829 self.advance();
21830 // Right-hand side: parse at the same precedence
21831 // tier as comparison (rung 5) so `x IS DISTINCT FROM a + b`
21832 // groups as `x IS DISTINCT FROM (a + b)`.
21833 let rhs = self.parse_expr(5)?;
21834 let op = if negated {
21835 BinOp::IsNotDistinctFrom
21836 } else {
21837 BinOp::IsDistinctFrom
21838 };
21839 expr = Expr::Binary {
21840 op,
21841 lhs: Box::new(expr),
21842 rhs: Box::new(rhs),
21843 };
21844 {
21845 return Ok(Some(expr));
21846 }
21847 }
21848 // v7.37.17 (17.6 siblings) — SQL:2016 / PG 16
21849 // `IS [NOT] JSON [VALUE|OBJECT|ARRAY|SCALAR]`.
21850 // Lowers onto pg_is_json(x, kind); NOT wraps the
21851 // call in a logical negation.
21852 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
21853 if s.eq_ignore_ascii_case("json"))
21854 {
21855 self.advance(); // JSON
21856 let kind = match self.peek() {
21857 Token::Ident(s) | Token::QuotedIdent(s)
21858 if matches!(
21859 s.to_ascii_lowercase().as_str(),
21860 "value" | "object" | "array" | "scalar"
21861 ) =>
21862 {
21863 let k = s.to_ascii_lowercase();
21864 self.advance();
21865 k
21866 }
21867 _ => "value".to_string(),
21868 };
21869 let call = Expr::FunctionCall {
21870 name: "pg_is_json".to_string(),
21871 args: alloc::vec![expr, Expr::Literal(Literal::String(kind)),],
21872 };
21873 expr = if negated {
21874 Expr::Unary {
21875 op: UnOp::Not,
21876 expr: Box::new(call),
21877 }
21878 } else {
21879 call
21880 };
21881 {
21882 return Ok(Some(expr));
21883 }
21884 }
21885 // v7.38 (read01 sweep) — SQL:2016 `x IS [NOT] [form]
21886 // NORMALIZED` (form ∈ NFC/NFD/NFKC/NFKD, default NFC).
21887 // Lowers onto is_normalized(x [, 'FORM']); NOT negates.
21888 {
21889 let form_kw = match self.peek() {
21890 Token::Ident(s) | Token::QuotedIdent(s)
21891 if matches!(
21892 s.to_ascii_uppercase().as_str(),
21893 "NFC" | "NFD" | "NFKC" | "NFKD"
21894 ) && matches!(
21895 self.tokens.get(self.pos + 1),
21896 Some(Token::Ident(n) | Token::QuotedIdent(n))
21897 if n.eq_ignore_ascii_case("normalized")
21898 ) =>
21899 {
21900 Some(s.to_ascii_uppercase())
21901 }
21902 _ => None,
21903 };
21904 let bare_normalized = form_kw.is_none()
21905 && matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
21906 if s.eq_ignore_ascii_case("normalized"));
21907 if form_kw.is_some() || bare_normalized {
21908 if form_kw.is_some() {
21909 self.advance(); // form keyword
21910 }
21911 self.advance(); // NORMALIZED
21912 let mut args = alloc::vec![expr];
21913 if let Some(f) = form_kw {
21914 args.push(Expr::Literal(Literal::String(f)));
21915 }
21916 let call = Expr::FunctionCall {
21917 name: "is_normalized".to_string(),
21918 args,
21919 };
21920 expr = if negated {
21921 Expr::Unary {
21922 op: UnOp::Not,
21923 expr: Box::new(call),
21924 }
21925 } else {
21926 call
21927 };
21928 {
21929 return Ok(Some(expr));
21930 }
21931 }
21932 }
21933 // `x IS [NOT] TRUE | FALSE | UNKNOWN` — the
21934 // three-valued boolean tests. IS TRUE/FALSE never
21935 // return NULL, so they lower to CASE forms whose
21936 // ELSE catches the NULL branch; IS UNKNOWN on a
21937 // boolean is exactly IS NULL.
21938 if matches!(self.peek(), Token::True | Token::False)
21939 || matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("unknown"))
21940 {
21941 let tok = self.advance();
21942 let test = match tok {
21943 Token::True => Some(true),
21944 Token::False => Some(false),
21945 _ => None, // UNKNOWN
21946 };
21947 // v7.39 (round 328, V45) — kept as what the user
21948 // wrote. These used to be lowered here into `CASE` /
21949 // `IS NULL`; the semantics were right but the AST no
21950 // longer knew the form, so `CHECK ((a > 1) IS TRUE)`
21951 // was echoed back as
21952 // `CHECK ((CASE WHEN (a > 1) THEN TRUE ELSE FALSE END))`.
21953 expr = Expr::BoolTest {
21954 expr: Box::new(expr),
21955 value: test,
21956 negated,
21957 };
21958 {
21959 return Ok(Some(expr));
21960 }
21961 }
21962 if !matches!(self.peek(), Token::Null) {
21963 return Err(self.err(format!(
21964 "expected NULL, DISTINCT, JSON, TRUE, FALSE or UNKNOWN after IS{}, got {:?}",
21965 if negated { " NOT" } else { "" },
21966 self.peek()
21967 )));
21968 }
21969 self.advance();
21970 expr = Expr::IsNull {
21971 expr: Box::new(expr),
21972 negated,
21973 };
21974 {
21975 return Ok(Some(expr));
21976 }
21977 }
21978 }
21979 // BETWEEN / IN / LIKE / ILIKE / SIMILAR: comparison rung (5).
21980 if min_prec <= 5 {
21981 // `x [NOT] BETWEEN a AND b`, `x [NOT] IN (...)`, `x [NOT] LIKE p`.
21982 // Look one token ahead so a stray `NOT` not followed by any of
21983 // these flows through to the early return below untouched.
21984 let negated = if matches!(self.peek(), Token::Not) {
21985 let next = self.tokens.get(self.pos + 1);
21986 matches!(next, Some(Token::Between | Token::In | Token::Like))
21987 || matches!(next, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("ilike")
21988 || (self.mysql_dialect
21989 && (s.eq_ignore_ascii_case("regexp") || s.eq_ignore_ascii_case("rlike")))
21990 || s.eq_ignore_ascii_case("similar"))
21991 } else {
21992 false
21993 };
21994 if negated {
21995 self.advance();
21996 }
21997 if matches!(self.peek(), Token::Between) {
21998 expr = self.parse_between_tail(expr, negated)?;
21999 {
22000 return Ok(Some(expr));
22001 }
22002 }
22003 if matches!(self.peek(), Token::In) {
22004 if self.suppress_in_tail && !negated {
22005 // POSITION(sub IN str) — IN belongs to the
22006 // enclosing function syntax; stop here.
22007 {
22008 return Ok(None);
22009 }
22010 }
22011 expr = self.parse_in_tail(expr, negated)?;
22012 {
22013 return Ok(Some(expr));
22014 }
22015 }
22016 // v7.39 (read01 regexp.c) — `x [NOT] SIMILAR TO p [ESCAPE e]`
22017 // lowers onto the internal __similar_to(expr, pat[, esc]) call
22018 // (the SQL→regex transform runs inside, in the backtracking-
22019 // friendly shape SPG's matcher needs).
22020 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("similar"))
22021 && matches!(self.tokens.get(self.pos + 1), Some(Token::To))
22022 {
22023 self.advance(); // SIMILAR
22024 self.advance(); // TO
22025 let pattern = self.parse_expr(6)?;
22026 let mut args = alloc::vec![expr, pattern];
22027 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("escape")) {
22028 self.advance();
22029 args.push(self.parse_expr(6)?);
22030 }
22031 let call = Expr::FunctionCall {
22032 name: "__similar_to".to_string(),
22033 args,
22034 };
22035 expr = maybe_not(call, negated);
22036 {
22037 return Ok(Some(expr));
22038 }
22039 }
22040 if matches!(self.peek(), Token::Like) {
22041 self.advance();
22042 // `x [NOT] LIKE ANY/ALL (ARRAY[...])` — quantified LIKE.
22043 if let Some(q) = self.try_like_any_all(&expr, negated, false)? {
22044 expr = q;
22045 {
22046 return Ok(Some(expr));
22047 }
22048 }
22049 // Pattern at the same precedence as other comparison RHSes —
22050 // 5 leaves AND/OR alone so `a LIKE 'x%' AND b` parses right.
22051 let mut pattern = self.parse_expr(6)?;
22052 // `ESCAPE 'c'` — rewrite a literal pattern to the
22053 // default backslash escape at parse time. Custom
22054 // escapes on non-literal patterns would need
22055 // matcher support; error honestly.
22056 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("escape")) {
22057 self.advance();
22058 let esc = self.parse_expr(6)?;
22059 pattern = Self::rewrite_like_escape(pattern, esc).map_err(|m| self.err(m))?;
22060 }
22061 expr = Expr::Like {
22062 expr: Box::new(expr),
22063 pattern: Box::new(pattern),
22064 negated,
22065 case_insensitive: false,
22066 };
22067 {
22068 return Ok(Some(expr));
22069 }
22070 }
22071 // v7.25 (round-17) — ILIKE: case-insensitive LIKE. The
22072 // keyword reaches us as a plain identifier.
22073 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("ilike")) {
22074 self.advance();
22075 if let Some(q) = self.try_like_any_all(&expr, negated, true)? {
22076 expr = q;
22077 {
22078 return Ok(Some(expr));
22079 }
22080 }
22081 let pattern = self.parse_expr(6)?;
22082 expr = Expr::Like {
22083 expr: Box::new(expr),
22084 pattern: Box::new(pattern),
22085 negated,
22086 case_insensitive: true,
22087 };
22088 {
22089 return Ok(Some(expr));
22090 }
22091 }
22092 // v7.39 (round 380) — MySQL's REGEXP / RLIKE regex-match
22093 // operator (RLIKE is the alias). It is a keyword, not `~`, and
22094 // matches case-insensitively under the default collation, so it
22095 // lowers onto the same `regexp_like(expr, pattern, 'i')` the
22096 // `~*` operator uses, wrapped in NOT when negated.
22097 if self.mysql_dialect
22098 && matches!(self.peek(), Token::Ident(s)
22099 if s.eq_ignore_ascii_case("regexp") || s.eq_ignore_ascii_case("rlike"))
22100 {
22101 self.advance();
22102 let pattern = self.parse_expr(6)?;
22103 let call = Expr::FunctionCall {
22104 name: String::from("regexp_like"),
22105 args: alloc::vec![
22106 expr,
22107 pattern,
22108 Expr::Literal(Literal::String(String::from("i"))),
22109 ],
22110 };
22111 return Ok(Some(maybe_not(call, negated)));
22112 }
22113 }
22114 let _ = expr;
22115 Ok(None)
22116 }
22117
22118 fn parse_expr_inner(&mut self, min_prec: u8) -> Result<Expr, ParseError> {
22119 let mut lhs = self.parse_unary()?;
22120 let mut chain_len = 0usize;
22121 loop {
22122 // OPERATOR([schema.]op) reduces to its underlying
22123 // operator token before the normal dispatch.
22124 let explicit = self.peek_explicit_operator();
22125 let dispatch = match &explicit {
22126 Some((_, tok)) => self.binop_here(tok),
22127 None => self.binop_here(self.peek()),
22128 };
22129 let Some((op, prec)) = dispatch else {
22130 // v7.39 (round 539) — `OPERATOR(pg_catalog.~)` and the rest
22131 // of the symbol family. `binop_here` answers None for them
22132 // because they lower onto function calls rather than a
22133 // BinOp, and the fallback below reads `self.peek()` — the
22134 // word OPERATOR, not the operator. `pg_dump` writes every
22135 // catalog predicate this way, so its first query failed
22136 // and no dump ran:
22137 //
22138 // AND c.relname OPERATOR(pg_catalog.~) '^(t)$'
22139 //
22140 // Collapsing the wrapper to the operator it names puts the
22141 // token where the fallback already looks.
22142 if let Some((next, op_tok)) = explicit {
22143 self.tokens.splice(self.pos..next, [op_tok]);
22144 }
22145 if let Some(e) = self.try_symbol_operator(&lhs, min_prec)? {
22146 lhs = e;
22147 chain_len += 1;
22148 if chain_len > MAX_BINARY_CHAIN {
22149 return Err(self.err(alloc::format!(
22150 "more than {MAX_BINARY_CHAIN} chained binary operators"
22151 )));
22152 }
22153 continue;
22154 }
22155 break;
22156 };
22157 if prec < min_prec {
22158 break;
22159 }
22160 // v7.30.2 (mailrs round-25 ask 2) — the chain builds
22161 // iteratively but evaluates and drops recursively;
22162 // depth beyond the budget overflows worker stacks.
22163 chain_len += 1;
22164 if chain_len > MAX_BINARY_CHAIN {
22165 return Err(self.err(alloc::format!(
22166 "more than {MAX_BINARY_CHAIN} chained binary operators; rewrite long OR-equality chains as IN (…)"
22167 )));
22168 }
22169 match explicit {
22170 Some((end_pos, _)) => self.pos = end_pos,
22171 None => {
22172 self.advance();
22173 }
22174 }
22175 // v7.10.12 — `x <op> ANY(arr)` / `x <op> ALL(arr)`.
22176 // ANY is a bare ident; ALL is a reserved Token. Both
22177 // require an immediate `(` to disambiguate from
22178 // identifier columns named `any` / `all`.
22179 let any_kind = match self.peek() {
22180 Token::All if matches!(self.tokens.get(self.pos + 1), Some(Token::LParen)) => {
22181 Some(false)
22182 }
22183 Token::Ident(s) | Token::QuotedIdent(s)
22184 if (s.eq_ignore_ascii_case("any")
22185 || s.eq_ignore_ascii_case("some")
22186 || s.eq_ignore_ascii_case("all"))
22187 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen)) =>
22188 {
22189 Some(!s.eq_ignore_ascii_case("all"))
22190 }
22191 _ => None,
22192 };
22193 if let Some(is_any) = any_kind {
22194 lhs = self.parse_any_all_rhs(lhs, op, is_any)?;
22195 continue;
22196 }
22197 let rhs = self.parse_expr(prec + 1)?;
22198 lhs = Expr::Binary {
22199 lhs: Box::new(lhs),
22200 op,
22201 rhs: Box::new(rhs),
22202 };
22203 }
22204 Ok(lhs)
22205 }
22206
22207 /// `x <op> ANY (…)` / `ALL (…)`, both the quantified-subquery form
22208 /// and the array form.
22209 ///
22210 /// `#[inline(never)]` and out of `parse_expr_inner`, which sits on the
22211 /// frame chain `MAX_NEST_DEPTH` is tuned against: a debug build gives
22212 /// this block's `Expr` temporaries and four `format!` sites slots in
22213 /// that frame on every level of `((((1))))`, which never reaches it.
22214 #[inline(never)]
22215 fn parse_any_all_rhs(
22216 &mut self,
22217 lhs: Expr,
22218 op: BinOp,
22219 is_any: bool,
22220 ) -> Result<Expr, ParseError> {
22221 self.advance(); // ident
22222 self.advance(); // (
22223 // `x op ANY (SELECT …)` — the quantified-subquery
22224 // form. `= ANY` is exactly IN; the other operators
22225 // lower onto EXISTS over the subquery as a derived
22226 // table, comparing against its single projection
22227 // aliased __v (x's columns resolve correlated).
22228 // ALL is the negated-EXISTS complement; a NULL
22229 // element makes PG return NULL where this lowering
22230 // returns true — the NOT NULL column case (the
22231 // practical one) is exact.
22232 if matches!(self.peek(), Token::Select) || self.peek_is_with_kw() {
22233 // v7.39 (round 153) — `ANY (WITH … SELECT …)` is
22234 // legal PG too (round-151 sibling). Out-of-line
22235 // (#[inline(never)] helper) — this sits on
22236 // parse_expr's recursive frame and the two-armed
22237 // SELECT temporary blew the nesting-budget stack.
22238 let mut sub = self.parse_any_all_select_body()?;
22239 if !matches!(self.peek(), Token::RParen) {
22240 return Err(self.err(alloc::format!(
22241 "expected ')' after ANY/ALL subquery, got {:?}",
22242 self.peek()
22243 )));
22244 }
22245 self.advance();
22246 if sub.items.len() != 1 {
22247 return Err(self.err(alloc::format!(
22248 "ANY/ALL subquery must return one column, got {}",
22249 sub.items.len()
22250 )));
22251 }
22252 if is_any && matches!(op, BinOp::Eq) {
22253 return Ok(Expr::InSubquery {
22254 expr: Box::new(lhs),
22255 subquery: Box::new(sub),
22256 negated: false,
22257 });
22258 }
22259 // The engine's subquery resolvers materialise
22260 // the single-column result into an ARRAY the
22261 // existing AnyAll three-valued eval consumes.
22262 return Ok(Expr::AnyAll {
22263 expr: Box::new(lhs),
22264 op,
22265 array: Box::new(Expr::ScalarSubquery(Box::new(sub))),
22266 is_any,
22267 });
22268 }
22269 let arr = self.parse_expr(0)?;
22270 if !matches!(self.peek(), Token::RParen) {
22271 return Err(self.err(alloc::format!(
22272 "expected ')' after ANY/ALL argument, got {:?}",
22273 self.peek()
22274 )));
22275 }
22276 self.advance();
22277 Ok(Expr::AnyAll {
22278 expr: Box::new(lhs),
22279 op,
22280 array: Box::new(arr),
22281 is_any,
22282 })
22283 }
22284
22285 /// v7.39 (read01 geo_ops.c) — prefix `@@` (center-of). Out-of-line
22286 /// from `parse_unary` (see the frame-budget note at MAX_NEST_DEPTH).
22287 #[inline(never)]
22288 fn parse_prefix_center(&mut self) -> Result<Expr, ParseError> {
22289 self.advance();
22290 let e = self.parse_expr(9)?;
22291 Ok(build_center_call(e))
22292 }
22293
22294 /// v7.39 (round 508) — a prefix operator that IS a function: `@ x` is
22295 /// `abs(x)`, `# p` is `npoints(p)`, `@-@ p` is `length(p)`. Binds like
22296 /// unary minus.
22297 ///
22298 /// `#[inline(never)]` for the same reason as its neighbours: parse_unary
22299 /// sits on the recursive frame chain MAX_NEST_DEPTH is tuned against, so
22300 /// the Expr-sized local stays out of that frame.
22301 #[inline(never)]
22302 fn parse_prefix_call(&mut self, name: &str) -> Result<Expr, ParseError> {
22303 self.advance();
22304 let e = self.parse_expr(9)?;
22305 Ok(Expr::FunctionCall {
22306 name: alloc::string::String::from(name),
22307 args: alloc::vec![e],
22308 })
22309 }
22310
22311 /// v7.39 (read01 geo_ops.c) — prefix `?|` (vertical) / `?-`
22312 /// (horizontal). Out-of-line from `parse_unary` (frame budget).
22313 #[inline(never)]
22314 fn parse_prefix_geom_axis(&mut self, vertical: bool) -> Result<Expr, ParseError> {
22315 self.advance();
22316 let e = self.parse_expr(9)?;
22317 Ok(Expr::FunctionCall {
22318 name: alloc::string::String::from(if vertical {
22319 "isvertical"
22320 } else {
22321 "ishorizontal"
22322 }),
22323 args: alloc::vec![e],
22324 })
22325 }
22326
22327 /// v7.39 (round 355, M13) — `BINARY <expr>`, lowered onto the same
22328 /// cast the `CAST(x AS BINARY)` spelling produces. It binds tightly:
22329 /// MariaDB reads `BINARY 1 + 1` as `(BINARY 1) + 1` = 2.
22330 #[inline(never)]
22331 fn parse_binary_prefix(&mut self) -> Result<Expr, ParseError> {
22332 self.advance();
22333 let e = self.parse_expr(9)?;
22334 Ok(Expr::Cast {
22335 expr: Box::new(e),
22336 target: CastTarget::Named("binary".to_string()),
22337 })
22338 }
22339
22340 /// The prefix operators that share one shape: take the token, parse
22341 /// an operand at `prec`, wrap it.
22342 ///
22343 /// `#[inline(never)]`, and one function instead of five arms, for the
22344 /// reason the neighbouring `parse_prefix_*` helpers give: `parse_unary`
22345 /// sits on the frame chain `MAX_NEST_DEPTH` is tuned against, and a
22346 /// debug build gives EVERY arm's locals a slot in the frame, whichever
22347 /// arm runs. `((((1))))` reaches none of these arms and was carrying
22348 /// five `Expr`-sized locals per level for them anyway.
22349 #[inline(never)]
22350 fn parse_unary_op(&mut self, op: UnOp, prec: u8) -> Result<Expr, ParseError> {
22351 self.advance();
22352 let e = self.parse_expr(prec)?;
22353 Ok(Expr::Unary {
22354 op,
22355 expr: Box::new(e),
22356 })
22357 }
22358
22359 /// Unary minus. Out-of-line for the frame reason on `parse_unary_op`,
22360 /// and separate from it because of the literal folding below and the
22361 /// `format!` temporaries that folding needs.
22362 #[inline(never)]
22363 fn parse_prefix_minus(&mut self) -> Result<Expr, ParseError> {
22364 self.advance();
22365 // v7.39 (round 549) — fold the sign into an integer literal that
22366 // only fits once it is negative.
22367 //
22368 // `9223372036854775808` is one past i64::MAX, so the lexer hands
22369 // it over as a NUMERIC and `-` on a numeric stays numeric. PG
22370 // folds the sign first, so `-9223372036854775808` is a bigint
22371 // there — and `-9223372036854775808 - 1` raises "bigint out of
22372 // range" where SPG quietly answered -9223372036854775809, a value
22373 // no bigint can hold. The arithmetic itself was already checked;
22374 // only the literal's type was wrong.
22375 if let Token::Numeric(lit) = self.peek()
22376 && let Ok(folded) = alloc::format!("-{lit}").parse::<i64>()
22377 {
22378 self.advance();
22379 return Ok(Expr::Literal(Literal::Integer(folded)));
22380 }
22381 // Unary minus binds tighter than `*`/`/` (now at prec 7 after
22382 // `<->` slotted into 5 and arithmetic shifted up).
22383 let e = self.parse_expr(9)?;
22384 Ok(Expr::Unary {
22385 op: UnOp::Neg,
22386 expr: Box::new(e),
22387 })
22388 }
22389
22390 /// tsquery `!!` prefix negation, lowered to the catalog function.
22391 /// Binds like unary minus. Out-of-line for the frame reason on
22392 /// `parse_unary_op`.
22393 #[inline(never)]
22394 fn parse_prefix_tsquery_not(&mut self) -> Result<Expr, ParseError> {
22395 self.advance();
22396 let e = self.parse_expr(9)?;
22397 Ok(Expr::FunctionCall {
22398 name: String::from("tsquery_not"),
22399 args: alloc::vec![e],
22400 })
22401 }
22402
22403 fn parse_unary(&mut self) -> Result<Expr, ParseError> {
22404 match self.peek() {
22405 // NOT binds tighter than AND / XOR / OR but looser than
22406 // comparisons — its operand takes everything ≥ the comparison
22407 // rung (4), leaving AND (3) / XOR (2) / OR (1) outside so
22408 // `NOT a AND b` groups as `(NOT a) AND b`. (v7.39 round 407:
22409 // was rung 3, behaviour-identical when 3 was unused; AND now
22410 // occupies 3, so this must be 4 to keep NOT tighter than AND.)
22411 Token::Not => self.parse_unary_op(UnOp::Not, 4),
22412 // v7.39 (round 355, M13) — MySQL's `BINARY <expr>` prefix.
22413 // The body is out-of-line: `parse_unary` is one of the three
22414 // frames the parser's MAX_NEST_DEPTH is tuned against, and an
22415 // inline arm here overflowed the native stack in
22416 // `nesting_budget_errors_cleanly` — the guard test caught it,
22417 // exactly as the eval-side cliff did in rounds 346 and 351.
22418 Token::Ident(w) if self.mysql_dialect && w.eq_ignore_ascii_case("binary") => {
22419 self.parse_binary_prefix()
22420 }
22421 // v7.39 (round 353, M10) — MySQL's `!`. It binds TIGHTER than
22422 // arithmetic, unlike NOT: MariaDB answers 1 for `!1 + 1`
22423 // (`(!1)+1`) and 0 for `NOT 1 + 1` (`NOT (1+1)`), measured.
22424 Token::Bang => self.parse_unary_op(UnOp::Not, 9),
22425 Token::Minus => self.parse_prefix_minus(),
22426 // v7.39 (round 507) — unary `+`, which SPG did not have. `+1`
22427 // worked only because the lexer reads it as one signed literal;
22428 // `+ 1`, `+a`, `+(1)` and `1 + +1` were syntax errors, and both
22429 // PG18 and MariaDB take all of them. Binds like unary minus.
22430 Token::Plus => self.parse_unary_op(UnOp::Plus, 9),
22431 // Bitwise NOT binds like unary minus.
22432 Token::Tilde => self.parse_unary_op(UnOp::BitNot, 9),
22433 // v7.39 (read01 geo_ops.c) — prefix `@@` is PG's geometric
22434 // "center of" operator; desugars to center(x). The whole arm
22435 // is out-of-line: parse_unary sits on the per-nesting-level
22436 // frame chain that MAX_NEST_DEPTH is tuned against, so no
22437 // Expr-sized local may live in this frame.
22438 Token::TsMatch => self.parse_prefix_center(),
22439 // v7.39 (round 508) — the prefix operators that are named
22440 // functions in disguise: `@ x` is abs, `# p` is npoints, `@-@ p`
22441 // is length. Out-of-line for the same nesting-frame reason as
22442 // parse_prefix_center — parse_unary sits on the recursive cycle
22443 // MAX_NEST_DEPTH is tuned against, so no Expr-sized local may
22444 // live in this frame.
22445 Token::At => self.parse_prefix_call("abs"),
22446 Token::Hash => self.parse_prefix_call("npoints"),
22447 Token::AtMinusAt => self.parse_prefix_call("length"),
22448 // v7.39 (read01 geo_ops.c) — prefix `?|` / `?-`: "is vertical" /
22449 // "is horizontal" (lseg / line); desugars to the existing
22450 // isvertical()/ishorizontal() functions. Out-of-line for the
22451 // same nesting-frame reason as parse_prefix_center.
22452 Token::JsonKeysAny => self.parse_prefix_geom_axis(true),
22453 Token::GeomHoriz => self.parse_prefix_geom_axis(false),
22454 Token::DoubleBang => self.parse_prefix_tsquery_not(),
22455 _ => self.parse_atom(),
22456 }
22457 }
22458
22459 /// Parse a parenthesised scalar subquery body after the caller has consumed
22460 /// `(` and confirmed the next token is SELECT (or WITH, when `is_with`).
22461 /// v7.37 D.43 — `#[inline(never)]` keeps the large `Statement` local and the
22462 /// SELECT/WITH parse machinery off `parse_atom`'s stack frame; parse_atom sits
22463 /// on the recursive `((…))` cycle whose depth budget is tuned to that frame.
22464 /// v7.39 (read01 round 105) — is the current position `( <subquery-start>`,
22465 /// i.e. an `ARRAY(<subquery>)` and not `ARRAY[...]`? A subquery starts with
22466 /// SELECT, VALUES, or WITH (WITH lexes as a bare ident).
22467 /// `#[inline(never)]`: keeps this guard's locals off parse_atom's frame,
22468 /// which sits on the recursive nesting-budget cycle (a few extra bytes there
22469 /// tips the deep-nesting test into a stack overflow).
22470 #[inline(never)]
22471 fn array_subquery_ahead(&self) -> bool {
22472 if !matches!(self.peek(), Token::LParen) {
22473 return false;
22474 }
22475 matches!(
22476 self.tokens.get(self.pos + 1),
22477 Some(Token::Select | Token::Values)
22478 ) || matches!(
22479 self.tokens.get(self.pos + 1),
22480 Some(Token::Ident(w) | Token::QuotedIdent(w)) if w.eq_ignore_ascii_case("with")
22481 )
22482 }
22483
22484 /// v7.10.10 — `ARRAY[expr, …]` literal body. The `array` ident is consumed
22485 /// and the current token is `[`. `#[inline(never)]` so its `Vec`/loop
22486 /// locals stay off parse_atom's recursive frame (round 105).
22487 #[inline(never)]
22488 fn parse_array_literal_body(&mut self) -> Result<Expr, ParseError> {
22489 self.advance(); // consume `[`
22490 let mut items: Vec<Expr> = Vec::new();
22491 if !matches!(self.peek(), Token::RBracket) {
22492 loop {
22493 // Inside `ARRAY[...]`, a nested `[...]` is a sub-array
22494 // (`ARRAY[[1,2],[3,4]]`), not a pgvector literal.
22495 if matches!(self.peek(), Token::LBracket) {
22496 items.push(self.parse_array_bracket_body()?);
22497 } else {
22498 items.push(self.parse_expr(0)?);
22499 }
22500 match self.peek() {
22501 Token::Comma => {
22502 self.advance();
22503 }
22504 Token::RBracket => break,
22505 other => {
22506 return Err(self.err(alloc::format!(
22507 "expected ',' or ']' in ARRAY literal, got {other:?}"
22508 )));
22509 }
22510 }
22511 }
22512 }
22513 self.advance(); // consume `]`
22514 Ok(Expr::Array(items))
22515 }
22516
22517 /// v7.39 (read01 round 105) — parse `ARRAY(<subquery>)`. The `array` ident
22518 /// is already consumed; the current token is `(`. Desugars to a scalar
22519 /// subquery `SELECT array_agg(c) FROM (<subquery>) AS t(c)`, which collects
22520 /// the subquery's single-column rows in order — reusing the existing
22521 /// ScalarSubquery machinery rather than adding an AST node. `#[inline(never)]`
22522 /// keeps the large `Statement` local off parse_atom's recursive frame.
22523 #[inline(never)]
22524 fn parse_array_subquery(&mut self) -> Result<Expr, ParseError> {
22525 self.advance(); // consume `(`
22526 let is_with = matches!(self.peek(), Token::Ident(w) | Token::QuotedIdent(w)
22527 if w.eq_ignore_ascii_case("with"));
22528 let sub = if is_with {
22529 self.advance(); // WITH
22530 self.parse_with_cte_then_select()?
22531 } else {
22532 self.parse_select_stmt()?
22533 };
22534 if !matches!(self.peek(), Token::RParen) {
22535 return Err(self.err(alloc::format!(
22536 "expected ')' to close ARRAY(subquery), got {:?}",
22537 self.peek()
22538 )));
22539 }
22540 self.advance(); // consume `)`
22541 // Reuse the parser to build the array_agg wrapper from the subquery's
22542 // canonical text — avoids hand-constructing the derived-table AST.
22543 let wrapper = alloc::format!(
22544 "SELECT array_agg(\"__spg_arr_c\") FROM ({sub}) AS \"__spg_arr_t\"(\"__spg_arr_c\")"
22545 );
22546 let stmt = parse_statement(&wrapper)
22547 .map_err(|e| self.err(alloc::format!("ARRAY(subquery): {}", e.message)))?;
22548 let Statement::Select(sel) = stmt else {
22549 return Err(self.err("ARRAY(subquery) did not desugar to a SELECT".into()));
22550 };
22551 Ok(Expr::ScalarSubquery(alloc::boxed::Box::new(sel)))
22552 }
22553
22554 #[inline(never)]
22555 fn parse_paren_scalar_subquery(&mut self, is_with: bool) -> Result<Expr, ParseError> {
22556 let inner = if is_with {
22557 self.advance(); // WITH
22558 self.parse_with_cte_then_select()?
22559 } else {
22560 self.parse_select_stmt()?
22561 };
22562 match self.advance() {
22563 Token::RParen => {
22564 let Statement::Select(s) = inner else {
22565 return Err(ParseError {
22566 message: "scalar subquery body must be a SELECT".into(),
22567 token_pos: self.consumed_pos(),
22568 });
22569 };
22570 Ok(Expr::ScalarSubquery(Box::new(s)))
22571 }
22572 other => Err(ParseError {
22573 message: format!("expected ')' after scalar subquery, got {other:?}"),
22574 token_pos: self.consumed_pos(),
22575 }),
22576 }
22577 }
22578
22579 /// `B'1010'` / `X'1F'` bit-string (PG) or binary-string (MySQL)
22580 /// literals. The lexer splits them into an ident + string; recombine
22581 /// here. Out-of-line and returning `Option` so `parse_atom` — the
22582 /// recursive frame the 768 KiB stack budget is tuned against — pays no
22583 /// frame for the `body` / `bits` strings and their char loops (the
22584 /// round-367 frame cliff, M20).
22585 #[inline(never)]
22586 fn try_parse_bit_string_literal(&mut self) -> Option<Result<Expr, ParseError>> {
22587 let is_hex = match self.peek() {
22588 Token::Ident(p) if p.eq_ignore_ascii_case("x") => true,
22589 Token::Ident(p) if p.eq_ignore_ascii_case("b") => false,
22590 _ => return None,
22591 };
22592 if !matches!(self.tokens.get(self.pos + 1), Some(Token::String(_))) {
22593 return None;
22594 }
22595 // v7.39.3 — where the LITERAL starts, because the errors below
22596 // are about the literal and both engines point at it. `err`
22597 // reports the CURRENT token, which by then is the one after the
22598 // string: `SELECT x'123'` pointed at Eof, so the MySQL wire's
22599 // `near '…'` snippet — which runs from the reported position to
22600 // the end — came out empty where MySQL 9.7.2 says `near
22601 // 'x'123''`.
22602 let lit_pos = self.pos;
22603 self.advance();
22604 let Token::String(body) = self.advance() else {
22605 unreachable!("guarded above");
22606 };
22607 // v7.39 (round 367, M20) — in the MySQL dialect `X'…'` and `b'…'`
22608 // are BINARY STRINGS, not PG bit strings. `X'41'` is the byte 0x41
22609 // (hex pairs, even count required — MariaDB errors on an odd
22610 // count); `b'1010'` packs its bits big-endian, left-padded to a
22611 // byte. Lower both onto the bytea cast.
22612 if self.mysql_dialect {
22613 if is_hex {
22614 if body.len() % 2 == 1 {
22615 return Some(Err(self.err_at(
22616 lit_pos,
22617 alloc::format!("invalid hex string literal X'{body}': odd digit count"),
22618 )));
22619 }
22620 for c in body.chars() {
22621 if !c.is_ascii_hexdigit() {
22622 return Some(Err(self.err_at(
22623 lit_pos,
22624 alloc::format!("invalid hexadecimal digit {c:?} in X'…'"),
22625 )));
22626 }
22627 }
22628 return Some(self.finish_postfix_casts(hex_literal_to_bytea_expr(&body)));
22629 }
22630 if let Some(bad) = body.chars().find(|c| *c != '0' && *c != '1') {
22631 return Some(Err(self.err_at(
22632 lit_pos,
22633 alloc::format!("invalid binary digit {bad:?} in b'…'"),
22634 )));
22635 }
22636 return Some(self.finish_postfix_casts(bits_literal_to_bytea_expr(&body)));
22637 }
22638 let bits = if is_hex {
22639 let mut out = String::with_capacity(body.len() * 4);
22640 for c in body.chars() {
22641 let Some(d) = c.to_digit(16) else {
22642 // v7.39.3 — PostgreSQL 18.6's own sentence, and its
22643 // own quoting: `"g" is not a valid hexadecimal
22644 // digit` (measured, with the caret on the literal).
22645 return Some(Err(self.err_at(
22646 lit_pos,
22647 alloc::format!("\"{c}\" is not a valid hexadecimal digit"),
22648 )));
22649 };
22650 out.push_str(&alloc::format!("{d:04b}"));
22651 }
22652 out
22653 } else {
22654 if let Some(bad) = body.chars().find(|c| *c != '0' && *c != '1') {
22655 return Some(Err(self.err_at(
22656 lit_pos,
22657 alloc::format!("\"{bad}\" is not a valid binary digit"),
22658 )));
22659 }
22660 body
22661 };
22662 // Route through the postfix-cast loop so a chained cast like
22663 // `B'1010'::int` attaches onto the implicit `::bit` cast instead
22664 // of erroring at the `::`.
22665 // v7.39 (read01 varbit.c) — a distinct internal target: a B'...'
22666 // literal keeps its exact length, while an explicit `::bit` cast is
22667 // bit(1) with pad/truncate semantics (PG).
22668 Some(self.finish_postfix_casts(Expr::Cast {
22669 expr: Box::new(Expr::Literal(Literal::String(bits))),
22670 target: CastTarget::Named("__bit_literal".to_string()),
22671 }))
22672 }
22673
22674 fn parse_atom(&mut self) -> Result<Expr, ParseError> {
22675 if let Some(res) = self.try_parse_bit_string_literal() {
22676 return res;
22677 }
22678 let tok_pos = self.pos;
22679 match self.advance() {
22680 Token::Integer(n) => Ok(Expr::Literal(Literal::Integer(n))),
22681 Token::Float(x) => Ok(Expr::Literal(Literal::Float(x))),
22682 // v7.38 (read01) — dotted / over-i64 literal → exact NUMERIC (PG),
22683 // carrying the source mantissa + scale so no precision is lost. A
22684 // literal too wide for i128 falls back to double precision.
22685 // Out-of-line (#[inline(never)]) — this arm sits on the
22686 // parse_expr recursion chain; its expansion locals must not
22687 // widen the recursive frame (debug frame-cliff discipline).
22688 Token::Numeric(s) => match numeric_token_to_literal(s) {
22689 Ok(lit) => Ok(Expr::Literal(lit)),
22690 Err(msg) => Err(self.err(msg)),
22691 },
22692 Token::String(s) => Ok(Expr::Literal(Literal::String(s))),
22693 // v7.39 (round 367, M20) — a MySQL `0x…` binary-string literal
22694 // (the lexer only emits this token in the MySQL dialect). Lower
22695 // onto the existing bytea cast; out-of-line to keep this arm off
22696 // the parse recursion frame.
22697 Token::HexBytes(s) => Ok(hex_literal_to_bytea_expr(&s)),
22698 Token::True => Ok(Expr::Literal(Literal::Bool(true))),
22699 Token::False => Ok(Expr::Literal(Literal::Bool(false))),
22700 Token::Null => Ok(Expr::Literal(Literal::Null)),
22701 // v6.1.1 — `$N` placeholder. The actual Value lookup
22702 // happens in the engine eval path against the prepared-
22703 // statement bind buffer.
22704 Token::Placeholder(n) => Ok(Expr::Placeholder(n)),
22705 Token::LParen => {
22706 // v4.10: `(SELECT ...)` in expression position is a
22707 // scalar subquery; otherwise it's a parenthesised
22708 // expression. Peek for SELECT keyword to dispatch.
22709 // v7.37 D.43 — also accept `(WITH [RECURSIVE] … SELECT …)`; WITH
22710 // lexes as Ident("with") (not a reserved token). The subquery body
22711 // is parsed in `parse_paren_scalar_subquery` (marked #[inline(never)]
22712 // so its large `Statement` local stays out of parse_atom's stack
22713 // frame — parse_atom is on the recursive `((…))` cycle and the
22714 // nesting budget is tuned to its frame size).
22715 let is_with = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
22716 if s.eq_ignore_ascii_case("with"));
22717 if matches!(self.peek(), Token::Select) || is_with {
22718 self.parse_paren_scalar_subquery(is_with)
22719 } else {
22720 let e = self.parse_expr(0)?;
22721 // `(a, b, …)` — a row constructor. Valid only
22722 // in front of a comparison operator or [NOT]
22723 // IN; both expand at parse time (lexicographic
22724 // comparison / OR'd row equalities).
22725 if matches!(self.peek(), Token::Comma) {
22726 let mut row = alloc::vec![e];
22727 while matches!(self.peek(), Token::Comma) {
22728 self.advance();
22729 row.push(self.parse_expr(0)?);
22730 }
22731 if !matches!(self.peek(), Token::RParen) {
22732 return Err(self.err(alloc::format!(
22733 "expected ')' after row constructor, got {:?}",
22734 self.peek()
22735 )));
22736 }
22737 self.advance();
22738 // A bare `(a, b, …)` row constructor can carry postfix
22739 // (`::text`, `.field`) just like `ROW(a, b, …)`; the
22740 // early return here skips parse_atom's tail postfix
22741 // pass, so fold casts in explicitly. For the
22742 // comparison / predicate forms nothing postfix follows,
22743 // so this is a no-op.
22744 return self
22745 .parse_row_comparison_tail(row)
22746 .and_then(|e| self.finish_postfix_casts(e));
22747 }
22748 match self.advance() {
22749 Token::RParen => Ok(e),
22750 other => Err(ParseError {
22751 message: format!("expected ')', got {other:?}"),
22752 token_pos: self.consumed_pos(),
22753 }),
22754 }
22755 }
22756 }
22757 Token::LBracket => self.parse_vector_literal_body(),
22758 Token::Extract => self.parse_extract_atom(),
22759 Token::Interval => self.parse_interval_atom(),
22760 // `LEFT` / `RIGHT` are reserved-keyword tokens because the
22761 // grammar dedicates arms for `LEFT [OUTER] JOIN` /
22762 // `RIGHT [OUTER] JOIN`. When followed by `(` we're in
22763 // expression position calling the PG `left(string, n)` /
22764 // `right(string, n)` function; rebuild the AST as a regular
22765 // function call so the engine's apply_function dispatch picks
22766 // it up. Delegated to a #[inline(never)] helper so its locals
22767 // don't bloat this recursive `parse_atom` frame (the nesting
22768 // budget in `enter_nested` is tuned to parse_atom's size).
22769 Token::Left if matches!(self.peek(), Token::LParen) => {
22770 self.parse_lr_string_function_call("left")
22771 }
22772 Token::Right if matches!(self.peek(), Token::LParen) => {
22773 self.parse_lr_string_function_call("right")
22774 }
22775 // v4.10: EXISTS / NOT EXISTS. EXISTS isn't a reserved
22776 // token; we match on the bare ident. NOT is a token
22777 // (consumed in the comparison rung), but `EXISTS (...)`
22778 // at the top of an expression starts here.
22779 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("exists") => {
22780 self.parse_exists_atom(false)
22781 }
22782 // v7.13.0 — `CASE [<operand>] WHEN <cond> THEN <val>
22783 // [WHEN ...] [ELSE <val>] END` (mailrs round-5 G9).
22784 // CASE is a bare ident; we dispatch on lowercase match.
22785 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("case") => {
22786 self.parse_case_atom()
22787 }
22788 // v7.37.17 (17.6 siblings) — PG typed datetime literals:
22789 // `DATE '2003-01-02'` / `TIMESTAMP '…'` / `TIMESTAMPTZ
22790 // '…'`. Lower onto the ::cast node so the existing
22791 // runtime text→date/timestamp paths do the parsing. The
22792 // string must follow immediately, else the ident stays a
22793 // plain column reference.
22794 Token::Ident(s)
22795 if typed_literal_cast_target(&s.to_ascii_lowercase()).is_some()
22796 && matches!(self.peek(), Token::String(_)) =>
22797 {
22798 let target =
22799 typed_literal_cast_target(&s.to_ascii_lowercase()).expect("guard checked");
22800 let Token::String(lit) = self.advance() else {
22801 unreachable!("peek guaranteed a string token");
22802 };
22803 Ok(Expr::Cast {
22804 expr: Box::new(Expr::Literal(Literal::String(lit))),
22805 target,
22806 })
22807 }
22808 // v7.39 (round 221) — the SQL-standard long spellings:
22809 // `TIME [WITHOUT|WITH] TIME ZONE '…'` / `TIMESTAMP [WITHOUT|WITH]
22810 // TIME ZONE '…'`. Consume the modifier and lower to the same
22811 // typed-literal cast (`timetz` / `timestamptz` for WITH).
22812 Token::Ident(s)
22813 if (s.eq_ignore_ascii_case("time") || s.eq_ignore_ascii_case("timestamp"))
22814 && matches!(self.peek(), Token::Ident(w) if w.eq_ignore_ascii_case("with")
22815 || w.eq_ignore_ascii_case("without"))
22816 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(t)) if t.eq_ignore_ascii_case("time"))
22817 && matches!(self.tokens.get(self.pos + 2), Some(Token::Ident(z)) if z.eq_ignore_ascii_case("zone"))
22818 && matches!(self.tokens.get(self.pos + 3), Some(Token::String(_))) =>
22819 {
22820 let with_tz = matches!(self.peek(), Token::Ident(w) if w.eq_ignore_ascii_case("with"));
22821 self.advance(); // WITH / WITHOUT
22822 self.advance(); // TIME
22823 self.advance(); // ZONE
22824 let Token::String(lit) = self.advance() else {
22825 unreachable!("guard checked a string token");
22826 };
22827 let base = s.to_ascii_lowercase();
22828 let target = match (base.as_str(), with_tz) {
22829 ("time", true) => CastTarget::Named(alloc::string::String::from("timetz")),
22830 ("time", false) => CastTarget::Named(alloc::string::String::from("time")),
22831 (_, true) => CastTarget::Timestamptz,
22832 (_, false) => CastTarget::Timestamp,
22833 };
22834 Ok(Expr::Cast {
22835 expr: Box::new(Expr::Literal(Literal::String(lit))),
22836 target,
22837 })
22838 }
22839 // v7.39 (read01 round 105) — `ARRAY(<subquery>)` constructor:
22840 // gathers the subquery's single-column rows (in its row order)
22841 // into an array. Desugared to `array_agg` over the subquery as a
22842 // derived table; out-of-line to keep parse_atom's frame small (it
22843 // sits on the recursive nesting-budget cycle).
22844 Token::Ident(s) | Token::QuotedIdent(s)
22845 if s.eq_ignore_ascii_case("array") && self.array_subquery_ahead() =>
22846 {
22847 self.parse_array_subquery()
22848 }
22849 // v7.10.10 — `ARRAY[expr, expr, …]` constructor. ARRAY
22850 // is not a reserved token; we match by case-insensitive
22851 // ident. The opening `[` must follow immediately. v7.39 (read01
22852 // round 105) — the body moved out-of-line so its `Vec`/loop locals
22853 // leave parse_atom's frame (which sits on the nesting-budget cycle).
22854 Token::Ident(s) | Token::QuotedIdent(s)
22855 if s.eq_ignore_ascii_case("array") && matches!(self.peek(), Token::LBracket) =>
22856 {
22857 self.parse_array_literal_body()
22858 }
22859 // v7.17.0 Phase 2.2 — MySQL `MATCH(col, ...) AGAINST
22860 // ('term' [IN BOOLEAN MODE | IN NATURAL LANGUAGE MODE])`.
22861 // We special-case before the generic ident dispatch so
22862 // the AGAINST clause never reaches the function-call
22863 // loop (which would mis-read `(cols) AGAINST` as a
22864 // call with no trailing modifier). The shape is
22865 // rewritten to a Boolean OR over per-column
22866 // `to_tsvector('simple', col) @@ plainto_tsquery('simple',
22867 // term)` so the existing FTS evaluator handles
22868 // semantics — the fulltext-GIN built at CREATE TABLE
22869 // time is currently a "real index that survives dump
22870 // round-trip"; the planner hook that actually uses
22871 // it for posting-list intersection lands in a later
22872 // sub-phase (Phase 2.2b) without touching this surface.
22873 Token::Ident(s) | Token::QuotedIdent(s)
22874 if s.eq_ignore_ascii_case("match") && matches!(self.peek(), Token::LParen) =>
22875 {
22876 self.parse_match_against_atom()
22877 }
22878 Token::Ident(s) | Token::QuotedIdent(s) => self.finish_ident_atom(s),
22879 // v7.37.43-T4 — PG-unreserved keywords are legal column /
22880 // alias names in expression context too. `release` appears
22881 // in sentori `0003_partition_events.sql` as both a column
22882 // reference (SELECT … release …) and an INSERT column list
22883 // entry. Mirrors `expect_ident_like`'s expansion of the
22884 // identifier set.
22885 other if unreserved_keyword_text(&other).is_some() => {
22886 let s = unreserved_keyword_text(&other).unwrap();
22887 self.finish_ident_atom(s)
22888 }
22889 // v7.39 (round 331, V50) — `@@var` in an EXPRESSION. It parsed
22890 // only inside `SET` before, so `SELECT @@autocommit` — which
22891 // every MySQL connector asks at handshake — was a parse error.
22892 // MariaDB accepts the bare, `@@session.` and `@@global.`
22893 // spellings alike and answers from the session's own value.
22894 Token::SessionVar(v) => {
22895 // v7.39 (round 430) — ONE `@` is a MySQL USER variable, which
22896 // has nothing to do with a `@@` engine setting: its own
22897 // per-session namespace, and an unset one reads NULL instead
22898 // of raising. Stripping every `@` (as this did) made `@x` and
22899 // `@@x` the same node, so `SELECT @x` answered "Unknown
22900 // system variable".
22901 Ok(variable_ref_atom(&v))
22902 }
22903 other => Err(ParseError {
22904 message: format!("unexpected token {other:?} in expression"),
22905 token_pos: tok_pos,
22906 }),
22907 }
22908 // After parsing the atom, fold any postfix `::vector` casts.
22909 .and_then(|atom| self.finish_postfix_casts(atom))
22910 }
22911
22912 /// Postfix operators on an atom: `::TYPE` cast and `IS [NOT] NULL`.
22913 /// Both bind tighter than any binary op.
22914 /// Shared cast-target parser for postfix `::TYPE` and the
22915 /// standard `CAST(expr AS TYPE)` form (v7.25, round-17).
22916 /// If the next tokens are `( N )`, consume them and return the canonical
22917 /// `base(N)` name so a temporal cast (`::timestamp(2)`) carries its
22918 /// fractional-seconds precision into `CastTarget::Named`; otherwise `None`.
22919 fn consume_temporal_typmod(&mut self, base: &str) -> Option<alloc::string::String> {
22920 if !matches!(self.peek(), Token::LParen) {
22921 return None;
22922 }
22923 self.advance(); // (
22924 let n = match self.advance() {
22925 Token::Integer(n) => n,
22926 _ => return Some(base.to_string()), // malformed → drop precision
22927 };
22928 if matches!(self.peek(), Token::RParen) {
22929 self.advance();
22930 }
22931 Some(alloc::format!("{base}({n})"))
22932 }
22933
22934 fn parse_cast_target(&mut self) -> Result<CastTarget, ParseError> {
22935 // r1052 — `::pg_catalog.regproc` and friends: pg_dump
22936 // schema-qualifies every cast target, and `pg_catalog.X` names
22937 // exactly the builtin type X. Consume the qualifier and let
22938 // the ordinary target parse decide.
22939 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("pg_catalog"))
22940 && matches!(self.tokens.get(self.pos + 1), Some(Token::Dot))
22941 {
22942 self.advance();
22943 self.advance();
22944 }
22945 let target = match self.advance() {
22946 Token::Ident(s) => match s.to_ascii_lowercase().as_str() {
22947 "int" | "integer" | "int4" => {
22948 if matches!(self.peek(), Token::LBracket)
22949 && matches!(self.tokens.get(self.pos + 1), Some(Token::RBracket))
22950 {
22951 self.advance();
22952 self.advance();
22953 CastTarget::IntArray
22954 } else {
22955 CastTarget::Int
22956 }
22957 }
22958 "bigint" | "int8" => {
22959 if matches!(self.peek(), Token::LBracket)
22960 && matches!(self.tokens.get(self.pos + 1), Some(Token::RBracket))
22961 {
22962 self.advance();
22963 self.advance();
22964 CastTarget::BigIntArray
22965 } else {
22966 CastTarget::BigInt
22967 }
22968 }
22969 "float" | "double" => CastTarget::Float,
22970 "text" => {
22971 // v7.10.11 — `::TEXT[]` widens to TextArray.
22972 if matches!(self.peek(), Token::LBracket)
22973 && matches!(self.tokens.get(self.pos + 1), Some(Token::RBracket))
22974 {
22975 self.advance();
22976 self.advance();
22977 CastTarget::TextArray
22978 } else {
22979 CastTarget::Text
22980 }
22981 }
22982 "bool" | "boolean" => CastTarget::Bool,
22983 "vector" => CastTarget::Vector,
22984 "date" => CastTarget::Date,
22985 // v7.38 (read01) — `::timestamp(N)` carries its fractional-
22986 // seconds precision through the Named path (the engine rounds
22987 // the sub-second field); bare `::timestamp` keeps the fast arm.
22988 "timestamp" | "datetime" => match self.consume_temporal_typmod("timestamp") {
22989 Some(named) => CastTarget::Named(named),
22990 None => CastTarget::Timestamp,
22991 },
22992 "timestamptz" => match self.consume_temporal_typmod("timestamptz") {
22993 Some(named) => CastTarget::Named(named),
22994 None => CastTarget::Timestamptz,
22995 },
22996 "interval" => CastTarget::Interval,
22997 "json" => CastTarget::Json,
22998 "jsonb" => CastTarget::Jsonb,
22999 // v7.39 (round 694) — these have dedicated CastTarget
23000 // variants, so they never reached the postfix `[]` handling
23001 // further down and `::regtype[]` was a SYNTAX error at the
23002 // `]`. PG has an array type for every scalar; take the
23003 // suffix here and hand the canonical `<ty>_array` name to
23004 // the engine, the same shape every other array cast uses.
23005 "regtype" if self.peek_postfix_array_brackets() => {
23006 self.advance();
23007 self.advance();
23008 CastTarget::Named(alloc::string::String::from("regtype_array"))
23009 }
23010 "regclass" if self.peek_postfix_array_brackets() => {
23011 self.advance();
23012 self.advance();
23013 CastTarget::Named(alloc::string::String::from("regclass_array"))
23014 }
23015 "regtype" => CastTarget::RegType,
23016 "regclass" => CastTarget::RegClass,
23017 // v7.12.0 — `::tsvector` / `::tsquery`.
23018 // Engine decodes the LHS text via the PG
23019 // external form parser.
23020 // v7.39 (round 352, M8) — MySQL's own cast targets.
23021 // `CAST(x AS SIGNED)` / `UNSIGNED`, with the optional
23022 // `INTEGER` / `INT` tail MariaDB also accepts. PG has no
23023 // such type, so they are taken only in that dialect and
23024 // fall through to the "type does not exist" arm otherwise.
23025 "signed" | "unsigned" if self.mysql_dialect => {
23026 if matches!(self.peek(), Token::Ident(k)
23027 if k.eq_ignore_ascii_case("integer") || k.eq_ignore_ascii_case("int"))
23028 {
23029 self.advance();
23030 }
23031 CastTarget::Named(s.to_ascii_lowercase())
23032 }
23033 // v7.39 (round 352, M8) — `CAST(x AS CHAR)` is UNBOUNDED
23034 // in MySQL: MariaDB answers '123' where the SQL-standard
23035 // reading (PG's, and SPG's) is `char(1)` and answers '1'.
23036 // Truncating a number to its first digit is a wrong answer
23037 // with no error, so the MySQL session gets MySQL's reading.
23038 "char" if self.mysql_dialect && !matches!(self.peek(), Token::LParen) => {
23039 CastTarget::Text
23040 }
23041 "tsvector" => CastTarget::TsVector,
23042 "tsquery" => CastTarget::TsQuery,
23043 // v7.17.0 — `::uuid`. Engine decodes the LHS
23044 // text via `spg_storage::parse_uuid_str`.
23045 "uuid" => CastTarget::Uuid,
23046 // v7.18 — `::bytea`. Engine decodes the LHS
23047 // text via the PG hex form (`'\xdeadbeef'`)
23048 // or escape form (`'\\x05\\x00'`). Closes
23049 // mailrs D-pre #3 reverse-acceptance gap.
23050 "bytea" => CastTarget::Bytea,
23051 // v7.37.5 ship triage — generic typed-cast escape.
23052 // Anything the long-tail PG type ident table knows
23053 // about(network/bit/geometry/multirange/etc.)flows
23054 // through `CastTarget::Named(canonical)`; the engine
23055 // resolves via `column_type_to_data_type` and dispatches
23056 // through the typed `coerce_value` path. Truly
23057 // unrecognised idents still hit the error arm below
23058 // because the engine rejects them.
23059 other => {
23060 // Optional `(N[, M])` precision args — `::numeric(10,2)`,
23061 // `::varchar(255)`, etc. Capture into the canonical
23062 // `name(p,s)` form so `type_name_to_data_type` can
23063 // reconstruct the `DataType::Numeric { precision,
23064 // scale }` (and similar param-carrying types).
23065 let mut name = other.to_string();
23066 // v7.39 (round 281) — `::bit varying(3)` is two
23067 // words; fold the tail in so the typmod reaches the
23068 // type resolver instead of tripping the parser.
23069 if name.eq_ignore_ascii_case("bit")
23070 && matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("varying"))
23071 {
23072 self.advance();
23073 name = alloc::string::String::from("varbit");
23074 }
23075 // v7.39 (round 613) — `::character varying` is the same
23076 // two-word shape and had no fold, so the `varying` was
23077 // left behind and the cast became a bare `character`,
23078 // which is `char(1)`: `'ab'::CHARACTER VARYING` answered
23079 // `a` where PG answers `ab`. Silently, and for a spelling
23080 // pg_dump writes.
23081 if name.eq_ignore_ascii_case("character")
23082 && matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("varying"))
23083 {
23084 self.advance();
23085 name = alloc::string::String::from("varchar");
23086 }
23087 if matches!(self.peek(), Token::LParen) {
23088 let mut buf = alloc::string::String::from("(");
23089 let mut depth = 0usize;
23090 loop {
23091 match self.advance() {
23092 Token::LParen => {
23093 depth += 1;
23094 if depth > 1 {
23095 buf.push('(');
23096 }
23097 }
23098 Token::RParen => {
23099 depth -= 1;
23100 if depth == 0 {
23101 buf.push(')');
23102 break;
23103 }
23104 buf.push(')');
23105 }
23106 Token::Comma => buf.push(','),
23107 Token::Integer(n) => buf.push_str(&alloc::format!("{n}")),
23108 // v7.39 (round 273) — a minus used to fall
23109 // into the catch-all below and vanish, so
23110 // `::numeric(10,-2)` reached the engine as
23111 // the text `numeric(10,2)` and silently
23112 // rounded to two DECIMALS instead of to
23113 // hundreds. A dropped token is not a
23114 // no-op when it carries a sign.
23115 Token::Minus => buf.push('-'),
23116 Token::Eof => break,
23117 _ => {}
23118 }
23119 }
23120 name.push_str(&buf);
23121 }
23122 // Optional postfix `[]` widens to the array form —
23123 // `::BOOL[]`, `::NUMERIC[]`, `::SMALLINT[]`, etc.
23124 // The engine's `type_name_to_data_type` recognises
23125 // the canonical `<ty>_array` form.
23126 if matches!(self.peek(), Token::LBracket)
23127 && matches!(self.tokens.get(self.pos + 1), Some(Token::RBracket))
23128 {
23129 self.advance();
23130 self.advance();
23131 name.push_str("_array");
23132 }
23133 CastTarget::Named(name)
23134 }
23135 },
23136 Token::Interval => CastTarget::Interval,
23137 // v7.39 — a quoted type name: `::"char"` is PG's 1-byte
23138 // "char" (oid 18, SPG Char1 — distinct from bare `char`
23139 // = char(1)); other quoted names resolve like idents.
23140 Token::QuotedIdent(q) => {
23141 if q.eq_ignore_ascii_case("char") {
23142 CastTarget::Named("char1".into())
23143 } else {
23144 CastTarget::Named(q.to_ascii_lowercase())
23145 }
23146 }
23147 other => {
23148 return Err(ParseError {
23149 message: format!("expected type ident after `::`, got {other:?}"),
23150 token_pos: self.consumed_pos(),
23151 });
23152 }
23153 };
23154 // v7.37.5 ship triage — postfix `[]` widens a scalar cast
23155 // target to its array sibling. Closed-enum arms (Bool /
23156 // SmallInt / Numeric / Float / Date / …) didn't carry the
23157 // explicit widening that Text / Int / BigInt did, so
23158 // `::BOOL[]` / `::NUMERIC[]` etc. surfaced as a parse
23159 // error. The widening here mirrors the per-arm Text /
23160 // Int / BigInt logic above + folds the new ζ-A first-class
23161 // types through `CastTarget::Named("<ty>_array")`.
23162 if matches!(self.peek(), Token::LBracket)
23163 && matches!(self.tokens.get(self.pos + 1), Some(Token::RBracket))
23164 {
23165 let widened = match &target {
23166 CastTarget::Bool => Some(CastTarget::Named("bool_array".to_string())),
23167 CastTarget::Date => Some(CastTarget::Named("date_array".to_string())),
23168 // v7.39 (round 326, V43) — the two temporal types stay
23169 // distinct. Both used to widen to `timestamptz_array`, so
23170 // `::timestamp[]` named the wrong target in its own error
23171 // message and lost the zone-less identity on the way.
23172 CastTarget::Timestamp => Some(CastTarget::Named("timestamp_array".to_string())),
23173 CastTarget::Timestamptz => Some(CastTarget::Named("timestamptz_array".to_string())),
23174 CastTarget::Uuid => Some(CastTarget::Named("uuid_array".to_string())),
23175 CastTarget::Json | CastTarget::Jsonb => {
23176 Some(CastTarget::Named("jsonb_array".to_string()))
23177 }
23178 CastTarget::Bytea => Some(CastTarget::Named("bytea_array".to_string())),
23179 CastTarget::Interval => Some(CastTarget::Named("interval_array".to_string())),
23180 CastTarget::Float => Some(CastTarget::Named("float_array".to_string())),
23181 CastTarget::Named(name) => {
23182 let mut a = name.clone();
23183 a.push_str("_array");
23184 Some(CastTarget::Named(a))
23185 }
23186 // Int / BigInt / Text / Vector / TsVector / TsQuery /
23187 // RegType / RegClass / TextArray / IntArray /
23188 // BigIntArray already finalised — leave as is.
23189 _ => None,
23190 };
23191 if let Some(w) = widened {
23192 self.advance();
23193 self.advance();
23194 return Ok(w);
23195 }
23196 }
23197 Ok(target)
23198 }
23199
23200 fn finish_postfix_casts(&mut self, mut expr: Expr) -> Result<Expr, ParseError> {
23201 loop {
23202 // v7.38 (read01, T9) — composite field access `(expr).field`.
23203 // A bare `a.b` is consumed as a qualified column inside the ident
23204 // atom, so a Dot only survives to this postfix position when the
23205 // base was a parenthesised expression (`(e).id`, `(row(1,2)).f1`).
23206 // `.*` whole-row expansion is not handled here (projection-level).
23207 if matches!(self.peek(), Token::Dot)
23208 && matches!(
23209 self.tokens.get(self.pos + 1),
23210 Some(Token::Ident(_) | Token::QuotedIdent(_))
23211 )
23212 {
23213 self.advance(); // .
23214 let field = match self.advance() {
23215 Token::Ident(s) | Token::QuotedIdent(s) => s,
23216 other => {
23217 return Err(
23218 self.err(format!("expected a field name after '.', got {other:?}"))
23219 );
23220 }
23221 };
23222 expr = Expr::FieldAccess {
23223 base: Box::new(expr),
23224 field,
23225 };
23226 continue;
23227 }
23228 if matches!(self.peek(), Token::DoubleColon) {
23229 self.advance();
23230 // v7.9.25 / v7.9.26 — broaden the postfix `::` cast
23231 // target set to include INTERVAL (reserved Token),
23232 // TIMESTAMPTZ, and PG catalog regtype / regclass.
23233 // mailrs follow-up H3a + H3b.
23234 let target = self.parse_cast_target()?;
23235 expr = Expr::Cast {
23236 expr: Box::new(expr),
23237 target,
23238 };
23239 continue;
23240 }
23241 // v7.10.12 — `arr[i]` subscript. PG 1-based; engine
23242 // returns NULL for out-of-range. Multiple subscripts
23243 // chain: `a[i][j]` parses left-to-right.
23244 if matches!(self.peek(), Token::LBracket) {
23245 self.advance();
23246 // `[lo:hi]` / `[:hi]` / `[lo:]` — array slice. A
23247 // bare index stays a subscript.
23248 let lo = if matches!(self.peek(), Token::Colon) {
23249 None
23250 } else {
23251 Some(self.parse_expr(0)?)
23252 };
23253 if matches!(self.peek(), Token::Colon) {
23254 self.advance();
23255 let hi = if matches!(self.peek(), Token::RBracket) {
23256 None
23257 } else {
23258 Some(Box::new(self.parse_expr(0)?))
23259 };
23260 if !matches!(self.peek(), Token::RBracket) {
23261 return Err(self.err(alloc::format!(
23262 "expected ']' after array slice, got {:?}",
23263 self.peek()
23264 )));
23265 }
23266 self.advance();
23267 expr = Expr::ArraySlice {
23268 target: Box::new(expr),
23269 lo: lo.map(Box::new),
23270 hi,
23271 };
23272 continue;
23273 }
23274 let index = lo.expect("non-colon branch parsed an index");
23275 if !matches!(self.peek(), Token::RBracket) {
23276 return Err(self.err(alloc::format!(
23277 "expected ']' after array index, got {:?}",
23278 self.peek()
23279 )));
23280 }
23281 self.advance();
23282 expr = Expr::ArraySubscript {
23283 target: Box::new(expr),
23284 index: Box::new(index),
23285 };
23286 continue;
23287 }
23288 // `expr AT TIME ZONE zone` — lowers to PG's own function
23289 // form timezone(zone, expr); the scalar implements the
23290 // offset shift (named zones error there — no tzdata).
23291 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("at"))
23292 && matches!(self.tokens.get(self.pos + 1),
23293 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("time"))
23294 && matches!(self.tokens.get(self.pos + 2),
23295 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("zone"))
23296 {
23297 self.advance(); // AT
23298 self.advance(); // TIME
23299 self.advance(); // ZONE
23300 // Zone at comparison precedence so AND/OR stay out.
23301 let zone = self.parse_expr(6)?;
23302 expr = Expr::FunctionCall {
23303 name: "timezone".to_string(),
23304 args: alloc::vec![zone, expr],
23305 };
23306 continue;
23307 }
23308 // `expr COLLATE "name"` — SPG's single text ordering IS
23309 // byte order, i.e. the C collation. The byte-order
23310 // spellings absorb as no-ops; a locale collation would
23311 // silently sort differently from PG, so it errors
23312 // honestly instead.
23313 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("collate")) {
23314 self.advance();
23315 let mut cname = match self.advance() {
23316 Token::Ident(s) | Token::QuotedIdent(s) | Token::String(s) => s,
23317 other => {
23318 return Err(self.err(alloc::format!(
23319 "expected collation name after COLLATE, got {other:?}"
23320 )));
23321 }
23322 };
23323 // v7.39 (round 539) — a SCHEMA-QUALIFIED collation, which
23324 // is how `pg_dump` writes the default one:
23325 // `… COLLATE pg_catalog.default`. Reading a single token
23326 // left the SCHEMA as the name, so the clause was refused
23327 // as an unsupported locale collation and no dump ran.
23328 if matches!(self.peek(), Token::Dot) {
23329 // v7.39.2 — the qualifier is DROPPED (SPG is single
23330 // schema) but it is checked first. PostgreSQL 18.6
23331 // answers `schema "nosuch_schema" does not exist` for
23332 // one it has never heard of, and dropping it unread
23333 // meant `COLLATE nosuch_schema."C"` succeeded here —
23334 // a name that names nothing, accepted.
23335 let schema = cname.to_ascii_lowercase();
23336 if !matches!(
23337 schema.as_str(),
23338 "pg_catalog" | "public" | "information_schema"
23339 ) {
23340 return Err(self.err(alloc::format!("schema \"{cname}\" does not exist")));
23341 }
23342 self.advance();
23343 cname = match self.advance() {
23344 Token::Ident(s) | Token::QuotedIdent(s) | Token::String(s) => s,
23345 // `default` lexes as a KEYWORD, and it is the name
23346 // pg_dump writes — the same trap round 535 hit with
23347 // TABLE / INDEX / FULL.
23348 Token::Default => alloc::string::String::from("default"),
23349 other => {
23350 return Err(self.err(alloc::format!(
23351 "expected collation name after COLLATE, got {other:?}"
23352 )));
23353 }
23354 };
23355 }
23356 let lc = cname.to_ascii_lowercase();
23357 // v7.39 (round 371, M4 P4b) — a per-expression MySQL
23358 // collation override. `… COLLATE utf8mb4_bin` (any `_bin`
23359 // family / `binary`) forces byte-wise, which is exactly
23360 // what `BINARY expr` does — lower onto that so every fold
23361 // site (comparison, LIKE, ORDER BY) suppresses via
23362 // `is_binary_coerced`. A `_ci` family override folds, and
23363 // under the MySQL dialect the default already folds, so it
23364 // absorbs as a no-op; likewise the C / byte-order spellings.
23365 // v7.39.2 — against MySQL's own list, not against the
23366 // shape of the name. `nosuch_bin` took this shortcut and
23367 // became a BINARY cast; `nosuch_ci` took the one below
23368 // and was absorbed as a no-op. Either way the client
23369 // named a collation that does not exist and was told
23370 // nothing. An unknown name now falls through to the
23371 // node, and the engine refuses it.
23372 let real = crate::charset::is_mysql_collation(&lc);
23373 if self.mysql_dialect && real && (lc.ends_with("_bin") || lc == "binary") {
23374 expr = Expr::Cast {
23375 expr: alloc::boxed::Box::new(expr),
23376 target: CastTarget::Named("binary".to_string()),
23377 };
23378 continue;
23379 }
23380 let mysql_ci = self.mysql_dialect
23381 && ((real && lc.ends_with("_ci"))
23382 || matches!(lc.as_str(), "case_insensitive" | "nocase"));
23383 // v7.39 (round 691/692) — inside an ORDER BY key EVERY name
23384 // goes to the lowering channel, the byte-order spellings
23385 // included. Round 691 recorded only the names the old
23386 // allow-list rejected, which left `ORDER BY a COLLATE "C"`
23387 // absorbed as a no-op — and once a column could declare a
23388 // collation, absorbing the clause meant the COLUMN's
23389 // collation won where the query had asked for bytes.
23390 if self.in_order_by_key && !mysql_ci {
23391 self.order_key_collation = Some(cname);
23392 continue;
23393 }
23394 // v7.39.2 — the clause becomes a NODE rather than being
23395 // refused or absorbed.
23396 //
23397 // What stood here refused the locale names and SILENTLY
23398 // DROPPED the byte-order ones, so `'a' COLLATE "C" < 'B'`
23399 // answered `t` where PostgreSQL 18.6 answers `f`: the one
23400 // family it let through is the one where dropping it
23401 // changes the answer. Absorbing is only correct when the
23402 // collation asked for is the one the comparison would use
23403 // anyway, and that depends on the DATABASE — which the
23404 // parser cannot see. So it rides along and the engine,
23405 // which can, decides.
23406 //
23407 // `collate_derive` already modelled `Explicit(name)` and
23408 // had no way to be handed one.
23409 // v7.39.2 — a MySQL spelling does not exist on the
23410 // PostgreSQL wire, and THIS is where the wire is known.
23411 //
23412 // The check lived in the evaluator first and asked
23413 // `ctx.mysql_dialect`, which the INSERT path builds as a
23414 // hard-coded `false` — so `INSERT … VALUES (_utf8mb4'x')`
23415 // in a MySQL session was refused for a collation that
23416 // does not exist on a wire it was not on. Making that
23417 // context truthful would change INSERT-time evaluation
23418 // in other ways as a side effect; the parser already
23419 // gates the introducer on the same flag and is the
23420 // honest place to ask.
23421 if !self.mysql_dialect
23422 && (lc.ends_with("_ci")
23423 || lc.ends_with("_cs")
23424 || lc.ends_with("_bin")
23425 || lc == "binary"
23426 || matches!(lc.as_str(), "case_insensitive" | "nocase"))
23427 {
23428 return Err(self.err(alloc::format!(
23429 "collation \"{cname}\" for encoding \"UTF8\" does not exist"
23430 )));
23431 }
23432 // v7.39.3 — the node is built for EVERY name, `_ci`
23433 // included.
23434 //
23435 // A MySQL `_ci` spelling used to be absorbed here on the
23436 // reasoning that a MySQL session folds anyway, so the
23437 // clause asked for what it would have got. That stopped
23438 // being true when the fold learned to read the session's
23439 // collation NAME: under `SET NAMES utf8mb4 COLLATE
23440 // utf8mb4_bin`, `'AB' COLLATE utf8mb4_general_ci = 'ab'`
23441 // is 1 on MySQL 9.7.2 and was 0 here, because the clause
23442 // that would have made it 1 had been dropped in the
23443 // parser. Absorbing is only ever correct when the
23444 // collation asked for is the one the comparison would use
23445 // anyway, and the parser cannot know that — the same
23446 // reasoning already written above for the byte-order
23447 // spellings, applied to the family it had exempted.
23448 expr = Expr::Collate {
23449 expr: alloc::boxed::Box::new(expr),
23450 collation: cname,
23451 };
23452 continue;
23453 }
23454 return Ok(expr);
23455 }
23456 }
23457
23458 /// v7.39 (round 696) — a comma-separated list of bare names, stopping at
23459 /// the first token that is not one. Schema qualifiers collapse to the
23460 /// last part, which is what every other name path here does (SPG is
23461 /// single-schema).
23462 fn take_comma_separated_names(&mut self) -> Vec<String> {
23463 let mut out = Vec::new();
23464 while let Token::Ident(n) | Token::QuotedIdent(n) = self.peek().clone() {
23465 self.advance();
23466 let mut last = n;
23467 while matches!(self.peek(), Token::Dot) {
23468 self.advance();
23469 if let Token::Ident(t) | Token::QuotedIdent(t) = self.advance() {
23470 last = t;
23471 }
23472 }
23473 out.push(last);
23474 if matches!(self.peek(), Token::Comma) {
23475 self.advance();
23476 } else {
23477 break;
23478 }
23479 }
23480 out
23481 }
23482
23483 /// v7.39 (round 694) — is the next token pair a postfix `[]`?
23484 ///
23485 /// The general cast-target path tests this inline; the types with their
23486 /// own `CastTarget` variant need it as a guard on their match arm,
23487 /// which is what this exists for.
23488 fn peek_postfix_array_brackets(&self) -> bool {
23489 matches!(self.peek(), Token::LBracket)
23490 && matches!(self.tokens.get(self.pos + 1), Some(Token::RBracket))
23491 }
23492
23493 /// Parse the operator tail after a `(a, b, …)` row constructor
23494 /// and expand at parse time. `=` is the conjunction of element
23495 /// equalities; `<>` its negation; the order operators expand
23496 /// lexicographically; `[NOT] IN ( (row), … )` ORs the row
23497 /// equalities. Anything else (a bare row value, a subquery
23498 /// RHS) errors honestly — SPG has no composite runtime value.
23499 fn parse_row_comparison_tail(&mut self, row: Vec<Expr>) -> Result<Expr, ParseError> {
23500 fn row_eq(lhs: &[Expr], rhs: &[Expr]) -> Expr {
23501 let mut it = lhs.iter().zip(rhs.iter()).map(|(l, r)| Expr::Binary {
23502 lhs: Box::new(l.clone()),
23503 op: BinOp::Eq,
23504 rhs: Box::new(r.clone()),
23505 });
23506 let first = it.next().expect("row has at least two elements");
23507 it.fold(first, |acc, e| Expr::Binary {
23508 lhs: Box::new(acc),
23509 op: BinOp::And,
23510 rhs: Box::new(e),
23511 })
23512 }
23513 // Lexicographic (a,b) OP (c,d):
23514 // a STRICT c OR (a = c AND (b OP d)) — recursing right.
23515 fn row_lex(lhs: &[Expr], rhs: &[Expr], strict: BinOp, last: BinOp) -> Expr {
23516 if lhs.len() == 1 {
23517 return Expr::Binary {
23518 lhs: Box::new(lhs[0].clone()),
23519 op: last,
23520 rhs: Box::new(rhs[0].clone()),
23521 };
23522 }
23523 let head_strict = Expr::Binary {
23524 lhs: Box::new(lhs[0].clone()),
23525 op: strict,
23526 rhs: Box::new(rhs[0].clone()),
23527 };
23528 let head_eq = Expr::Binary {
23529 lhs: Box::new(lhs[0].clone()),
23530 op: BinOp::Eq,
23531 rhs: Box::new(rhs[0].clone()),
23532 };
23533 Expr::Binary {
23534 lhs: Box::new(head_strict),
23535 op: BinOp::Or,
23536 rhs: Box::new(Expr::Binary {
23537 lhs: Box::new(head_eq),
23538 op: BinOp::And,
23539 rhs: Box::new(row_lex(&lhs[1..], &rhs[1..], strict, last)),
23540 }),
23541 }
23542 }
23543 let negated_in = if matches!(self.peek(), Token::Not)
23544 && matches!(self.tokens.get(self.pos + 1), Some(Token::In))
23545 {
23546 self.advance();
23547 true
23548 } else {
23549 false
23550 };
23551 if matches!(self.peek(), Token::In) {
23552 self.advance();
23553 if !matches!(self.peek(), Token::LParen) {
23554 return Err(self.err(alloc::format!(
23555 "expected '(' after row IN, got {:?}",
23556 self.peek()
23557 )));
23558 }
23559 self.advance();
23560 // `(a, b) [NOT] IN (SELECT x, y)` — a multi-column subquery,
23561 // not a list of literal rows. Row-vs-list decomposes to
23562 // OR-of-AND above, but the subquery's rows are only known at
23563 // runtime, so keep it as a RowInSubquery node.
23564 if matches!(self.peek(), Token::Select) {
23565 let inner = self.parse_select_stmt()?;
23566 if !matches!(self.peek(), Token::RParen) {
23567 return Err(self.err(alloc::format!(
23568 "expected ')' after row IN-subquery, got {:?}",
23569 self.peek()
23570 )));
23571 }
23572 self.advance();
23573 let Statement::Select(s) = inner else {
23574 unreachable!("parse_select_stmt always returns Statement::Select")
23575 };
23576 return Ok(Expr::RowInSubquery {
23577 row,
23578 subquery: Box::new(s),
23579 negated: negated_in,
23580 });
23581 }
23582 let mut alternatives: Vec<Expr> = Vec::new();
23583 loop {
23584 // Optional ROW keyword before the paren row.
23585 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("row"))
23586 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
23587 {
23588 self.advance();
23589 }
23590 if !matches!(self.peek(), Token::LParen) {
23591 return Err(self.err(alloc::format!(
23592 "expected '(' to open a row inside IN, got {:?}",
23593 self.peek()
23594 )));
23595 }
23596 self.advance();
23597 let mut rhs = alloc::vec![self.parse_expr(0)?];
23598 while matches!(self.peek(), Token::Comma) {
23599 self.advance();
23600 rhs.push(self.parse_expr(0)?);
23601 }
23602 if !matches!(self.peek(), Token::RParen) {
23603 return Err(self.err(alloc::format!(
23604 "expected ')' after row inside IN, got {:?}",
23605 self.peek()
23606 )));
23607 }
23608 self.advance();
23609 if rhs.len() != row.len() {
23610 return Err(self.err(alloc::format!(
23611 "row IN arity mismatch: left has {}, right has {}",
23612 row.len(),
23613 rhs.len()
23614 )));
23615 }
23616 alternatives.push(row_eq(&row, &rhs));
23617 if matches!(self.peek(), Token::Comma) {
23618 self.advance();
23619 continue;
23620 }
23621 break;
23622 }
23623 if !matches!(self.peek(), Token::RParen) {
23624 return Err(self.err(alloc::format!(
23625 "expected ')' to close row IN list, got {:?}",
23626 self.peek()
23627 )));
23628 }
23629 self.advance();
23630 let mut it = alternatives.into_iter();
23631 let first = it.next().expect("IN list has at least one row");
23632 let combined = it.fold(first, |acc, e| Expr::Binary {
23633 lhs: Box::new(acc),
23634 op: BinOp::Or,
23635 rhs: Box::new(e),
23636 });
23637 return Ok(maybe_not(combined, negated_in));
23638 }
23639 // SQL-standard `(S1, E1) OVERLAPS (S2, E2)` — true when the
23640 // two periods share at least one time point. Each pair is
23641 // normalised with least/greatest (PG accepts the endpoints
23642 // in either order), then lowered to the standard
23643 // `start1 < end2 AND start2 < end1` form.
23644 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("overlaps")) {
23645 if row.len() != 2 {
23646 return Err(self.err(alloc::format!(
23647 "OVERLAPS needs (start, end) pairs; left side has {} elements",
23648 row.len()
23649 )));
23650 }
23651 self.advance();
23652 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("row"))
23653 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
23654 {
23655 self.advance();
23656 }
23657 if !matches!(self.peek(), Token::LParen) {
23658 return Err(self.err(alloc::format!(
23659 "expected '(' after OVERLAPS, got {:?}",
23660 self.peek()
23661 )));
23662 }
23663 self.advance();
23664 let r0 = self.parse_expr(0)?;
23665 if !matches!(self.peek(), Token::Comma) {
23666 return Err(self.err(alloc::format!(
23667 "OVERLAPS needs (start, end) on the right, got {:?}",
23668 self.peek()
23669 )));
23670 }
23671 self.advance();
23672 let r1 = self.parse_expr(0)?;
23673 if !matches!(self.peek(), Token::RParen) {
23674 return Err(self.err(alloc::format!(
23675 "expected ')' after OVERLAPS pair, got {:?}",
23676 self.peek()
23677 )));
23678 }
23679 self.advance();
23680 let pair_fn = |name: &str, a: &Expr, b: &Expr| Expr::FunctionCall {
23681 name: String::from(name),
23682 args: alloc::vec![a.clone(), b.clone()],
23683 };
23684 let lt = |lhs: Expr, rhs: Expr| Expr::Binary {
23685 lhs: Box::new(lhs),
23686 op: BinOp::Lt,
23687 rhs: Box::new(rhs),
23688 };
23689 return Ok(Expr::Binary {
23690 lhs: Box::new(lt(
23691 pair_fn("least", &row[0], &row[1]),
23692 pair_fn("greatest", &r0, &r1),
23693 )),
23694 op: BinOp::And,
23695 rhs: Box::new(lt(
23696 pair_fn("least", &r0, &r1),
23697 pair_fn("greatest", &row[0], &row[1]),
23698 )),
23699 });
23700 }
23701 // `(a, b, …) IS [NOT] NULL` — the SQL row null predicate. Per
23702 // PG, `IS NULL` is true only when EVERY field is NULL, and
23703 // `IS NOT NULL` is true only when every field is non-NULL — the
23704 // latter is NOT the negation of the former (a mixed row is
23705 // neither). Desugar to an AND chain of per-field `IS [NOT] NULL`,
23706 // which reproduces exactly that all-fields semantics.
23707 if matches!(self.peek(), Token::Is) {
23708 self.advance();
23709 let negated = if matches!(self.peek(), Token::Not) {
23710 self.advance();
23711 true
23712 } else {
23713 false
23714 };
23715 if !matches!(self.peek(), Token::Null) {
23716 return Err(self.err(alloc::format!(
23717 "expected NULL after row IS [NOT], got {:?}",
23718 self.peek()
23719 )));
23720 }
23721 self.advance();
23722 let mut it = row.iter().map(|e| Expr::IsNull {
23723 expr: Box::new(e.clone()),
23724 negated,
23725 });
23726 let first = it.next().expect("row has at least two elements");
23727 return Ok(it.fold(first, |acc, e| Expr::Binary {
23728 lhs: Box::new(acc),
23729 op: BinOp::And,
23730 rhs: Box::new(e),
23731 }));
23732 }
23733 let op = match self.peek() {
23734 Token::Eq => BinOp::Eq,
23735 Token::NotEq => BinOp::NotEq,
23736 Token::Lt => BinOp::Lt,
23737 Token::LtEq => BinOp::LtEq,
23738 Token::Gt => BinOp::Gt,
23739 Token::GtEq => BinOp::GtEq,
23740 // v7.38 (read01, composite) — a bare `(a, b, …)` not followed by a
23741 // comparison / [NOT] IN / IS [NOT] NULL / OVERLAPS is a row (record)
23742 // constructor value, identical to the `ROW(a, b, …)` keyword form:
23743 // `(1,'a')::text` → `(1,a)`, `SELECT (1,2,3)` → `(1,2,3)`. Postfix
23744 // (`::text`, `.field`) applies at the caller just as it does for the
23745 // ROW(...) node. All the comparison / predicate forms returned above.
23746 _ => {
23747 return Ok(Expr::FunctionCall {
23748 name: String::from("row"),
23749 args: row,
23750 });
23751 }
23752 };
23753 self.advance();
23754 // Optional ROW keyword before the paren row.
23755 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("row"))
23756 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
23757 {
23758 self.advance();
23759 }
23760 if !matches!(self.peek(), Token::LParen) {
23761 return Err(self.err(alloc::format!(
23762 "expected '(' to open the right-hand row, got {:?}",
23763 self.peek()
23764 )));
23765 }
23766 self.advance();
23767 // `(a, b) <op> (SELECT x, y)` — compare against a single-row
23768 // subquery. Kept as a node (the subquery's row is a runtime value);
23769 // the literal-RHS form below still decomposes at parse time.
23770 if matches!(self.peek(), Token::Select) {
23771 let inner = self.parse_select_stmt()?;
23772 if !matches!(self.peek(), Token::RParen) {
23773 return Err(self.err(alloc::format!(
23774 "expected ')' after row comparison subquery, got {:?}",
23775 self.peek()
23776 )));
23777 }
23778 self.advance();
23779 let Statement::Select(s) = inner else {
23780 unreachable!("parse_select_stmt always returns Statement::Select")
23781 };
23782 return Ok(Expr::RowCmpSubquery {
23783 row,
23784 op,
23785 subquery: Box::new(s),
23786 });
23787 }
23788 let mut rhs = alloc::vec![self.parse_expr(0)?];
23789 while matches!(self.peek(), Token::Comma) {
23790 self.advance();
23791 rhs.push(self.parse_expr(0)?);
23792 }
23793 if !matches!(self.peek(), Token::RParen) {
23794 return Err(self.err(alloc::format!(
23795 "expected ')' after right-hand row, got {:?}",
23796 self.peek()
23797 )));
23798 }
23799 self.advance();
23800 if rhs.len() != row.len() {
23801 // v7.39 (round 239) — PG's wording (42601).
23802 return Err(self.err("unequal number of entries in row expressions".to_string()));
23803 }
23804 Ok(match op {
23805 BinOp::Eq => row_eq(&row, &rhs),
23806 BinOp::NotEq => maybe_not(row_eq(&row, &rhs), true),
23807 BinOp::Lt => row_lex(&row, &rhs, BinOp::Lt, BinOp::Lt),
23808 BinOp::LtEq => row_lex(&row, &rhs, BinOp::Lt, BinOp::LtEq),
23809 BinOp::Gt => row_lex(&row, &rhs, BinOp::Gt, BinOp::Gt),
23810 BinOp::GtEq => row_lex(&row, &rhs, BinOp::Gt, BinOp::GtEq),
23811 _ => unreachable!("op restricted above"),
23812 })
23813 }
23814
23815 /// `LIKE p ESCAPE 'c'` — rewrite the pattern so the custom
23816 /// escape character becomes the matcher's default backslash:
23817 /// `c%` (escaped wildcard) → `\%`, `cc` (literal escape char)
23818 /// → the char itself, and any pre-existing backslash escapes
23819 /// itself so it stays literal. Both operands must be string
23820 /// literals — a runtime pattern would need matcher support.
23821 fn rewrite_like_escape(pattern: Expr, esc: Expr) -> Result<Expr, String> {
23822 let (Expr::Literal(Literal::String(p)), Expr::Literal(Literal::String(e))) =
23823 (&pattern, &esc)
23824 else {
23825 return Err(
23826 "LIKE ... ESCAPE requires string-literal pattern and escape \
23827 (runtime escape characters are not supported yet)"
23828 .into(),
23829 );
23830 };
23831 // v7.38 (read01 P6.18) — PG accepts `ESCAPE ''` to mean "no escape
23832 // character" (every `%`/`_` is a wildcard, nothing is escaped). Only a
23833 // multi-character escape is an error.
23834 let esc_ch: Option<char> = {
23835 let mut ch_iter = e.chars();
23836 match (ch_iter.next(), ch_iter.next()) {
23837 (Some(c), None) => Some(c),
23838 (None, _) => None,
23839 (Some(_), Some(_)) => {
23840 return Err(alloc::format!(
23841 "ESCAPE must be a single character, got {e:?}"
23842 ));
23843 }
23844 }
23845 };
23846 let mut out = String::with_capacity(p.len() + 4);
23847 let mut chars = p.chars();
23848 while let Some(c) = chars.next() {
23849 if Some(c) == esc_ch {
23850 match chars.next() {
23851 // Escaped wildcard / escaped escape → keep the
23852 // next char literal via backslash.
23853 Some(next) => {
23854 out.push('\\');
23855 out.push(next);
23856 }
23857 None => {
23858 return Err("LIKE pattern ends with the escape character".into());
23859 }
23860 }
23861 } else if c == '\\' && esc_ch != Some('\\') {
23862 // A raw backslash is literal under a custom (or absent) escape
23863 // — escape it for the backslash-based matcher.
23864 out.push('\\');
23865 out.push('\\');
23866 } else {
23867 out.push(c);
23868 }
23869 }
23870 Ok(Expr::Literal(Literal::String(out)))
23871 }
23872
23873 /// `x [NOT] LIKE ANY/ALL (ARRAY[p1, p2, …])` — quantified pattern
23874 /// match. Desugars to an OR (ANY) / AND (ALL) chain of per-element
23875 /// `x [NOT] LIKE pi`, which reproduces PG's three-valued semantics
23876 /// exactly (a NULL pattern makes an element NULL; `false OR NULL` =
23877 /// NULL, `true AND NULL` = NULL, …). ANY over an empty array is
23878 /// FALSE, ALL over empty is TRUE. Returns `None` when the token after
23879 /// LIKE is not `ANY(`/`ALL(`, so the caller falls back to a plain
23880 /// pattern. Only a literal `ARRAY[...]` is accepted today — a runtime
23881 /// array expression errors honestly rather than silently mismatching.
23882 fn try_like_any_all(
23883 &mut self,
23884 base: &Expr,
23885 negated: bool,
23886 case_insensitive: bool,
23887 ) -> Result<Option<Expr>, ParseError> {
23888 let is_any = match self.peek() {
23889 Token::All if matches!(self.tokens.get(self.pos + 1), Some(Token::LParen)) => false,
23890 Token::Ident(s)
23891 if s.eq_ignore_ascii_case("any")
23892 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen)) =>
23893 {
23894 true
23895 }
23896 _ => return Ok(None),
23897 };
23898 self.advance(); // ANY / ALL
23899 self.advance(); // '('
23900 let arr = self.parse_expr(0)?;
23901 if !matches!(self.peek(), Token::RParen) {
23902 return Err(self.err(format!(
23903 "expected ')' after LIKE {} argument, got {:?}",
23904 if is_any { "ANY" } else { "ALL" },
23905 self.peek()
23906 )));
23907 }
23908 self.advance(); // ')'
23909 let Expr::Array(items) = arr else {
23910 return Err(self.err(
23911 "LIKE ANY/ALL currently requires a literal ARRAY[...] of patterns".to_string(),
23912 ));
23913 };
23914 let mut clauses = items.into_iter().map(|p| Expr::Like {
23915 expr: Box::new(base.clone()),
23916 pattern: Box::new(p),
23917 negated,
23918 case_insensitive,
23919 });
23920 let Some(first) = clauses.next() else {
23921 // ANY(empty) = FALSE, ALL(empty) = TRUE.
23922 return Ok(Some(Expr::Literal(Literal::Bool(!is_any))));
23923 };
23924 let op = if is_any { BinOp::Or } else { BinOp::And };
23925 let combined = clauses.fold(first, |acc, c| Expr::Binary {
23926 lhs: Box::new(acc),
23927 op,
23928 rhs: Box::new(c),
23929 });
23930 Ok(Some(combined))
23931 }
23932
23933 /// `x BETWEEN low AND high` → `(x >= low) AND (x <= high)`, wrapped in
23934 /// `NOT` when `negated`. Bounds parse at precedence 5 so the trailing
23935 /// `AND` is not swallowed.
23936 fn parse_between_tail(&mut self, expr: Expr, negated: bool) -> Result<Expr, ParseError> {
23937 self.advance(); // BETWEEN
23938 // SYMMETRIC — the bounds may arrive in either order; both
23939 // orientations OR together. ASYMMETRIC is the default and
23940 // absorbs as noise.
23941 let symmetric = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("symmetric"))
23942 {
23943 self.advance();
23944 true
23945 } else {
23946 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("asymmetric")) {
23947 self.advance();
23948 }
23949 false
23950 };
23951 let low = self.parse_expr(6)?;
23952 if !matches!(self.peek(), Token::And) {
23953 return Err(self.err(format!(
23954 "expected AND after BETWEEN low bound, got {:?}",
23955 self.peek()
23956 )));
23957 }
23958 self.advance();
23959 let high = self.parse_expr(6)?;
23960 let target = Box::new(expr);
23961 let range = |lo: Expr, hi: Expr| Expr::Binary {
23962 lhs: Box::new(Expr::Binary {
23963 lhs: target.clone(),
23964 op: BinOp::GtEq,
23965 rhs: Box::new(lo),
23966 }),
23967 op: BinOp::And,
23968 rhs: Box::new(Expr::Binary {
23969 lhs: target.clone(),
23970 op: BinOp::LtEq,
23971 rhs: Box::new(hi),
23972 }),
23973 };
23974 let combined = if symmetric {
23975 Expr::Binary {
23976 lhs: Box::new(range(low.clone(), high.clone())),
23977 op: BinOp::Or,
23978 rhs: Box::new(range(high, low)),
23979 }
23980 } else {
23981 range(low, high)
23982 };
23983 Ok(maybe_not(combined, negated))
23984 }
23985
23986 /// `x IN (a, b, c)` → chained OR of equalities. Empty list collapses
23987 /// to FALSE (TRUE under NOT IN), matching standard SQL semantics.
23988 /// v4.11: parse `WITH name AS (SELECT ...) [, ...] SELECT ...`.
23989 /// Caller already consumed the leading `WITH` ident.
23990 /// v7.38 (read01) — recursive-CTE well-formedness. PG rejects ORDER BY
23991 /// / LIMIT / OFFSET anywhere in a recursive query, and a recursive
23992 /// self-reference that appears more than once in a single term.
23993 fn validate_recursive_cte(&self, cte: &crate::ast::Cte) -> Result<(), ParseError> {
23994 use crate::ast::{CteBody, SelectStatement};
23995 if !cte.recursive {
23996 return Ok(());
23997 }
23998 let CteBody::Select(body) = &cte.body else {
23999 return Ok(());
24000 };
24001 // A recursive CTE body is `base UNION [ALL] recursive [UNION …]`;
24002 // check the anchor and every peer term.
24003 let has_order = |s: &SelectStatement| !s.order_by.is_empty();
24004 let has_limit = |s: &SelectStatement| s.limit.is_some() || s.offset.is_some();
24005 if has_order(body) || body.unions.iter().any(|(_, u)| has_order(u)) {
24006 return Err(self.err(String::from(
24007 "ORDER BY in a recursive query is not implemented",
24008 )));
24009 }
24010 if has_limit(body) || body.unions.iter().any(|(_, u)| has_limit(u)) {
24011 return Err(self.err(String::from(
24012 "LIMIT in a recursive query is not implemented",
24013 )));
24014 }
24015 let self_refs = |s: &SelectStatement| -> usize {
24016 let Some(from) = &s.from else {
24017 return 0;
24018 };
24019 let mut n = usize::from(from.primary.name.eq_ignore_ascii_case(&cte.name));
24020 for j in &from.joins {
24021 if j.table.name.eq_ignore_ascii_case(&cte.name) {
24022 n += 1;
24023 }
24024 }
24025 n
24026 };
24027 if body.unions.iter().any(|(_, u)| self_refs(u) > 1) {
24028 return Err(self.err(alloc::format!(
24029 "recursive reference to query \"{}\" must not appear more than once",
24030 cte.name
24031 )));
24032 }
24033 // v7.39 (round 145, parse_cte.c) — the remaining well-formedness rules
24034 // apply only when the body actually references itself (a non-self-
24035 // referencing CTE under WITH RECURSIVE may use any set-op shape).
24036 let anchor_refs = self_refs(body);
24037 let union_refs = body.unions.iter().any(|(_, u)| self_refs(u) > 0);
24038 if anchor_refs > 0 || union_refs {
24039 // Shape: the top level must be UNION [ALL] arms only. A self-ref
24040 // under INTERSECT / EXCEPT (or with no set-op at all) is PG's
24041 // "does not have the form" error — SPG used to compute a value.
24042 if body.unions.is_empty()
24043 || body.unions.iter().any(|(k, _)| {
24044 !matches!(
24045 k,
24046 crate::ast::UnionKind::Distinct | crate::ast::UnionKind::All
24047 )
24048 })
24049 {
24050 return Err(self.err(alloc::format!(
24051 "recursive query \"{}\" does not have the form non-recursive-term \
24052 UNION [ALL] recursive-term",
24053 cte.name
24054 )));
24055 }
24056 if anchor_refs > 0 {
24057 return Err(self.err(alloc::format!(
24058 "recursive reference to query \"{}\" must not appear within its non-recursive term",
24059 cte.name
24060 )));
24061 }
24062 }
24063 let is_self = |t: &crate::ast::TableRef| t.name.eq_ignore_ascii_case(&cte.name);
24064 for (_, u) in &body.unions {
24065 if self_refs(u) == 0 {
24066 continue;
24067 }
24068 // The self-reference must not sit on the nullable side of an outer
24069 // join (LEFT: right side; RIGHT: everything before it; FULL: both).
24070 if let Some(from) = &u.from {
24071 for (i, j) in from.joins.iter().enumerate() {
24072 let left_has_self = is_self(&from.primary)
24073 || from.joins[..i].iter().any(|pj| is_self(&pj.table));
24074 let violated = match j.kind {
24075 crate::ast::JoinKind::Left => is_self(&j.table),
24076 crate::ast::JoinKind::Right => left_has_self,
24077 crate::ast::JoinKind::FullOuter => is_self(&j.table) || left_has_self,
24078 _ => false,
24079 };
24080 if violated {
24081 return Err(self.err(alloc::format!(
24082 "recursive reference to query \"{}\" must not appear within an outer join",
24083 cte.name
24084 )));
24085 }
24086 }
24087 }
24088 // No aggregates at the top level of the recursive term (SPG used
24089 // to run them and surface a misleading downstream error).
24090 let mut items_and_having: Vec<&Expr> = Vec::new();
24091 for it in &u.items {
24092 if let crate::ast::SelectItem::Expr { expr, .. } = it {
24093 items_and_having.push(expr);
24094 }
24095 }
24096 if let Some(h) = &u.having {
24097 items_and_having.push(h);
24098 }
24099 for e in items_and_having {
24100 if expr_has_toplevel_aggregate(e) {
24101 return Err(self.err(String::from(
24102 "aggregate functions are not allowed in a recursive query's recursive term",
24103 )));
24104 }
24105 }
24106 }
24107 // A self-reference inside a sublink expression (EXISTS / IN / scalar
24108 // subquery) anywhere in the body is rejected; a plain FROM derived
24109 // table is legal in PG and untouched here.
24110 let mut all_terms: Vec<&SelectStatement> = alloc::vec![body];
24111 all_terms.extend(body.unions.iter().map(|(_, u)| u));
24112 for term in all_terms {
24113 if select_has_self_ref_in_sublink(term, &cte.name) {
24114 return Err(self.err(alloc::format!(
24115 "recursive reference to query \"{}\" must not appear within a subquery",
24116 cte.name
24117 )));
24118 }
24119 }
24120 Ok(())
24121 }
24122
24123 /// v7.38 (read01 U16) — desugar a CTE's SEARCH / CYCLE clause into
24124 /// extra body columns, mirroring PG's `rewriteSearchAndCycle`. Runs
24125 /// right after parse so the engine sees a plain recursive CTE with the
24126 /// tracking columns already projected. DEPTH FIRST and CYCLE are
24127 /// supported; BREADTH FIRST needs numeric-composite ordering SPG's
24128 /// text-rendered rows can't provide, and errors honestly.
24129 fn desugar_cte_search_cycle(&self, cte: &mut crate::ast::Cte) -> Result<(), ParseError> {
24130 use crate::ast::{BinOp, ColumnName, CteBody, Expr, Literal, SelectItem, UnOp};
24131 if cte.search.is_none() && cte.cycle.is_none() {
24132 return Ok(());
24133 }
24134 let cte_name = cte.name.clone();
24135 let col_names = cte.column_overrides.clone();
24136 if col_names.is_empty() {
24137 return Err(
24138 self.err("SEARCH / CYCLE requires an explicit WITH name(cols) column list".into())
24139 );
24140 }
24141 let search = cte.search.take();
24142 let cycle = cte.cycle.take();
24143 let mut extra_cols: Vec<String> = Vec::new();
24144 let col_ref = |name: &str| {
24145 Expr::Column(ColumnName {
24146 qualifier: Some(cte_name.clone()),
24147 name: name.to_string(),
24148 })
24149 };
24150 // Position of a SEARCH/CYCLE column within the CTE's column list.
24151 let pos_of = |name: &str| -> Result<usize, ParseError> {
24152 col_names
24153 .iter()
24154 .position(|c| c.eq_ignore_ascii_case(name))
24155 .ok_or_else(|| {
24156 self.err(format!("SEARCH/CYCLE column {name:?} is not a CTE column"))
24157 })
24158 };
24159 let row_of = |items: &[SelectItem], positions: &[usize]| -> Result<Expr, ParseError> {
24160 let mut args = Vec::with_capacity(positions.len());
24161 for &p in positions {
24162 match items.get(p) {
24163 Some(SelectItem::Expr { expr, .. }) => args.push(expr.clone()),
24164 _ => {
24165 return Err(self.err(
24166 "SEARCH/CYCLE column maps to a non-expression select item".into(),
24167 ));
24168 }
24169 }
24170 }
24171 Ok(Expr::FunctionCall {
24172 name: "row".into(),
24173 args,
24174 })
24175 };
24176 let CteBody::Select(body) = &mut cte.body else {
24177 return Err(self.err("SEARCH / CYCLE requires a SELECT CTE body".into()));
24178 };
24179 if body.unions.is_empty() {
24180 return Err(self.err("SEARCH / CYCLE requires a recursive (UNION) CTE".into()));
24181 }
24182 let rec = body.unions.len() - 1; // recursive term = last UNION peer
24183
24184 if let Some(srch) = search {
24185 // v7.38 (T31) — SEARCH's SET column is ORDER BY'd, and PG's key is a
24186 // `record[]` (DEPTH) or `(depth, keys…)` record (BREADTH). SPG has
24187 // no typed `record[]`, but element-wise array ORDER BY is correct
24188 // (`[1,2] < [1,10] < [2]`), so a SINGLE scalar BY column maps
24189 // exactly onto a typed array: DEPTH is the root→node path
24190 // `array_append(parent, key)`, BREADTH is `[depth, key]`. This
24191 // orders numerically (multi-digit keys included), matching PG.
24192 //
24193 // A multi-column BY would need a record[] to keep the per-node key
24194 // tuple orderable, which SPG can't express — error honestly there
24195 // rather than mis-order.
24196 if srch.by_columns.len() != 1 {
24197 return Err(self.err(
24198 "SEARCH … BY with multiple columns needs typed record[] ordering \
24199 SPG doesn't have yet; a single BY column is supported"
24200 .into(),
24201 ));
24202 }
24203 let key_pos = pos_of(&srch.by_columns[0])?;
24204 let base_key = match body.items.get(key_pos) {
24205 Some(SelectItem::Expr { expr, .. }) => expr.clone(),
24206 _ => {
24207 return Err(
24208 self.err("SEARCH BY column maps to a non-expression select item".into())
24209 );
24210 }
24211 };
24212 let rec_key = match body.unions[rec].1.items.get(key_pos) {
24213 Some(SelectItem::Expr { expr, .. }) => expr.clone(),
24214 _ => {
24215 return Err(
24216 self.err("SEARCH BY column maps to a non-expression select item".into())
24217 );
24218 }
24219 };
24220 if srch.depth_first {
24221 // base: ARRAY[key]; rec: array_append(cte.set, key).
24222 body.items.push(SelectItem::Expr {
24223 expr: Expr::Array(alloc::vec![base_key]),
24224 alias: Some(srch.set_column.clone()),
24225 });
24226 body.unions[rec].1.items.push(SelectItem::Expr {
24227 expr: Expr::FunctionCall {
24228 name: "array_append".into(),
24229 args: alloc::vec![col_ref(&srch.set_column), rec_key],
24230 },
24231 alias: Some(srch.set_column.clone()),
24232 });
24233 } else {
24234 // BREADTH: [depth, key]; depth starts at 0 and increments. The
24235 // leading depth element dominates the element-wise comparison,
24236 // so shallower rows sort first, then by key — PG's (depth, key).
24237 body.items.push(SelectItem::Expr {
24238 expr: Expr::Array(alloc::vec![Expr::Literal(Literal::Integer(0)), base_key,]),
24239 alias: Some(srch.set_column.clone()),
24240 });
24241 // rec depth = cte.set[1] + 1.
24242 let parent_depth = Expr::ArraySubscript {
24243 target: Box::new(col_ref(&srch.set_column)),
24244 index: Box::new(Expr::Literal(Literal::Integer(1))),
24245 };
24246 body.unions[rec].1.items.push(SelectItem::Expr {
24247 expr: Expr::Array(alloc::vec![
24248 Expr::Binary {
24249 lhs: Box::new(parent_depth),
24250 op: BinOp::Add,
24251 rhs: Box::new(Expr::Literal(Literal::Integer(1))),
24252 },
24253 rec_key,
24254 ]),
24255 alias: Some(srch.set_column.clone()),
24256 });
24257 }
24258 extra_cols.push(srch.set_column);
24259 }
24260
24261 if let Some(cyc) = cycle {
24262 let positions: Vec<usize> = cyc
24263 .columns
24264 .iter()
24265 .map(|c| pos_of(c))
24266 .collect::<Result<_, _>>()?;
24267 // v7.38 (read01, T9) — ROW(cols) is now a first-class composite, so
24268 // cast it to text for the cycle path: membership only needs equality,
24269 // and the record text form gives SPG a TextArray path (SPG has no
24270 // typed record[] array). Cycle detection is unaffected.
24271 let base_row = Expr::Cast {
24272 expr: Box::new(row_of(&body.items, &positions)?),
24273 target: CastTarget::Text,
24274 };
24275 let rec_row = Expr::Cast {
24276 expr: Box::new(row_of(&body.unions[rec].1.items, &positions)?),
24277 target: CastTarget::Text,
24278 };
24279 let mark = cyc.mark_value.clone().unwrap_or(Literal::Bool(true));
24280 let dflt = cyc.default_value.clone().unwrap_or(Literal::Bool(false));
24281 // base: <default> AS mark, ARRAY[ROW(cols)] AS path.
24282 body.items.push(SelectItem::Expr {
24283 expr: Expr::Literal(dflt.clone()),
24284 alias: Some(cyc.mark_column.clone()),
24285 });
24286 body.items.push(SelectItem::Expr {
24287 expr: Expr::Array(alloc::vec![base_row]),
24288 alias: Some(cyc.path_column.clone()),
24289 });
24290 // rec mark: ROW(cols) already in the path → cycle.
24291 let hit = Expr::AnyAll {
24292 expr: Box::new(rec_row.clone()),
24293 op: BinOp::Eq,
24294 array: Box::new(col_ref(&cyc.path_column)),
24295 is_any: true,
24296 };
24297 let mark_expr = if cyc.mark_value.is_some() || cyc.default_value.is_some() {
24298 Expr::Case {
24299 operand: None,
24300 branches: alloc::vec![(hit, Expr::Literal(mark))],
24301 else_branch: Some(Box::new(Expr::Literal(dflt))),
24302 }
24303 } else {
24304 hit
24305 };
24306 body.unions[rec].1.items.push(SelectItem::Expr {
24307 expr: mark_expr,
24308 alias: Some(cyc.mark_column.clone()),
24309 });
24310 // rec path: array_append(cte.path, ROW(cols)).
24311 body.unions[rec].1.items.push(SelectItem::Expr {
24312 expr: Expr::FunctionCall {
24313 name: "array_append".into(),
24314 args: alloc::vec![col_ref(&cyc.path_column), rec_row],
24315 },
24316 alias: Some(cyc.path_column.clone()),
24317 });
24318 // rec WHERE: AND NOT cte.mark — stop expanding a cycled row.
24319 let stop = Expr::Unary {
24320 op: UnOp::Not,
24321 expr: Box::new(col_ref(&cyc.mark_column)),
24322 };
24323 let w = &mut body.unions[rec].1.where_;
24324 *w = Some(match w.take() {
24325 Some(prev) => Expr::Binary {
24326 lhs: Box::new(prev),
24327 op: BinOp::And,
24328 rhs: Box::new(stop),
24329 },
24330 None => stop,
24331 });
24332 extra_cols.push(cyc.mark_column);
24333 extra_cols.push(cyc.path_column);
24334 }
24335 cte.column_overrides.extend(extra_cols);
24336 Ok(())
24337 }
24338
24339 /// v7.38 (read01 U16) — `SEARCH { DEPTH | BREADTH } FIRST BY col [,
24340 /// col…] SET seqcol`. Returns None when the next token isn't SEARCH.
24341 fn parse_cte_search_clause(&mut self) -> Result<Option<crate::ast::SearchClause>, ParseError> {
24342 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("search")) {
24343 return Ok(None);
24344 }
24345 self.advance(); // SEARCH
24346 let depth_first = match self.peek() {
24347 Token::Ident(s) if s.eq_ignore_ascii_case("depth") => true,
24348 Token::Ident(s) if s.eq_ignore_ascii_case("breadth") => false,
24349 other => {
24350 return Err(self.err(format!(
24351 "expected DEPTH or BREADTH after SEARCH, got {other:?}"
24352 )));
24353 }
24354 };
24355 self.advance();
24356 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("first")) {
24357 return Err(self.err(format!(
24358 "expected FIRST after SEARCH mode, got {:?}",
24359 self.peek()
24360 )));
24361 }
24362 self.advance();
24363 if !self.peek_is_by() {
24364 return Err(self.err(format!(
24365 "expected BY after SEARCH … FIRST, got {:?}",
24366 self.peek()
24367 )));
24368 }
24369 self.advance();
24370 let mut by_columns = alloc::vec![self.expect_ident_like()?];
24371 while matches!(self.peek(), Token::Comma) {
24372 self.advance();
24373 by_columns.push(self.expect_ident_like()?);
24374 }
24375 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("set")) {
24376 return Err(self.err(format!(
24377 "expected SET in SEARCH clause, got {:?}",
24378 self.peek()
24379 )));
24380 }
24381 self.advance();
24382 let set_column = self.expect_ident_like()?;
24383 Ok(Some(crate::ast::SearchClause {
24384 depth_first,
24385 by_columns,
24386 set_column,
24387 }))
24388 }
24389
24390 /// v7.38 (read01 U16) — `CYCLE col [, col…] SET markcol [TO v DEFAULT w]
24391 /// USING pathcol`. Returns None when the next token isn't CYCLE.
24392 fn parse_cte_cycle_clause(&mut self) -> Result<Option<crate::ast::CycleClause>, ParseError> {
24393 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("cycle")) {
24394 return Ok(None);
24395 }
24396 self.advance(); // CYCLE
24397 let mut columns = alloc::vec![self.expect_ident_like()?];
24398 while matches!(self.peek(), Token::Comma) {
24399 self.advance();
24400 columns.push(self.expect_ident_like()?);
24401 }
24402 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("set")) {
24403 return Err(self.err(format!(
24404 "expected SET in CYCLE clause, got {:?}",
24405 self.peek()
24406 )));
24407 }
24408 self.advance();
24409 let mark_column = self.expect_ident_like()?;
24410 let mut mark_value = None;
24411 let mut default_value = None;
24412 if matches!(self.peek(), Token::To) {
24413 self.advance();
24414 mark_value = Some(self.parse_cycle_literal()?);
24415 if !matches!(self.peek(), Token::Default) {
24416 return Err(self.err(format!(
24417 "expected DEFAULT after CYCLE … TO, got {:?}",
24418 self.peek()
24419 )));
24420 }
24421 self.advance();
24422 default_value = Some(self.parse_cycle_literal()?);
24423 }
24424 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
24425 return Err(self.err(format!(
24426 "expected USING in CYCLE clause, got {:?}",
24427 self.peek()
24428 )));
24429 }
24430 self.advance();
24431 let path_column = self.expect_ident_like()?;
24432 Ok(Some(crate::ast::CycleClause {
24433 columns,
24434 mark_column,
24435 mark_value,
24436 default_value,
24437 path_column,
24438 }))
24439 }
24440
24441 /// The mark / default value in a CYCLE `TO v DEFAULT w` — a bare
24442 /// literal (string / bool / number) in PG.
24443 fn parse_cycle_literal(&mut self) -> Result<crate::ast::Literal, ParseError> {
24444 match self.parse_expr(0)? {
24445 Expr::Literal(l) => Ok(l),
24446 other => Err(self.err(format!(
24447 "CYCLE mark/default value must be a literal, got {other:?}"
24448 ))),
24449 }
24450 }
24451
24452 fn parse_with_cte_then_select(&mut self) -> Result<Statement, ParseError> {
24453 // v4.22: WITH RECURSIVE — optional keyword right after WITH.
24454 // Comes through as an identifier; consume it if present and
24455 // mark every CTE in the clause as recursive (PG semantics —
24456 // the flag is per-WITH, not per-CTE).
24457 let mut recursive = false;
24458 if let Token::Ident(s) | Token::QuotedIdent(s) = self.peek()
24459 && s.eq_ignore_ascii_case("recursive")
24460 {
24461 self.advance();
24462 recursive = true;
24463 }
24464 let mut ctes = Vec::new();
24465 loop {
24466 let name = self.expect_ident_like()?;
24467 // v4.22: optional column-name list — `WITH t(a,b,c) AS ...`.
24468 // PG uses these to rename the body's output columns; we
24469 // do the same below by overriding `columns[i].name`.
24470 let column_overrides: Vec<String> = if matches!(self.peek(), Token::LParen) {
24471 self.advance();
24472 let mut names = Vec::new();
24473 loop {
24474 names.push(self.expect_ident_like()?);
24475 if matches!(self.peek(), Token::Comma) {
24476 self.advance();
24477 continue;
24478 }
24479 break;
24480 }
24481 if !matches!(self.peek(), Token::RParen) {
24482 return Err(self.err(format!(
24483 "expected ')' to close CTE column list, got {:?}",
24484 self.peek()
24485 )));
24486 }
24487 self.advance();
24488 names
24489 } else {
24490 Vec::new()
24491 };
24492 // AS is a reserved Token::As (used by SELECT-item / FROM
24493 // aliasing) — handle it specially rather than as a bare
24494 // ident.
24495 if !matches!(self.peek(), Token::As) {
24496 return Err(self.err(format!(
24497 "expected AS after CTE name {name:?}, got {:?}",
24498 self.peek()
24499 )));
24500 }
24501 self.advance();
24502 // v7.37.17 (17.6 siblings) — PG 12+ `AS [NOT]
24503 // MATERIALIZED` optimizer hints. SPG materialises every
24504 // CTE, so both spellings are accepted and absorbed.
24505 if matches!(self.peek(), Token::Not) {
24506 self.advance(); // NOT
24507 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
24508 if s.eq_ignore_ascii_case("materialized"))
24509 {
24510 self.advance();
24511 } else {
24512 return Err(self.err(format!(
24513 "expected MATERIALIZED after AS NOT, got {:?}",
24514 self.peek()
24515 )));
24516 }
24517 } else if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
24518 if s.eq_ignore_ascii_case("materialized"))
24519 {
24520 self.advance();
24521 }
24522 if !matches!(self.peek(), Token::LParen) {
24523 return Err(self.err(format!(
24524 "expected '(' after AS in WITH clause, got {:?}",
24525 self.peek()
24526 )));
24527 }
24528 self.advance();
24529 // v7.37.43-T4.4 — accept INSERT / UPDATE / DELETE (with
24530 // RETURNING) as the CTE body in addition to SELECT.
24531 // PG writable CTE semantics. UPDATE / DELETE come in as
24532 // bare Idents (lexer keeps SELECT / INSERT as reserved
24533 // tokens but treats the rest of DML as case-insensitive
24534 // idents).
24535 let is_update_kw = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("update"));
24536 let is_delete_kw = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("delete"));
24537 let is_merge_kw = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("merge"));
24538 let body = match self.peek() {
24539 Token::Select => {
24540 let inner = self.parse_select_stmt()?;
24541 let Statement::Select(s) = inner else {
24542 unreachable!("parse_select_stmt returns Select");
24543 };
24544 crate::ast::CteBody::Select(s)
24545 }
24546 // v7.39 (round 869) — `WITH x AS (TABLE t)`. PG spells
24547 // `SELECT * FROM t` this way and accepts it wherever a
24548 // SELECT goes, so the CTE body dispatch needs its own
24549 // arm: this match is keyed on the FIRST token, and
24550 // `Token::Table` fell through to a tail that then
24551 // rejected what it got. `parse_table_shorthand` has
24552 // returned a desugared SelectStatement since the
24553 // shorthand landed — only the routing was missing.
24554 // Round 868 found this by putting the shorthand in a
24555 // subquery; every earlier check used a top-level form.
24556 // v7.39 (round 869) — `WITH x AS (TABLE t)`. PG spells
24557 // `SELECT * FROM t` this way and accepts it wherever a
24558 // SELECT goes, so the CTE body dispatch needs its own
24559 // arm: this match is keyed on the FIRST token, and
24560 // `Token::Table` fell through to a tail that rejected
24561 // what it got. `parse_table_shorthand` has returned a
24562 // desugared SelectStatement since the shorthand landed —
24563 // only the routing was missing, here and in the derived
24564 // table's second-token gate. Round 868 found both by
24565 // putting the shorthand in a subquery; every earlier
24566 // check had used a top-level form.
24567 Token::Table
24568 if matches!(
24569 self.tokens.get(self.pos + 1),
24570 Some(Token::Ident(_) | Token::QuotedIdent(_))
24571 ) =>
24572 {
24573 let mut head = self.parse_table_shorthand()?;
24574 self.parse_setop_chain_into(&mut head)?;
24575 self.parse_select_tail_into(&mut head)?;
24576 crate::ast::CteBody::Select(head)
24577 }
24578 // v7.37.17 (17.6 siblings) — VALUES as a CTE body:
24579 // WITH t(a) AS (VALUES (1), (2)) … lowers through
24580 // the shared rows helper onto a Select body.
24581 Token::Values => {
24582 self.advance(); // VALUES
24583 let mut head = self.parse_values_rows_body()?;
24584 // A VALUES seed can head a set-operation chain —
24585 // WITH RECURSIVE t(n) AS (VALUES(1) UNION ALL
24586 // SELECT n+1 FROM t …). Attach any trailing
24587 // UNION / INTERSECT / EXCEPT peers so the
24588 // recursive-CTE body parses like the SELECT seed.
24589 self.parse_setop_chain_into(&mut head)?;
24590 crate::ast::CteBody::Select(head)
24591 }
24592 Token::Insert => {
24593 let inner = self.parse_one_statement()?;
24594 let Statement::Insert(s) = inner else {
24595 unreachable!("Token::Insert routes to Insert");
24596 };
24597 crate::ast::CteBody::Insert(alloc::boxed::Box::new(s))
24598 }
24599 _ if is_update_kw => {
24600 let inner = self.parse_one_statement()?;
24601 let Statement::Update(s) = inner else {
24602 return Err(
24603 self.err(format!("expected UPDATE inside WITH (…), got {inner:?}"))
24604 );
24605 };
24606 crate::ast::CteBody::Update(alloc::boxed::Box::new(s))
24607 }
24608 _ if is_delete_kw => {
24609 let inner = self.parse_one_statement()?;
24610 let Statement::Delete(s) = inner else {
24611 return Err(
24612 self.err(format!("expected DELETE inside WITH (…), got {inner:?}"))
24613 );
24614 };
24615 crate::ast::CteBody::Delete(alloc::boxed::Box::new(s))
24616 }
24617 // v7.39 (round 149) — PG 17 allows MERGE as a
24618 // data-modifying CTE body.
24619 _ if is_merge_kw => {
24620 let inner = self.parse_one_statement()?;
24621 let Statement::Merge(s) = inner else {
24622 return Err(
24623 self.err(format!("expected MERGE inside WITH (…), got {inner:?}"))
24624 );
24625 };
24626 crate::ast::CteBody::Merge(alloc::boxed::Box::new(s))
24627 }
24628 // v7.39 (round 151) — a CTE body may itself be
24629 // WITH-headed (PG grammar: PreparableStmt carries its
24630 // own with_clause). The nested statement keeps its own
24631 // ctes; the modifying-CTE-at-top-level rule is enforced
24632 // at execution.
24633 Token::Ident(s) if s.eq_ignore_ascii_case("with") => {
24634 self.advance(); // WITH
24635 match self.parse_with_cte_then_select()? {
24636 Statement::Select(s) => crate::ast::CteBody::Select(s),
24637 Statement::Insert(s) => {
24638 crate::ast::CteBody::Insert(alloc::boxed::Box::new(s))
24639 }
24640 Statement::Update(s) => {
24641 crate::ast::CteBody::Update(alloc::boxed::Box::new(s))
24642 }
24643 Statement::Delete(s) => {
24644 crate::ast::CteBody::Delete(alloc::boxed::Box::new(s))
24645 }
24646 Statement::Merge(s) => {
24647 crate::ast::CteBody::Merge(alloc::boxed::Box::new(s))
24648 }
24649
24650 other => {
24651 return Err(self.err(format!(
24652 "WITH body must be SELECT / INSERT / UPDATE / DELETE / MERGE, got {other:?}"
24653 )));
24654 }
24655 }
24656 }
24657 other => {
24658 return Err(self.err(format!(
24659 "WITH body must be SELECT / INSERT / UPDATE / DELETE / MERGE, got {other:?}"
24660 )));
24661 }
24662 };
24663 if !matches!(self.peek(), Token::RParen) {
24664 return Err(self.err(format!(
24665 "expected ')' after CTE body, got {:?}",
24666 self.peek()
24667 )));
24668 }
24669 self.advance();
24670 // v7.38 (read01 U16) — optional SEARCH / CYCLE on a recursive
24671 // CTE, desugared into extra body columns by the engine.
24672 let search = self.parse_cte_search_clause()?;
24673 let cycle = self.parse_cte_cycle_clause()?;
24674 let mut cte = crate::ast::Cte {
24675 name,
24676 body,
24677 recursive,
24678 column_overrides,
24679 search,
24680 cycle,
24681 };
24682 self.validate_recursive_cte(&cte)?;
24683 self.desugar_cte_search_cycle(&mut cte)?;
24684 ctes.push(cte);
24685 if matches!(self.peek(), Token::Comma) {
24686 self.advance();
24687 continue;
24688 }
24689 break;
24690 }
24691 // v7.37.43-T4.4 — the outer body may be SELECT (classical),
24692 // or INSERT / UPDATE / DELETE (writable CTE outer). Attach
24693 // the parsed CTEs to whichever statement the body produces.
24694 let outer_is_update = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("update"));
24695 let outer_is_delete = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("delete"));
24696 let outer_is_merge = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("merge"));
24697 match self.peek() {
24698 Token::Select => {
24699 let body_stmt = self.parse_select_stmt()?;
24700 let Statement::Select(mut body) = body_stmt else {
24701 unreachable!()
24702 };
24703 body.ctes = ctes;
24704 Ok(Statement::Select(body))
24705 }
24706 Token::Insert => {
24707 let body_stmt = self.parse_one_statement()?;
24708 let Statement::Insert(mut body) = body_stmt else {
24709 unreachable!()
24710 };
24711 body.ctes = ctes;
24712 Ok(Statement::Insert(body))
24713 }
24714 _ if outer_is_update => {
24715 let body_stmt = self.parse_one_statement()?;
24716 let Statement::Update(mut body) = body_stmt else {
24717 return Err(self.err(format!("expected UPDATE after WITH clause")));
24718 };
24719 body.ctes = ctes;
24720 Ok(Statement::Update(body))
24721 }
24722 _ if outer_is_delete => {
24723 let body_stmt = self.parse_one_statement()?;
24724 let Statement::Delete(mut body) = body_stmt else {
24725 return Err(self.err(format!("expected DELETE after WITH clause")));
24726 };
24727 body.ctes = ctes;
24728 Ok(Statement::Delete(body))
24729 }
24730 // v7.39 (round 149) — PG 15 allows a WITH clause on MERGE;
24731 // WITH RECURSIVE is rejected with PG's exact message
24732 // (parse analysis, transformWithClause).
24733 _ if outer_is_merge => {
24734 if recursive {
24735 return Err(self.err(String::from(
24736 "WITH RECURSIVE is not supported for MERGE statement",
24737 )));
24738 }
24739 let body_stmt = self.parse_one_statement()?;
24740 let Statement::Merge(mut body) = body_stmt else {
24741 return Err(self.err(format!("expected MERGE after WITH clause")));
24742 };
24743 body.ctes = ctes;
24744 Ok(Statement::Merge(body))
24745 }
24746 other => Err(self.err(format!(
24747 "expected SELECT / INSERT / UPDATE / DELETE / MERGE after WITH clause, got {other:?}"
24748 ))),
24749 }
24750 }
24751
24752 /// v4.10: parse `EXISTS (SELECT ...)`. Caller (`parse_atom`)
24753 /// already consumed the leading `EXISTS` ident via
24754 /// `self.advance()`.
24755 /// v7.13.0 — parse the rest of a `CASE … END` expression after
24756 /// the leading `CASE` ident has been consumed (mailrs round-5
24757 /// G9). Supports both the searched form
24758 /// (`CASE WHEN cond THEN val …`) and the simple form
24759 /// (`CASE operand WHEN val THEN val …`).
24760 fn parse_case_atom(&mut self) -> Result<Expr, ParseError> {
24761 // Disambiguate searched vs simple form: if the next token
24762 // is `WHEN`, we're in the searched form. Otherwise the
24763 // intervening expression is the operand.
24764 let operand = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("when")) {
24765 None
24766 } else {
24767 Some(Box::new(self.parse_expr(0)?))
24768 };
24769 let mut branches: Vec<(Expr, Expr)> = Vec::new();
24770 loop {
24771 match self.peek() {
24772 Token::Ident(s) if s.eq_ignore_ascii_case("when") => {
24773 self.advance();
24774 let cond = self.parse_expr(0)?;
24775 match self.peek() {
24776 Token::Ident(t) if t.eq_ignore_ascii_case("then") => {
24777 self.advance();
24778 }
24779 other => {
24780 return Err(self.err(alloc::format!(
24781 "expected THEN after CASE WHEN <expr>, got {other:?}"
24782 )));
24783 }
24784 }
24785 let value = self.parse_expr(0)?;
24786 branches.push((cond, value));
24787 }
24788 _ => break,
24789 }
24790 }
24791 if branches.is_empty() {
24792 return Err(self.err("CASE requires at least one WHEN … THEN … branch".into()));
24793 }
24794 let else_branch = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("else"))
24795 {
24796 self.advance();
24797 Some(Box::new(self.parse_expr(0)?))
24798 } else {
24799 None
24800 };
24801 match self.peek() {
24802 Token::Ident(s) if s.eq_ignore_ascii_case("end") => {
24803 self.advance();
24804 }
24805 other => {
24806 return Err(self.err(alloc::format!(
24807 "expected END to close CASE expression, got {other:?}"
24808 )));
24809 }
24810 }
24811 Ok(Expr::Case {
24812 operand,
24813 branches,
24814 else_branch,
24815 })
24816 }
24817
24818 /// v7.39 (round 151) — nested `WITH … SELECT …` in a subquery /
24819 /// query-source position (EXISTS / IN / INSERT source / CTE body /
24820 /// view body). Caller consumed the WITH keyword. Only a SELECT
24821 /// outer is grammatical here; the data-modifying-CTE-at-top-level
24822 /// rule (PG 0A000) is enforced at execution, where the SQLSTATE
24823 /// maps correctly.
24824 fn parse_nested_with_select(&mut self) -> Result<crate::ast::SelectStatement, ParseError> {
24825 let inner = self.parse_with_cte_then_select()?;
24826 match inner {
24827 Statement::Select(s) => Ok(s),
24828 other => Err(self.err(format!(
24829 "expected SELECT after WITH in a subquery, got {other:?}"
24830 ))),
24831 }
24832 }
24833
24834 /// True when the next token is the (unquoted) WITH keyword. WITH is
24835 /// reserved in PG, so a bare `with` can never be a column reference
24836 /// in these positions; a quoted `"with"` stays an identifier.
24837 fn peek_is_with_kw(&self) -> bool {
24838 matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with"))
24839 }
24840
24841 /// v7.39 (round 153) — the `ANY / ALL ( [WITH …] SELECT … )` body.
24842 /// `#[inline(never)]` keeps the large SelectStatement temporaries
24843 /// off parse_expr's recursive frame (the nesting-budget stack
24844 /// cliff — see the round-153 gate regression).
24845 #[inline(never)]
24846 fn parse_any_all_select_body(&mut self) -> Result<crate::ast::SelectStatement, ParseError> {
24847 if self.peek_is_with_kw() {
24848 self.advance();
24849 self.parse_nested_with_select()
24850 } else {
24851 match self.parse_select_stmt()? {
24852 Statement::Select(s) => Ok(s),
24853 other => Err(self.err(alloc::format!(
24854 "expected SELECT inside ANY/ALL, got {other:?}"
24855 ))),
24856 }
24857 }
24858 }
24859
24860 fn parse_exists_atom(&mut self, negated: bool) -> Result<Expr, ParseError> {
24861 if !matches!(self.peek(), Token::LParen) {
24862 return Err(self.err(format!("expected '(' after EXISTS, got {:?}", self.peek())));
24863 }
24864 self.advance();
24865 // v7.39 (round 151) — `EXISTS (WITH … SELECT …)` is legal PG.
24866 let s = if self.peek_is_with_kw() {
24867 self.advance();
24868 self.parse_nested_with_select()?
24869 } else {
24870 let inner = self.parse_select_stmt()?;
24871 let Statement::Select(s) = inner else {
24872 unreachable!("parse_select_stmt returns Select")
24873 };
24874 s
24875 };
24876 if !matches!(self.peek(), Token::RParen) {
24877 return Err(self.err(format!(
24878 "expected ')' after EXISTS-subquery, got {:?}",
24879 self.peek()
24880 )));
24881 }
24882 self.advance();
24883 Ok(Expr::Exists {
24884 subquery: Box::new(s),
24885 negated,
24886 })
24887 }
24888
24889 fn parse_in_tail(&mut self, expr: Expr, negated: bool) -> Result<Expr, ParseError> {
24890 self.advance(); // IN
24891 if !matches!(self.peek(), Token::LParen) {
24892 return Err(self.err(format!("expected '(' after IN, got {:?}", self.peek())));
24893 }
24894 self.advance();
24895 // v4.10: `IN (SELECT ...)` — subquery branch. v7.39 (round 151)
24896 // also accepts a WITH-headed subquery (`IN (WITH … SELECT …)`).
24897 if matches!(self.peek(), Token::Select) || self.peek_is_with_kw() {
24898 let s = if self.peek_is_with_kw() {
24899 self.advance();
24900 self.parse_nested_with_select()?
24901 } else {
24902 let inner = self.parse_select_stmt()?;
24903 let Statement::Select(s) = inner else {
24904 unreachable!("parse_select_stmt always returns Statement::Select")
24905 };
24906 s
24907 };
24908 if !matches!(self.peek(), Token::RParen) {
24909 return Err(self.err(format!(
24910 "expected ')' after IN-subquery, got {:?}",
24911 self.peek()
24912 )));
24913 }
24914 self.advance();
24915 return Ok(Expr::InSubquery {
24916 expr: Box::new(expr),
24917 subquery: Box::new(s),
24918 negated,
24919 });
24920 }
24921 let mut elements = Vec::new();
24922 if !matches!(self.peek(), Token::RParen) {
24923 loop {
24924 elements.push(self.parse_expr(0)?);
24925 match self.peek() {
24926 Token::Comma => {
24927 self.advance();
24928 }
24929 Token::RParen => break,
24930 other => {
24931 return Err(
24932 self.err(format!("expected ',' or ')' in IN list, got {other:?}"))
24933 );
24934 }
24935 }
24936 }
24937 }
24938 self.advance(); // ')'
24939 // v7.30.2 (mailrs round-25) — flat InList node instead of a
24940 // left-deep OR-Eq chain: chain depth scaled with the element
24941 // count and overflowed the stack (eval + drop are recursive).
24942 if elements.is_empty() {
24943 return Ok(maybe_not(Expr::Literal(Literal::Bool(false)), negated));
24944 }
24945 Ok(Expr::InList {
24946 expr: Box::new(expr),
24947 list: elements,
24948 negated,
24949 })
24950 }
24951
24952 /// Parse a pgvector array literal `[ x1, x2, ... ]`. The opening `[` is
24953 /// already consumed by the caller. Elements must be numeric literals
24954 /// (with optional unary `-`); any compound expression is rejected at
24955 /// parse time so the runtime never needs to evaluate inside a vector.
24956 /// `EXTRACT(<field> FROM <source>)`. The dispatching `parse_atom`
24957 /// has already consumed the `EXTRACT` token before calling us —
24958 /// we pick up at the opening `(`.
24959 /// v7.17.0 Phase 2.2 — MySQL `MATCH(col [, col ...]) AGAINST
24960 /// (expr [IN BOOLEAN MODE | IN NATURAL LANGUAGE MODE
24961 /// [WITH QUERY EXPANSION]])`. Rewritten in-place to a
24962 /// per-column OR-fold of
24963 /// `to_tsvector('simple', col) @@ plainto_tsquery('simple',
24964 /// term)` so the existing FTS evaluator handles semantics.
24965 ///
24966 /// The mode modifier is accepted-and-ignored at v7.17 — all
24967 /// modes map to the same `plainto_tsquery` rewrite. Boolean-
24968 /// mode operators (`+foo -bar`) would need their own parser
24969 /// (Phase 2.2c); customers who hit them today already get a
24970 /// correct lexeme-match against the bare term, only without
24971 /// the +/- precedence the customer asked for.
24972 fn parse_match_against_atom(&mut self) -> Result<Expr, ParseError> {
24973 // Already at `MATCH`-consumed position; the dispatcher
24974 // confirmed the next token is `(`.
24975 if !matches!(self.peek(), Token::LParen) {
24976 return Err(self.err(alloc::format!(
24977 "expected '(' after MATCH, got {:?}",
24978 self.peek()
24979 )));
24980 }
24981 self.advance();
24982 let mut cols: Vec<Expr> = Vec::new();
24983 loop {
24984 cols.push(self.parse_expr(0)?);
24985 match self.peek() {
24986 Token::Comma => {
24987 self.advance();
24988 }
24989 Token::RParen => break,
24990 other => {
24991 return Err(self.err(alloc::format!(
24992 "expected ',' or ')' in MATCH column list, got {other:?}"
24993 )));
24994 }
24995 }
24996 }
24997 self.advance(); // ')'
24998 // Expect AGAINST.
24999 match self.peek() {
25000 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("against") => {
25001 self.advance();
25002 }
25003 other => {
25004 return Err(self.err(alloc::format!(
25005 "expected AGAINST after MATCH column list, got {other:?}"
25006 )));
25007 }
25008 }
25009 if !matches!(self.peek(), Token::LParen) {
25010 return Err(self.err(alloc::format!(
25011 "expected '(' after AGAINST, got {:?}",
25012 self.peek()
25013 )));
25014 }
25015 self.advance();
25016 // Read AGAINST's argument as a single primary token —
25017 // string literal, placeholder, or column-ref ident. We
25018 // can't call `parse_expr` / `parse_unary` here because
25019 // the postfix chain inside `parse_atom` would greedily
25020 // fold a trailing `IN BOOLEAN MODE` as `expr IN (...)`
25021 // and fail at "expected '(' after IN". Customers always
25022 // write a literal or bound parameter in AGAINST, so this
25023 // restriction is non-blocking; the error path explains
25024 // the limit if a more complex expression shows up.
25025 let term = match self.advance() {
25026 Token::String(s) => Expr::Literal(crate::ast::Literal::String(s)),
25027 Token::Placeholder(n) => Expr::Placeholder(n),
25028 Token::Ident(s) | Token::QuotedIdent(s) => Expr::Column(crate::ast::ColumnName {
25029 qualifier: None,
25030 name: s,
25031 }),
25032 other => {
25033 return Err(self.err(alloc::format!(
25034 "MATCH ... AGAINST(<term>) expects a string literal, \
25035 bound parameter, or column ref, got {other:?}"
25036 )));
25037 }
25038 };
25039 // Optional mode tail — accept-and-ignore at v7.17:
25040 // IN NATURAL LANGUAGE MODE [WITH QUERY EXPANSION]
25041 // IN BOOLEAN MODE
25042 // WITH QUERY EXPANSION
25043 loop {
25044 match self.peek() {
25045 // IN lexes as a reserved Token::In, not an ident,
25046 // so it gets its own arm.
25047 Token::In => {
25048 self.advance();
25049 }
25050 Token::Ident(s) | Token::QuotedIdent(s)
25051 if s.eq_ignore_ascii_case("natural")
25052 || s.eq_ignore_ascii_case("language")
25053 || s.eq_ignore_ascii_case("boolean")
25054 || s.eq_ignore_ascii_case("mode")
25055 || s.eq_ignore_ascii_case("with")
25056 || s.eq_ignore_ascii_case("query")
25057 || s.eq_ignore_ascii_case("expansion") =>
25058 {
25059 self.advance();
25060 }
25061 _ => break,
25062 }
25063 }
25064 if !matches!(self.peek(), Token::RParen) {
25065 return Err(self.err(alloc::format!(
25066 "expected ')' to close AGAINST, got {:?}",
25067 self.peek()
25068 )));
25069 }
25070 self.advance();
25071 // Build per-column `to_tsvector('simple', col) @@
25072 // plainto_tsquery('simple', term)` and OR-fold.
25073 let simple_lit = || Expr::Literal(crate::ast::Literal::String(String::from("simple")));
25074 let plainto = Expr::FunctionCall {
25075 name: String::from("plainto_tsquery"),
25076 args: alloc::vec![simple_lit(), term.clone()],
25077 };
25078 let mut folded: Option<Expr> = None;
25079 for col in cols {
25080 let to_tsv = Expr::FunctionCall {
25081 name: String::from("to_tsvector"),
25082 args: alloc::vec![simple_lit(), col],
25083 };
25084 let leaf = Expr::Binary {
25085 lhs: Box::new(to_tsv),
25086 op: crate::ast::BinOp::TsMatch,
25087 rhs: Box::new(plainto.clone()),
25088 };
25089 folded = Some(match folded {
25090 None => leaf,
25091 Some(prev) => Expr::Binary {
25092 lhs: Box::new(prev),
25093 op: crate::ast::BinOp::Or,
25094 rhs: Box::new(leaf),
25095 },
25096 });
25097 }
25098 match folded {
25099 Some(e) => Ok(e),
25100 None => Err(self.err(String::from(
25101 "MATCH(...) AGAINST(...) requires at least one column",
25102 ))),
25103 }
25104 }
25105
25106 fn parse_extract_atom(&mut self) -> Result<Expr, ParseError> {
25107 if !matches!(self.peek(), Token::LParen) {
25108 return Err(self.err(format!("expected '(' after EXTRACT, got {:?}", self.peek())));
25109 }
25110 self.advance();
25111 let field_name = self.expect_ident_like()?;
25112 let field = match field_name.to_ascii_lowercase().as_str() {
25113 // PG accepts the plural spellings (years/months/…/millenniums) as
25114 // aliases for the singular fields — its datetime unit table has both.
25115 // (quarter has no plural; dow/doy/isoyear/epoch/julian likewise.)
25116 "year" | "years" => ExtractField::Year,
25117 "month" | "months" => ExtractField::Month,
25118 "day" | "days" => ExtractField::Day,
25119 "hour" | "hours" => ExtractField::Hour,
25120 "minute" | "minutes" => ExtractField::Minute,
25121 "second" | "seconds" => ExtractField::Second,
25122 "microsecond" | "microseconds" => ExtractField::Microsecond,
25123 "epoch" => ExtractField::Epoch,
25124 "dow" => ExtractField::Dow,
25125 "isodow" => ExtractField::Isodow,
25126 "doy" => ExtractField::Doy,
25127 "week" | "weeks" => ExtractField::Week,
25128 "isoyear" => ExtractField::Isoyear,
25129 "quarter" => ExtractField::Quarter,
25130 "decade" | "decades" => ExtractField::Decade,
25131 "century" | "centuries" => ExtractField::Century,
25132 "millennium" | "millenniums" | "millennia" => ExtractField::Millennium,
25133 "julian" => ExtractField::Julian,
25134 "millisecond" | "milliseconds" => ExtractField::Millisecond,
25135 "timezone" => ExtractField::Timezone,
25136 "timezone_hour" => ExtractField::TimezoneHour,
25137 "timezone_minute" => ExtractField::TimezoneMinute,
25138 // v7.39 (round 253) — PG resolves EXTRACT fields at runtime and
25139 // reports an unknown one with the source type (22023); carry the
25140 // raw name so eval can word it.
25141 other => ExtractField::Other(alloc::string::String::from(other)),
25142 };
25143 if !matches!(self.peek(), Token::From) {
25144 return Err(self.err(format!(
25145 "expected FROM after EXTRACT field, got {:?}",
25146 self.peek()
25147 )));
25148 }
25149 self.advance();
25150 let source = self.parse_expr(0)?;
25151 if !matches!(self.peek(), Token::RParen) {
25152 return Err(self.err(format!(
25153 "expected ')' to close EXTRACT, got {:?}",
25154 self.peek()
25155 )));
25156 }
25157 self.advance();
25158 Ok(Expr::Extract {
25159 field,
25160 source: Box::new(source),
25161 })
25162 }
25163
25164 /// `INTERVAL '<n> <unit> [<n> <unit> ...]'` — the `INTERVAL` keyword
25165 /// is already consumed; we expect a single string literal next and
25166 /// resolve it into `Literal::Interval` at parse time so the engine
25167 /// never has to re-tokenise inside the string.
25168 /// The unquoted count of a MySQL `INTERVAL <n> <UNIT>`, when the
25169 /// tokens ahead really are one. A quoted count (`INTERVAL '2' DAY`)
25170 /// is the SQL-standard form and is left to the path below.
25171 fn peek_unquoted_interval_count(&self) -> Option<(alloc::string::String, usize)> {
25172 // A negative count lexes as `-` then the number (`INTERVAL -1 DAY`).
25173 let (offset, sign) = match self.peek() {
25174 Token::Minus => (1, "-"),
25175 _ => (0, ""),
25176 };
25177 let Some(Token::Integer(n)) = self.tokens.get(self.pos + offset) else {
25178 return None;
25179 };
25180 self.tokens
25181 .get(self.pos + offset + 1)
25182 .filter(|t| mysql_interval_unit(t).is_some())?;
25183 Some((alloc::format!("{sign}{n}"), offset + 1))
25184 }
25185
25186 /// v7.39 (round 422) — is the parenthesised group starting at the CURRENT
25187 /// `(` a single quantity followed by a time unit (`INTERVAL (1+1) DAY`),
25188 /// rather than the argument list of MySQL's `INTERVAL(N, N1, …)` function?
25189 ///
25190 /// Scans `self.tokens` by index and consumes NOTHING. Round 409 decided
25191 /// this by parsing the group and then restoring `self.pos` — which could
25192 /// never have worked, because `advance()` DESTROYS the token it returns
25193 /// (`mem::replace(.., Eof)`); the restore yielded a stream of Eof. It was
25194 /// inert only because both branches errored back then.
25195 fn interval_paren_is_quantity(&self) -> bool {
25196 let mut depth = 0usize;
25197 let mut saw_top_level_comma = false;
25198 let mut i = self.pos;
25199 while let Some(tok) = self.tokens.get(i) {
25200 match tok {
25201 Token::LParen => depth += 1,
25202 Token::RParen => {
25203 depth = depth.saturating_sub(1);
25204 if depth == 0 {
25205 return !saw_top_level_comma
25206 && mysql_interval_unit(self.tokens.get(i + 1).unwrap_or(&Token::Eof))
25207 .is_some();
25208 }
25209 }
25210 // A comma directly inside the outermost parens means the
25211 // argument list of the INTERVAL() function.
25212 Token::Comma if depth == 1 => saw_top_level_comma = true,
25213 Token::Eof => return false,
25214 _ => {}
25215 }
25216 i += 1;
25217 }
25218 false
25219 }
25220
25221 fn parse_interval_atom(&mut self) -> Result<Expr, ParseError> {
25222 // v7.39 (round 409) — MySQL's `INTERVAL(N, N1, N2, …)` function
25223 // (the index of the last Ni ≤ N), distinct from the interval literal.
25224 // `INTERVAL (` is ambiguous with `INTERVAL (expr) UNIT`, so the shape
25225 // is decided by a non-destructive lookahead (round 422) before either
25226 // branch consumes anything. MySQL only.
25227 if self.mysql_dialect
25228 && matches!(self.peek(), Token::LParen)
25229 && !self.interval_paren_is_quantity()
25230 {
25231 self.advance(); // (
25232 let mut args = Vec::new();
25233 if !matches!(self.peek(), Token::RParen) {
25234 loop {
25235 args.push(self.parse_expr(0)?);
25236 if matches!(self.peek(), Token::Comma) {
25237 self.advance();
25238 continue;
25239 }
25240 break;
25241 }
25242 }
25243 if !matches!(self.peek(), Token::RParen) {
25244 return Err(self.err(alloc::format!(
25245 "expected ')' after INTERVAL() arguments, got {:?}",
25246 self.peek()
25247 )));
25248 }
25249 self.advance(); // )
25250 return Ok(Expr::FunctionCall {
25251 name: alloc::string::String::from("interval"),
25252 args,
25253 });
25254 }
25255 // v7.39 (round 350, M7) — MySQL's `INTERVAL <n> <UNIT>`, with the
25256 // number UNQUOTED: `DATE_ADD(d, INTERVAL 1 MONTH)`,
25257 // `d + INTERVAL 90 MINUTE`, `INTERVAL -1 DAY`. It is how MySQL
25258 // writes every date arithmetic there is, and it did not parse at
25259 // all. PG rejects the unquoted form outright (`syntax error at or
25260 // near "1"`, measured), so it is taken only in the MySQL dialect —
25261 // PG's own `INTERVAL '1' DAY` is untouched below.
25262 if self.mysql_dialect
25263 && let Some((text, consume)) = self.peek_unquoted_interval_count()
25264 {
25265 for _ in 0..consume {
25266 self.advance(); // the optional `-` and the number
25267 }
25268 let Some(unit) = mysql_interval_unit(self.peek()) else {
25269 return Err(self.err(alloc::format!(
25270 "expected an interval unit after INTERVAL {text}, got {:?}",
25271 self.peek()
25272 )));
25273 };
25274 self.advance(); // the unit
25275 let (months, days, micros) = scale_mysql_interval(&text, unit)
25276 .ok_or_else(|| self.err(alloc::format!("cannot read INTERVAL {text} {unit}")))?;
25277 return Ok(Expr::Literal(Literal::Interval {
25278 months,
25279 days,
25280 micros,
25281 // The canonical rendering, so Display round-trips into a
25282 // form both dialects read back.
25283 text: alloc::format!("{text} {unit}"),
25284 }));
25285 }
25286 // v7.39 (round 422) — MySQL's interval QUANTITY may be any expression,
25287 // not just a literal: `DATE_ADD(d, INTERVAL n DAY)`,
25288 // `d + INTERVAL n*2 DAY`, `INTERVAL (1+1) DAY`, `INTERVAL ABS(-5) DAY`.
25289 // Those cannot fold into a compile-time `Literal::Interval`, so they
25290 // lower onto the existing `make_interval(y, mo, w, d, h, mi, s)`
25291 // builtin, which builds the value at run time (and yields NULL for a
25292 // NULL quantity, as MariaDB does). The literal path above still folds
25293 // the constant case — it is cheaper and round-trips through Display.
25294 //
25295 // Guarded off a String operand so PG's own `INTERVAL '1 day'` (and
25296 // MySQL's quoted spelling) keep the qualifier path below.
25297 if self.mysql_dialect && !matches!(self.peek(), Token::String(_)) {
25298 let qty = self.parse_expr(0)?;
25299 let Some(unit) = mysql_interval_unit(self.peek()) else {
25300 return Err(self.err(alloc::format!(
25301 "expected an interval unit after INTERVAL <expr>, got {:?}",
25302 self.peek()
25303 )));
25304 };
25305 self.advance(); // the unit
25306 return Ok(make_interval_call(qty, unit));
25307 }
25308 let tok = self.advance();
25309 let Token::String(text) = tok else {
25310 return Err(self.err(format!(
25311 "expected string literal after INTERVAL, got {tok:?}"
25312 )));
25313 };
25314 // v7.39 (read01 round 102) — SQL-standard trailing field qualifier
25315 // `<FIELD> [TO <FIELD>]` (`INTERVAL '2' YEAR`, `INTERVAL '1-6' YEAR TO
25316 // MONTH`, `INTERVAL '1 2:03:04' DAY TO SECOND`). It sets which field a
25317 // bare number means and the leading/trailing precision.
25318 let field1 = interval_field_of(self.peek());
25319 let qualifier = if let Some(f1) = field1 {
25320 self.advance();
25321 let f2 = if matches!(self.peek(), Token::To) {
25322 self.advance();
25323 let Some(f) = interval_field_of(self.peek()) else {
25324 return Err(self.err(format!(
25325 "expected an interval field after TO, got {:?}",
25326 self.peek()
25327 )));
25328 };
25329 self.advance();
25330 Some(f)
25331 } else {
25332 None
25333 };
25334 Some((f1, f2))
25335 } else {
25336 None
25337 };
25338 let (months, days, micros) = match qualifier {
25339 Some(q) => interpret_qualified_interval(&text, q),
25340 None => parse_interval_text(&text),
25341 }
25342 .ok_or_else(|| ParseError {
25343 message: format!(
25344 "cannot parse INTERVAL {text:?}; \
25345 expected `<n> <unit> [<n> <unit> ...]` with units \
25346 microsecond[s], millisecond[s], second[s], minute[s], \
25347 hour[s], day[s], week[s], month[s], year[s]"
25348 ),
25349 token_pos: self.consumed_pos(),
25350 })?;
25351 Ok(Expr::Literal(Literal::Interval {
25352 months,
25353 days,
25354 micros,
25355 text,
25356 }))
25357 }
25358
25359 /// v7.38 (read01, T10) — parse a bracketed sub-array `[e, e, …]` inside an
25360 /// `ARRAY[...]` constructor, recursing on further nested `[...]` so
25361 /// `ARRAY[[1,2],[3,4]]` (and deeper) becomes nested `Expr::Array` rather
25362 /// than a pgvector literal.
25363 fn parse_array_bracket_body(&mut self) -> Result<Expr, ParseError> {
25364 self.advance(); // consume `[`
25365 let mut items: Vec<Expr> = Vec::new();
25366 if !matches!(self.peek(), Token::RBracket) {
25367 loop {
25368 if matches!(self.peek(), Token::LBracket) {
25369 items.push(self.parse_array_bracket_body()?);
25370 } else {
25371 items.push(self.parse_expr(0)?);
25372 }
25373 match self.peek() {
25374 Token::Comma => {
25375 self.advance();
25376 }
25377 Token::RBracket => break,
25378 other => {
25379 return Err(self.err(alloc::format!(
25380 "expected ',' or ']' in array literal, got {other:?}"
25381 )));
25382 }
25383 }
25384 }
25385 }
25386 self.advance(); // consume `]`
25387 Ok(Expr::Array(items))
25388 }
25389
25390 fn parse_vector_literal_body(&mut self) -> Result<Expr, ParseError> {
25391 let mut elems = Vec::new();
25392 if matches!(self.peek(), Token::RBracket) {
25393 self.advance();
25394 return Ok(Expr::Literal(Literal::Vector(elems)));
25395 }
25396 loop {
25397 let e = self.parse_expr(0)?;
25398 let x = extract_numeric_literal(&e).ok_or_else(|| ParseError {
25399 message: format!("vector element must be a numeric literal, got {e:?}"),
25400 token_pos: self.pos,
25401 })?;
25402 elems.push(x);
25403 match self.peek() {
25404 Token::Comma => {
25405 self.advance();
25406 }
25407 Token::RBracket => {
25408 self.advance();
25409 break;
25410 }
25411 other => {
25412 return Err(self.err(format!("expected ',' or ']' in vector, got {other:?}")));
25413 }
25414 }
25415 }
25416 Ok(Expr::Literal(Literal::Vector(elems)))
25417 }
25418
25419 /// Atom that started with an identifier: could be `t.col`, `col`, or
25420 /// `func(arg, ...)`. Detect each shape by looking at the next token.
25421 /// v4.12: parse `(PARTITION BY expr, ... ORDER BY expr [DESC]
25422 /// [, ...])`. Caller has already consumed `OVER`. Either clause
25423 /// is optional; an empty `()` is also legal (PG semantics).
25424 /// v6.4.2 — consume an optional `IGNORE NULLS` / `RESPECT NULLS`
25425 /// modifier between `name(args)` and `OVER (...)`. Default is
25426 /// `Respect`. Unrecognised idents leave the stream unchanged.
25427 fn parse_null_treatment_modifier(&mut self) -> NullTreatment {
25428 let Token::Ident(s) = self.peek().clone() else {
25429 return NullTreatment::Respect;
25430 };
25431 let is_ignore = s.eq_ignore_ascii_case("ignore");
25432 let is_respect = s.eq_ignore_ascii_case("respect");
25433 if !is_ignore && !is_respect {
25434 return NullTreatment::Respect;
25435 }
25436 // Lookahead for NULLS — only consume both tokens together.
25437 // pos+1 must hold a "nulls" ident.
25438 if self.pos + 1 < self.tokens.len()
25439 && let Token::Ident(s2) = &self.tokens[self.pos + 1]
25440 && s2.eq_ignore_ascii_case("nulls")
25441 {
25442 self.advance();
25443 self.advance();
25444 return if is_ignore {
25445 NullTreatment::Ignore
25446 } else {
25447 NullTreatment::Respect
25448 };
25449 }
25450 NullTreatment::Respect
25451 }
25452
25453 /// v7.32 (mailrs round-29) — `agg(args) FILTER (WHERE cond)`.
25454 /// `FILTER` is an unreserved keyword, so it arrives as an `Ident`
25455 /// (same shape as the `OVER` tail). Consumes the whole clause and
25456 /// returns the predicate; returns `None` when no `FILTER` follows.
25457 fn parse_filter_clause(&mut self) -> Result<Option<Box<Expr>>, ParseError> {
25458 let (Token::Ident(s) | Token::QuotedIdent(s)) = self.peek() else {
25459 return Ok(None);
25460 };
25461 if !s.eq_ignore_ascii_case("filter") {
25462 return Ok(None);
25463 }
25464 self.advance(); // FILTER
25465 if !matches!(self.peek(), Token::LParen) {
25466 return Err(self.err(format!("expected '(' after FILTER, got {:?}", self.peek())));
25467 }
25468 self.advance(); // (
25469 if !matches!(self.peek(), Token::Where) {
25470 return Err(self.err(format!(
25471 "expected WHERE inside FILTER (...), got {:?}",
25472 self.peek()
25473 )));
25474 }
25475 self.advance(); // WHERE
25476 let cond = self.parse_expr(0)?;
25477 if !matches!(self.peek(), Token::RParen) {
25478 return Err(self.err(format!(
25479 "expected ')' to close FILTER (WHERE ...), got {:?}",
25480 self.peek()
25481 )));
25482 }
25483 self.advance(); // )
25484 Ok(Some(Box::new(cond)))
25485 }
25486
25487 /// v7.39 (round 354, M12) — consume a `SEPARATOR '<s>'` tail and push
25488 /// the separator as the aggregate's second argument, which is the
25489 /// shape `string_agg` already takes. Returns whether one was there.
25490 fn consume_group_concat_separator(&mut self, args: &mut Vec<Expr>) -> Result<bool, ParseError> {
25491 if !matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("separator")) {
25492 return Ok(false);
25493 }
25494 self.advance();
25495 let Token::String(sep) = self.peek().clone() else {
25496 return Err(self.err(alloc::format!(
25497 "expected a string literal after SEPARATOR, got {:?}",
25498 self.peek()
25499 )));
25500 };
25501 self.advance();
25502 args.push(Expr::Literal(Literal::String(sep)));
25503 Ok(true)
25504 }
25505
25506 /// v7.32 (round-29) — `WITHIN GROUP ( ORDER BY <sort_spec> )` tail
25507 /// for ordered-set aggregates. `WITHIN` is unreserved (arrives as an
25508 /// `Ident`); `GROUP` and `ORDER`/`BY` are keywords. Returns the sort
25509 /// keys, or an empty vec when no `WITHIN GROUP` follows.
25510 fn parse_within_group_clause(&mut self) -> Result<Vec<OrderBy>, ParseError> {
25511 let (Token::Ident(s) | Token::QuotedIdent(s)) = self.peek() else {
25512 return Ok(Vec::new());
25513 };
25514 if !s.eq_ignore_ascii_case("within") {
25515 return Ok(Vec::new());
25516 }
25517 self.advance(); // WITHIN
25518 if !matches!(self.peek(), Token::Group) {
25519 return Err(self.err(format!(
25520 "expected GROUP after WITHIN, got {:?}",
25521 self.peek()
25522 )));
25523 }
25524 self.advance(); // GROUP
25525 if !matches!(self.peek(), Token::LParen) {
25526 return Err(self.err(format!(
25527 "expected '(' after WITHIN GROUP, got {:?}",
25528 self.peek()
25529 )));
25530 }
25531 self.advance(); // (
25532 if !matches!(self.peek(), Token::Order) {
25533 return Err(self.err(format!(
25534 "expected ORDER BY inside WITHIN GROUP (...), got {:?}",
25535 self.peek()
25536 )));
25537 }
25538 self.advance(); // ORDER
25539 if !self.peek_is_by() {
25540 return Err(self.err(format!("expected BY after ORDER, got {:?}", self.peek())));
25541 }
25542 self.advance(); // BY
25543 let mut keys: Vec<OrderBy> = Vec::new();
25544 loop {
25545 // v7.39 (round 691) — save/restore, the discipline this parser
25546 // already uses around `pending_sample_preds`, so a subquery inside
25547 // a key neither inherits nor leaks the channel.
25548 let saved_flag = core::mem::replace(&mut self.in_order_by_key, true);
25549 let saved_coll = self.order_key_collation.take();
25550 let parsed = self.parse_expr(0);
25551 self.in_order_by_key = saved_flag;
25552 let collation = core::mem::replace(&mut self.order_key_collation, saved_coll);
25553 let expr = parsed?;
25554 let desc = if matches!(self.peek(), Token::Desc) {
25555 self.advance();
25556 true
25557 } else if matches!(self.peek(), Token::Asc) {
25558 self.advance();
25559 false
25560 } else {
25561 false
25562 };
25563 let nulls_first = self.parse_optional_nulls_placement()?;
25564 keys.push(OrderBy {
25565 expr,
25566 desc,
25567 nulls_first,
25568 collation,
25569 });
25570 if matches!(self.peek(), Token::Comma) {
25571 self.advance();
25572 } else {
25573 break;
25574 }
25575 }
25576 if !matches!(self.peek(), Token::RParen) {
25577 return Err(self.err(format!(
25578 "expected ')' to close WITHIN GROUP (ORDER BY ...), got {:?}",
25579 self.peek()
25580 )));
25581 }
25582 self.advance(); // )
25583 Ok(keys)
25584 }
25585
25586 /// No frame clause is supported.
25587 #[allow(clippy::type_complexity)] // (partitions, ordered-keys-with-desc) is the natural shape
25588 fn parse_over_clause(
25589 &mut self,
25590 ) -> Result<
25591 (
25592 Vec<Expr>,
25593 Vec<(Expr, bool, Option<bool>)>,
25594 Option<WindowFrame>,
25595 ),
25596 ParseError,
25597 > {
25598 // `OVER w` — a named-window reference. The WINDOW clause
25599 // parses after the select list, so the name rides out as a
25600 // marker in partition_by; parse_bare_select substitutes the
25601 // definition once the clause is known.
25602 if let Token::Ident(w) | Token::QuotedIdent(w) = self.peek() {
25603 let name = w.clone();
25604 self.advance();
25605 return Ok((
25606 alloc::vec![Expr::Column(crate::ast::ColumnName {
25607 qualifier: Some("__named_window__".to_string()),
25608 name,
25609 })],
25610 Vec::new(),
25611 None,
25612 ));
25613 }
25614 if !matches!(self.peek(), Token::LParen) {
25615 return Err(self.err(format!("expected '(' after OVER, got {:?}", self.peek())));
25616 }
25617 self.advance();
25618 let mut partition_by = Vec::new();
25619 let mut order_by = Vec::new();
25620 // v7.39 (round 229) — `OVER (w1 …)`: a *copy* of an existing named
25621 // window, refined in place. PG's rules (probed against 18.4) differ
25622 // from the bare `OVER w1` form, so the reference rides out under its
25623 // own marker and `substitute_named_windows` applies them. The base
25624 // name is any leading identifier that isn't a window-spec keyword.
25625 let base_window = match self.peek() {
25626 Token::Ident(s) | Token::QuotedIdent(s)
25627 if !s.eq_ignore_ascii_case("partition")
25628 && !s.eq_ignore_ascii_case("rows")
25629 && !s.eq_ignore_ascii_case("range")
25630 && !s.eq_ignore_ascii_case("groups") =>
25631 {
25632 let n = s.clone();
25633 self.advance();
25634 Some(n)
25635 }
25636 _ => None,
25637 };
25638 // PARTITION BY ?
25639 // v7.37.6-B promoted PARTITION to a reserved keyword
25640 // (Token::Partition); pre-7.37.6-B catalogs lexed it as
25641 // `Token::Ident("partition")`. Accept both so older sources
25642 // and the new lexer surface land on the same path.
25643 let is_partition_kw = match self.peek() {
25644 Token::Partition => true,
25645 Token::Ident(s) | Token::QuotedIdent(s) => s.eq_ignore_ascii_case("partition"),
25646 _ => false,
25647 };
25648 if is_partition_kw {
25649 self.advance();
25650 if !self.peek_is_by() {
25651 return Err(self.err(format!(
25652 "expected BY after PARTITION, got {:?}",
25653 self.peek()
25654 )));
25655 }
25656 self.advance();
25657 loop {
25658 partition_by.push(self.parse_expr(0)?);
25659 if matches!(self.peek(), Token::Comma) {
25660 self.advance();
25661 continue;
25662 }
25663 break;
25664 }
25665 }
25666 // ORDER BY ?
25667 if matches!(self.peek(), Token::Order) {
25668 self.advance();
25669 if !self.peek_is_by() {
25670 return Err(self.err(format!("expected BY after ORDER, got {:?}", self.peek())));
25671 }
25672 self.advance();
25673 loop {
25674 let e = self.parse_expr(0)?;
25675 let desc = if matches!(self.peek(), Token::Desc) {
25676 self.advance();
25677 true
25678 } else if matches!(self.peek(), Token::Asc) {
25679 self.advance();
25680 false
25681 } else {
25682 false
25683 };
25684 // v7.24.1 — NULLS FIRST/LAST inside OVER (…).
25685 let nulls_first = self.parse_optional_nulls_placement()?;
25686 order_by.push((e, desc, nulls_first));
25687 if matches!(self.peek(), Token::Comma) {
25688 self.advance();
25689 continue;
25690 }
25691 break;
25692 }
25693 }
25694 // v4.20: optional explicit frame, `ROWS ...` / `RANGE ...`.
25695 // Both keywords come through the lexer as identifiers; match
25696 // case-insensitively.
25697 let mut frame: Option<WindowFrame> = None;
25698 if let Token::Ident(s) | Token::QuotedIdent(s) = self.peek() {
25699 let kind = if s.eq_ignore_ascii_case("rows") {
25700 Some(FrameKind::Rows)
25701 } else if s.eq_ignore_ascii_case("range") {
25702 Some(FrameKind::Range)
25703 } else if s.eq_ignore_ascii_case("groups") {
25704 // v7.37.19 (19.11) — PG 11+ GROUPS frame mode.
25705 Some(FrameKind::Groups)
25706 } else {
25707 None
25708 };
25709 if let Some(kind) = kind {
25710 self.advance();
25711 frame = Some(self.parse_frame_tail(kind)?);
25712 }
25713 }
25714 if !matches!(self.peek(), Token::RParen) {
25715 return Err(self.err(format!(
25716 "expected ')' to close OVER clause, got {:?}",
25717 self.peek()
25718 )));
25719 }
25720 self.advance();
25721 if let Some(base) = base_window {
25722 // A copy may refine but never override the base's partitioning
25723 // (PG rejects it outright, before looking the name up).
25724 if !partition_by.is_empty() {
25725 return Err(self.err(alloc::format!(
25726 "cannot override PARTITION BY clause of window \"{base}\""
25727 )));
25728 }
25729 partition_by = alloc::vec![Expr::Column(crate::ast::ColumnName {
25730 qualifier: Some("__named_window_ref__".to_string()),
25731 name: base,
25732 })];
25733 }
25734 Ok((partition_by, order_by, frame))
25735 }
25736
25737 /// v4.20: parse the tail of an explicit frame, given the `ROWS`
25738 /// or `RANGE` keyword was just consumed. Accepts both
25739 /// `BETWEEN <bound> AND <bound>` and the single-bound shorthand
25740 /// (`ROWS UNBOUNDED PRECEDING`, `ROWS 5 PRECEDING`, etc.) which
25741 /// PG normalises to `BETWEEN <bound> AND CURRENT ROW`.
25742 fn parse_frame_tail(&mut self, kind: FrameKind) -> Result<WindowFrame, ParseError> {
25743 let (start, end) = if matches!(self.peek(), Token::Between) {
25744 self.advance();
25745 let start = self.parse_frame_bound()?;
25746 if !matches!(self.peek(), Token::And) {
25747 return Err(self.err(format!("expected AND in frame spec, got {:?}", self.peek())));
25748 }
25749 self.advance();
25750 let end = self.parse_frame_bound()?;
25751 (start, Some(end))
25752 } else {
25753 (self.parse_frame_bound()?, None)
25754 };
25755 let exclude = self.parse_frame_exclusion()?;
25756 Ok(WindowFrame {
25757 kind,
25758 start,
25759 end,
25760 exclude,
25761 })
25762 }
25763
25764 /// Optional `EXCLUDE {CURRENT ROW | GROUP | TIES | NO OTHERS}`
25765 /// after a frame spec. NO OTHERS is the default no-op.
25766 fn parse_frame_exclusion(&mut self) -> Result<FrameExclusion, ParseError> {
25767 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("exclude")) {
25768 return Ok(FrameExclusion::NoOthers);
25769 }
25770 self.advance(); // EXCLUDE
25771 match self.peek() {
25772 Token::Ident(s) if s.eq_ignore_ascii_case("current") => {
25773 self.advance();
25774 if !matches!(self.peek(), Token::Ident(r) if r.eq_ignore_ascii_case("row")) {
25775 return Err(self.err(format!(
25776 "expected ROW after EXCLUDE CURRENT, got {:?}",
25777 self.peek()
25778 )));
25779 }
25780 self.advance();
25781 Ok(FrameExclusion::CurrentRow)
25782 }
25783 // v7.39 (read01 round 109) — GROUP is a reserved keyword token, so
25784 // `EXCLUDE GROUP` arrives as `Token::Group`, not `Ident("group")`.
25785 // Without this arm it fell to the catch-all, whose message
25786 // self-contradictingly listed GROUP as expected.
25787 Token::Ident(s) if s.eq_ignore_ascii_case("group") => {
25788 self.advance();
25789 Ok(FrameExclusion::Group)
25790 }
25791 Token::Group => {
25792 self.advance();
25793 Ok(FrameExclusion::Group)
25794 }
25795 Token::Ident(s) if s.eq_ignore_ascii_case("ties") => {
25796 self.advance();
25797 Ok(FrameExclusion::Ties)
25798 }
25799 Token::Ident(s) if s.eq_ignore_ascii_case("no") => {
25800 self.advance();
25801 if !matches!(self.peek(), Token::Ident(r) if r.eq_ignore_ascii_case("others")) {
25802 return Err(self.err(format!(
25803 "expected OTHERS after EXCLUDE NO, got {:?}",
25804 self.peek()
25805 )));
25806 }
25807 self.advance();
25808 Ok(FrameExclusion::NoOthers)
25809 }
25810 other => Err(self.err(format!(
25811 "expected CURRENT ROW / GROUP / TIES / NO OTHERS after EXCLUDE, got {other:?}"
25812 ))),
25813 }
25814 }
25815
25816 /// Parse one frame bound: `UNBOUNDED PRECEDING`, `<n> PRECEDING`,
25817 /// `<interval> PRECEDING`, `CURRENT ROW`, `<n>/<interval> FOLLOWING`,
25818 /// `UNBOUNDED FOLLOWING`.
25819 fn parse_frame_bound(&mut self) -> Result<FrameBound, ParseError> {
25820 // Interval-typed offset for a value-based RANGE frame over a
25821 // DATE / TIMESTAMP ORDER BY column (PG time-series windows),
25822 // spelled `INTERVAL '1 day' PRECEDING` or `'1 day'::interval
25823 // PRECEDING`.
25824 if let Some((months, days, micros)) = self.try_take_interval_offset()? {
25825 let dir = self.expect_ident_like()?;
25826 return if dir.eq_ignore_ascii_case("preceding") {
25827 Ok(FrameBound::IntervalPreceding {
25828 months,
25829 days,
25830 micros,
25831 })
25832 } else if dir.eq_ignore_ascii_case("following") {
25833 Ok(FrameBound::IntervalFollowing {
25834 months,
25835 days,
25836 micros,
25837 })
25838 } else {
25839 Err(self.err(format!(
25840 "expected PRECEDING or FOLLOWING after interval offset, got {dir:?}"
25841 )))
25842 };
25843 }
25844 // Number-led: "<n> PRECEDING" / "<n> FOLLOWING".
25845 if let Token::Integer(n) = *self.peek() {
25846 self.advance();
25847 let n: u64 = u64::try_from(n).map_err(|_| {
25848 self.err(format!(
25849 "invalid frame offset {n} — expected non-negative integer"
25850 ))
25851 })?;
25852 let dir = self.expect_ident_like()?;
25853 return if dir.eq_ignore_ascii_case("preceding") {
25854 Ok(FrameBound::OffsetPreceding(n))
25855 } else if dir.eq_ignore_ascii_case("following") {
25856 Ok(FrameBound::OffsetFollowing(n))
25857 } else {
25858 Err(self.err(format!(
25859 "expected PRECEDING or FOLLOWING after offset, got {dir:?}"
25860 )))
25861 };
25862 }
25863 let first = self.expect_ident_like()?;
25864 if first.eq_ignore_ascii_case("unbounded") {
25865 let dir = self.expect_ident_like()?;
25866 return if dir.eq_ignore_ascii_case("preceding") {
25867 Ok(FrameBound::UnboundedPreceding)
25868 } else if dir.eq_ignore_ascii_case("following") {
25869 Ok(FrameBound::UnboundedFollowing)
25870 } else {
25871 Err(self.err(format!(
25872 "expected PRECEDING or FOLLOWING after UNBOUNDED, got {dir:?}"
25873 )))
25874 };
25875 }
25876 if first.eq_ignore_ascii_case("current") {
25877 let row = self.expect_ident_like()?;
25878 if !row.eq_ignore_ascii_case("row") {
25879 return Err(self.err(format!("expected ROW after CURRENT, got {row:?}")));
25880 }
25881 return Ok(FrameBound::CurrentRow);
25882 }
25883 Err(self.err(format!(
25884 "expected frame bound (UNBOUNDED/CURRENT/<n>), got {first:?}"
25885 )))
25886 }
25887
25888 /// Detect and consume a leading interval offset in a frame bound —
25889 /// `INTERVAL '1 day'` or `'1 day'::interval` — returning its folded
25890 /// `(months, days, micros)`. Leaves the cursor on the trailing
25891 /// PRECEDING / FOLLOWING keyword. Returns `None` (without advancing)
25892 /// when the next tokens are not an interval offset.
25893 fn try_take_interval_offset(&mut self) -> Result<Option<(i32, i32, i64)>, ParseError> {
25894 // Shape A — `INTERVAL '1 day'`.
25895 if matches!(self.peek(), Token::Interval) {
25896 self.advance(); // INTERVAL
25897 let atom = self.parse_interval_atom()?;
25898 if let Expr::Literal(Literal::Interval {
25899 months,
25900 days,
25901 micros,
25902 ..
25903 }) = atom
25904 {
25905 return Ok(Some((months, days, micros)));
25906 }
25907 return Err(self.err("expected an interval literal in frame offset".to_string()));
25908 }
25909 // Shape B — `'1 day'::interval`. Look ahead for the exact
25910 // string / `::` / interval-target triple before committing.
25911 if let Token::String(text) = self.peek() {
25912 let target_is_interval = match self.tokens.get(self.pos + 2) {
25913 Some(Token::Interval) => true,
25914 Some(Token::Ident(s)) => s.eq_ignore_ascii_case("interval"),
25915 _ => false,
25916 };
25917 let is_cast = matches!(self.tokens.get(self.pos + 1), Some(Token::DoubleColon))
25918 && target_is_interval;
25919 if is_cast {
25920 let text = text.clone();
25921 self.advance(); // string
25922 self.advance(); // ::
25923 self.advance(); // interval
25924 let parts = parse_interval_text(&text).ok_or_else(|| {
25925 self.err(format!("cannot parse INTERVAL {text:?} in frame offset"))
25926 })?;
25927 return Ok(Some(parts));
25928 }
25929 }
25930 Ok(None)
25931 }
25932
25933 fn finish_ident_atom(&mut self, first: String) -> Result<Expr, ParseError> {
25934 // v7.39.2 — MySQL's charset INTRODUCER: `_utf8mb4'x'`, `N'y'`,
25935 // `_binary'z'`. All three were `ERROR 1064 syntax error` here
25936 // and all three answer the literal on MySQL 9.7.2.
25937 //
25938 // It is not only syntax, which is why it waited for
25939 // `Expr::Collate`: measured, `_binary'A' = 'a'` is 0 on MySQL
25940 // because `_binary` makes the comparison byte-wise, while
25941 // `_utf8mb4'A' = _utf8mb4'a'` is 1. Accepting the syntax and
25942 // dropping the charset would have turned a hard error into a
25943 // silently wrong comparison — worse than the error it replaced.
25944 //
25945 // An UNKNOWN charset is NOT an introducer: MySQL answers
25946 // `Unknown column '_nosuch'`, because it parses as a column
25947 // reference followed by a string. So the table decides, and it
25948 // is the same table `SET NAMES` reads.
25949 //
25950 // A space is allowed between the two (`_utf8mb4 'x'`), which
25951 // falls out of asking the token stream rather than the bytes.
25952 if self.mysql_dialect
25953 && let Token::String(_) = self.peek()
25954 {
25955 let lower = first.to_ascii_lowercase();
25956 let charset = if lower == "n" {
25957 // `N'…'` is the national character set, which MySQL
25958 // documents as utf8 — utf8mb3 in 9.7.2's spelling.
25959 //
25960 // utf8mb3 and utf8mb4 both fold case in their default
25961 // collations, so nothing SPG can be asked distinguishes
25962 // the two here: an ablation swapping this to utf8mb4
25963 // reddens no pin. Recorded rather than implied — the
25964 // spelling follows MySQL's documentation, not a
25965 // measurement.
25966 Some("utf8mb3")
25967 } else {
25968 // No filter here: the lookup below IS the test for
25969 // "is this a charset". An ablation that removed a filter
25970 // in this spot reddened nothing, which is how the two
25971 // were found to be one check written twice.
25972 lower.strip_prefix('_')
25973 };
25974 if let Some(cs) = charset
25975 && let Some(collation) = crate::charset::charset_default_collation(cs)
25976 {
25977 let Token::String(body) = self.advance() else {
25978 unreachable!("peeked a string");
25979 };
25980 return Ok(Expr::Collate {
25981 expr: Box::new(Expr::Literal(Literal::String(body))),
25982 collation: String::from(collation),
25983 });
25984 }
25985 }
25986 if matches!(self.peek(), Token::Dot) {
25987 self.advance();
25988 let name = self.expect_ident_like()?;
25989 // v7.14.0 — schema-qualified function call
25990 // `<schema>.<fn>(args)`. PG dumps emit
25991 // `pg_catalog.set_config(...)` in the preamble. SPG
25992 // is single-namespace: drop the schema prefix and
25993 // route the dispatch on the bare function name.
25994 if matches!(self.peek(), Token::LParen) {
25995 return self.finish_ident_atom(name);
25996 }
25997 return Ok(Expr::Column(ColumnName {
25998 qualifier: Some(first),
25999 name,
26000 }));
26001 }
26002 if matches!(self.peek(), Token::LParen) {
26003 self.advance();
26004 // `COUNT(*)` — special-cased here because `*` isn't a normal
26005 // expression token. Lower-case match on `first` since the lexer
26006 // folds identifiers.
26007 if first.eq_ignore_ascii_case("count") && matches!(self.peek(), Token::Star) {
26008 self.advance();
26009 if !matches!(self.peek(), Token::RParen) {
26010 return Err(self.err(format!(
26011 "expected ')' after COUNT(*), got {:?}",
26012 self.peek()
26013 )));
26014 }
26015 self.advance();
26016 // v7.32 (round-29) — `COUNT(*) FILTER (WHERE …)`.
26017 let filter = self.parse_filter_clause()?;
26018 // v4.12: COUNT(*) OVER (...) — same window tail.
26019 let null_treatment = self.parse_null_treatment_modifier();
26020 if let Token::Ident(s) | Token::QuotedIdent(s) = self.peek()
26021 && s.eq_ignore_ascii_case("over")
26022 {
26023 self.advance();
26024 let (partition_by, order_by, frame) = self.parse_over_clause()?;
26025 return Ok(Expr::WindowFunction {
26026 name: "count_star".into(),
26027 args: Vec::new(),
26028 partition_by,
26029 order_by,
26030 frame,
26031 null_treatment,
26032 filter,
26033 });
26034 }
26035 if let Some(filter) = filter {
26036 return Ok(Expr::AggregateOrdered {
26037 call: Box::new(Expr::FunctionCall {
26038 name: "count_star".into(),
26039 args: Vec::new(),
26040 }),
26041 order_by: Vec::new(),
26042 distinct: false,
26043 filter: Some(filter),
26044 });
26045 }
26046 return Ok(Expr::FunctionCall {
26047 name: "count_star".into(),
26048 args: Vec::new(),
26049 });
26050 }
26051 // Function call. PG-style: zero-or-more comma-separated args.
26052 let mut args = Vec::new();
26053 // v7.38 (read01, T14) — named-argument notation `argname => value`.
26054 // Names are collected in lock-step with `args` and resolved to
26055 // positional order after the loop (the AST stays positional).
26056 let mut arg_names: Vec<Option<String>> = Vec::new();
26057 let mut agg_order_by: Vec<OrderBy> = Vec::new();
26058 // v7.39 (round 354, M12) — whether a `SEPARATOR '<s>'` tail was
26059 // seen, so the value arguments before it can be folded.
26060 let mut saw_separator = false;
26061 // v7.25 (round-17) — `COUNT(DISTINCT x)` and friends.
26062 // v7.32 (round-29) — accept the dual `ALL` quantifier too
26063 // (the default; ORMs emit `COUNT(ALL x)` / `SUM(ALL x)`).
26064 let agg_distinct = if matches!(self.peek(), Token::Distinct) {
26065 self.advance();
26066 true
26067 } else if matches!(self.peek(), Token::All) {
26068 self.advance();
26069 false
26070 } else {
26071 false
26072 };
26073 // v7.37.17 (17.6 siblings) — MySQL TIMESTAMPADD /
26074 // TIMESTAMPDIFF take a bare unit keyword as the first
26075 // argument (MINUTE, DAY, ...), and GET_FORMAT takes a
26076 // bare type keyword (DATE / TIME / DATETIME); lower them
26077 // onto string literals so the evaluator sees plain text.
26078 if ((first.eq_ignore_ascii_case("timestampadd")
26079 || first.eq_ignore_ascii_case("timestampdiff"))
26080 && matches!(self.peek(), Token::Ident(u) if matches!(
26081 u.to_ascii_lowercase().as_str(),
26082 "microsecond" | "second" | "minute" | "hour" | "day"
26083 | "week" | "month" | "quarter" | "year"
26084 )))
26085 || (first.eq_ignore_ascii_case("get_format")
26086 && matches!(self.peek(), Token::Ident(u) if matches!(
26087 u.to_ascii_lowercase().as_str(),
26088 "date" | "time" | "datetime" | "timestamp"
26089 )))
26090 {
26091 if let Token::Ident(u) = self.peek() {
26092 args.push(Expr::Literal(Literal::String(u.to_ascii_lowercase())));
26093 }
26094 self.advance();
26095 if matches!(self.peek(), Token::Comma) {
26096 self.advance();
26097 }
26098 }
26099 // `ROW(a, b, …)` keyword constructor. Followed by a
26100 // comparison operator or [NOT] IN it joins the paren
26101 // row-constructor machinery (fieldwise parse-time
26102 // expansion); bare, it stays a `row` call the evaluator
26103 // renders as PG record text.
26104 if first.eq_ignore_ascii_case("row") {
26105 let mut row_items = Vec::new();
26106 if !matches!(self.peek(), Token::RParen) {
26107 loop {
26108 row_items.push(self.parse_expr(0)?);
26109 match self.peek() {
26110 Token::Comma => {
26111 self.advance();
26112 }
26113 Token::RParen => break,
26114 other => {
26115 return Err(self.err(format!(
26116 "expected ',' or ')' in ROW(...), got {other:?}"
26117 )));
26118 }
26119 }
26120 }
26121 }
26122 self.advance(); // ')'
26123 let comparison_follows = matches!(
26124 self.peek(),
26125 Token::Eq
26126 | Token::NotEq
26127 | Token::Lt
26128 | Token::LtEq
26129 | Token::Gt
26130 | Token::GtEq
26131 | Token::In
26132 ) || (matches!(self.peek(), Token::Not)
26133 && matches!(self.tokens.get(self.pos + 1), Some(Token::In)))
26134 || matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("overlaps"));
26135 if comparison_follows && !row_items.is_empty() {
26136 return self.parse_row_comparison_tail(row_items);
26137 }
26138 return Ok(Expr::FunctionCall {
26139 name: String::from("row"),
26140 args: row_items,
26141 });
26142 }
26143 // v7.39 (read01 xml.c) — `XMLPARSE(DOCUMENT|CONTENT expr)`:
26144 // the parse-mode keyword introduces the source text. SPG
26145 // carries XML as text, so both modes lower to __xmlparse(expr)
26146 // which validates well-formedness and returns Value::Xml.
26147 if first.eq_ignore_ascii_case("xmlparse")
26148 && matches!(self.peek(), Token::Ident(kw)
26149 if kw.eq_ignore_ascii_case("document")
26150 || kw.eq_ignore_ascii_case("content"))
26151 {
26152 let mode = match self.advance() {
26153 Token::Ident(kw) => kw.to_ascii_lowercase(),
26154 _ => unreachable!("peeked an ident"),
26155 };
26156 let src = self.parse_expr(0)?;
26157 if !matches!(self.peek(), Token::RParen) {
26158 return Err(self.err(format!(
26159 "expected ')' to close XMLPARSE, got {:?}",
26160 self.peek()
26161 )));
26162 }
26163 self.advance();
26164 return Ok(Expr::FunctionCall {
26165 name: String::from("__xmlparse"),
26166 args: alloc::vec![src, Expr::Literal(Literal::String(mode))],
26167 });
26168 }
26169 // SQL/XML `XMLELEMENT(NAME ident [, content …])` — the NAME
26170 // keyword introduces the element name (a bare or quoted
26171 // identifier), then optional content expressions. Lower to a
26172 // plain `xmlelement(name_text, content …)` call.
26173 if first.eq_ignore_ascii_case("xmlelement")
26174 && matches!(self.peek(), Token::Ident(kw) if kw.eq_ignore_ascii_case("name"))
26175 {
26176 self.advance(); // consume NAME
26177 let elem_name = match self.peek().clone() {
26178 Token::Ident(n) | Token::QuotedIdent(n) => {
26179 self.advance();
26180 n
26181 }
26182 other => {
26183 return Err(self.err(format!(
26184 "expected element name after XMLELEMENT NAME, got {other:?}"
26185 )));
26186 }
26187 };
26188 let mut args = alloc::vec![Expr::Literal(Literal::String(elem_name))];
26189 while matches!(self.peek(), Token::Comma) {
26190 self.advance();
26191 args.push(self.parse_expr(0)?);
26192 }
26193 if !matches!(self.peek(), Token::RParen) {
26194 return Err(self.err(format!(
26195 "expected ')' to close XMLELEMENT, got {:?}",
26196 self.peek()
26197 )));
26198 }
26199 self.advance();
26200 return Ok(Expr::FunctionCall {
26201 name: String::from("xmlelement"),
26202 args,
26203 });
26204 }
26205 // SQL/XML `XMLFOREST(value [AS name], …)` — each `value AS name`
26206 // becomes a `<name>value</name>` element; a bare column infers its
26207 // own name. Lower to `xmlforest(name1, val1, name2, val2, …)`.
26208 // v7.39.2 — MySQL's two CONVERT forms, neither of which parsed.
26209 // `CONVERT(expr USING cs)` was a syntax error at USING, and
26210 // `CONVERT(expr, CHAR)` was read as PostgreSQL's three-argument
26211 // `convert(bytea, src, dest)` and answered `column "char" does
26212 // not exist`. Both are casts in MySQL: measured on 9.7.2,
26213 // `CONVERT(0x41 USING utf8mb4)` and `CONVERT(0x41, CHAR)` are
26214 // both 'A', and `CONVERT(123, CHAR)` is '123'.
26215 //
26216 // The charset is checked against the same table the introducers
26217 // use, so an unknown one is refused rather than quietly ignored.
26218 if self.mysql_dialect
26219 && first.eq_ignore_ascii_case("convert")
26220 && !matches!(self.peek(), Token::RParen)
26221 {
26222 let save = self.pos;
26223 let inner = self.parse_expr(0)?;
26224 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
26225 self.advance();
26226 let cs = match self.peek().clone() {
26227 Token::Ident(n) | Token::QuotedIdent(n) => {
26228 self.advance();
26229 n
26230 }
26231 other => {
26232 return Err(self.err(alloc::format!(
26233 "expected a charset after USING, got {other:?}"
26234 )));
26235 }
26236 };
26237 let lc = cs.to_ascii_lowercase();
26238 if lc != "binary" && crate::charset::charset_default_collation(&lc).is_none() {
26239 return Err(self.err(alloc::format!("unknown character set: '{cs}'")));
26240 }
26241 if !matches!(self.peek(), Token::RParen) {
26242 return Err(self.err(alloc::format!(
26243 "expected ')' after CONVERT … USING, got {:?}",
26244 self.peek()
26245 )));
26246 }
26247 self.advance();
26248 let target = if lc == "binary" {
26249 CastTarget::Named("binary".to_string())
26250 } else {
26251 CastTarget::Text
26252 };
26253 return self.finish_postfix_casts(Expr::Cast {
26254 expr: alloc::boxed::Box::new(inner),
26255 target,
26256 });
26257 }
26258 if matches!(self.peek(), Token::Comma) {
26259 self.advance();
26260 // A type name here is MySQL's cast form; anything else
26261 // (three string arguments) is PostgreSQL's `convert`,
26262 // which keeps its own path.
26263 if let Ok(target) = self.parse_cast_target()
26264 && matches!(self.peek(), Token::RParen)
26265 {
26266 self.advance();
26267 return self.finish_postfix_casts(Expr::Cast {
26268 expr: alloc::boxed::Box::new(inner),
26269 target,
26270 });
26271 }
26272 }
26273 self.pos = save;
26274 }
26275 if first.eq_ignore_ascii_case("xmlforest") && !matches!(self.peek(), Token::RParen) {
26276 let mut args: Vec<Expr> = Vec::new();
26277 loop {
26278 let val = self.parse_expr(0)?;
26279 let name = if matches!(self.peek(), Token::As) {
26280 self.advance();
26281 match self.peek().clone() {
26282 Token::Ident(n) | Token::QuotedIdent(n) => {
26283 self.advance();
26284 n
26285 }
26286 other => {
26287 return Err(self.err(format!(
26288 "expected name after AS in XMLFOREST, got {other:?}"
26289 )));
26290 }
26291 }
26292 } else if let Expr::Column(c) = &val {
26293 c.name.clone()
26294 } else {
26295 return Err(
26296 self.err("XMLFOREST element without a column name needs AS".into())
26297 );
26298 };
26299 args.push(Expr::Literal(Literal::String(name)));
26300 args.push(val);
26301 if matches!(self.peek(), Token::Comma) {
26302 self.advance();
26303 } else {
26304 break;
26305 }
26306 }
26307 if !matches!(self.peek(), Token::RParen) {
26308 return Err(self.err(format!(
26309 "expected ')' to close XMLFOREST, got {:?}",
26310 self.peek()
26311 )));
26312 }
26313 self.advance();
26314 return Ok(Expr::FunctionCall {
26315 name: String::from("xmlforest"),
26316 args,
26317 });
26318 }
26319 // SQL-standard `POSITION(sub IN str)` — lowers onto
26320 // strpos(str, sub). IN is the argument separator here,
26321 // so the needle parses with the IN-tail suppressed.
26322 if first.eq_ignore_ascii_case("position") && !matches!(self.peek(), Token::RParen) {
26323 let saved = self.suppress_in_tail;
26324 self.suppress_in_tail = true;
26325 let needle = self.parse_expr(0);
26326 self.suppress_in_tail = saved;
26327 let needle = needle?;
26328 if matches!(self.peek(), Token::In) {
26329 self.advance();
26330 let haystack = self.parse_expr(0)?;
26331 if !matches!(self.peek(), Token::RParen) {
26332 return Err(self.err(format!(
26333 "expected ')' to close POSITION, got {:?}",
26334 self.peek()
26335 )));
26336 }
26337 self.advance();
26338 return Ok(Expr::FunctionCall {
26339 name: String::from("strpos"),
26340 args: alloc::vec![haystack, needle],
26341 });
26342 }
26343 // position(sub, str) comma form (incl. bytea) —
26344 // hand the parsed first arg to the generic list.
26345 args.push(needle);
26346 if matches!(self.peek(), Token::Comma) {
26347 self.advance();
26348 }
26349 }
26350 // SQL-standard `TRIM([BOTH|LEADING|TRAILING] [chars]
26351 // FROM str)` — lowers onto btrim / ltrim / rtrim. The
26352 // plain comma forms TRIM(str) / TRIM(str, chars) keep
26353 // riding the generic argument list below.
26354 if first.eq_ignore_ascii_case("trim") {
26355 let mode = match self.peek() {
26356 Token::Ident(k) if k.eq_ignore_ascii_case("both") => {
26357 self.advance();
26358 Some("btrim")
26359 }
26360 Token::Ident(k) if k.eq_ignore_ascii_case("leading") => {
26361 self.advance();
26362 Some("ltrim")
26363 }
26364 Token::Ident(k) if k.eq_ignore_ascii_case("trailing") => {
26365 self.advance();
26366 Some("rtrim")
26367 }
26368 _ => None,
26369 };
26370 if mode.is_some() || matches!(self.peek(), Token::From) {
26371 // TRIM([mode] FROM str) — no strip-chars.
26372 let chars = if matches!(self.peek(), Token::From) {
26373 None
26374 } else {
26375 Some(self.parse_expr(0)?)
26376 };
26377 if !matches!(self.peek(), Token::From) {
26378 return Err(self.err(format!(
26379 "expected FROM in TRIM([BOTH|LEADING|TRAILING] [chars] FROM str), got {:?}",
26380 self.peek()
26381 )));
26382 }
26383 self.advance();
26384 let target = self.parse_expr(0)?;
26385 if !matches!(self.peek(), Token::RParen) {
26386 return Err(
26387 self.err(format!("expected ')' to close TRIM, got {:?}", self.peek()))
26388 );
26389 }
26390 self.advance();
26391 let mut trim_args = alloc::vec![target];
26392 if let Some(c) = chars {
26393 trim_args.push(c);
26394 }
26395 return Ok(Expr::FunctionCall {
26396 name: String::from(mode.unwrap_or("btrim")),
26397 args: trim_args,
26398 });
26399 }
26400 }
26401 if !matches!(self.peek(), Token::RParen) {
26402 loop {
26403 // v7.38 (read01, T14) — `argname => value` names this arg.
26404 // v7.39 (read01 round 77) — `argname := value` is the same
26405 // thing, and it is the spelling PG's own docs lead with. It
26406 // was simply never lexed here, so every `f(x := 1)` died in
26407 // the parser regardless of what `f` was.
26408 let this_name = match (&self.tokens[self.pos], self.tokens.get(self.pos + 1)) {
26409 (
26410 Token::Ident(n) | Token::QuotedIdent(n),
26411 Some(Token::FatArrow | Token::ColonEq),
26412 ) => {
26413 let name = n.clone();
26414 self.advance(); // name
26415 self.advance(); // => / :=
26416 Some(name)
26417 }
26418 _ => None,
26419 };
26420 // v7.39 (read01 round 100) — `VARIADIC <array>` spreads an
26421 // array's elements into a variadic call's trailing args
26422 // (`concat_ws(',', VARIADIC ARRAY[…])`). VARIADIC isn't
26423 // reserved, so it arrives as a bare ident before the arg.
26424 let is_variadic = this_name.is_none()
26425 && matches!(&self.tokens[self.pos], Token::Ident(s) if s.eq_ignore_ascii_case("variadic"));
26426 if is_variadic {
26427 self.advance();
26428 }
26429 let arg = self.parse_expr(0)?;
26430 args.push(match &this_name {
26431 // The callee's parameter names decide the slot, and a
26432 // user function's live in the catalog. Carry the name
26433 // to eval rather than guessing here.
26434 Some(n) => Expr::NamedArg {
26435 name: n.clone(),
26436 expr: Box::new(arg),
26437 },
26438 None if is_variadic => Expr::Variadic(Box::new(arg)),
26439 None => arg,
26440 });
26441 arg_names.push(this_name);
26442 // v7.25 (round-17) — standard `CAST(expr AS type)`.
26443 // The `::` cast already worked; this lowers the
26444 // function form onto the same Expr::Cast node.
26445 if first.eq_ignore_ascii_case("cast")
26446 && args.len() == 1
26447 && matches!(self.peek(), Token::As)
26448 {
26449 self.advance();
26450 let target = self.parse_cast_target()?;
26451 if !matches!(self.peek(), Token::RParen) {
26452 return Err(self.err(format!(
26453 "expected ')' to close CAST, got {:?}",
26454 self.peek()
26455 )));
26456 }
26457 self.advance();
26458 return Ok(Expr::Cast {
26459 expr: Box::new(args.pop().expect("one arg")),
26460 target,
26461 });
26462 }
26463 // v7.38 (read01 P6.-) — `normalize(text, FORM)` where FORM is
26464 // a bare keyword NFC / NFD / NFKC / NFKD. PG parses these as
26465 // keywords; SPG's lexer makes them plain idents (so they'd be
26466 // read as column refs). Lower the keyword to the string form
26467 // the evaluator already accepts.
26468 if first.eq_ignore_ascii_case("normalize")
26469 && args.len() == 1
26470 && matches!(self.peek(), Token::Comma)
26471 {
26472 let form = match self.tokens.get(self.pos + 1) {
26473 Some(Token::Ident(f) | Token::QuotedIdent(f)) => {
26474 let up = f.to_ascii_uppercase();
26475 matches!(up.as_str(), "NFC" | "NFD" | "NFKC" | "NFKD").then_some(up)
26476 }
26477 _ => None,
26478 };
26479 if let Some(up) = form {
26480 self.advance(); // comma
26481 self.advance(); // form keyword
26482 args.push(Expr::Literal(Literal::String(up)));
26483 }
26484 }
26485 // v7.37.7 C.1.8 — PG `substring(str FROM pos FOR len)` syntactic
26486 // form. Desugars to the comma-list shape evaluator already
26487 // handles. Triggered after the first arg when the function
26488 // name is substring / substr and the next token is FROM
26489 // (a reserved keyword in PG; SPG also reserves it).
26490 // v7.39 (read01 regexp.c) — `substring(str SIMILAR pat
26491 // ESCAPE esc)` (SQL:1999 three-part form) desugars to the
26492 // internal __substring_similar(str, pat, esc) call.
26493 if (first.eq_ignore_ascii_case("substring")
26494 || first.eq_ignore_ascii_case("substr"))
26495 && args.len() == 1
26496 && matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("similar"))
26497 {
26498 self.advance(); // SIMILAR
26499 let pattern = self.parse_expr(0)?;
26500 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("escape"))
26501 {
26502 return Err(self.err(format!(
26503 "expected ESCAPE in substring(... SIMILAR ...), got {:?}",
26504 self.peek()
26505 )));
26506 }
26507 self.advance(); // ESCAPE
26508 let esc = self.parse_expr(0)?;
26509 if !matches!(self.peek(), Token::RParen) {
26510 return Err(self.err(format!(
26511 "expected ')' to close substring(... SIMILAR ...), got {:?}",
26512 self.peek()
26513 )));
26514 }
26515 self.advance();
26516 args.push(pattern);
26517 args.push(esc);
26518 return Ok(Expr::FunctionCall {
26519 name: "__substring_similar".to_string(),
26520 args,
26521 });
26522 }
26523 if (first.eq_ignore_ascii_case("substring")
26524 || first.eq_ignore_ascii_case("substr"))
26525 && args.len() == 1
26526 && matches!(self.peek(), Token::From | Token::For)
26527 {
26528 // `substring(str FROM pos [FOR len])`, or the FOR-only
26529 // `substring(str FOR len)` which PG treats as FROM 1.
26530 if matches!(self.peek(), Token::From) {
26531 self.advance();
26532 let start = self.parse_expr(0)?;
26533 args.push(start);
26534 } else {
26535 args.push(Expr::Literal(Literal::Integer(1)));
26536 }
26537 if matches!(self.peek(), Token::For) {
26538 self.advance();
26539 let length = self.parse_expr(0)?;
26540 args.push(length);
26541 }
26542 if !matches!(self.peek(), Token::RParen) {
26543 return Err(self.err(format!(
26544 "expected ')' to close substring(... FROM ... [FOR ...]), got {:?}",
26545 self.peek()
26546 )));
26547 }
26548 self.advance();
26549 return Ok(Expr::FunctionCall {
26550 name: first.to_ascii_lowercase(),
26551 args,
26552 });
26553 }
26554 // PG `overlay(str PLACING repl FROM n [FOR len])`
26555 // syntactic form. Desugars to the `overlay(str,
26556 // repl, n[, len])` comma-list shape the evaluator
26557 // already implements. `PLACING` is not a reserved
26558 // token in SPG, so it arrives as a bare Ident.
26559 if first.eq_ignore_ascii_case("overlay")
26560 && args.len() == 1
26561 && matches!(self.peek(), Token::Ident(kw) if kw == "placing")
26562 {
26563 self.advance(); // consume PLACING
26564 args.push(self.parse_expr(0)?); // replacement
26565 if !matches!(self.peek(), Token::From) {
26566 return Err(self.err(format!(
26567 "expected FROM in overlay(... PLACING ... FROM ...), got {:?}",
26568 self.peek()
26569 )));
26570 }
26571 self.advance();
26572 args.push(self.parse_expr(0)?); // start position
26573 if matches!(self.peek(), Token::For) {
26574 self.advance();
26575 args.push(self.parse_expr(0)?); // length
26576 }
26577 if !matches!(self.peek(), Token::RParen) {
26578 return Err(self.err(format!(
26579 "expected ')' to close overlay(... PLACING ... FROM ... [FOR ...]), got {:?}",
26580 self.peek()
26581 )));
26582 }
26583 self.advance();
26584 return Ok(Expr::FunctionCall {
26585 name: String::from("overlay"),
26586 args,
26587 });
26588 }
26589 // `TRIM(chars FROM str)` — the keyword-less
26590 // spelling lands here after the chars parse
26591 // (the keyword forms return earlier).
26592 if first.eq_ignore_ascii_case("trim")
26593 && args.len() == 1
26594 && matches!(self.peek(), Token::From)
26595 {
26596 self.advance();
26597 let target = self.parse_expr(0)?;
26598 if !matches!(self.peek(), Token::RParen) {
26599 return Err(self.err(format!(
26600 "expected ')' to close TRIM(chars FROM str), got {:?}",
26601 self.peek()
26602 )));
26603 }
26604 self.advance();
26605 let chars = args.pop().expect("one arg");
26606 return Ok(Expr::FunctionCall {
26607 name: String::from("btrim"),
26608 args: alloc::vec![target, chars],
26609 });
26610 }
26611 // v7.24 (round-16 A) — aggregate-internal
26612 // ordering: `array_agg(x ORDER BY y DESC NULLS
26613 // LAST)`. Keys close the argument list.
26614 if matches!(self.peek(), Token::Order) {
26615 self.advance();
26616 if !self.peek_is_by() {
26617 return Err(self.err(format!(
26618 "expected BY after ORDER in aggregate args, got {:?}",
26619 self.peek()
26620 )));
26621 }
26622 self.advance();
26623 loop {
26624 // v7.39 (round 691) — save/restore, the discipline this parser
26625 // already uses around `pending_sample_preds`, so a subquery inside
26626 // a key neither inherits nor leaks the channel.
26627 let saved_flag = core::mem::replace(&mut self.in_order_by_key, true);
26628 let saved_coll = self.order_key_collation.take();
26629 let parsed = self.parse_expr(0);
26630 self.in_order_by_key = saved_flag;
26631 let collation =
26632 core::mem::replace(&mut self.order_key_collation, saved_coll);
26633 let expr = parsed?;
26634 let desc = if matches!(self.peek(), Token::Desc) {
26635 self.advance();
26636 true
26637 } else if matches!(self.peek(), Token::Asc) {
26638 self.advance();
26639 false
26640 } else {
26641 false
26642 };
26643 let nulls_first = self.parse_optional_nulls_placement()?;
26644 agg_order_by.push(OrderBy {
26645 expr,
26646 desc,
26647 nulls_first,
26648 collation,
26649 });
26650 if matches!(self.peek(), Token::Comma) {
26651 self.advance();
26652 } else {
26653 break;
26654 }
26655 }
26656 // v7.39 (round 354, M12) — `SEPARATOR '<s>'` may
26657 // follow the ORDER BY inside GROUP_CONCAT.
26658 if self.consume_group_concat_separator(&mut args)? {
26659 saw_separator = true;
26660 }
26661 if !matches!(self.peek(), Token::RParen) {
26662 return Err(self.err(format!(
26663 "expected ')' after aggregate ORDER BY, got {:?}",
26664 self.peek()
26665 )));
26666 }
26667 break;
26668 }
26669 // v7.39 (round 354, M12) — …or directly after the
26670 // arguments (`GROUP_CONCAT(t SEPARATOR '|')`). MySQL's
26671 // own spelling of what PG passes as string_agg's second
26672 // argument; it was a parse error, so every MySQL query
26673 // that names its own separator failed outright.
26674 if self.consume_group_concat_separator(&mut args)? {
26675 saw_separator = true;
26676 break;
26677 }
26678 match self.peek() {
26679 Token::Comma => {
26680 self.advance();
26681 }
26682 Token::RParen => break,
26683 other => {
26684 return Err(self.err(format!(
26685 "expected ',' or ')' in function args, got {other:?}"
26686 )));
26687 }
26688 }
26689 }
26690 }
26691 // v7.39 (round 354, M12) — MySQL's GROUP_CONCAT concatenates
26692 // its value arguments PER ROW: `GROUP_CONCAT(n, ':', t)` is
26693 // `3:c,1:a,…` (measured on MariaDB 11), NOT a second argument
26694 // meaning a separator — that is what the explicit SEPARATOR
26695 // tail is for. Fold them into one `concat(...)` so the
26696 // aggregate keeps its single value argument.
26697 if self.mysql_dialect && first.eq_ignore_ascii_case("group_concat") {
26698 let values = args.len() - usize::from(saw_separator);
26699 if values > 1 {
26700 let sep_arg = if saw_separator { args.pop() } else { None };
26701 let folded = Expr::FunctionCall {
26702 name: "concat".to_string(),
26703 args: core::mem::take(&mut args),
26704 };
26705 args.push(folded);
26706 if let Some(sep) = sep_arg {
26707 args.push(sep);
26708 }
26709 }
26710 }
26711 self.advance(); // consume ')'
26712 // v7.39 (read01 round 77) — named arguments are NOT reordered here
26713 // any more. The parser has no catalog, so it could only ever resolve
26714 // the handful of `make_*` builtins whose parameter names were baked
26715 // into a table right here — every user function got
26716 // "does not support named arguments", though the catalog has been
26717 // storing its parameter names all along. Reordering happens in eval,
26718 // in one place, for builtins and user functions alike.
26719 // v7.32 (round-29) — ordered-set aggregate tail
26720 // `name(direct_args) WITHIN GROUP (ORDER BY …)`
26721 // (percentile_cont / percentile_disc / mode). The sort spec
26722 // lands in the same `order_by` slot a decorated aggregate
26723 // uses; the executor dispatches on the function name. WITHIN
26724 // GROUP and an intra-argument ORDER BY are mutually
26725 // exclusive (PG rejects both).
26726 let within_group_order = self.parse_within_group_clause()?;
26727 if !within_group_order.is_empty() && !agg_order_by.is_empty() {
26728 return Err(self.err(
26729 "an aggregate may not carry both an in-argument ORDER BY and WITHIN GROUP"
26730 .into(),
26731 ));
26732 }
26733 let within_group_seen = !within_group_order.is_empty();
26734 let agg_order_by = if within_group_order.is_empty() {
26735 agg_order_by
26736 } else {
26737 within_group_order
26738 };
26739 // v7.32 (round-29) — `name(args) FILTER (WHERE …)`.
26740 let filter = self.parse_filter_clause()?;
26741 // v4.12: window-function tail — `name(args) OVER (...)`.
26742 // Promotes the just-parsed FunctionCall into a
26743 // WindowFunction node carrying partition + order.
26744 // v6.4.2: also accepts `name(args) IGNORE NULLS OVER (...)`
26745 // / `RESPECT NULLS OVER (...)` between the closing paren
26746 // and `OVER`.
26747 let null_treatment = self.parse_null_treatment_modifier();
26748 if let Token::Ident(s) | Token::QuotedIdent(s) = self.peek()
26749 && s.eq_ignore_ascii_case("over")
26750 {
26751 self.advance();
26752 // v7.39 (round 230) — PG implements neither modifier for a
26753 // windowed call and says so (0A000). Both used to be parsed
26754 // and then silently dropped here, so `count(DISTINCT v)
26755 // OVER (…)` quietly answered the non-distinct count.
26756 if agg_distinct {
26757 return Err(
26758 self.err("DISTINCT is not implemented for window functions".to_string())
26759 );
26760 }
26761 if !agg_order_by.is_empty() {
26762 // PG separates the two shapes that land here: a
26763 // WITHIN GROUP call is an ordered-set aggregate and gets
26764 // its own message naming the aggregate; a plain
26765 // `agg(x ORDER BY y)` gets the generic one.
26766 let msg = if within_group_seen {
26767 alloc::format!("OVER is not supported for ordered-set aggregate {first}")
26768 } else {
26769 "aggregate ORDER BY is not implemented for window functions".to_string()
26770 };
26771 return Err(self.err(msg));
26772 }
26773 let (partition_by, order_by, frame) = self.parse_over_clause()?;
26774 return Ok(Expr::WindowFunction {
26775 name: first,
26776 args,
26777 partition_by,
26778 order_by,
26779 frame,
26780 null_treatment,
26781 filter,
26782 });
26783 }
26784 if !agg_order_by.is_empty() || agg_distinct || filter.is_some() {
26785 return Ok(Expr::AggregateOrdered {
26786 call: Box::new(Expr::FunctionCall { name: first, args }),
26787 order_by: agg_order_by,
26788 distinct: agg_distinct,
26789 filter,
26790 });
26791 }
26792 // v7.39 (round 522) — PG declares `date_add` / `date_subtract`
26793 // over TIMESTAMPTZ and has no timestamp overload, so a
26794 // timestamp argument is coerced on the way in and the answer
26795 // is timestamptz — measured: `pg_typeof(date_add(TIMESTAMP
26796 // '2020-01-01', INTERVAL '1 hour'))` is `timestamp with time
26797 // zone`. SPG answered `timestamp without time zone`, dropping
26798 // the offset from every rendering.
26799 //
26800 // Writing the coercion PG performs makes the existing
26801 // argument-driven typing (the one `date_trunc` uses) reach the
26802 // right answer, rather than teaching the type layer a second
26803 // rule. MySQL's DATE_ADD is a different function that returns
26804 // DATE or DATETIME, so this is PG-dialect only.
26805 //
26806 // Out-of-line because this sits on the RECURSIVE descent
26807 // frame: an inline block with locals here costs every nesting
26808 // level, and the suite's deep-nesting sentinel overflowed the
26809 // 512 KiB parser stack the moment one was added (round 430's
26810 // lesson, in the same shape).
26811 if !self.mysql_dialect {
26812 lift_date_add_arg_to_timestamptz(&first, &mut args);
26813 }
26814 return Ok(Expr::FunctionCall { name: first, args });
26815 }
26816 // v7.9.20 — SQL-standard parenless keyword expressions
26817 // (PG treats these as functions called without parens).
26818 // Resolve to a synthetic FunctionCall so the engine's
26819 // eval path reuses the existing function-call routing.
26820 // mailrs G3.
26821 let lc = first.to_ascii_lowercase();
26822 if matches!(
26823 lc.as_str(),
26824 "current_date"
26825 | "current_time"
26826 | "current_timestamp"
26827 | "localtimestamp"
26828 | "localtime"
26829 // v7.37.17 (17.6 siblings) — session-identity SQL-
26830 // standard parenless keywords. current_user /
26831 // session_user / user were already caught by the
26832 // pgwire canned-response shortcut but bare-select
26833 // in the embedded engine went through Expr::Column
26834 // and errored. Adding them here so the parser
26835 // resolves to a synthetic FunctionCall that reuses
26836 // the existing eval/functions.rs dispatch.
26837 | "current_user"
26838 | "session_user"
26839 | "current_role"
26840 | "current_catalog"
26841 | "current_schema"
26842 | "current_database"
26843 // v7.39 (read01 round 51) — PG 16's system_user is parenless too.
26844 | "system_user"
26845 ) {
26846 return Ok(Expr::FunctionCall {
26847 name: lc,
26848 args: Vec::new(),
26849 });
26850 }
26851 Ok(Expr::Column(ColumnName {
26852 qualifier: None,
26853 name: first,
26854 }))
26855 }
26856}
26857
26858/// v7.39 (round 522) — write the coercion PG's `date_add` /
26859/// `date_subtract` signature performs.
26860///
26861/// PG declares both over TIMESTAMPTZ and has no timestamp overload, so a
26862/// timestamp argument is cast on the way in and the answer is
26863/// timestamptz — measured: `pg_typeof(date_add(TIMESTAMP '2020-01-01',
26864/// INTERVAL '1 hour'))` is `timestamp with time zone`. SPG answered
26865/// `timestamp without time zone`, dropping the offset from every
26866/// rendering of the result.
26867///
26868/// Writing the cast the signature implies lets the existing
26869/// argument-driven typing (the one `date_trunc` uses) reach the right
26870/// answer instead of teaching the type layer a second rule. MySQL's
26871/// DATE_ADD is a different function returning DATE or DATETIME, so the
26872/// caller applies this in PG dialect only.
26873///
26874/// A free function, and not a block at the call site, because the caller
26875/// is on the recursive-descent frame chain.
26876#[inline(never)]
26877fn lift_date_add_arg_to_timestamptz(name: &str, args: &mut alloc::vec::Vec<Expr>) {
26878 if args.len() != 2
26879 || !(name.eq_ignore_ascii_case("date_add") || name.eq_ignore_ascii_case("date_subtract"))
26880 {
26881 return;
26882 }
26883 let base = args.remove(0);
26884 args.insert(
26885 0,
26886 Expr::Cast {
26887 expr: Box::new(base),
26888 target: CastTarget::Timestamptz,
26889 },
26890 );
26891}
26892
26893/// v6.8.2 — walk an expression tree and return the first column
26894/// reference's bare name. Used by `parse_create_index_stmt_after_create`
26895/// to derive `CreateIndexStatement.column` from an expression
26896/// key (so downstream planner code resolving a primary column
26897/// position keeps working with expression indexes). Returns
26898/// `None` when the expression has no column ref at all — caller
26899/// surfaces that as a parse error.
26900fn extract_first_column(expr: &Expr) -> Option<String> {
26901 match expr {
26902 Expr::Column(cn) => Some(cn.name.clone()),
26903 Expr::FunctionCall { args, .. } => args.iter().find_map(extract_first_column),
26904 Expr::Binary { lhs, rhs, .. } => {
26905 extract_first_column(lhs).or_else(|| extract_first_column(rhs))
26906 }
26907 Expr::Unary { expr: e, .. } => extract_first_column(e),
26908 // v7.39 (read01 round 93) — a cast wraps its operand: a common
26909 // expression-index key is `lower(col::text)`, where the column
26910 // sits under the `::text` cast inside the function arg. Without
26911 // descending here the key was rejected as "references no column".
26912 Expr::Cast { expr: e, .. } => extract_first_column(e),
26913 // v7.39.2 — and a COLLATE wraps its operand the same way.
26914 // `CREATE INDEX rc ON t (c COLLATE "C" DESC)` stopped naming a
26915 // column the moment the clause became a node instead of being
26916 // absorbed, and the key was rejected as referencing none. This
26917 // is the shape the wildcard below silently produces, which is
26918 // why it is spelled out.
26919 Expr::Collate { expr: e, .. } => extract_first_column(e),
26920 _ => None,
26921 }
26922}
26923
26924fn maybe_not(expr: Expr, negated: bool) -> Expr {
26925 if negated {
26926 Expr::Unary {
26927 op: UnOp::Not,
26928 expr: Box::new(expr),
26929 }
26930 } else {
26931 expr
26932 }
26933}
26934
26935/// v7.39 (round 353, M9/M10) — three operator TOKENS mean different
26936/// things in the two dialects, and SPG read all three PG's way:
26937///
26938/// | token | PG (and SPG) | MySQL, measured |
26939/// |---|---|---|
26940/// | `\|\|` | string concatenation | **OR** — `1 \|\| 0` is 1, not '10' |
26941/// | `&&` | inet / array overlap | **AND** |
26942/// | `<=>` | pgvector cosine distance | **NULL-safe equal** |
26943///
26944/// `1 || 0` answering the string '10' on a MySQL session is a wrong
26945/// answer with no error, which is why they are routed here rather than
26946/// left to the shared table.
26947impl Parser {
26948 fn binop_here(&self, tok: &Token) -> Option<(BinOp, u8)> {
26949 if self.mysql_dialect {
26950 // v7.39 (round 353, M9) — `DIV` is MySQL's truncating integer
26951 // division (`5 DIV 2` is 2, `-7 DIV 2` is -3 — toward zero —
26952 // and `5 DIV 0` is NULL). It is a plain ident to the lexer.
26953 if let Token::Ident(w) = tok
26954 && w.eq_ignore_ascii_case("div")
26955 {
26956 return Some((BinOp::IntDiv, 8));
26957 }
26958 // v7.39 (round 394) — `MOD` is MySQL's modulo operator, a synonym
26959 // for `%` (`10 MOD 3` is 1, `5.5 MOD 2` is 1.5). A plain ident to
26960 // the lexer; the `MOD(x, y)` function form is unaffected (MOD
26961 // there sits in operand position, not infix).
26962 if let Token::Ident(w) = tok
26963 && w.eq_ignore_ascii_case("mod")
26964 {
26965 return Some((BinOp::Mod, 8));
26966 }
26967 // v7.39 (round 407) — `XOR` is MySQL's logical exclusive-or, a
26968 // plain ident to the lexer. Its precedence sits between OR (1)
26969 // and AND (3) — hence rung 2, the slot freed by moving AND up.
26970 if let Token::Ident(w) = tok
26971 && w.eq_ignore_ascii_case("xor")
26972 {
26973 return Some((BinOp::LogicalXor, 2));
26974 }
26975 match tok {
26976 Token::Concat => return Some((BinOp::Or, 1)),
26977 // MySQL's `&&` is logical AND, sharing AND's rung (3).
26978 Token::InetOverlap => return Some((BinOp::And, 3)),
26979 // MySQL's `<=>` is NULL-safe equal, at the comparison rung (5).
26980 Token::CosineDistance => return Some((BinOp::IsNotDistinctFrom, 5)),
26981 _ => {}
26982 }
26983 }
26984 binop_from(tok)
26985 }
26986}
26987
26988// v7.39 (round 407) — precedence ladder. To open a rung for MySQL's `XOR`
26989// (which sits strictly between OR and AND), every level from AND upward was
26990// shifted +1: the ladder is now OR=1, XOR=2, AND=3, IS=4, comparison=5,
26991// distance=6, additive/concat/bitwise=7, multiplicative/JSON=8, prefix=9.
26992// XOR only exists in the MySQL dialect (binop_here); PG never sees it, and
26993// the *relative* order of every PG operator is unchanged by the shift.
26994fn binop_from(tok: &Token) -> Option<(BinOp, u8)> {
26995 let pair = match tok {
26996 Token::Or => (BinOp::Or, 1),
26997 Token::And => (BinOp::And, 3),
26998 Token::Eq => (BinOp::Eq, 5),
26999 Token::NotEq => (BinOp::NotEq, 5),
27000 Token::Lt => (BinOp::Lt, 5),
27001 Token::LtEq => (BinOp::LtEq, 5),
27002 Token::Gt => (BinOp::Gt, 5),
27003 Token::GtEq => (BinOp::GtEq, 5),
27004 // pgvector distance ops all sit on the same rung — tighter than
27005 // comparisons (5) so `col <-> v < threshold` parses correctly.
27006 Token::L2Distance => (BinOp::L2Distance, 6),
27007 // v7.39 (read01 geo_ops.c) — geometric predicates ride the
27008 // comparison rung.
27009 Token::GeomParallel => (BinOp::GeomParallel, 5),
27010 // v7.39 (read01 rangetypes.c) — range `&<` / `&>` on the
27011 // comparison rung.
27012 Token::OverLeft => (BinOp::OverLeft, 5),
27013 Token::OverRight => (BinOp::OverRight, 5),
27014 Token::GeomPerp => (BinOp::GeomPerp, 5),
27015 Token::GeomSameAs => (BinOp::GeomSameAs, 5),
27016 Token::ClosestPoint => (BinOp::ClosestPoint, 6),
27017 Token::GeomHoriz => (BinOp::GeomHoriz, 5),
27018 Token::InnerProduct => (BinOp::InnerProduct, 6),
27019 Token::CosineDistance => (BinOp::CosineDistance, 6),
27020 Token::Plus => (BinOp::Add, 7),
27021 Token::Minus => (BinOp::Sub, 7),
27022 // v7.39 (round 760, F31-B1) — the generic-operator rung. PG
27023 // binds every "other" operator (`||`, `|`, `&`, `#`, the
27024 // pgvector distances above) BETWEEN additive (7) and the
27025 // comparisons (5): `'a' || 1 + 1` is `'a' || 2` → `a2`,
27026 // `a & b + 1` is `a & (b + 1)`, and `flags & $1 = 0` stays
27027 // `(flags & $1) = 0`. They shared rung 7 with `+ -` since v1
27028 // ("matches PG conceptually" — the round-753 audit measured it
27029 // false; the old rung errored on `'a' || 1 + 1` with
27030 // `text + integer`). Same-level chains left-fold, as PG does.
27031 Token::Concat => (BinOp::Concat, 6),
27032 Token::Pipe => (BinOp::BitOr, 6),
27033 Token::Amp => (BinOp::BitAnd, 6),
27034 Token::Star => (BinOp::Mul, 8),
27035 Token::Slash => (BinOp::Div, 8),
27036 Token::Percent => (BinOp::Mod, 8),
27037 // v4.14: JSON path ops bind tighter than comparisons (5)
27038 // and additive (7) so `doc->'k' = 'v'` parses correctly.
27039 // Same rung as the multiplicative ops.
27040 Token::JsonGet => (BinOp::JsonGet, 8),
27041 Token::JsonGetText => (BinOp::JsonGetText, 8),
27042 Token::JsonGetPath => (BinOp::JsonGetPath, 8),
27043 Token::JsonGetPathText => (BinOp::JsonGetPathText, 8),
27044 Token::JsonContains => (BinOp::JsonContains, 8),
27045 Token::JsonPathExists => (BinOp::JsonPathExists, 8),
27046 Token::JsonContainedBy => (BinOp::JsonContainedBy, 8),
27047 Token::JsonKeyExists => (BinOp::JsonKeyExists, 8),
27048 Token::JsonKeysAny => (BinOp::JsonKeysAny, 8),
27049 Token::JsonKeysAll => (BinOp::JsonKeysAll, 8),
27050 Token::JsonDeletePath => (BinOp::JsonDeletePath, 8),
27051 // v7.12.2 — `@@` binds at the comparison rung (looser than
27052 // arithmetic, tighter than AND / OR). PG places `@@` at
27053 // the same precedence as `=` / `<`, so we follow.
27054 Token::TsMatch => (BinOp::TsMatch, 5),
27055 // v7.17.0 Phase 3.P0-47 — PG INET / CIDR containment + overlap.
27056 // PG places these at the comparison rung (same level as `=`),
27057 // so we follow.
27058 Token::InetContainedBy => (BinOp::InetContainedBy, 5),
27059 Token::InetContainedByEq => (BinOp::InetContainedByEq, 5),
27060 Token::InetContains => (BinOp::InetContains, 5),
27061 Token::InetContainsEq => (BinOp::InetContainsEq, 5),
27062 Token::InetOverlap => (BinOp::InetOverlap, 5),
27063 // v7.39 (round 508) — the geometric and pattern-order predicates
27064 // ride the comparison rung, as every other predicate does.
27065 Token::Intersects => (BinOp::Intersects, 5),
27066 Token::IsBelow => (BinOp::IsBelow, 5),
27067 Token::IsAbove => (BinOp::IsAbove, 5),
27068 Token::PatternLt => (BinOp::PatternLt, 5),
27069 Token::PatternLtEq => (BinOp::PatternLtEq, 5),
27070 Token::PatternGt => (BinOp::PatternGt, 5),
27071 Token::PatternGtEq => (BinOp::PatternGtEq, 5),
27072 // `@@@` is the old spelling of `@@` and means exactly it.
27073 Token::TsMatchOld => (BinOp::TsMatch, 5),
27074 _ => return None,
27075 };
27076 Some(pair)
27077}
27078
27079#[allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
27080// `as f32` here is intentional: vector elements widen / narrow into f32 on
27081// purpose. i64 → f32 loses precision past 2^24, f64 → f32 loses precision
27082// past ~15 decimal digits — both are acceptable for a fixed-precision
27083// pgvector column.
27084/// v7.17.0 Phase 1.3 — words that would otherwise be eaten as an
27085/// implicit table alias and break trailing clauses. WITH lands
27086/// here so `… FROM t WITH NO DATA` doesn't consume WITH as the
27087/// alias for `t`; same for ON / WHERE / HAVING / GROUP / ORDER /
27088/// LIMIT / OFFSET / UNION / EXCEPT / INTERSECT / RETURNING / SET
27089/// / VALUES / FOR / LATERAL — all of which would otherwise be
27090/// silently swallowed by `parse_optional_alias`.
27091fn is_alias_stopword(s: &str) -> bool {
27092 matches!(
27093 s.to_ascii_lowercase().as_str(),
27094 "with"
27095 | "on"
27096 | "where"
27097 | "having"
27098 | "group"
27099 | "order"
27100 | "limit"
27101 | "offset"
27102 | "union"
27103 | "except"
27104 | "intersect"
27105 | "returning"
27106 | "set"
27107 | "values"
27108 | "for"
27109 | "window"
27110 | "tablesample"
27111 | "lateral"
27112 | "left"
27113 | "right"
27114 | "inner"
27115 | "outer"
27116 | "full"
27117 | "cross"
27118 | "join"
27119 | "natural"
27120 | "using"
27121 | "fetch"
27122 )
27123}
27124
27125fn extract_numeric_literal(e: &Expr) -> Option<f32> {
27126 match e {
27127 Expr::Literal(Literal::Integer(n)) => Some(*n as f32),
27128 Expr::Literal(Literal::Float(x)) => Some(*x as f32),
27129 // v7.38 (read01) — a dotted literal is now NUMERIC, so a vector element
27130 // like `2.5` arrives as Literal::Numeric; widen it into f32. (`no_std`,
27131 // so scale the divisor by hand instead of `f32::powi`.)
27132 Expr::Literal(Literal::Numeric { unscaled, scale }) => {
27133 let mut div = 1.0f32;
27134 for _ in 0..*scale {
27135 div *= 10.0;
27136 }
27137 Some(*unscaled as f32 / div)
27138 }
27139 Expr::Unary {
27140 op: UnOp::Neg,
27141 expr,
27142 } => extract_numeric_literal(expr).map(|x| -x),
27143 _ => None,
27144 }
27145}
27146
27147/// Parse the text inside `INTERVAL '...'` into `(months, micros)`. Accepts
27148/// one or more `<n> <unit>` pairs separated by whitespace. `<n>` may be
27149/// negative. Returns `None` if any pair fails to parse or no pair is found.
27150///
27151/// Recognised units (case-insensitive, optional trailing `s`):
27152/// `microsecond`, `millisecond`, `second`, `minute`, `hour`, `day`, `week`,
27153/// `month`, `year`. `week` widens to 7 days; `year` widens to 12 months.
27154/// v7.37.5 β — returns `(months, days, micros)`. `days` is preserved
27155/// as its own dimension so `INTERVAL '1 day'` ≠ `INTERVAL '24 hours'`
27156/// (PG-canonical: DST and month-boundary semantics depend on this).
27157/// `week` rolls into `days` (× 7). Sub-day units flow into `micros`.
27158/// ISO 8601 duration input for INTERVAL: `P1Y2M3DT4H5M6S`. Before the `T`,
27159/// `M` is months; after it, `M` is minutes. Returns `(months, days, micros)`.
27160#[allow(clippy::cast_possible_truncation)]
27161fn parse_iso8601_interval(rest: &str) -> Option<(i32, i32, i64)> {
27162 let mut months: i64 = 0;
27163 let mut days: i64 = 0;
27164 let mut micros: i64 = 0;
27165 let mut in_time = false;
27166 let mut num = alloc::string::String::new();
27167 for ch in rest.chars() {
27168 if ch.is_ascii_digit() || ch == '.' || ch == '-' || ch == '+' {
27169 num.push(ch);
27170 continue;
27171 }
27172 if ch == 'T' || ch == 't' {
27173 if !num.is_empty() {
27174 return None;
27175 }
27176 in_time = true;
27177 continue;
27178 }
27179 let n: f64 = num.parse().ok()?;
27180 num.clear();
27181 match (ch, in_time) {
27182 ('Y' | 'y', false) => months += (n * 12.0) as i64,
27183 ('M', false) => months += n as i64,
27184 ('W' | 'w', false) => days += (n * 7.0) as i64,
27185 ('D' | 'd', false) => days += n as i64,
27186 ('H' | 'h', true) => micros += (n * 3_600_000_000.0) as i64,
27187 ('M', true) => micros += (n * 60_000_000.0) as i64,
27188 ('S' | 's', true) => micros += (n * 1_000_000.0) as i64,
27189 _ => return None,
27190 }
27191 }
27192 if !num.is_empty() {
27193 return None;
27194 }
27195 Some((
27196 i32::try_from(months).ok()?,
27197 i32::try_from(days).ok()?,
27198 micros,
27199 ))
27200}
27201
27202/// PG year-month shorthand for INTERVAL: `1-2` = 1 year 2 mons (an optional
27203/// leading `-` negates the whole value). Rejects date-like strings.
27204fn parse_year_month_interval(s: &str) -> Option<(i32, i32, i64)> {
27205 let (neg, body) = match s.strip_prefix('-') {
27206 Some(b) => (true, b),
27207 None => (false, s),
27208 };
27209 let (y, m) = body.split_once('-')?;
27210 let years: i32 = y.parse().ok()?;
27211 let mons: i32 = m.parse().ok()?;
27212 if years < 0 || mons < 0 {
27213 return None;
27214 }
27215 let total = years.checked_mul(12)?.checked_add(mons)?;
27216 Some((if neg { -total } else { total }, 0, 0))
27217}
27218
27219/// Parse a clock-time interval token `HH:MM[:SS[.ffffff]]` (optionally signed)
27220/// into microseconds. Used for the `3 days 14:30:45` / bare `14:30:45` forms.
27221fn parse_interval_clock(tok: &str) -> Option<i64> {
27222 let (neg, body) = match tok.strip_prefix('-') {
27223 Some(r) => (true, r),
27224 None => (false, tok.strip_prefix('+').unwrap_or(tok)),
27225 };
27226 let mut it = body.split(':');
27227 let h: i64 = it.next()?.parse().ok()?;
27228 let m: i64 = it.next()?.parse().ok()?;
27229 let s_tok = it.next().unwrap_or("0");
27230 if it.next().is_some() {
27231 return None;
27232 }
27233 let sec_us: i64 = if let Some((sec, frac)) = s_tok.split_once('.') {
27234 let sec: i64 = sec.parse().ok()?;
27235 let mut f = alloc::string::String::from(frac);
27236 while f.len() < 6 {
27237 f.push('0');
27238 }
27239 f.truncate(6);
27240 let fus: i64 = f.parse().ok()?;
27241 sec.checked_mul(1_000_000)?.checked_add(fus)?
27242 } else {
27243 s_tok.parse::<i64>().ok()?.checked_mul(1_000_000)?
27244 };
27245 let total = h
27246 .checked_mul(3_600_000_000)?
27247 .checked_add(m.checked_mul(60_000_000)?)?
27248 .checked_add(sec_us)?;
27249 Some(if neg { -total } else { total })
27250}
27251
27252/// v7.39 (read01 round 77) — one canonical name per interval unit, covering
27253/// every spelling PG accepts (measured against live PG18.4, not guessed):
27254/// `min` / `mins` / `m` are minutes, `mon` / `mons` are months, `y` is years.
27255/// Before this, the unit table matched long names only, with an ad-hoc
27256/// `strip_suffix('s')` in front of it — so `'15 min'` (and `hrs`, `secs`,
27257/// `yrs`, every abbreviation anyone actually types) was "cannot parse as
27258/// INTERVAL", and it had also grown arms for the debris that stripping leaves
27259/// behind (`centurie`, `millenniu`). Two parallel unit matches (integer and
27260/// fractional) both read from this one table now.
27261fn canonical_interval_unit(raw: &str) -> Option<&'static str> {
27262 let u = raw.to_ascii_lowercase();
27263 Some(match u.as_str() {
27264 "microsecond" | "microseconds" | "us" | "usec" | "usecs" | "usecond" | "useconds" => {
27265 "microsecond"
27266 }
27267 "millisecond" | "milliseconds" | "ms" | "msec" | "msecs" | "msecond" | "mseconds" => {
27268 "millisecond"
27269 }
27270 "second" | "seconds" | "sec" | "secs" | "s" => "second",
27271 "minute" | "minutes" | "min" | "mins" | "m" => "minute",
27272 "hour" | "hours" | "hr" | "hrs" | "h" => "hour",
27273 "day" | "days" | "d" => "day",
27274 "week" | "weeks" | "w" => "week",
27275 "month" | "months" | "mon" | "mons" => "month",
27276 "year" | "years" | "yr" | "yrs" | "y" => "year",
27277 "decade" | "decades" | "dec" | "decs" => "decade",
27278 "century" | "centuries" | "cent" | "c" => "century",
27279 "millennium" | "millenniums" | "millennia" | "mil" | "mils" => "millennium",
27280 _ => return None,
27281 })
27282}
27283
27284/// v7.39 (read01 round 102) — the six SQL-standard interval fields that can
27285/// qualify an `INTERVAL '…' <FIELD> [TO <FIELD>]` literal.
27286#[derive(Debug, Clone, Copy, PartialEq, Eq)]
27287pub(crate) enum IntervalField {
27288 Year,
27289 Month,
27290 Day,
27291 Hour,
27292 Minute,
27293 Second,
27294}
27295
27296/// Recognise an interval field keyword (bare ident, case-insensitive). Plural
27297/// spellings aren't standard for the qualifier position, so only the singular
27298/// forms are accepted.
27299/// v7.39 (round 350, M7) — MySQL's interval units, measured against
27300/// MariaDB 11. QUARTER is three months and WEEK seven days; MICROSECOND
27301/// is the finest. (The compound spellings — `DAY_HOUR` and friends, which
27302/// take a `'1 2'` style literal — are not read here; they stay a parse
27303/// error rather than being silently misread.)
27304/// v7.39 (round 430) — lower a `@name` / `@@name` reference to its call.
27305///
27306/// ONE `@` is a MySQL USER variable: its own per-session namespace, nothing
27307/// to do with a `@@` engine setting, and an unset one reads NULL rather
27308/// than raising. (The parser used to strip every `@`, so `@x` and `@@x`
27309/// were the same node and `SELECT @x` answered "Unknown system variable".)
27310/// For `@@`, the `session.` / `global.` scope is KEPT: a global read must
27311/// not see a session override — measured, after `SET autocommit=0`,
27312/// `@@global.autocommit` is still 1.
27313///
27314/// Out-of-line and NOT a method: `parse_atom` is the giant recursive frame
27315/// the parser's nesting budget is tuned against, and building these
27316/// `String` + `Vec` locals inside it overflowed the guard's stack (the same
27317/// wall `parse_left_right_atom` and friends were factored out for).
27318#[inline(never)]
27319fn variable_ref_atom(raw: &str) -> Expr {
27320 let user_var = !raw.starts_with("@@");
27321 let bare = raw.trim_start_matches('@').to_ascii_lowercase();
27322 Expr::FunctionCall {
27323 name: String::from(if user_var {
27324 "__spg_user_var"
27325 } else {
27326 "__spg_session_var"
27327 }),
27328 args: alloc::vec![Expr::Literal(Literal::String(bare))],
27329 }
27330}
27331
27332fn mysql_interval_unit(tok: &Token) -> Option<&'static str> {
27333 let Token::Ident(s) = tok else { return None };
27334 Some(match () {
27335 () if s.eq_ignore_ascii_case("microsecond") => "microsecond",
27336 () if s.eq_ignore_ascii_case("second") => "second",
27337 () if s.eq_ignore_ascii_case("minute") => "minute",
27338 () if s.eq_ignore_ascii_case("hour") => "hour",
27339 () if s.eq_ignore_ascii_case("day") => "day",
27340 () if s.eq_ignore_ascii_case("week") => "week",
27341 () if s.eq_ignore_ascii_case("month") => "month",
27342 () if s.eq_ignore_ascii_case("quarter") => "quarter",
27343 () if s.eq_ignore_ascii_case("year") => "year",
27344 () => return None,
27345 })
27346}
27347
27348/// v7.39 (round 422) — lower `INTERVAL <expr> <unit>` onto the existing
27349/// `make_interval(years, months, weeks, days, hours, mins, secs)` builtin,
27350/// which constructs the value at run time. Only the slot the unit names
27351/// carries the quantity; QUARTER and MICROSECOND scale it into the nearest
27352/// slot the builtin has (months and fractional seconds respectively).
27353fn make_interval_call(qty: Expr, unit: &str) -> Expr {
27354 let zero = || Expr::Literal(Literal::Integer(0));
27355 let scaled = |op: crate::ast::BinOp, by: Expr| Expr::Binary {
27356 lhs: alloc::boxed::Box::new(qty.clone()),
27357 op,
27358 rhs: alloc::boxed::Box::new(by),
27359 };
27360 // (years, months, weeks, days, hours, mins, secs)
27361 let mut args = alloc::vec![zero(), zero(), zero(), zero(), zero(), zero(), zero()];
27362 match unit {
27363 "year" => args[0] = qty,
27364 "quarter" => {
27365 args[1] = scaled(crate::ast::BinOp::Mul, Expr::Literal(Literal::Integer(3)));
27366 }
27367 "month" => args[1] = qty,
27368 "week" => args[2] = qty,
27369 "day" => args[3] = qty,
27370 "hour" => args[4] = qty,
27371 "minute" => args[5] = qty,
27372 "second" => args[6] = qty,
27373 // The builtin's seconds slot takes a fraction, so microseconds ride
27374 // it scaled down; the divisor is a NUMERIC literal so the division
27375 // stays exact rather than going through a float.
27376 "microsecond" => {
27377 args[6] = scaled(
27378 crate::ast::BinOp::Div,
27379 Expr::Literal(Literal::Numeric {
27380 unscaled: 1_000_000,
27381 scale: 0,
27382 }),
27383 );
27384 }
27385 _ => args[3] = qty,
27386 }
27387 Expr::FunctionCall {
27388 name: alloc::string::String::from("make_interval"),
27389 args,
27390 }
27391}
27392
27393/// `(count, unit)` → `(months, days, micros)`.
27394fn scale_mysql_interval(count: &str, unit: &str) -> Option<(i32, i32, i64)> {
27395 let n: i64 = count.trim().parse().ok()?;
27396 Some(match unit {
27397 "microsecond" => (0, 0, n),
27398 "second" => (0, 0, n.checked_mul(1_000_000)?),
27399 "minute" => (0, 0, n.checked_mul(60_000_000)?),
27400 "hour" => (0, 0, n.checked_mul(3_600_000_000)?),
27401 "day" => (0, i32::try_from(n).ok()?, 0),
27402 "week" => (0, i32::try_from(n.checked_mul(7)?).ok()?, 0),
27403 "month" => (i32::try_from(n).ok()?, 0, 0),
27404 "quarter" => (i32::try_from(n.checked_mul(3)?).ok()?, 0, 0),
27405 "year" => (i32::try_from(n.checked_mul(12)?).ok()?, 0, 0),
27406 _ => return None,
27407 })
27408}
27409
27410fn interval_field_of(tok: &Token) -> Option<IntervalField> {
27411 let Token::Ident(s) = tok else { return None };
27412 Some(match () {
27413 () if s.eq_ignore_ascii_case("year") => IntervalField::Year,
27414 () if s.eq_ignore_ascii_case("month") => IntervalField::Month,
27415 () if s.eq_ignore_ascii_case("day") => IntervalField::Day,
27416 () if s.eq_ignore_ascii_case("hour") => IntervalField::Hour,
27417 () if s.eq_ignore_ascii_case("minute") => IntervalField::Minute,
27418 () if s.eq_ignore_ascii_case("second") => IntervalField::Second,
27419 () => return None,
27420 })
27421}
27422
27423/// v7.39 (read01 round 102) — interpret an interval literal under a field
27424/// qualifier. Returns `(months, days, micros)`.
27425///
27426/// * A single field applied to a bare number sets which unit the number means,
27427/// truncated to that field's precision (`INTERVAL '1.5' HOUR` → `01:00:00`);
27428/// SECOND keeps its fraction (`'90.5' SECOND` → `00:01:30.5`).
27429/// * `YEAR TO MONTH` reads the `Y-M` form (`'1-6'` → 1 year 6 months).
27430/// * Every other range, and any literal a single field can't read as a plain
27431/// number (`'2 days' DAY`), falls back to the unqualified parse — SPG's
27432/// interval-text parser already reads the `D H:MM:SS` / `H:MM` forms exactly
27433/// like PG, and the qualifier there only bounds precision.
27434fn interpret_qualified_interval(
27435 text: &str,
27436 (f1, f2): (IntervalField, Option<IntervalField>),
27437) -> Option<(i32, i32, i64)> {
27438 if let Some(f2) = f2 {
27439 if f1 == IntervalField::Year && f2 == IntervalField::Month {
27440 if let Some(m) = parse_year_month_literal(text) {
27441 return Some((m, 0, 0));
27442 }
27443 }
27444 return parse_interval_text(text);
27445 }
27446 // Single field: reinterpret a bare number; otherwise the default parse.
27447 let trimmed = text.trim();
27448 if let Ok(val) = trimmed.parse::<f64>() {
27449 // no_std: f64 has no trunc/round; cast toward zero + round-half-away.
27450 #[allow(clippy::cast_possible_truncation)]
27451 let whole = val as i64;
27452 #[allow(clippy::cast_possible_truncation)]
27453 let secs_micros = {
27454 let m = val * 1_000_000.0;
27455 if m >= 0.0 {
27456 (m + 0.5) as i64
27457 } else {
27458 (m - 0.5) as i64
27459 }
27460 };
27461 return Some(match f1 {
27462 IntervalField::Year => (i32::try_from(whole).ok()?.checked_mul(12)?, 0, 0),
27463 IntervalField::Month => (i32::try_from(whole).ok()?, 0, 0),
27464 IntervalField::Day => (0, i32::try_from(whole).ok()?, 0),
27465 IntervalField::Hour => (0, 0, whole.checked_mul(3_600_000_000)?),
27466 IntervalField::Minute => (0, 0, whole.checked_mul(60_000_000)?),
27467 IntervalField::Second => (0, 0, secs_micros),
27468 });
27469 }
27470 parse_interval_text(text)
27471}
27472
27473/// Parse the `Y-M` (optionally signed) year-to-month literal into total months.
27474fn parse_year_month_literal(text: &str) -> Option<i32> {
27475 let t = text.trim();
27476 let (neg, body) = match t.strip_prefix('-') {
27477 Some(r) => (true, r),
27478 None => (false, t.strip_prefix('+').unwrap_or(t)),
27479 };
27480 let mut it = body.split('-');
27481 let years: i32 = it.next()?.trim().parse().ok()?;
27482 let months: i32 = match it.next() {
27483 Some(m) => m.trim().parse().ok()?,
27484 None => 0,
27485 };
27486 if it.next().is_some() {
27487 return None;
27488 }
27489 let total = years.checked_mul(12)?.checked_add(months)?;
27490 Some(if neg { -total } else { total })
27491}
27492
27493pub fn parse_interval_text(s: &str) -> Option<(i32, i32, i64)> {
27494 // v7.38.19 — the two infinities, answered as the three extreme
27495 // fields PostgreSQL itself puts on the wire for them:
27496 //
27497 // COPY (SELECT 'infinity'::interval) TO STDOUT (FORMAT binary)
27498 // … 7fffffffffffffff 7fffffff 7fffffff
27499 //
27500 // So no caller has to know the spelling — every one of them already
27501 // reads the three numbers, and `IntervalKind::from_fields` names
27502 // what they mean.
27503 //
27504 // `inf` is NOT one of them, measured: `'inf'::interval` is *invalid
27505 // input syntax* on PostgreSQL 18.4 while `'inf'::float8` is
27506 // infinity. Interval takes the full word, in any case.
27507 {
27508 let word = s.trim();
27509 let word = word.strip_prefix('@').map_or(word, str::trim);
27510 let (neg, body) = match word.strip_prefix('-') {
27511 Some(rest) => (true, rest.trim_start()),
27512 None => (false, word.strip_prefix('+').map_or(word, str::trim_start)),
27513 };
27514 if body.eq_ignore_ascii_case("infinity") {
27515 return Some(if neg {
27516 (i32::MIN, i32::MIN, i64::MIN)
27517 } else {
27518 (i32::MAX, i32::MAX, i64::MAX)
27519 });
27520 }
27521 }
27522 // v7.39 (read01 timestamp.c) — PG's postgres_verbose forms: a leading
27523 // `@` is decorative; a trailing `ago` negates the whole interval.
27524 let mut trimmed = s.trim();
27525 trimmed = trimmed.strip_prefix('@').map_or(trimmed, str::trim);
27526 let mut negate = false;
27527 if let Some(rest) = trimmed
27528 .strip_suffix("ago")
27529 .filter(|r| r.ends_with(char::is_whitespace) || r.is_empty())
27530 {
27531 negate = true;
27532 trimmed = rest.trim();
27533 }
27534 let finish = |v: Option<(i32, i32, i64)>| -> Option<(i32, i32, i64)> {
27535 let (mo, d, us) = v?;
27536 if negate {
27537 Some((mo.checked_neg()?, d.checked_neg()?, us.checked_neg()?))
27538 } else {
27539 Some((mo, d, us))
27540 }
27541 };
27542 let s = trimmed;
27543 // ISO 8601 duration (`P1Y2M3DT4H`) and PG's year-month shorthand (`1-2`)
27544 // are single tokens, not the `<n> <unit>` pair form handled below.
27545 if let Some(rest) = trimmed.strip_prefix(['P', 'p']) {
27546 return finish(parse_iso8601_interval(rest));
27547 }
27548 if !trimmed.contains(char::is_whitespace) && trimmed.contains('-') {
27549 if let Some(iv) = parse_year_month_interval(trimmed) {
27550 return finish(Some(iv));
27551 }
27552 }
27553 // v7.39 (GUC knife 3, differential) — PG accepts a bare number as
27554 // SECONDS: `INTERVAL '0'` = 00:00:00, `INTERVAL '5'` = 00:00:05,
27555 // fractions kept to the microsecond (`'1.5'` = 00:00:01.5).
27556 if !trimmed.is_empty() && !trimmed.contains(char::is_whitespace) {
27557 if let Ok(n) = trimmed.parse::<i64>() {
27558 return finish(Some((0, 0, n.checked_mul(1_000_000)?)));
27559 }
27560 if let Ok(f) = trimmed.parse::<f64>() {
27561 if f.is_finite() {
27562 #[allow(clippy::cast_possible_truncation)]
27563 return finish(Some((0, 0, (f * 1_000_000.0) as i64)));
27564 }
27565 }
27566 }
27567 // v7.39 (round 243) — PG accepts the number and unit run together
27568 // (`'15h 2m 12s'`); split each token at the digit→letter boundary so
27569 // the `<n> <unit>` pair loop below sees them as two.
27570 let raw_parts: Vec<&str> = s.split_whitespace().collect();
27571 let mut parts: Vec<&str> = Vec::with_capacity(raw_parts.len());
27572 for p in raw_parts {
27573 let boundary = p
27574 .char_indices()
27575 .find(|(i, c)| {
27576 *i > 0
27577 && c.is_ascii_alphabetic()
27578 && p[..*i]
27579 .chars()
27580 .all(|d| d.is_ascii_digit() || matches!(d, '.' | '-' | '+'))
27581 && p[..*i].chars().any(|d| d.is_ascii_digit())
27582 })
27583 .map(|(i, _)| i);
27584 match boundary {
27585 Some(i) => {
27586 parts.push(&p[..i]);
27587 parts.push(&p[i..]);
27588 }
27589 None => parts.push(p),
27590 }
27591 }
27592 // A bare clock-time token `HH:MM[:SS[.ffffff]]` carries the time-of-day
27593 // part (PG: `3 days 14:30:45`, or `14:30:45` alone). Extract it; whatever
27594 // remains is the `<n> <unit>` pair form handled below.
27595 let mut clock_us: i64 = 0;
27596 let mut had_clock = false;
27597 if let Some(pos) = parts.iter().position(|p| p.contains(':')) {
27598 clock_us = parse_interval_clock(parts[pos])?;
27599 parts.remove(pos);
27600 had_clock = true;
27601 }
27602 // v7.39 (read01 timestamp.c) — a lone bare number alongside a clock
27603 // time is DAYS (PG: '3 4:05:06' = 3 days 04:05:06).
27604 let mut lone_days: i32 = 0;
27605 if had_clock && parts.len() == 1 {
27606 if let Ok(n) = parts[0].parse::<i64>() {
27607 lone_days = i32::try_from(n).ok()?;
27608 parts.clear();
27609 }
27610 }
27611 if !parts.len().is_multiple_of(2) || (parts.is_empty() && !had_clock && lone_days == 0) {
27612 return None;
27613 }
27614 let mut months: i32 = 0;
27615 let mut days: i32 = lone_days;
27616 let mut micros: i64 = clock_us;
27617 let mut i = 0;
27618 while i < parts.len() {
27619 let unit_stripped = canonical_interval_unit(parts[i + 1])?;
27620 if let Ok(n) = parts[i].parse::<i64>() {
27621 match unit_stripped {
27622 "microsecond" => micros = micros.checked_add(n)?,
27623 "millisecond" => micros = micros.checked_add(n.checked_mul(1_000)?)?,
27624 "second" => micros = micros.checked_add(n.checked_mul(1_000_000)?)?,
27625 "minute" => micros = micros.checked_add(n.checked_mul(60_000_000)?)?,
27626 "hour" => micros = micros.checked_add(n.checked_mul(3_600_000_000)?)?,
27627 "day" => {
27628 let n32 = i32::try_from(n).ok()?;
27629 days = days.checked_add(n32)?;
27630 }
27631 "week" => {
27632 let n32 = i32::try_from(n).ok()?;
27633 days = days.checked_add(n32.checked_mul(7)?)?;
27634 }
27635 "month" => {
27636 let n32 = i32::try_from(n).ok()?;
27637 months = months.checked_add(n32)?;
27638 }
27639 "year" => {
27640 let n32 = i32::try_from(n).ok()?;
27641 months = months.checked_add(n32.checked_mul(12)?)?;
27642 }
27643 // v7.39 (read01 timestamp.c) — the larger calendar units.
27644 "decade" => {
27645 let n32 = i32::try_from(n).ok()?;
27646 months = months.checked_add(n32.checked_mul(120)?)?;
27647 }
27648 "century" => {
27649 let n32 = i32::try_from(n).ok()?;
27650 months = months.checked_add(n32.checked_mul(1200)?)?;
27651 }
27652 "millennium" => {
27653 let n32 = i32::try_from(n).ok()?;
27654 months = months.checked_add(n32.checked_mul(12000)?)?;
27655 }
27656 _ => return None,
27657 }
27658 } else if let Ok(f) = parts[i].parse::<f64>() {
27659 // Fractional units cascade down to the next-finer field the way
27660 // PG does: `1.5 days` -> `1 day 12:00:00`, `1.5 months` ->
27661 // `1 mon 15 days` (30-day month), `1.5 years` -> `1 year 6 mons`.
27662 // no_std: f64 has no trunc/fract/round methods, so do them with
27663 // casts (toward-zero) + explicit round-half-away-from-zero.
27664 #[allow(clippy::cast_possible_truncation)]
27665 fn round_i64(x: f64) -> i64 {
27666 if x >= 0.0 {
27667 (x + 0.5) as i64
27668 } else {
27669 (x - 0.5) as i64
27670 }
27671 }
27672 #[allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
27673 fn add_days_frac(days: &mut i32, micros: &mut i64, d: f64) -> Option<()> {
27674 const DAY_US: f64 = 86_400_000_000.0;
27675 let whole = d as i64; // truncates toward zero
27676 let frac = d - whole as f64;
27677 *days = days.checked_add(i32::try_from(whole).ok()?)?;
27678 *micros = micros.checked_add(round_i64(frac * DAY_US))?;
27679 Some(())
27680 }
27681 #[allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
27682 match unit_stripped {
27683 "microsecond" => micros = micros.checked_add(round_i64(f))?,
27684 "millisecond" => micros = micros.checked_add(round_i64(f * 1_000.0))?,
27685 "second" => micros = micros.checked_add(round_i64(f * 1_000_000.0))?,
27686 "minute" => micros = micros.checked_add(round_i64(f * 60_000_000.0))?,
27687 "hour" => micros = micros.checked_add(round_i64(f * 3_600_000_000.0))?,
27688 "day" => add_days_frac(&mut days, &mut micros, f)?,
27689 "week" => add_days_frac(&mut days, &mut micros, f * 7.0)?,
27690 "month" => {
27691 let whole = f as i64;
27692 months = months.checked_add(i32::try_from(whole).ok()?)?;
27693 add_days_frac(&mut days, &mut micros, (f - whole as f64) * 30.0)?;
27694 }
27695 "year" => {
27696 let m = f * 12.0;
27697 let whole = m as i64;
27698 months = months.checked_add(i32::try_from(whole).ok()?)?;
27699 add_days_frac(&mut days, &mut micros, (m - whole as f64) * 30.0)?;
27700 }
27701 _ => return None,
27702 }
27703 } else {
27704 return None;
27705 }
27706 i += 2;
27707 }
27708 finish(Some((months, days, micros)))
27709}
27710
27711/// v7.37 — map a scalar type keyword to its [`CastTarget`] for the PG
27712/// `TYPE 'literal'` typed-literal syntax (`time '10:30'` == `'10:30'::time`).
27713/// `interval` is intentionally absent (handled by its own parser arm).
27714/// Returns `None` for names that aren't sensible as a bare typed literal, so
27715/// the caller falls back to treating the ident as a column reference.
27716fn typed_literal_cast_target(ident: &str) -> Option<CastTarget> {
27717 Some(match ident {
27718 "date" => CastTarget::Date,
27719 "timestamp" | "datetime" => CastTarget::Timestamp,
27720 "timestamptz" => CastTarget::Timestamptz,
27721 "bool" | "boolean" => CastTarget::Bool,
27722 "int" | "integer" | "int4" => CastTarget::Int,
27723 "bigint" | "int8" => CastTarget::BigInt,
27724 "float8" | "double precision" => CastTarget::Float,
27725 "uuid" => CastTarget::Uuid,
27726 "bytea" => CastTarget::Bytea,
27727 "json" => CastTarget::Json,
27728 "jsonb" => CastTarget::Jsonb,
27729 // Types without a dedicated CastTarget variant flow through the
27730 // generic Named path (engine resolves via column_type_to_data_type).
27731 "time" | "timetz" | "smallint" | "int2" | "numeric" | "decimal"
27732 | "real" | "float4" | "inet" | "cidr" | "macaddr" | "macaddr8"
27733 | "money" | "bit" | "varbit"
27734 // Geometric types accept the `TYPE 'literal'` prefix spelling too.
27735 | "point" | "line" | "lseg" | "box" | "path" | "polygon" | "circle"
27736 // Range / multirange types likewise.
27737 | "int4range" | "int8range" | "numrange" | "daterange" | "tsrange"
27738 | "tstzrange" | "int4multirange" | "int8multirange" | "nummultirange"
27739 | "datemultirange" | "tsmultirange" | "tstzmultirange"
27740 // v7.39 (read01 round 18) — oid / name / jsonpath literal prefixes.
27741 | "oid" | "name" | "jsonpath" | "pg_lsn" | "varchar" | "text" | "xid" | "xid8" => {
27742 CastTarget::Named(alloc::string::String::from(ident))
27743 }
27744 _ => return None,
27745 })
27746}
27747
27748/// v7.12.4 — map a bare type-name identifier (the form that
27749/// appears in a function arg list or RETURNS clause) to a
27750/// [`ColumnTypeName`]. Returns `None` for unknown / extension
27751/// types so the caller can preserve them as
27752/// [`FunctionArgType::Raw`] / [`FunctionReturn::Other`].
27753///
27754/// Subset of the full column-type grammar — we deliberately
27755/// don't parse parameterised forms (`VARCHAR(n)`, `NUMERIC(p,s)`)
27756/// here because function-arg types in v7.12.4 are mostly the
27757/// bare form (`text`, `int`, `bytea`, …).
27758/// v7.39 (round 315, V19) — does this whole phrase name a type, rather
27759/// than being `name TYPE`?
27760///
27761/// The multi-word spellings SQL allows for a bare argument type, each
27762/// verified accepted by PG 18.4 as `CREATE FUNCTION f(<phrase>)`.
27763///
27764/// NOTE this list also exists in `spg-storage`, which computes the
27765/// signature key from the rendered argument text and has to reach the
27766/// same verdict. The two crates are siblings — neither depends on the
27767/// other — and each already carries its own table of type spellings
27768/// (`map_type_ident_to_column_type_name` here, `normalize_type_name`
27769/// there), so this follows the structure rather than inventing new
27770/// duplication. Recorded as V49.
27771pub fn is_multiword_type_phrase(phrase: &str) -> bool {
27772 let t = phrase.trim().to_ascii_lowercase();
27773 let base = t.split_once('(').map_or(t.as_str(), |(h, _)| h).trim();
27774 matches!(
27775 base,
27776 "double precision"
27777 | "character varying"
27778 | "bit varying"
27779 | "timestamp with time zone"
27780 | "timestamp without time zone"
27781 | "time with time zone"
27782 | "time without time zone"
27783 | "national character"
27784 | "national character varying"
27785 )
27786}
27787
27788fn map_type_ident_to_column_type_name(ident: &str) -> Option<ColumnTypeName> {
27789 Some(match ident.to_ascii_lowercase().as_str() {
27790 "smallint" | "tinyint" => ColumnTypeName::SmallInt,
27791 "int" | "integer" | "mediumint" => ColumnTypeName::Int,
27792 "bigint" => ColumnTypeName::BigInt,
27793 "float" | "double" => ColumnTypeName::Float,
27794 // v7.39 (round 269) — real is 32-bit.
27795 "real" | "float4" => ColumnTypeName::Real,
27796 "text" => ColumnTypeName::Text,
27797 "bool" | "boolean" => ColumnTypeName::Bool,
27798 "date" => ColumnTypeName::Date,
27799 "timestamp" | "datetime" => ColumnTypeName::Timestamp,
27800 "timestamptz" => ColumnTypeName::Timestamptz,
27801 "json" => ColumnTypeName::Json,
27802 "jsonb" => ColumnTypeName::Jsonb,
27803 "bytea" | "bytes" => ColumnTypeName::Bytes,
27804 "tsvector" => ColumnTypeName::TsVector,
27805 "tsquery" => ColumnTypeName::TsQuery,
27806 "uuid" => ColumnTypeName::Uuid,
27807 "interval" => ColumnTypeName::Interval,
27808 "time" => ColumnTypeName::Time,
27809 "year" => ColumnTypeName::Year,
27810 "timetz" => ColumnTypeName::TimeTz,
27811 "money" => ColumnTypeName::Money,
27812 _ => return None,
27813 })
27814}
27815
27816/// v7.12.4 — parse a PL/pgSQL function body (the bytes between
27817/// `$$ ... $$`). Returns the parsed `BEGIN ... END;` block.
27818///
27819/// v7.12.4 grammar (strict subset — IF / LOOP / DECLARE / RAISE
27820/// / embedded SQL land in v7.12.5+):
27821///
27822/// ```text
27823/// body := [ws] block [ws]
27824/// block := BEGIN stmt ( ; stmt )* [ ; ] END [ ; ]
27825/// stmt := assign | return
27826/// assign := assign_target := expr
27827/// assign_target := ( NEW | OLD ) . ident | ident
27828/// return := RETURN ( NEW | OLD | NULL | expr )
27829/// ```
27830///
27831/// `expr` is parsed by recursing into the regular `Parser` — so a
27832/// PL/pgSQL `NEW.search_vector := to_tsvector('english',
27833/// NEW.subject || ' ' || NEW.sender)` body shape works without
27834/// the body parser knowing what `to_tsvector` is.
27835///
27836/// Errors here cause the caller to fall back to
27837/// `FunctionBody::Raw` — keeping the CREATE FUNCTION DDL itself
27838/// successful, but the executor will refuse to invoke the
27839/// function with an "unparseable body" error.
27840/// v7.12.4 — public alias for [`parse_plpgsql_body`] re-exported
27841/// from the crate root as `spg_sql::parse_function_body`.
27842pub fn parse_function_body(body: &str) -> Result<PlPgSqlBlock, ParseError> {
27843 parse_plpgsql_body(body)
27844}
27845
27846fn parse_plpgsql_body(body: &str) -> Result<PlPgSqlBlock, ParseError> {
27847 // Use the regular lexer on the body text. The trailing
27848 // `END;` may or may not have a semicolon; the lexer treats
27849 // both forms identically.
27850 let tokens = lexer::tokenize(body).map_err(|e| ParseError {
27851 message: alloc::format!("plpgsql body lex error: {e}"),
27852 token_pos: 0,
27853 })?;
27854 let mut parser = Parser::new(tokens);
27855 parser.parse_plpgsql_block()
27856}
27857
27858/// v7.39 (GUC) — the textual body of a SET value, for list joining.
27859fn set_value_text(v: &crate::ast::SetValue) -> alloc::string::String {
27860 match v {
27861 crate::ast::SetValue::String(s)
27862 | crate::ast::SetValue::Ident(s)
27863 | crate::ast::SetValue::Number(s) => s.clone(),
27864 crate::ast::SetValue::Default => "DEFAULT".into(),
27865 }
27866}
27867
27868/// v7.39 (round 145, parse_cte.c / parse_agg.c) — true when an expression
27869/// contains an aggregate call at ITS OWN query level (recursion stops at
27870/// sublink boundaries — a sublink's aggregates belong to the sublink).
27871/// Backs the "aggregate functions are not allowed in a recursive query's
27872/// recursive term" well-formedness check.
27873fn expr_has_toplevel_aggregate(e: &Expr) -> bool {
27874 const AGG_NAMES: &[&str] = &[
27875 "count",
27876 "sum",
27877 "min",
27878 "max",
27879 "avg",
27880 "string_agg",
27881 "array_agg",
27882 "bool_and",
27883 "bool_or",
27884 "every",
27885 "any_value",
27886 "json_agg",
27887 "jsonb_agg",
27888 "json_object_agg",
27889 "jsonb_object_agg",
27890 "bit_and",
27891 "bit_or",
27892 "bit_xor",
27893 "var_pop",
27894 "var_samp",
27895 "variance",
27896 "stddev",
27897 "stddev_pop",
27898 "stddev_samp",
27899 "range_agg",
27900 "range_intersect_agg",
27901 "percentile_cont",
27902 "percentile_disc",
27903 "mode",
27904 "corr",
27905 "covar_pop",
27906 "covar_samp",
27907 ];
27908 match e {
27909 Expr::AggregateOrdered { .. } => true,
27910 Expr::FunctionCall { name, args } => {
27911 AGG_NAMES.contains(&name.to_ascii_lowercase().as_str())
27912 || args.iter().any(expr_has_toplevel_aggregate)
27913 }
27914 Expr::NamedArg { expr, .. }
27915 | Expr::Variadic(expr)
27916 | Expr::Unary { expr, .. }
27917 | Expr::Cast { expr, .. }
27918 | Expr::IsNull { expr, .. }
27919 | Expr::FieldAccess { base: expr, .. }
27920 | Expr::Extract { source: expr, .. } => expr_has_toplevel_aggregate(expr),
27921 Expr::Binary { lhs, rhs, .. } => {
27922 expr_has_toplevel_aggregate(lhs) || expr_has_toplevel_aggregate(rhs)
27923 }
27924 Expr::Like { expr, pattern, .. } => {
27925 expr_has_toplevel_aggregate(expr) || expr_has_toplevel_aggregate(pattern)
27926 }
27927 Expr::Array(items) => items.iter().any(expr_has_toplevel_aggregate),
27928 Expr::InList { expr, list, .. } => {
27929 expr_has_toplevel_aggregate(expr) || list.iter().any(expr_has_toplevel_aggregate)
27930 }
27931 Expr::ArraySubscript { target, index } => {
27932 expr_has_toplevel_aggregate(target) || expr_has_toplevel_aggregate(index)
27933 }
27934 Expr::ArraySlice { target, lo, hi } => {
27935 expr_has_toplevel_aggregate(target)
27936 || lo.as_deref().is_some_and(expr_has_toplevel_aggregate)
27937 || hi.as_deref().is_some_and(expr_has_toplevel_aggregate)
27938 }
27939 Expr::AnyAll { expr, array, .. } => {
27940 expr_has_toplevel_aggregate(expr) || expr_has_toplevel_aggregate(array)
27941 }
27942 Expr::Case {
27943 operand,
27944 branches,
27945 else_branch,
27946 } => {
27947 operand.as_deref().is_some_and(expr_has_toplevel_aggregate)
27948 || branches
27949 .iter()
27950 .any(|(w, t)| expr_has_toplevel_aggregate(w) || expr_has_toplevel_aggregate(t))
27951 || else_branch
27952 .as_deref()
27953 .is_some_and(expr_has_toplevel_aggregate)
27954 }
27955 // The outer-level operands of a sublink can aggregate; the sublink's
27956 // own body cannot leak its aggregates up here.
27957 Expr::InSubquery { expr, .. } => expr_has_toplevel_aggregate(expr),
27958 Expr::RowInSubquery { row, .. } | Expr::RowCmpSubquery { row, .. } => {
27959 row.iter().any(expr_has_toplevel_aggregate)
27960 }
27961 _ => false,
27962 }
27963}
27964
27965/// v7.39 (round 145, parse_cte.c) — true when any sublink expression
27966/// (EXISTS / IN / scalar subquery) inside this SELECT term references the
27967/// named table anywhere in its subtree. A plain FROM derived table is NOT a
27968/// sublink and is legal in a recursive term, so it is not walked here.
27969fn select_has_self_ref_in_sublink(s: &crate::ast::SelectStatement, name: &str) -> bool {
27970 let mut exprs: Vec<&Expr> = Vec::new();
27971 for it in &s.items {
27972 if let crate::ast::SelectItem::Expr { expr, .. } = it {
27973 exprs.push(expr);
27974 }
27975 }
27976 if let Some(w) = &s.where_ {
27977 exprs.push(w);
27978 }
27979 if let Some(h) = &s.having {
27980 exprs.push(h);
27981 }
27982 if let Some(g) = &s.group_by {
27983 exprs.extend(g.iter());
27984 }
27985 if let Some(from) = &s.from {
27986 for j in &from.joins {
27987 if let Some(on) = &j.on {
27988 exprs.push(on);
27989 }
27990 }
27991 }
27992 exprs.into_iter().any(|e| expr_sublink_mentions(e, name))
27993}
27994
27995/// Does this expression contain a sublink whose subquery mentions `name`?
27996fn expr_sublink_mentions(e: &Expr, name: &str) -> bool {
27997 match e {
27998 Expr::ScalarSubquery(sub) => select_mentions_table(sub, name),
27999 Expr::Exists { subquery, .. } => select_mentions_table(subquery, name),
28000 Expr::InSubquery { expr, subquery, .. } => {
28001 expr_sublink_mentions(expr, name) || select_mentions_table(subquery, name)
28002 }
28003 Expr::RowInSubquery { row, subquery, .. } => {
28004 row.iter().any(|x| expr_sublink_mentions(x, name))
28005 || select_mentions_table(subquery, name)
28006 }
28007 Expr::RowCmpSubquery { row, subquery, .. } => {
28008 row.iter().any(|x| expr_sublink_mentions(x, name))
28009 || select_mentions_table(subquery, name)
28010 }
28011 Expr::NamedArg { expr, .. }
28012 | Expr::Variadic(expr)
28013 | Expr::Unary { expr, .. }
28014 | Expr::Cast { expr, .. }
28015 | Expr::IsNull { expr, .. }
28016 | Expr::FieldAccess { base: expr, .. }
28017 | Expr::Extract { source: expr, .. } => expr_sublink_mentions(expr, name),
28018 Expr::Binary { lhs, rhs, .. } => {
28019 expr_sublink_mentions(lhs, name) || expr_sublink_mentions(rhs, name)
28020 }
28021 Expr::Like { expr, pattern, .. } => {
28022 expr_sublink_mentions(expr, name) || expr_sublink_mentions(pattern, name)
28023 }
28024 Expr::FunctionCall { args, .. } | Expr::Array(args) => {
28025 args.iter().any(|x| expr_sublink_mentions(x, name))
28026 }
28027 Expr::InList { expr, list, .. } => {
28028 expr_sublink_mentions(expr, name) || list.iter().any(|x| expr_sublink_mentions(x, name))
28029 }
28030 Expr::ArraySubscript { target, index } => {
28031 expr_sublink_mentions(target, name) || expr_sublink_mentions(index, name)
28032 }
28033 Expr::ArraySlice { target, lo, hi } => {
28034 expr_sublink_mentions(target, name)
28035 || lo
28036 .as_deref()
28037 .is_some_and(|x| expr_sublink_mentions(x, name))
28038 || hi
28039 .as_deref()
28040 .is_some_and(|x| expr_sublink_mentions(x, name))
28041 }
28042 Expr::AnyAll { expr, array, .. } => {
28043 expr_sublink_mentions(expr, name) || expr_sublink_mentions(array, name)
28044 }
28045 Expr::Case {
28046 operand,
28047 branches,
28048 else_branch,
28049 } => {
28050 operand
28051 .as_deref()
28052 .is_some_and(|x| expr_sublink_mentions(x, name))
28053 || branches
28054 .iter()
28055 .any(|(w, t)| expr_sublink_mentions(w, name) || expr_sublink_mentions(t, name))
28056 || else_branch
28057 .as_deref()
28058 .is_some_and(|x| expr_sublink_mentions(x, name))
28059 }
28060 _ => false,
28061 }
28062}
28063
28064/// Does this SELECT (in full — FROM tables, derived tables, its own
28065/// sublinks, and union arms) mention the named table?
28066fn select_mentions_table(s: &crate::ast::SelectStatement, name: &str) -> bool {
28067 if let Some(from) = &s.from {
28068 if from.primary.name.eq_ignore_ascii_case(name) {
28069 return true;
28070 }
28071 if let Some(sub) = &from.primary.lateral_subquery
28072 && select_mentions_table(sub, name)
28073 {
28074 return true;
28075 }
28076 for j in &from.joins {
28077 if j.table.name.eq_ignore_ascii_case(name) {
28078 return true;
28079 }
28080 if let Some(sub) = &j.table.lateral_subquery
28081 && select_mentions_table(sub, name)
28082 {
28083 return true;
28084 }
28085 }
28086 }
28087 if select_has_self_ref_in_sublink(s, name) {
28088 return true;
28089 }
28090 s.unions.iter().any(|(_, u)| select_mentions_table(u, name))
28091}
28092
28093/// v7.39 (round 284) — fold a constant `LIMIT` / `OFFSET` expression to a
28094/// row count, the way PG evaluates one before applying it.
28095///
28096/// `None` = not a constant (a column, a subquery, a function call).
28097/// `Some(Err(msg))` = PG rejects it, and the message is PG's; `{L}` in the
28098/// message stands in for LIMIT / OFFSET, which the caller substitutes.
28099/// All wordings were read off live PG 18.4.
28100fn fold_limit_constant(e: &crate::ast::Expr) -> Option<Result<i128, alloc::string::String>> {
28101 use crate::ast::{BinOp, Expr, Literal, UnOp};
28102 match e {
28103 Expr::Literal(Literal::Integer(n)) => Some(Ok(i128::from(*n))),
28104 Expr::Literal(Literal::Numeric { unscaled, scale }) => {
28105 Some(Ok(round_scaled_half_away(*unscaled, *scale)))
28106 }
28107 // PG coerces a string by its CONTENT, and fails on the value.
28108 Expr::Literal(Literal::String(t)) => Some(t.trim().parse::<i64>().map_or_else(
28109 |_| {
28110 Err(alloc::format!(
28111 "invalid input syntax for type bigint: \"{t}\""
28112 ))
28113 },
28114 |n| Ok(i128::from(n)),
28115 )),
28116 Expr::Literal(Literal::Bool(_)) => Some(Err(
28117 "argument of {L} must be type bigint, not type boolean".into(),
28118 )),
28119 Expr::Unary {
28120 op: UnOp::Neg,
28121 expr,
28122 } => match fold_limit_constant(expr)? {
28123 Ok(v) => Some(Ok(-v)),
28124 e @ Err(_) => Some(e),
28125 },
28126 Expr::Binary { lhs, op, rhs } => {
28127 let a = match fold_limit_constant(lhs)? {
28128 Ok(v) => v,
28129 e @ Err(_) => return Some(e),
28130 };
28131 let b = match fold_limit_constant(rhs)? {
28132 Ok(v) => v,
28133 e @ Err(_) => return Some(e),
28134 };
28135 let out = match op {
28136 BinOp::Add => a.checked_add(b),
28137 BinOp::Sub => a.checked_sub(b),
28138 BinOp::Mul => a.checked_mul(b),
28139 BinOp::Div if b != 0 => a.checked_div(b),
28140 BinOp::Div => return Some(Err("division by zero".into())),
28141 BinOp::Mod if b != 0 => a.checked_rem(b),
28142 BinOp::Mod => return Some(Err("division by zero".into())),
28143 _ => return None,
28144 };
28145 // PG evaluates the arithmetic in the operand's own type, so an
28146 // int-by-int product that leaves int range fails there — before
28147 // the row count is ever looked at.
28148 match out {
28149 Some(v) if v > i128::from(i32::MAX) || v < i128::from(i32::MIN) => {
28150 Some(Err("integer out of range".into()))
28151 }
28152 Some(v) => Some(Ok(v)),
28153 None => Some(Err("integer out of range".into())),
28154 }
28155 }
28156 _ => None,
28157 }
28158}
28159
28160/// Round `unscaled / 10^scale` half away from zero — PG's numeric→bigint
28161/// cast, which is what makes `LIMIT 2.5` keep three rows.
28162fn round_scaled_half_away(unscaled: i128, scale: u16) -> i128 {
28163 if scale == 0 {
28164 return unscaled;
28165 }
28166 let Some(div) = 10i128.checked_pow(u32::from(scale)) else {
28167 return 0;
28168 };
28169 let neg = unscaled < 0;
28170 let mag = unscaled.unsigned_abs() as i128;
28171 let rounded = (mag + div / 2) / div;
28172 if neg { -rounded } else { rounded }
28173}
28174
28175#[cfg(test)]
28176mod tests {
28177 use super::*;
28178 use alloc::string::ToString;
28179
28180 fn parse(s: &str) -> Statement {
28181 parse_statement(s).expect("parse ok")
28182 }
28183
28184 // v7.37.43-T4 sentori cutover acceptance — `release`, `index`,
28185 // `tables`, `partition`, etc. are unreserved keywords per PG's
28186 // `pg_get_keywords()` and MUST be usable as column / table /
28187 // alias names. Pre-T4 every drop-in user whose schema had one
28188 // of these as a column name (sentori events.release, mailrs
28189 // messages.index in some forks) blew the parser up at CREATE
28190 // TABLE time with "expected identifier, got Release". The
28191 // generalisation lives in `unreserved_keyword_text` + the
28192 // `expect_ident_like` and `parse_atom` arms that consult it.
28193 #[test]
28194 fn release_usable_as_column_name_in_create_table() {
28195 let stmt =
28196 parse("CREATE TABLE events (id INT PRIMARY KEY, release TEXT NOT NULL, payload TEXT)");
28197 if let Statement::CreateTable(t) = stmt {
28198 let names: alloc::vec::Vec<&str> = t.columns.iter().map(|c| c.name.as_str()).collect();
28199 assert_eq!(names, alloc::vec!["id", "release", "payload"]);
28200 } else {
28201 panic!("expected CreateTable");
28202 }
28203 }
28204
28205 #[test]
28206 fn release_usable_as_column_ref_in_select_projection() {
28207 // The sentori `0003_partition_events.sql` INSERT-SELECT
28208 // walk references `release` in both column lists; the
28209 // projection-side use exercises `parse_atom`'s relaxed
28210 // identifier set.
28211 parse("SELECT id, release, payload FROM events WHERE id = 1");
28212 }
28213
28214 #[test]
28215 fn release_usable_as_column_ref_in_insert_column_list() {
28216 // INSERT INTO t (id, release, payload) VALUES (…)
28217 parse("INSERT INTO events (id, release, payload) VALUES (1, '1.0.0', 'data')");
28218 }
28219
28220 #[test]
28221 fn alter_column_drop_not_null_uses_keyword_drop_token() {
28222 // Sentori `0013_audit_tombstone.sql` issues
28223 // `ALTER TABLE … ALTER COLUMN x DROP NOT NULL`. The lexer
28224 // emits Token::Drop (not Ident("drop")); the parser must
28225 // accept both in the ALTER COLUMN sub-dispatch.
28226 parse("ALTER TABLE audit_logs ALTER COLUMN org_id DROP NOT NULL");
28227 }
28228
28229 #[test]
28230 fn create_index_accepts_parenthesised_expression_key() {
28231 // sentori `0040_events_bundle_idx.sql` shape — JSONB
28232 // expression index. Pre-T4 the parser bailed at the
28233 // inner `(` with "expected column ident or expression,
28234 // got LParen". The Token::LParen arm in CREATE INDEX
28235 // routes through the expression parser instead.
28236 parse(
28237 "CREATE INDEX IF NOT EXISTS events_bundle_id_idx \
28238 ON events ((payload->'bundle'->>'id'))",
28239 );
28240 }
28241
28242 // v7.30.2 (mailrs round-25 ask 2) — nesting / chain budgets must
28243 // surface as parse errors, never stack overflows (embed hosts
28244 // abort on overflow).
28245 /// The nesting budget is a COUNT; what it has to fit inside is a
28246 /// number of BYTES, and only one of those two is stable across
28247 /// compiler versions. Round 847 measured 30,336 bytes per level
28248 /// after a toolchain move, which puts 64 levels at 1.94 MB and
28249 /// overflows a 2 MiB thread — `nesting_budget_errors_cleanly`
28250 /// aborted instead of erroring, which is precisely the outcome it
28251 /// exists to rule out.
28252 ///
28253 /// So the budget is metered rather than assumed. The ceiling leaves
28254 /// the depth SPG advertises fitting in a default 2 MiB thread with
28255 /// room to spare, in the debug build, where frames are widest.
28256 #[test]
28257 fn nesting_frame_cost_stays_under_ceiling() {
28258 // Room for MAX_NEST_DEPTH levels inside 1.2 MB, so a 2 MiB
28259 // thread keeps a margin for whatever called the parser.
28260 const CEILING: usize = 1_200_000 / MAX_NEST_DEPTH;
28261
28262 frame_meter::reset();
28263 let depth = frame_meter::SAMPLE_HI + 8;
28264 let sql = format!("SELECT {}1{}", "(".repeat(depth), ")".repeat(depth));
28265 parse(&sql);
28266
28267 let per_level = frame_meter::bytes_per_level();
28268 {
28269 extern crate std;
28270 std::eprintln!("nesting frame: {per_level} bytes/level, ceiling {CEILING}");
28271 }
28272 assert!(
28273 per_level <= CEILING,
28274 "{per_level} bytes per nesting level exceeds {CEILING}; \
28275 {MAX_NEST_DEPTH} levels would want {} bytes. Out-line arms \
28276 in parse_expr_inner / parse_unary rather than lowering the \
28277 depth or widening the stack.",
28278 per_level * MAX_NEST_DEPTH
28279 );
28280 }
28281
28282 #[test]
28283 fn nesting_budget_errors_cleanly() {
28284 let depth = MAX_NEST_DEPTH + 50;
28285 let sql = format!("SELECT {}1{}", "(".repeat(depth), ")".repeat(depth));
28286 let err = parse_statement(&sql).expect_err("must reject");
28287 assert!(err.message.contains("nests deeper"), "{err:?}");
28288 // Within budget still parses.
28289 let sql = format!("SELECT {}1{}", "(".repeat(48), ")".repeat(48));
28290 parse(&sql);
28291 }
28292
28293 #[test]
28294 fn binary_chain_budget_errors_cleanly() {
28295 let sql = format!("SELECT 1{}", " + 1".repeat(MAX_BINARY_CHAIN + 50));
28296 let err = parse_statement(&sql).expect_err("must reject");
28297 assert!(err.message.contains("chained binary"), "{err:?}");
28298 // Within budget still parses (chain depth ≤ budget is safe
28299 // for recursive eval/drop on 2 MiB stacks).
28300 let sql = format!("SELECT 1{}", " + 1".repeat(200));
28301 parse(&sql);
28302 }
28303
28304 #[test]
28305 fn in_list_unaffected_by_chain_budget() {
28306 // Flat InList: 20k elements parse fine and stay flat.
28307 let items: alloc::vec::Vec<String> = (0..20_000).map(|k| k.to_string()).collect();
28308 let sql = format!("SELECT 1 WHERE 5 IN ({})", items.join(","));
28309 let Statement::Select(s) = parse(&sql) else {
28310 panic!("expected select")
28311 };
28312 let Some(Expr::InList { list, negated, .. }) = s.where_ else {
28313 panic!("expected flat InList, got {:?}", s.where_)
28314 };
28315 assert_eq!(list.len(), 20_000);
28316 assert!(!negated);
28317 }
28318
28319 fn lit_int(n: i64) -> Expr {
28320 Expr::Literal(Literal::Integer(n))
28321 }
28322
28323 fn col(name: &str) -> Expr {
28324 Expr::Column(ColumnName {
28325 qualifier: None,
28326 name: name.into(),
28327 })
28328 }
28329
28330 #[test]
28331 fn select_single_integer() {
28332 let s = parse("SELECT 1");
28333 let Statement::Select(s) = s else {
28334 panic!("expected SELECT")
28335 };
28336 assert_eq!(s.items.len(), 1);
28337 assert!(s.from.is_none());
28338 assert!(s.where_.is_none());
28339 }
28340
28341 #[test]
28342 fn select_multiple_literal_kinds() {
28343 let s = parse("SELECT 1, 'hi', NULL, TRUE, 1.5");
28344 let Statement::Select(s) = s else {
28345 panic!("expected SELECT")
28346 };
28347 assert_eq!(s.items.len(), 5);
28348 }
28349
28350 #[test]
28351 fn select_wildcard_from_table() {
28352 let s = parse("SELECT * FROM users");
28353 let Statement::Select(s) = s else {
28354 panic!("expected SELECT")
28355 };
28356 assert!(matches!(s.items[..], [SelectItem::Wildcard]));
28357 assert_eq!(s.from.as_ref().unwrap().primary.name, "users");
28358 }
28359
28360 #[test]
28361 fn select_with_table_alias() {
28362 let s = parse("SELECT * FROM users AS u");
28363 let Statement::Select(s) = s else {
28364 panic!("expected SELECT")
28365 };
28366 let t = &s.from.as_ref().unwrap().primary;
28367 assert_eq!(t.name, "users");
28368 assert_eq!(t.alias.as_deref(), Some("u"));
28369 }
28370
28371 #[test]
28372 fn select_with_where_eq() {
28373 let s = parse("SELECT a FROM t WHERE a = 1");
28374 let Statement::Select(s) = s else {
28375 panic!("expected SELECT")
28376 };
28377 let w = s.where_.unwrap();
28378 assert_eq!(
28379 w,
28380 Expr::Binary {
28381 lhs: Box::new(col("a")),
28382 op: BinOp::Eq,
28383 rhs: Box::new(lit_int(1)),
28384 }
28385 );
28386 }
28387
28388 #[test]
28389 fn arithmetic_precedence() {
28390 let s = parse("SELECT 1 + 2 * 3");
28391 let Statement::Select(s) = s else {
28392 panic!("expected SELECT")
28393 };
28394 let SelectItem::Expr { expr, .. } = &s.items[0] else {
28395 panic!("wildcard?")
28396 };
28397 assert_eq!(
28398 expr,
28399 &Expr::Binary {
28400 lhs: Box::new(lit_int(1)),
28401 op: BinOp::Add,
28402 rhs: Box::new(Expr::Binary {
28403 lhs: Box::new(lit_int(2)),
28404 op: BinOp::Mul,
28405 rhs: Box::new(lit_int(3)),
28406 }),
28407 }
28408 );
28409 }
28410
28411 #[test]
28412 fn parentheses_override_precedence() {
28413 let s = parse("SELECT (1 + 2) * 3");
28414 let Statement::Select(s) = s else {
28415 panic!("expected SELECT")
28416 };
28417 let SelectItem::Expr { expr, .. } = &s.items[0] else {
28418 panic!()
28419 };
28420 assert_eq!(
28421 expr,
28422 &Expr::Binary {
28423 lhs: Box::new(Expr::Binary {
28424 lhs: Box::new(lit_int(1)),
28425 op: BinOp::Add,
28426 rhs: Box::new(lit_int(2)),
28427 }),
28428 op: BinOp::Mul,
28429 rhs: Box::new(lit_int(3)),
28430 }
28431 );
28432 }
28433
28434 #[test]
28435 fn not_binds_below_comparison() {
28436 // `NOT a = 1` should parse as `NOT (a = 1)`.
28437 let s = parse("SELECT NOT a = 1 FROM t");
28438 let Statement::Select(s) = s else {
28439 panic!("expected SELECT")
28440 };
28441 let SelectItem::Expr { expr, .. } = &s.items[0] else {
28442 panic!()
28443 };
28444 assert_eq!(
28445 expr,
28446 &Expr::Unary {
28447 op: UnOp::Not,
28448 expr: Box::new(Expr::Binary {
28449 lhs: Box::new(col("a")),
28450 op: BinOp::Eq,
28451 rhs: Box::new(lit_int(1)),
28452 }),
28453 }
28454 );
28455 }
28456
28457 #[test]
28458 fn unary_minus_binds_above_multiplication() {
28459 // `-a * 2` should be `(-a) * 2`.
28460 let s = parse("SELECT -a * 2 FROM t");
28461 let Statement::Select(s) = s else {
28462 panic!("expected SELECT")
28463 };
28464 let SelectItem::Expr { expr, .. } = &s.items[0] else {
28465 panic!()
28466 };
28467 assert_eq!(
28468 expr,
28469 &Expr::Binary {
28470 lhs: Box::new(Expr::Unary {
28471 op: UnOp::Neg,
28472 expr: Box::new(col("a")),
28473 }),
28474 op: BinOp::Mul,
28475 rhs: Box::new(lit_int(2)),
28476 }
28477 );
28478 }
28479
28480 #[test]
28481 fn qualified_column() {
28482 let s = parse("SELECT t.col FROM t");
28483 let Statement::Select(s) = s else {
28484 panic!("expected SELECT")
28485 };
28486 let SelectItem::Expr { expr, .. } = &s.items[0] else {
28487 panic!()
28488 };
28489 assert_eq!(
28490 expr,
28491 &Expr::Column(ColumnName {
28492 qualifier: Some("t".into()),
28493 name: "col".into()
28494 })
28495 );
28496 }
28497
28498 #[test]
28499 fn select_item_alias_with_as() {
28500 let s = parse("SELECT a AS y FROM t");
28501 let Statement::Select(s) = s else {
28502 panic!("expected SELECT")
28503 };
28504 let SelectItem::Expr { alias, .. } = &s.items[0] else {
28505 panic!()
28506 };
28507 assert_eq!(alias.as_deref(), Some("y"));
28508 }
28509
28510 #[test]
28511 fn trailing_semicolon_accepted() {
28512 let s = parse("SELECT 1;");
28513 let Statement::Select(s) = s else {
28514 panic!("expected SELECT")
28515 };
28516 assert_eq!(s.items.len(), 1);
28517 }
28518
28519 #[test]
28520 fn boolean_chain_with_and_or_not() {
28521 // (NOT a) OR (b AND (NOT c))
28522 let s = parse("SELECT NOT a OR b AND NOT c FROM t");
28523 let Statement::Select(s) = s else {
28524 panic!("expected SELECT")
28525 };
28526 let SelectItem::Expr { expr, .. } = &s.items[0] else {
28527 panic!()
28528 };
28529 let expected = Expr::Binary {
28530 lhs: Box::new(Expr::Unary {
28531 op: UnOp::Not,
28532 expr: Box::new(col("a")),
28533 }),
28534 op: BinOp::Or,
28535 rhs: Box::new(Expr::Binary {
28536 lhs: Box::new(col("b")),
28537 op: BinOp::And,
28538 rhs: Box::new(Expr::Unary {
28539 op: UnOp::Not,
28540 expr: Box::new(col("c")),
28541 }),
28542 }),
28543 };
28544 assert_eq!(expr, &expected);
28545 }
28546
28547 #[test]
28548 fn empty_input_errors() {
28549 // v7.14.0 — pg_dump preambles emit several comment-only
28550 // / blank-line statements that collapse to Statement::
28551 // Empty rather than a parse error. The old "SELECT in
28552 // message" assertion is stale; verify the new contract:
28553 // empty / whitespace / comment-only input parses to
28554 // Statement::Empty.
28555 assert!(matches!(parse_statement("").unwrap(), Statement::Empty));
28556 assert!(matches!(
28557 parse_statement(" \n\t ").unwrap(),
28558 Statement::Empty
28559 ));
28560 // Sanity: malformed-but-non-empty still errors.
28561 assert!(parse_statement("SELECT FROM WHERE").is_err());
28562 }
28563
28564 #[test]
28565 fn unmatched_paren_errors() {
28566 assert!(parse_statement("SELECT (1 + 2").is_err());
28567 }
28568
28569 #[test]
28570 fn display_round_trip_simple_select() {
28571 let original = parse("SELECT a + 1 FROM t WHERE a > 0");
28572 let text = original.to_string();
28573 let again = parse_statement(&text).expect("re-parse");
28574 assert_eq!(original, again);
28575 }
28576
28577 // --- CREATE TABLE & INSERT (v0.3) ---------------------------------------
28578
28579 #[test]
28580 fn create_table_single_column() {
28581 let s = parse("CREATE TABLE foo (a INT)");
28582 let Statement::CreateTable(c) = s else {
28583 panic!("expected CreateTable")
28584 };
28585 assert_eq!(c.name, "foo");
28586 assert_eq!(c.columns.len(), 1);
28587 assert_eq!(c.columns[0].name, "a");
28588 assert_eq!(c.columns[0].ty, ColumnTypeName::Int);
28589 assert!(c.columns[0].nullable);
28590 }
28591
28592 #[test]
28593 fn create_table_multi_column_with_not_null_mix() {
28594 let s = parse("CREATE TABLE u (id INT NOT NULL, name TEXT, score FLOAT NOT NULL, ok BOOL)");
28595 let Statement::CreateTable(c) = s else {
28596 panic!()
28597 };
28598 assert_eq!(c.columns.len(), 4);
28599 assert_eq!(c.columns[0].ty, ColumnTypeName::Int);
28600 assert!(!c.columns[0].nullable);
28601 assert_eq!(c.columns[1].ty, ColumnTypeName::Text);
28602 assert!(c.columns[1].nullable);
28603 assert_eq!(c.columns[2].ty, ColumnTypeName::Float);
28604 assert!(!c.columns[2].nullable);
28605 assert_eq!(c.columns[3].ty, ColumnTypeName::Bool);
28606 }
28607
28608 #[test]
28609 fn create_table_bigint_supported() {
28610 let s = parse("CREATE TABLE accounts (id BIGINT NOT NULL)");
28611 let Statement::CreateTable(c) = s else {
28612 panic!()
28613 };
28614 assert_eq!(c.columns[0].ty, ColumnTypeName::BigInt);
28615 }
28616
28617 #[test]
28618 fn create_table_vector_default_is_f32() {
28619 let s = parse("CREATE TABLE t (v VECTOR(128))");
28620 let Statement::CreateTable(c) = s else {
28621 panic!()
28622 };
28623 assert_eq!(
28624 c.columns[0].ty,
28625 ColumnTypeName::Vector {
28626 dim: 128,
28627 encoding: VecEncoding::F32,
28628 },
28629 );
28630 }
28631
28632 #[test]
28633 fn create_table_vector_using_sq8() {
28634 // v6.0.1: `USING SQ8` selects scalar-quantised encoding.
28635 // Case-insensitive on both `USING` and the encoding name.
28636 for sql in [
28637 "CREATE TABLE t (v VECTOR(128) USING SQ8)",
28638 "CREATE TABLE t (v VECTOR(128) using sq8)",
28639 ] {
28640 let s = parse(sql);
28641 let Statement::CreateTable(c) = s else {
28642 panic!()
28643 };
28644 assert_eq!(
28645 c.columns[0].ty,
28646 ColumnTypeName::Vector {
28647 dim: 128,
28648 encoding: VecEncoding::Sq8,
28649 },
28650 "{sql}",
28651 );
28652 }
28653 }
28654
28655 #[test]
28656 fn create_table_vector_using_unknown_errors() {
28657 // v7.16.1 — the inline `USING <encoding>` shape on
28658 // CREATE TABLE column defs was withdrawn before
28659 // v7.14.0 in favour of `CREATE INDEX … USING hnsw
28660 // (col vector_<metric>_ops)`; the parser now rejects
28661 // USING at column-list position with a clearer
28662 // "expected ',' or ')'" message. Test asserts the
28663 // current rejection, not the old "unknown vector
28664 // encoding" string.
28665 let err = parse_statement("CREATE TABLE t (v VECTOR(8) USING PQ8)").unwrap_err();
28666 assert!(
28667 err.message.contains("USING")
28668 || err.message.contains("using")
28669 || err.message.contains("')'")
28670 || err.message.contains("','"),
28671 "expected USING/column-list rejection, got: {}",
28672 err.message
28673 );
28674 }
28675
28676 #[test]
28677 fn vector_using_sq8_display_roundtrips() {
28678 // The Display impl must produce text that re-parses to the
28679 // same AST. Guard for the v6.0.1 `USING SQ8` suffix.
28680 let s = parse("CREATE TABLE t (v VECTOR(64) USING SQ8)");
28681 let Statement::CreateTable(c) = s else {
28682 panic!()
28683 };
28684 assert_eq!(c.columns[0].ty.to_string(), "VECTOR(64) USING SQ8");
28685 }
28686
28687 #[test]
28688 fn parser_recognises_placeholders() {
28689 use crate::ast::{Expr, SelectItem, Statement};
28690 // $N in expression position parses as Expr::Placeholder(N).
28691 let s = parse("SELECT $1, $2 + 1 FROM t WHERE x = $3");
28692 let Statement::Select(sel) = s else { panic!() };
28693 assert!(matches!(
28694 sel.items[0],
28695 SelectItem::Expr {
28696 expr: Expr::Placeholder(1),
28697 alias: None
28698 }
28699 ));
28700 // $2 + 1
28701 let SelectItem::Expr {
28702 expr: Expr::Binary { lhs, rhs, .. },
28703 ..
28704 } = &sel.items[1]
28705 else {
28706 panic!()
28707 };
28708 assert!(matches!(**lhs, Expr::Placeholder(2)));
28709 assert!(matches!(**rhs, Expr::Literal(Literal::Integer(1))));
28710 // WHERE x = $3
28711 let Some(Expr::Binary { rhs, .. }) = sel.where_.as_ref() else {
28712 panic!()
28713 };
28714 assert!(matches!(**rhs, Expr::Placeholder(3)));
28715 }
28716
28717 #[test]
28718 fn parser_rejects_dollar_zero() {
28719 // $0 is not valid in PG; the lexer rejects it.
28720 assert!(parse_statement("SELECT $0").is_err());
28721 }
28722
28723 #[test]
28724 fn placeholder_display_roundtrips() {
28725 // The Display impl must produce text that re-lexes to the
28726 // same Placeholder token.
28727 let s = parse("SELECT $42 FROM t");
28728 let printed = s.to_string();
28729 assert!(printed.contains("$42"));
28730 let again = parse(&printed);
28731 assert_eq!(s, again);
28732 }
28733
28734 #[test]
28735 fn alter_index_rebuild_bare() {
28736 use crate::ast::{AlterIndexTarget, Statement};
28737 let s = parse("ALTER INDEX my_idx REBUILD");
28738 let Statement::AlterIndex(a) = s else {
28739 panic!("expected AlterIndex, got {s:?}")
28740 };
28741 assert_eq!(a.name, "my_idx");
28742 assert_eq!(a.target, AlterIndexTarget::Rebuild { encoding: None });
28743 }
28744
28745 #[test]
28746 fn alter_index_rebuild_with_encoding() {
28747 use crate::ast::{AlterIndexTarget, Statement};
28748 for (sql, want) in [
28749 (
28750 "ALTER INDEX my_idx REBUILD WITH (encoding = F32)",
28751 VecEncoding::F32,
28752 ),
28753 (
28754 "ALTER INDEX my_idx REBUILD WITH (encoding = sq8)",
28755 VecEncoding::Sq8,
28756 ),
28757 (
28758 "ALTER INDEX my_idx REBUILD WITH (encoding = HALF)",
28759 VecEncoding::F16,
28760 ),
28761 ] {
28762 let s = parse(sql);
28763 let Statement::AlterIndex(a) = s else {
28764 panic!("{sql}: expected AlterIndex")
28765 };
28766 assert_eq!(a.name, "my_idx");
28767 assert_eq!(
28768 a.target,
28769 AlterIndexTarget::Rebuild {
28770 encoding: Some(want)
28771 },
28772 "{sql}"
28773 );
28774 }
28775 }
28776
28777 #[test]
28778 fn alter_index_rebuild_unknown_encoding_errors() {
28779 let err = parse_statement("ALTER INDEX my_idx REBUILD WITH (encoding = PQ8)").unwrap_err();
28780 assert!(
28781 err.message.contains("unknown vector encoding"),
28782 "got: {}",
28783 err.message
28784 );
28785 }
28786
28787 #[test]
28788 fn alter_index_rebuild_display_roundtrips() {
28789 for (input, want) in [
28790 ("ALTER INDEX my_idx REBUILD", "ALTER INDEX my_idx REBUILD"),
28791 (
28792 "ALTER INDEX my_idx REBUILD WITH (encoding = SQ8)",
28793 "ALTER INDEX my_idx REBUILD WITH (encoding = SQ8)",
28794 ),
28795 (
28796 "ALTER INDEX my_idx REBUILD WITH (encoding = HALF)",
28797 "ALTER INDEX my_idx REBUILD WITH (encoding = HALF)",
28798 ),
28799 ] {
28800 let s = parse(input);
28801 assert_eq!(s.to_string(), want);
28802 }
28803 }
28804
28805 #[test]
28806 fn create_table_unknown_type_defers_to_engine() {
28807 // v4.9 picked XML as a parse-time "unsupported column
28808 // type" probe. v7.17.0 Phase 1.4 changed the contract:
28809 // an unknown type ident parses as Text + `user_type_ref`
28810 // so CREATE TABLE can resolve user-defined enum / domain
28811 // types — rejection of truly-unknown types moved to the
28812 // engine's catalog lookup. v7.37.5 ζ-A then promoted XML
28813 // to a first-class built-in, so this probe switched to a
28814 // synthetic name nothing in the lexer will ever recognise.
28815 let stmt = parse_statement("CREATE TABLE x (a my_user_type)").unwrap();
28816 let Statement::CreateTable(t) = stmt else {
28817 panic!("expected CreateTable");
28818 };
28819 assert_eq!(t.columns[0].user_type_ref.as_deref(), Some("my_user_type"));
28820 }
28821
28822 #[test]
28823 fn create_table_missing_table_keyword_errors() {
28824 assert!(parse_statement("CREATE x (a INT)").is_err());
28825 }
28826
28827 // v7.37.6-B(sentori Epic 2 P0)— `PARTITION BY RANGE` parent +
28828 // `PARTITION OF parent <bounds>` child parse + Display round-trip.
28829
28830 #[test]
28831 fn parse_create_table_partition_by_range() {
28832 use crate::ast::{PartitionBySpec, PartitionKindAst};
28833 let stmt = parse_statement(
28834 "CREATE TABLE events_partitioned (id BIGINT NOT NULL, ts TIMESTAMPTZ NOT NULL, \
28835 payload JSONB) PARTITION BY RANGE (ts)",
28836 )
28837 .unwrap();
28838 let Statement::CreateTable(t) = stmt else {
28839 panic!("expected CreateTable");
28840 };
28841 assert!(t.partition_of.is_none(), "parent has no partition_of");
28842 assert_eq!(t.columns.len(), 3);
28843 let by = t.partition_by.as_ref().expect("expected PARTITION BY");
28844 assert_eq!(
28845 by,
28846 &PartitionBySpec {
28847 kind: PartitionKindAst::Range,
28848 key_columns: alloc::vec!["ts".to_string()],
28849 }
28850 );
28851 // Display round-trip preserves the suffix. `quote_ident`
28852 // only adds double quotes when the ident needs escaping, so
28853 // a plain `ts` survives bare here.
28854 assert!(
28855 t.to_string().contains("PARTITION BY RANGE (ts)"),
28856 "Display lost PARTITION BY suffix: {t}"
28857 );
28858 }
28859
28860 #[test]
28861 fn parse_create_table_partition_of_range() {
28862 use crate::ast::{PartitionOfBoundsAst, PartitionOfSpec};
28863 let stmt = parse_statement(
28864 "CREATE TABLE events_2026_06 PARTITION OF events_partitioned \
28865 FOR VALUES FROM ('2026-06-01 00:00:00+00') TO ('2026-07-01 00:00:00+00')",
28866 )
28867 .unwrap();
28868 let Statement::CreateTable(t) = stmt else {
28869 panic!("expected CreateTable");
28870 };
28871 assert!(t.columns.is_empty(), "child inherits columns from parent");
28872 assert!(t.partition_by.is_none());
28873 let of = t.partition_of.as_ref().expect("expected PARTITION OF");
28874 assert_eq!(of.parent_name, "events_partitioned");
28875 let PartitionOfSpec { bounds, .. } = of.clone();
28876 match bounds {
28877 PartitionOfBoundsAst::Range { lower, upper } => {
28878 assert!(lower.to_string().contains("2026-06-01"));
28879 assert!(upper.to_string().contains("2026-07-01"));
28880 }
28881 other => panic!("expected Range, got {other:?}"),
28882 }
28883 // Display round-trip emits the FOR VALUES tail. `quote_ident`
28884 // skips quotes when not required, so the parent name appears
28885 // bare here.
28886 let s = t.to_string();
28887 assert!(
28888 s.contains("PARTITION OF events_partitioned"),
28889 "Display lost PARTITION OF: {s}"
28890 );
28891 assert!(s.contains("FOR VALUES FROM"), "Display lost FROM: {s}");
28892 assert!(s.contains(") TO ("), "Display lost TO: {s}");
28893 }
28894
28895 #[test]
28896 fn parse_create_table_partition_of_default() {
28897 use crate::ast::PartitionOfBoundsAst;
28898 let stmt =
28899 parse_statement("CREATE TABLE events_default PARTITION OF events_partitioned DEFAULT")
28900 .unwrap();
28901 let Statement::CreateTable(t) = stmt else {
28902 panic!("expected CreateTable");
28903 };
28904 let of = t.partition_of.as_ref().expect("expected PARTITION OF");
28905 assert_eq!(of.parent_name, "events_partitioned");
28906 assert!(matches!(of.bounds, PartitionOfBoundsAst::Default));
28907 assert!(
28908 t.to_string()
28909 .contains("PARTITION OF events_partitioned DEFAULT"),
28910 "Display lost DEFAULT: {t}"
28911 );
28912 }
28913
28914 #[test]
28915 fn parse_create_table_partition_by_list() {
28916 // v7.37.16 (16.1) — `PARTITION BY LIST (key)` parent + a
28917 // child with `FOR VALUES IN (lit, lit, …)`.
28918 use crate::ast::{PartitionBySpec, PartitionKindAst, PartitionOfBoundsAst};
28919 let parent =
28920 parse_statement("CREATE TABLE events_listed (region TEXT) PARTITION BY LIST (region)")
28921 .unwrap();
28922 let Statement::CreateTable(t) = parent else {
28923 panic!("expected CreateTable");
28924 };
28925 let Some(PartitionBySpec {
28926 kind,
28927 ref key_columns,
28928 }) = t.partition_by
28929 else {
28930 panic!("expected PARTITION BY");
28931 };
28932 assert_eq!(kind, PartitionKindAst::List);
28933 assert_eq!(*key_columns, vec!["region".to_string()]);
28934 assert!(t.to_string().contains("PARTITION BY LIST (region)"));
28935
28936 let child = parse_statement(
28937 "CREATE TABLE events_apac PARTITION OF events_listed \
28938 FOR VALUES IN ('jp', 'kr', 'tw')",
28939 )
28940 .unwrap();
28941 let Statement::CreateTable(c) = child else {
28942 panic!("expected CreateTable");
28943 };
28944 let of = c.partition_of.as_ref().expect("expected PARTITION OF");
28945 let PartitionOfBoundsAst::List { values } = &of.bounds else {
28946 panic!("expected List bounds, got {:?}", of.bounds);
28947 };
28948 assert_eq!(values.len(), 3);
28949 let disp = c.to_string();
28950 assert!(disp.contains("FOR VALUES IN ("), "Display lost IN: {disp}");
28951 }
28952
28953 #[test]
28954 fn parse_create_table_partition_by_hash() {
28955 // v7.37.16 (16.2) — `PARTITION BY HASH (key)` parent + a
28956 // child with `FOR VALUES WITH (MODULUS m, REMAINDER r)`.
28957 use crate::ast::{PartitionBySpec, PartitionKindAst, PartitionOfBoundsAst};
28958 let parent =
28959 parse_statement("CREATE TABLE orders_h (id BIGINT) PARTITION BY HASH (id)").unwrap();
28960 let Statement::CreateTable(t) = parent else {
28961 panic!("expected CreateTable");
28962 };
28963 let Some(PartitionBySpec {
28964 kind,
28965 ref key_columns,
28966 }) = t.partition_by
28967 else {
28968 panic!("expected PARTITION BY");
28969 };
28970 assert_eq!(kind, PartitionKindAst::Hash);
28971 assert_eq!(*key_columns, vec!["id".to_string()]);
28972 assert!(t.to_string().contains("PARTITION BY HASH (id)"));
28973
28974 let child = parse_statement(
28975 "CREATE TABLE orders_h_0 PARTITION OF orders_h \
28976 FOR VALUES WITH (MODULUS 4, REMAINDER 0)",
28977 )
28978 .unwrap();
28979 let Statement::CreateTable(c) = child else {
28980 panic!("expected CreateTable");
28981 };
28982 let of = c.partition_of.as_ref().expect("expected PARTITION OF");
28983 let PartitionOfBoundsAst::Hash { modulus, remainder } = of.bounds else {
28984 panic!("expected Hash bounds");
28985 };
28986 assert_eq!(modulus, 4);
28987 assert_eq!(remainder, 0);
28988 let disp = c.to_string();
28989 assert!(
28990 disp.contains("FOR VALUES WITH (MODULUS 4, REMAINDER 0)"),
28991 "Display lost HASH bounds: {disp}"
28992 );
28993
28994 // Validation: REMAINDER ≥ MODULUS is rejected at parse time.
28995 let bad = parse_statement(
28996 "CREATE TABLE orders_h_bad PARTITION OF orders_h \
28997 FOR VALUES WITH (MODULUS 4, REMAINDER 4)",
28998 );
28999 let msg = format!("{}", bad.unwrap_err());
29000 assert!(
29001 msg.contains("REMAINDER") && msg.contains("MODULUS"),
29002 "expected REMAINDER/MODULUS validation error: {msg}"
29003 );
29004 }
29005
29006 #[test]
29007 fn parse_create_table_partition_of_rejects_columns() {
29008 // v7.37.6-B contract: PARTITION OF children inherit columns
29009 // from the parent; an explicit list MUST surface as a parse
29010 // error rather than getting silently ignored.
29011 let err = parse_statement(
29012 "CREATE TABLE events_2026_06 PARTITION OF events_partitioned (id BIGINT) \
29013 FOR VALUES FROM ('a') TO ('b')",
29014 );
29015 assert!(err.is_err(), "expected parse error for explicit columns");
29016 let msg = format!("{}", err.unwrap_err());
29017 assert!(
29018 msg.contains("PARTITION OF") && msg.contains("column"),
29019 "error should mention PARTITION OF + columns: {msg}"
29020 );
29021 }
29022
29023 #[test]
29024 fn insert_single_value() {
29025 let s = parse("INSERT INTO foo VALUES (42)");
29026 let Statement::Insert(i) = s else {
29027 panic!("expected Insert")
29028 };
29029 assert_eq!(i.table, "foo");
29030 assert_eq!(i.rows.len(), 1);
29031 assert_eq!(i.rows[0].len(), 1);
29032 assert!(matches!(i.rows[0][0], Expr::Literal(Literal::Integer(42))));
29033 }
29034
29035 #[test]
29036 fn insert_multi_value_with_mixed_literals() {
29037 let s = parse("INSERT INTO foo VALUES (1, 'hi', 3.14, TRUE, NULL)");
29038 let Statement::Insert(i) = s else { panic!() };
29039 assert_eq!(i.rows.len(), 1);
29040 assert_eq!(i.rows[0].len(), 5);
29041 }
29042
29043 #[test]
29044 fn insert_missing_into_errors() {
29045 assert!(parse_statement("INSERT foo VALUES (1)").is_err());
29046 }
29047
29048 #[test]
29049 fn create_table_round_trip() {
29050 let original =
29051 parse("CREATE TABLE foo (id BIGINT NOT NULL, label TEXT, score FLOAT NOT NULL)");
29052 let text = original.to_string();
29053 let again = parse_statement(&text).expect("re-parse");
29054 assert_eq!(original, again);
29055 }
29056
29057 #[test]
29058 fn insert_round_trip_with_negation_and_string() {
29059 let original = parse("INSERT INTO t VALUES (-1, 'it''s', NULL)");
29060 let text = original.to_string();
29061 let again = parse_statement(&text).expect("re-parse");
29062 assert_eq!(original, again);
29063 }
29064
29065 #[test]
29066 fn unknown_keyword_at_statement_start_errors() {
29067 // v4.4: UPDATE is real SQL now. Use a fabricated keyword so
29068 // the top-level dispatch still has no branch to take.
29069 let err = parse_statement("FROBNICATE foo SET x = 1").unwrap_err();
29070 assert_eq!(err.message, "syntax error at or near \"FROBNICATE\"");
29071 }
29072
29073 // --- v0.8 CREATE INDEX --------------------------------------------------
29074
29075 #[test]
29076 fn create_index_basic() {
29077 let s = parse("CREATE INDEX idx_id ON users (id)");
29078 let Statement::CreateIndex(c) = s else {
29079 panic!("expected CreateIndex")
29080 };
29081 assert_eq!(c.name, "idx_id");
29082 assert_eq!(c.table, "users");
29083 assert_eq!(c.column, "id");
29084 }
29085
29086 #[test]
29087 fn create_index_missing_on_errors() {
29088 assert!(parse_statement("CREATE INDEX foo users (id)").is_err());
29089 }
29090
29091 #[test]
29092 fn create_index_missing_paren_errors() {
29093 assert!(parse_statement("CREATE INDEX foo ON users id").is_err());
29094 }
29095
29096 #[test]
29097 fn create_index_round_trip() {
29098 let original = parse("CREATE INDEX by_name ON users (name)");
29099 let again = parse_statement(&original.to_string()).unwrap();
29100 assert_eq!(original, again);
29101 }
29102
29103 // --- v7.9.29 CREATE UNIQUE INDEX [WHERE pred] (mailrs K1) -------------
29104
29105 #[test]
29106 fn create_unique_index_basic() {
29107 let s = parse("CREATE UNIQUE INDEX uq_x ON t (a)");
29108 let Statement::CreateIndex(c) = s else {
29109 panic!("expected CreateIndex");
29110 };
29111 assert!(c.is_unique);
29112 assert_eq!(c.column, "a");
29113 assert!(c.partial_predicate.is_none());
29114 }
29115
29116 #[test]
29117 fn create_unique_index_partial() {
29118 // mailrs's email_templates "one default per user" shape.
29119 let s = parse(
29120 "CREATE UNIQUE INDEX idx_email_templates_user_default \
29121 ON email_templates (user_address) WHERE is_default = true",
29122 );
29123 let Statement::CreateIndex(c) = s else {
29124 panic!("expected CreateIndex");
29125 };
29126 assert!(c.is_unique);
29127 assert_eq!(c.table, "email_templates");
29128 assert_eq!(c.column, "user_address");
29129 assert!(c.partial_predicate.is_some());
29130 }
29131
29132 #[test]
29133 fn create_unique_index_composite_with_predicate() {
29134 // mailrs's calendar_events instance: composite columns.
29135 let s = parse(
29136 "CREATE UNIQUE INDEX uq_calendar_events_instance \
29137 ON calendar_events (calendar_id, uid, recurrence_id) \
29138 WHERE recurrence_id IS NOT NULL",
29139 );
29140 let Statement::CreateIndex(c) = s else {
29141 panic!("expected CreateIndex");
29142 };
29143 assert!(c.is_unique);
29144 assert_eq!(c.column, "calendar_id");
29145 assert_eq!(
29146 c.extra_columns,
29147 vec!["uid".to_string(), "recurrence_id".to_string()]
29148 );
29149 assert!(c.partial_predicate.is_some());
29150 }
29151
29152 #[test]
29153 fn create_unique_index_using_btree_ok() {
29154 let s = parse("CREATE UNIQUE INDEX uq_x ON t USING btree (a)");
29155 assert!(matches!(s, Statement::CreateIndex(ref c) if c.is_unique));
29156 }
29157
29158 #[test]
29159 fn create_unique_index_using_hnsw_rejected() {
29160 let err =
29161 parse_statement("CREATE UNIQUE INDEX uq_v ON t USING hnsw (embedding)").unwrap_err();
29162 assert!(err.message.contains("UNIQUE"), "{}", err.message);
29163 }
29164
29165 #[test]
29166 fn create_unique_index_round_trip() {
29167 let original = parse(
29168 "CREATE UNIQUE INDEX uq_calendar_events_master \
29169 ON calendar_events (calendar_id, uid) WHERE recurrence_id IS NULL",
29170 );
29171 let again = parse_statement(&original.to_string()).unwrap();
29172 assert_eq!(original, again);
29173 }
29174
29175 #[test]
29176 fn create_unique_without_index_errors() {
29177 let err = parse_statement("CREATE UNIQUE TABLE t (a INT)").unwrap_err();
29178 // v7.39 (round 340, V56) — PG 18.4, verbatim.
29179 assert_eq!(err.message, "syntax error at or near \"TABLE\"");
29180 }
29181
29182 // --- v7.10.4 BYTES / BYTEA column type (Epic 1) ----------------------
29183
29184 #[test]
29185 fn create_table_bytea_column() {
29186 let s = parse("CREATE TABLE t (id INT NOT NULL, payload BYTEA NOT NULL)");
29187 let Statement::CreateTable(c) = s else {
29188 panic!("expected CreateTable");
29189 };
29190 assert_eq!(c.columns.len(), 2);
29191 assert_eq!(c.columns[1].ty, ColumnTypeName::Bytes);
29192 assert!(!c.columns[1].nullable);
29193 }
29194
29195 #[test]
29196 fn create_table_bytes_alias_column() {
29197 let s = parse("CREATE TABLE t (blob BYTES)");
29198 let Statement::CreateTable(c) = s else {
29199 panic!("expected CreateTable");
29200 };
29201 assert_eq!(c.columns[0].ty, ColumnTypeName::Bytes);
29202 }
29203
29204 #[test]
29205 fn bytea_round_trip_display() {
29206 let original = parse("CREATE TABLE t (a BYTEA NOT NULL)");
29207 let again = parse_statement(&original.to_string()).unwrap();
29208 assert_eq!(original, again);
29209 }
29210
29211 // --- v0.9 transactions -------------------------------------------------
29212
29213 #[test]
29214 fn begin_commit_rollback_parse_as_unit_variants() {
29215 let plain = crate::ast::TransactionModes::default();
29216 assert_eq!(parse("BEGIN"), Statement::Begin(plain));
29217 assert_eq!(parse("COMMIT"), Statement::Commit);
29218 // r1066 — PG synonyms pgbench's tpcb script relies on.
29219 assert_eq!(parse("END"), Statement::Commit);
29220 assert_eq!(parse("END TRANSACTION"), Statement::Commit);
29221 assert_eq!(parse("COMMIT WORK"), Statement::Commit);
29222 assert_eq!(parse("ROLLBACK"), Statement::Rollback);
29223 // Trailing semicolons accepted too.
29224 assert_eq!(parse("BEGIN;"), Statement::Begin(plain));
29225 // v7.39 (read01 round 118, B3) — an explicit ISOLATION LEVEL rides the
29226 // statement (with or without the WORK/TRANSACTION noise word).
29227 assert_eq!(
29228 parse("BEGIN ISOLATION LEVEL REPEATABLE READ"),
29229 Statement::Begin(crate::ast::TransactionModes {
29230 isolation: Some(IsolationLevel::RepeatableRead),
29231 read_only: None,
29232 })
29233 );
29234 assert_eq!(
29235 parse("START TRANSACTION ISOLATION LEVEL SERIALIZABLE"),
29236 Statement::Begin(crate::ast::TransactionModes {
29237 isolation: Some(IsolationLevel::Serializable),
29238 read_only: None,
29239 })
29240 );
29241 // v7.39 — this line used to read
29242 //
29243 // // A non-isolation mode keeps the session default (None).
29244 // assert_eq!(parse("BEGIN READ ONLY"), Statement::Begin(None));
29245 //
29246 // which pinned the defect rather than catching it: the READ ONLY
29247 // was thrown away, so the statement opened an ordinary read-write
29248 // transaction and every write inside it was accepted. The
29249 // isolation level is still absent here, because this statement
29250 // does not name one — that part was right.
29251 assert_eq!(
29252 parse("BEGIN READ ONLY"),
29253 Statement::Begin(crate::ast::TransactionModes {
29254 isolation: None,
29255 read_only: Some(true),
29256 })
29257 );
29258 assert_eq!(
29259 parse("START TRANSACTION READ WRITE"),
29260 Statement::Begin(crate::ast::TransactionModes {
29261 isolation: None,
29262 read_only: Some(false),
29263 })
29264 );
29265 assert_eq!(
29266 parse("BEGIN ISOLATION LEVEL SERIALIZABLE, READ ONLY"),
29267 Statement::Begin(crate::ast::TransactionModes {
29268 isolation: Some(IsolationLevel::Serializable),
29269 read_only: Some(true),
29270 })
29271 );
29272 }
29273
29274 // --- v1.2: pgvector distance ops + ::vector cast --------------------
29275
29276 #[test]
29277 fn inner_product_binop_parses() {
29278 let s = parse("SELECT v <#> [1.0, 2.0] FROM t");
29279 let Statement::Select(s) = s else { panic!() };
29280 let SelectItem::Expr { expr, .. } = &s.items[0] else {
29281 panic!()
29282 };
29283 assert!(matches!(
29284 expr,
29285 Expr::Binary {
29286 op: BinOp::InnerProduct,
29287 ..
29288 }
29289 ));
29290 }
29291
29292 #[test]
29293 fn cosine_distance_binop_parses() {
29294 let s = parse("SELECT v <=> [1.0, 2.0] FROM t");
29295 let Statement::Select(s) = s else { panic!() };
29296 let SelectItem::Expr { expr, .. } = &s.items[0] else {
29297 panic!()
29298 };
29299 assert!(matches!(
29300 expr,
29301 Expr::Binary {
29302 op: BinOp::CosineDistance,
29303 ..
29304 }
29305 ));
29306 }
29307
29308 #[test]
29309 fn vector_cast_postfix_wraps_string_literal() {
29310 let s = parse("SELECT '[1,2,3]'::vector FROM t");
29311 let Statement::Select(s) = s else { panic!() };
29312 let SelectItem::Expr { expr, .. } = &s.items[0] else {
29313 panic!()
29314 };
29315 assert!(matches!(
29316 expr,
29317 Expr::Cast {
29318 target: CastTarget::Vector,
29319 ..
29320 }
29321 ));
29322 }
29323
29324 #[test]
29325 fn unsupported_cast_target_errors() {
29326 // v7.37.5 ship triage promoted the parser to accept every
29327 // ident as a `CastTarget::Named(canonical)`; the engine
29328 // surfaces the "unsupported cast target" error at eval
29329 // time when `type_name_to_data_type` can't resolve it.
29330 // Parser-side error now requires a NON-ident after `::`
29331 // (e.g. a punctuation token).
29332 let err = parse_statement("SELECT 1::, FROM t").unwrap_err();
29333 assert_eq!(err.message, "syntax error at or near \",\"");
29334 }
29335
29336 #[test]
29337 fn tx_statements_round_trip() {
29338 for q in ["BEGIN", "COMMIT", "ROLLBACK"] {
29339 let original = parse(q);
29340 let again = parse_statement(&original.to_string()).unwrap();
29341 assert_eq!(original, again);
29342 }
29343 }
29344
29345 #[test]
29346 fn interval_text_parsing_units() {
29347 // v7.37.5 β — three-field shape `(months, days, micros)` so
29348 // `'1 day'` and `'24 hours'` no longer collide (PG parity).
29349 // Single unit.
29350 assert_eq!(parse_interval_text("1 day"), Some((0, 1, 0)));
29351 assert_eq!(
29352 parse_interval_text("24 hours"),
29353 Some((0, 0, 86_400_000_000))
29354 );
29355 assert_eq!(parse_interval_text("1 second"), Some((0, 0, 1_000_000)));
29356 assert_eq!(parse_interval_text("1 month"), Some((1, 0, 0)));
29357 assert_eq!(parse_interval_text("2 years"), Some((24, 0, 0)));
29358 assert_eq!(parse_interval_text("1 week"), Some((0, 7, 0)));
29359 // Compound spans accumulate per-dimension.
29360 assert_eq!(parse_interval_text("1 year 6 months"), Some((18, 0, 0)));
29361 assert_eq!(
29362 parse_interval_text("1 day 2 hours"),
29363 Some((0, 1, 7_200_000_000))
29364 );
29365 // Negative numbers carry through per-dimension.
29366 assert_eq!(parse_interval_text("-1 day"), Some((0, -1, 0)));
29367 // Bad shapes return None.
29368 assert_eq!(parse_interval_text(""), None);
29369 assert_eq!(parse_interval_text("garbage"), None);
29370 assert_eq!(parse_interval_text("1 fortnight"), None);
29371 // v7.39 (GUC knife 3) — PG reads a bare number as SECONDS
29372 // (`INTERVAL '1'` = 00:00:01), verified against the oracle.
29373 assert_eq!(parse_interval_text("1"), Some((0, 0, 1_000_000)));
29374 assert_eq!(parse_interval_text("0"), Some((0, 0, 0)));
29375 assert_eq!(parse_interval_text("1.5"), Some((0, 0, 1_500_000)));
29376 }
29377
29378 #[test]
29379 fn interval_literal_roundtrips_via_display() {
29380 let parsed = parse("SELECT INTERVAL '1 day 2 hours'");
29381 let s = parsed.to_string();
29382 // Display preserves the original text verbatim.
29383 assert!(s.contains("INTERVAL '1 day 2 hours'"), "got: {s}");
29384 // And re-parsing yields a structurally equal statement.
29385 let again = parse_statement(&s).unwrap();
29386 assert_eq!(parsed, again);
29387 }
29388
29389 // ── v6.1.2: CREATE / DROP PUBLICATION ────────────────────
29390
29391 #[test]
29392 fn parser_recognises_create_publication_bare() {
29393 let s = parse("CREATE PUBLICATION pub_a");
29394 let Statement::CreatePublication(p) = s else {
29395 panic!("expected CreatePublication, got {s:?}")
29396 };
29397 assert_eq!(p.name, "pub_a");
29398 assert_eq!(p.scope, PublicationScope::AllTables);
29399 }
29400
29401 #[test]
29402 fn parser_recognises_create_publication_for_all_tables() {
29403 let s = parse("CREATE PUBLICATION pub_a FOR ALL TABLES");
29404 let Statement::CreatePublication(p) = s else {
29405 panic!("expected CreatePublication, got {s:?}")
29406 };
29407 assert_eq!(p.name, "pub_a");
29408 assert_eq!(p.scope, PublicationScope::AllTables);
29409 }
29410
29411 #[test]
29412 fn parser_recognises_drop_publication() {
29413 let s = parse("DROP PUBLICATION pub_a");
29414 let Statement::DropPublication { name, .. } = s else {
29415 panic!("expected DropPublication, got {s:?}")
29416 };
29417 assert_eq!(name, "pub_a");
29418 }
29419
29420 #[test]
29421 fn parser_recognises_for_table_list() {
29422 let s = parse("CREATE PUBLICATION pub_a FOR TABLE t1, t2, t3");
29423 let Statement::CreatePublication(p) = s else {
29424 panic!("expected CreatePublication, got {s:?}")
29425 };
29426 assert_eq!(p.name, "pub_a");
29427 let PublicationScope::ForTables(ts) = p.scope else {
29428 panic!("expected ForTables scope")
29429 };
29430 assert_eq!(ts, alloc::vec!["t1", "t2", "t3"]);
29431 }
29432
29433 #[test]
29434 fn parser_rejects_bare_for_tables_and_takes_in_schema() {
29435 // v7.39 (round 754, F31-B5) — PG18-measured: the bare plural
29436 // is rejected (`invalid publication object list`; the old
29437 // test pinned an unverifiable "PG 19 accepts both" claim);
29438 // TABLES pairs with IN SCHEMA.
29439 let err = parse_statement("CREATE PUBLICATION pub_a FOR TABLES t1, t2")
29440 .expect_err("bare FOR TABLES must reject");
29441 assert!(
29442 alloc::format!("{err}").contains("invalid publication object list"),
29443 "got: {err}"
29444 );
29445 let s = parse("CREATE PUBLICATION pub_a FOR TABLES IN SCHEMA public");
29446 let Statement::CreatePublication(p) = s else {
29447 panic!("expected CreatePublication, got {s:?}")
29448 };
29449 let PublicationScope::TablesInSchema(schema) = p.scope else {
29450 panic!("expected TablesInSchema")
29451 };
29452 assert_eq!(schema, "public");
29453 }
29454
29455 #[test]
29456 fn parser_recognises_for_all_tables_except_list() {
29457 let s = parse("CREATE PUBLICATION p FOR ALL TABLES EXCEPT t1, t2");
29458 let Statement::CreatePublication(p) = s else {
29459 panic!()
29460 };
29461 let PublicationScope::AllTablesExcept(ts) = p.scope else {
29462 panic!("expected AllTablesExcept")
29463 };
29464 assert_eq!(ts, alloc::vec!["t1", "t2"]);
29465 }
29466
29467 #[test]
29468 fn parser_rejects_for_table_with_empty_list() {
29469 // `FOR TABLE` with nothing after is a parse error.
29470 let err = parse_statement("CREATE PUBLICATION p FOR TABLE")
29471 .expect_err("must error on empty list");
29472 // No specific message asserted — the call falls through to
29473 // expect_ident_like which yields "expected identifier, got …".
29474 assert!(!err.message.is_empty());
29475 }
29476
29477 #[test]
29478 fn parser_recognises_show_publications() {
29479 // v6.1.3 — SHOW PUBLICATIONS lands here. PUBLICATIONS is a
29480 // bare ident in this position, NOT a reserved keyword.
29481 let s = parse("SHOW PUBLICATIONS");
29482 assert!(matches!(s, Statement::ShowPublications));
29483 }
29484
29485 // ── v6.1.4: CREATE / DROP SUBSCRIPTION + SHOW SUBSCRIPTIONS ─
29486
29487 #[test]
29488 fn parser_recognises_create_subscription_single_publication() {
29489 let s = parse(
29490 "CREATE SUBSCRIPTION sub_a CONNECTION 'host=127.0.0.1 port=20002' PUBLICATION pub_a",
29491 );
29492 let Statement::CreateSubscription(c) = s else {
29493 panic!("expected CreateSubscription, got {s:?}")
29494 };
29495 assert_eq!(c.name, "sub_a");
29496 assert_eq!(c.conn_str, "host=127.0.0.1 port=20002");
29497 assert_eq!(c.publications, alloc::vec!["pub_a"]);
29498 }
29499
29500 #[test]
29501 fn parser_recognises_create_subscription_multi_publication() {
29502 let s = parse("CREATE SUBSCRIPTION sub_a CONNECTION 'host=h' PUBLICATION p1, p2, p3");
29503 let Statement::CreateSubscription(c) = s else {
29504 panic!()
29505 };
29506 assert_eq!(c.publications, alloc::vec!["p1", "p2", "p3"]);
29507 }
29508
29509 #[test]
29510 fn parser_rejects_create_subscription_missing_connection() {
29511 let err = parse_statement("CREATE SUBSCRIPTION s PUBLICATION p")
29512 .expect_err("must error on missing CONNECTION");
29513 assert_eq!(err.message, "syntax error at or near \"PUBLICATION\"");
29514 }
29515
29516 #[test]
29517 fn parser_rejects_create_subscription_missing_publication() {
29518 let err = parse_statement("CREATE SUBSCRIPTION s CONNECTION 'host=x'")
29519 .expect_err("must error on missing PUBLICATION");
29520 assert_eq!(err.message, "syntax error at end of input");
29521 }
29522
29523 #[test]
29524 fn parser_recognises_drop_subscription() {
29525 let s = parse("DROP SUBSCRIPTION sub_a");
29526 let Statement::DropSubscription { name, .. } = s else {
29527 panic!("expected DropSubscription, got {s:?}")
29528 };
29529 assert_eq!(name, "sub_a");
29530 }
29531
29532 #[test]
29533 fn parser_recognises_show_subscriptions() {
29534 let s = parse("SHOW SUBSCRIPTIONS");
29535 assert!(matches!(s, Statement::ShowSubscriptions));
29536 }
29537
29538 #[test]
29539 fn parser_recognises_wait_for_wal_position_no_timeout() {
29540 let s = parse("WAIT FOR WAL POSITION 12345");
29541 let Statement::WaitForWalPosition { pos, timeout_ms } = s else {
29542 panic!("expected WaitForWalPosition, got {s:?}")
29543 };
29544 assert_eq!(pos, 12345);
29545 assert!(timeout_ms.is_none());
29546 }
29547
29548 #[test]
29549 fn parser_recognises_wait_for_wal_position_with_timeout() {
29550 let s = parse("WAIT FOR WAL POSITION 67890 WITH TIMEOUT 5000");
29551 let Statement::WaitForWalPosition { pos, timeout_ms } = s else {
29552 panic!()
29553 };
29554 assert_eq!(pos, 67890);
29555 assert_eq!(timeout_ms, Some(5000));
29556 }
29557
29558 #[test]
29559 fn parser_rejects_wait_with_negative_position() {
29560 // The lexer treats `-` as a token; `expect_u64_literal`
29561 // only sees the Integer that follows, so the negative
29562 // arrives as a unary-minus expression at higher levels.
29563 // Bare `WAIT FOR WAL POSITION -1` thus surfaces as a
29564 // parse error one way or another.
29565 let err = parse_statement("WAIT FOR WAL POSITION -1").unwrap_err();
29566 assert!(!err.message.is_empty());
29567 }
29568
29569 #[test]
29570 fn parser_recognises_bare_analyze() {
29571 let s = parse("ANALYZE");
29572 assert!(matches!(s, Statement::Analyze(None)));
29573 }
29574
29575 #[test]
29576 fn parser_recognises_analyze_with_table() {
29577 let s = parse("ANALYZE users");
29578 let Statement::Analyze(Some(name)) = s else {
29579 panic!("expected Analyze, got {s:?}")
29580 };
29581 assert_eq!(name, "users");
29582 }
29583
29584 #[test]
29585 fn parser_recognises_analyze_with_quoted_table() {
29586 let s = parse("ANALYZE \"Mixed Case\"");
29587 let Statement::Analyze(Some(name)) = s else {
29588 panic!()
29589 };
29590 assert_eq!(name, "Mixed Case");
29591 }
29592
29593 #[test]
29594 fn parser_rejects_analyze_with_garbage_token() {
29595 let err = parse_statement("ANALYZE 42").expect_err("must error");
29596 assert!(!err.message.is_empty());
29597 }
29598
29599 #[test]
29600 fn analyze_display_roundtrips() {
29601 for sql in ["ANALYZE", "ANALYZE users"] {
29602 let s = parse(sql);
29603 let printed = s.to_string();
29604 let again = parse_statement(&printed)
29605 .unwrap_or_else(|e| panic!("re-parse failed for {printed:?}: {e}"));
29606 assert_eq!(s, again);
29607 }
29608 }
29609
29610 #[test]
29611 fn wait_for_display_roundtrips() {
29612 for sql in [
29613 "WAIT FOR WAL POSITION 12345",
29614 "WAIT FOR WAL POSITION 67890 WITH TIMEOUT 5000",
29615 ] {
29616 let s = parse(sql);
29617 let printed = s.to_string();
29618 let again = parse_statement(&printed)
29619 .unwrap_or_else(|e| panic!("re-parse failed for {printed:?}: {e}"));
29620 assert_eq!(s, again, "round-trip mismatch for {sql:?}");
29621 }
29622 }
29623
29624 #[test]
29625 fn subscription_ddl_display_roundtrips() {
29626 for sql in [
29627 "CREATE SUBSCRIPTION sub_a CONNECTION 'host=h port=20002' PUBLICATION pub_a",
29628 "CREATE SUBSCRIPTION sub_b CONNECTION 'host=h' PUBLICATION p1, p2",
29629 "DROP SUBSCRIPTION sub_a",
29630 "SHOW SUBSCRIPTIONS",
29631 ] {
29632 let s = parse(sql);
29633 let printed = s.to_string();
29634 let again = parse_statement(&printed)
29635 .unwrap_or_else(|e| panic!("re-parse failed for {printed:?}: {e}"));
29636 assert_eq!(s, again, "round-trip mismatch for {sql:?}");
29637 }
29638 }
29639
29640 #[test]
29641 fn parser_drop_dispatches_user_vs_publication() {
29642 // Pre-v6.1.2 DROP USER took the bare-ident path; v6.1.2
29643 // tokenises DROP. Both targets must still parse.
29644 let s = parse("DROP USER 'alice'");
29645 let Statement::DropUser { name, .. } = s else {
29646 panic!("expected DropUser, got {s:?}")
29647 };
29648 assert_eq!(name, "alice");
29649 // And DROP PUBLICATION lands the new variant.
29650 let s = parse("DROP PUBLICATION p1");
29651 assert!(matches!(s, Statement::DropPublication { .. }));
29652 }
29653
29654 #[test]
29655 fn publication_ddl_display_roundtrips() {
29656 // Every CREATE PUBLICATION variant must Display → parse →
29657 // same AST. v6.1.3 covers all three scope shapes.
29658 for sql in [
29659 "CREATE PUBLICATION pub_a",
29660 "CREATE PUBLICATION pub_a FOR ALL TABLES",
29661 "CREATE PUBLICATION pub_a FOR TABLE t1, t2",
29662 "CREATE PUBLICATION pub_a FOR ALL TABLES EXCEPT t1",
29663 "DROP PUBLICATION pub_a",
29664 "SHOW PUBLICATIONS",
29665 ] {
29666 let s = parse(sql);
29667 let printed = s.to_string();
29668 let again = parse_statement(&printed)
29669 .unwrap_or_else(|e| panic!("re-parse failed for {printed:?}: {e}"));
29670 assert_eq!(s, again, "round-trip mismatch for {sql:?}");
29671 }
29672 }
29673
29674 // --- v7.12.4: CREATE FUNCTION + CREATE TRIGGER + PL/pgSQL ---
29675
29676 #[test]
29677 fn create_function_returns_trigger_plpgsql_minimal() {
29678 let sql = "CREATE FUNCTION noop() RETURNS TRIGGER LANGUAGE plpgsql AS $$ BEGIN RETURN NEW; END; $$";
29679 let s = parse(sql);
29680 let Statement::CreateFunction(f) = s else {
29681 panic!("expected CreateFunction");
29682 };
29683 assert_eq!(f.name, "noop");
29684 assert!(!f.or_replace);
29685 assert!(f.args.is_empty());
29686 assert!(matches!(f.returns, FunctionReturn::Trigger));
29687 assert_eq!(f.language, "plpgsql");
29688 let FunctionBody::PlPgSql(block) = f.body else {
29689 panic!("expected PlPgSql body");
29690 };
29691 assert_eq!(block.statements.len(), 1);
29692 assert!(matches!(
29693 block.statements[0],
29694 PlPgSqlStmt::Return(ReturnTarget::New)
29695 ));
29696 }
29697
29698 #[test]
29699 fn create_function_or_replace_with_assignment() {
29700 // mailrs-shape trigger function: NEW.col := to_tsvector(...);
29701 // RETURN NEW.
29702 let sql = "CREATE OR REPLACE FUNCTION update_sv() RETURNS TRIGGER LANGUAGE plpgsql AS $$
29703BEGIN
29704 NEW.search_vector := to_tsvector('english', NEW.subject);
29705 RETURN NEW;
29706END;
29707$$";
29708 let s = parse(sql);
29709 let Statement::CreateFunction(f) = s else {
29710 panic!("expected CreateFunction");
29711 };
29712 assert!(f.or_replace);
29713 let FunctionBody::PlPgSql(block) = &f.body else {
29714 panic!("expected PlPgSql body");
29715 };
29716 assert_eq!(block.statements.len(), 2);
29717 // First statement: NEW.search_vector := to_tsvector(...)
29718 let PlPgSqlStmt::Assign { target, .. } = &block.statements[0] else {
29719 panic!("expected Assign as first stmt");
29720 };
29721 match target {
29722 AssignTarget::NewColumn(c) => assert_eq!(c, "search_vector"),
29723 other => panic!("expected NEW.col, got {other:?}"),
29724 }
29725 // Second statement: RETURN NEW
29726 assert!(matches!(
29727 block.statements[1],
29728 PlPgSqlStmt::Return(ReturnTarget::New)
29729 ));
29730 }
29731
29732 #[test]
29733 fn create_trigger_after_insert_or_update() {
29734 let sql = "CREATE TRIGGER tg AFTER INSERT OR UPDATE ON messages FOR EACH ROW EXECUTE FUNCTION update_sv()";
29735 let s = parse(sql);
29736 let Statement::CreateTrigger(t) = s else {
29737 panic!("expected CreateTrigger");
29738 };
29739 assert_eq!(t.name, "tg");
29740 assert_eq!(t.table, "messages");
29741 assert_eq!(t.timing, TriggerTiming::After);
29742 assert_eq!(t.events, vec![TriggerEvent::Insert, TriggerEvent::Update]);
29743 assert_eq!(t.for_each, TriggerForEach::Row);
29744 assert_eq!(t.function, "update_sv");
29745 }
29746
29747 #[test]
29748 fn create_trigger_before_delete_execute_procedure_alias() {
29749 // PG also accepts the legacy `EXECUTE PROCEDURE` spelling.
29750 let sql =
29751 "CREATE TRIGGER guard BEFORE DELETE ON t FOR EACH ROW EXECUTE PROCEDURE block_delete()";
29752 let s = parse(sql);
29753 let Statement::CreateTrigger(t) = s else {
29754 panic!("expected CreateTrigger");
29755 };
29756 assert_eq!(t.timing, TriggerTiming::Before);
29757 assert_eq!(t.events, vec![TriggerEvent::Delete]);
29758 }
29759
29760 #[test]
29761 fn drop_trigger_if_exists_round_trips() {
29762 // No parser support for DROP TRIGGER yet — added in v7.12.5
29763 // alongside the broader DROP …{IF EXISTS} cleanup. The
29764 // AST + Display impls are in place so we round-trip via
29765 // construction:
29766 let s = Statement::DropTrigger {
29767 name: "tg".into(),
29768 table: "messages".into(),
29769 if_exists: true,
29770 };
29771 assert_eq!(s.to_string(), "DROP TRIGGER IF EXISTS tg ON messages");
29772 }
29773
29774 #[test]
29775 fn trigger_ddl_display_roundtrips_through_parser() {
29776 // CREATE TRIGGER + its referenced CREATE FUNCTION must
29777 // Display → parse → same AST (modulo PL/pgSQL body
29778 // formatting which is parser-canonicalised).
29779 for sql in [
29780 "CREATE TRIGGER tg AFTER INSERT ON t FOR EACH ROW EXECUTE FUNCTION f()",
29781 "CREATE TRIGGER tg2 BEFORE UPDATE OR DELETE ON t FOR EACH ROW EXECUTE FUNCTION g()",
29782 ] {
29783 let s = parse(sql);
29784 let printed = s.to_string();
29785 let again = parse_statement(&printed)
29786 .unwrap_or_else(|e| panic!("re-parse failed for {printed:?}: {e}"));
29787 assert_eq!(s, again, "round-trip mismatch for {sql:?}");
29788 }
29789 }
29790}