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 let name = self.expect_ident_like()?;
11035 Ok(alloc::vec![crate::ast::AlterTableTarget::SetEngine(name)])
11036 }
11037 Token::Ident(s) if s.eq_ignore_ascii_case("convert") => {
11038 self.advance();
11039 // CONVERT TO CHARACTER SET <cs> [COLLATE <c>]
11040 if matches!(self.peek(), Token::To) {
11041 self.advance();
11042 }
11043 let kw = self.expect_ident_like()?;
11044 if !kw.eq_ignore_ascii_case("character") {
11045 return Err(self.err("expected CHARACTER after CONVERT TO".into()));
11046 }
11047 let set_kw = self.expect_ident_like()?;
11048 if !set_kw.eq_ignore_ascii_case("set") {
11049 return Err(self.err("expected SET after CHARACTER".into()));
11050 }
11051 let charset = self.expect_ident_like()?;
11052 let collate =
11053 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("collate")) {
11054 self.advance();
11055 Some(self.expect_ident_like()?)
11056 } else {
11057 None
11058 };
11059 Ok(alloc::vec![
11060 crate::ast::AlterTableTarget::ConvertToCharacterSet { charset, collate }
11061 ])
11062 }
11063 Token::Ident(s) if s.eq_ignore_ascii_case("set") => {
11064 self.advance();
11065 // v7.37.18 (18.7-18.15) — SET ( option = value, … )
11066 // storage parameters: paren-prefixed; consume.
11067 if matches!(self.peek(), Token::LParen) {
11068 self.consume_until_statement_boundary();
11069 return Ok(Vec::new());
11070 }
11071 let setting = self.expect_ident_like()?;
11072 if setting.eq_ignore_ascii_case("hot_tier_bytes") {
11073 if !matches!(self.peek(), Token::Eq) {
11074 return Err(self.err(alloc::format!(
11075 "expected '=' after hot_tier_bytes, got {:?}",
11076 self.peek()
11077 )));
11078 }
11079 self.advance();
11080 let n = self.expect_u64_literal()?;
11081 return Ok(alloc::vec![crate::ast::AlterTableTarget::SetHotTierBytes(n)]);
11082 }
11083 // v7.37.18 (18.7 / 18.8 / 18.11 / 18.13 / 18.14) —
11084 // accept-and-no-op for ALTER TABLE SET <subject>
11085 // forms that pg_dump emits but SPG either treats
11086 // as N/A (single-tenant, single-owner, no shared
11087 // tablespaces) or accepts the dump-side declaration
11088 // without runtime effect:
11089 // SET SCHEMA <name> (18.11)
11090 // SET TABLESPACE <name> (18.8)
11091 // SET LOGGED / UNLOGGED (18.7 alt-form)
11092 // SET WITHOUT CLUSTER (18.13)
11093 // SET WITHOUT OIDS (PG legacy)
11094 // SET (option = value, …) (storage parameters)
11095 // SET REPLICA IDENTITY {…} (18.14)
11096 if setting.eq_ignore_ascii_case("schema")
11097 || setting.eq_ignore_ascii_case("tablespace")
11098 || setting.eq_ignore_ascii_case("logged")
11099 || setting.eq_ignore_ascii_case("unlogged")
11100 || setting.eq_ignore_ascii_case("without")
11101 {
11102 self.consume_until_statement_boundary();
11103 return Ok(Vec::new());
11104 }
11105 if setting.eq_ignore_ascii_case("replica") {
11106 // SET REPLICA IDENTITY {DEFAULT|FULL|NOTHING|USING INDEX <name>}
11107 self.consume_until_statement_boundary();
11108 return Ok(Vec::new());
11109 }
11110 // SET (option=value, …) — storage parameters.
11111 if matches!(self.peek(), Token::LParen) {
11112 self.consume_until_statement_boundary();
11113 return Ok(Vec::new());
11114 }
11115 Err(self.err(alloc::format!(
11116 "ALTER TABLE SET: unknown setting {setting:?}; supported: \
11117 hot_tier_bytes / SCHEMA / TABLESPACE / LOGGED / UNLOGGED / \
11118 WITHOUT CLUSTER / WITHOUT OIDS / REPLICA IDENTITY / (storage_params)"
11119 )))
11120 }
11121 // v7.39 (round 647) — `ALTER TABLE c INHERIT p`. Carried now,
11122 // not ignored: round 645 gave SPG the inheritance the
11123 // v7.37.18 no-op said it did not have.
11124 Token::Ident(s) if s.eq_ignore_ascii_case("inherit") => {
11125 self.advance();
11126 let parent = self.expect_ident_like()?;
11127 self.consume_until_statement_boundary();
11128 Ok(alloc::vec![crate::ast::AlterTableTarget::Inherit {
11129 parent,
11130 detach: false
11131 }])
11132 }
11133 // `NO INHERIT <parent>`. Guarded to NOT match `NO FORCE ROW
11134 // LEVEL SECURITY`, which has its own RLS arm below — without
11135 // the guard this swallowed NO FORCE as a no-op.
11136 Token::Ident(s)
11137 if s.eq_ignore_ascii_case("no")
11138 && !matches!(
11139 self.tokens.get(self.pos + 1),
11140 Some(Token::Ident(t)) if t.eq_ignore_ascii_case("force")
11141 ) =>
11142 {
11143 self.advance();
11144 if matches!(self.peek(), Token::Ident(k) | Token::QuotedIdent(k)
11145 if k.eq_ignore_ascii_case("inherit"))
11146 {
11147 self.advance();
11148 let parent = self.expect_ident_like()?;
11149 self.consume_until_statement_boundary();
11150 return Ok(alloc::vec![crate::ast::AlterTableTarget::Inherit {
11151 parent,
11152 detach: true
11153 }]);
11154 }
11155 self.consume_until_statement_boundary();
11156 Ok(Vec::new())
11157 }
11158 // v7.37.18 (18.10) — ALTER TABLE OWNER TO <user>. SPG is
11159 // single-owner, so there is still nothing to record.
11160 //
11161 // v7.39 (round 652) — but the name now reaches the engine,
11162 // which refuses a role that does not exist as PG does. The
11163 // no-op was swallowing the whole statement, so a dump naming
11164 // a role this server never heard of restored clean and left
11165 // the table owned by whoever ran the restore.
11166 Token::Ident(s) if s.eq_ignore_ascii_case("owner") => {
11167 self.advance();
11168 if matches!(self.peek(), Token::To) {
11169 self.advance();
11170 }
11171 let role = self.expect_ident_like()?;
11172 Ok(alloc::vec![crate::ast::AlterTableTarget::OwnerTo {
11173 role
11174 }])
11175 }
11176 // v7.37.18 (18.13) — ALTER TABLE CLUSTER ON <index>.
11177 // PG sets a hint; SPG doesn't have clustered storage, so the
11178 // hint itself stays a no-op.
11179 //
11180 // v7.39 (round 652) — the index name is checked now. PG
11181 // errors on one that does not exist, and swallowing that let
11182 // a typo'd CLUSTER ON pass silently.
11183 Token::Ident(s) if s.eq_ignore_ascii_case("cluster") => {
11184 self.advance();
11185 // `ON` is a reserved token, not an ident.
11186 if !matches!(self.peek(), Token::On) {
11187 return Err(self.err(alloc::format!(
11188 "expected ON after CLUSTER, got {:?}",
11189 self.peek()
11190 )));
11191 }
11192 self.advance();
11193 let index = self.expect_ident_like()?;
11194 Ok(alloc::vec![crate::ast::AlterTableTarget::ClusterOn {
11195 index: Some(index)
11196 }])
11197 }
11198 // v7.39 (read01 round 49) — ALTER TABLE REPLICA IDENTITY
11199 // { DEFAULT | FULL | NOTHING | USING INDEX <name> }. PG records
11200 // what a logical decoder puts in the old-tuple image; SPG's
11201 // replication is SQL-text, so there is nothing to record.
11202 // Accept-and-no-op (it used to be a parse error).
11203 Token::Ident(s) if s.eq_ignore_ascii_case("replica") => {
11204 self.advance();
11205 // v7.39 (round 710) — `REPLICA IDENTITY USING INDEX <i>`
11206 // validates the index; DEFAULT / FULL / NOTHING stay no-op.
11207 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("identity"))
11208 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(k)) if k.eq_ignore_ascii_case("using"))
11209 {
11210 self.advance(); // IDENTITY
11211 self.advance(); // USING
11212 if matches!(self.peek(), Token::Index)
11213 || matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("index"))
11214 {
11215 self.advance();
11216 }
11217 let index = self.expect_ident_like()?;
11218 self.consume_until_statement_boundary();
11219 return Ok(alloc::vec![
11220 crate::ast::AlterTableTarget::ReplicaIdentityUsingIndex { index }
11221 ]);
11222 }
11223 self.consume_until_statement_boundary();
11224 Ok(Vec::new())
11225 }
11226 // v7.37.18 (18.15) — ALTER TABLE VALIDATE CONSTRAINT <name>.
11227 //
11228 // v7.39 (round 652) — it used to consume the statement and
11229 // return nothing, on the stated theory that SPG validated at
11230 // ADD CONSTRAINT time so there was never anything left to
11231 // validate. Measured against PG18, ADD CONSTRAINT did not
11232 // scan the existing rows at all — the comment described a
11233 // property SPG did not have, which is why nobody looked. Both
11234 // halves are real now: ADD scans unless told NOT VALID, and
11235 // this scans what NOT VALID skipped.
11236 Token::Ident(s) if s.eq_ignore_ascii_case("validate") => {
11237 self.advance();
11238 self.expect_keyword_ident("constraint")?;
11239 let name = self.expect_ident_like()?;
11240 Ok(alloc::vec![
11241 crate::ast::AlterTableTarget::ValidateConstraint { name }
11242 ])
11243 }
11244 // v7.37.18 (18.18) — RESET ( option [, …] ). Inverse of
11245 // SET (option = value, …). PG uses it to clear per-table
11246 // storage params like fillfactor or autovacuum_*. SPG
11247 // engine-manages those parameters; accept-and-no-op.
11248 Token::Ident(s) if s.eq_ignore_ascii_case("reset") => {
11249 self.advance();
11250 self.consume_until_statement_boundary();
11251 Ok(Vec::new())
11252 }
11253 // v7.37.18 (18.18) — OF <type_name> / NOT OF. Composite-
11254 // type-of binding (PG 9.0+). SPG composite types
11255 // (v7.37.5 ζ-B sub-commit) follow CREATE TYPE; ALTER
11256 // TABLE OF is rare and inverse of CREATE TABLE OF.
11257 // Accept-and-no-op until a customer dump round-trips it.
11258 Token::Ident(s) if s.eq_ignore_ascii_case("of") => {
11259 self.advance();
11260 // v7.39 (round 710) — the type name is validated now.
11261 let type_name = self.expect_ident_like()?;
11262 self.consume_until_statement_boundary();
11263 Ok(alloc::vec![crate::ast::AlterTableTarget::OfType {
11264 type_name
11265 }])
11266 }
11267 // v7.37.18 (18.18) — `NOT OF` lexes NOT as Token::Not
11268 // (reserved keyword) rather than Token::Ident("not"),
11269 // so it needs its own arm. Accept-and-no-op same as OF.
11270 Token::Not => {
11271 self.advance();
11272 self.consume_until_statement_boundary();
11273 Ok(Vec::new())
11274 }
11275 // v7.39 (RLS) — FORCE ROW LEVEL SECURITY (sets relforcerowsecurity).
11276 Token::Ident(s) if s.eq_ignore_ascii_case("force") => {
11277 self.advance();
11278 self.expect_row_level_security()?;
11279 Ok(alloc::vec![crate::ast::AlterTableTarget::SetRowSecurity {
11280 enabled: None,
11281 force: Some(true),
11282 }])
11283 }
11284 // v7.39 (RLS) — NO FORCE ROW LEVEL SECURITY.
11285 Token::Ident(s)
11286 if s.eq_ignore_ascii_case("no")
11287 && matches!(
11288 self.tokens.get(self.pos + 1),
11289 Some(Token::Ident(t)) if t.eq_ignore_ascii_case("force")
11290 ) =>
11291 {
11292 self.advance(); // NO
11293 self.advance(); // FORCE
11294 self.expect_row_level_security()?;
11295 Ok(alloc::vec![crate::ast::AlterTableTarget::SetRowSecurity {
11296 enabled: None,
11297 force: Some(false),
11298 }])
11299 }
11300 // v7.39 (RLS) — ENABLE/DISABLE ROW LEVEL SECURITY
11301 // (sets relrowsecurity). The guard requires the next token to be
11302 // `ROW` so the ENABLE/DISABLE TRIGGER arm still matches its case.
11303 Token::Ident(s)
11304 if (s.eq_ignore_ascii_case("enable") || s.eq_ignore_ascii_case("disable"))
11305 && matches!(
11306 self.tokens.get(self.pos + 1),
11307 Some(Token::Ident(t)) if t.eq_ignore_ascii_case("row")
11308 ) =>
11309 {
11310 let enabled = s.eq_ignore_ascii_case("enable");
11311 self.advance(); // ENABLE/DISABLE
11312 self.expect_row_level_security()?;
11313 Ok(alloc::vec![crate::ast::AlterTableTarget::SetRowSecurity {
11314 enabled: Some(enabled),
11315 force: None,
11316 }])
11317 }
11318 Token::Ident(s) if s.eq_ignore_ascii_case("add") => {
11319 self.advance();
11320 // v7.39 (round 431) — MySQL's `ALTER TABLE t ADD [UNIQUE]
11321 // {INDEX|KEY} [name] (cols)`, which every ORM migration
11322 // emits. The same grammar CREATE TABLE already accepts
11323 // inline (`KEY idx (a)`, prefix lengths and all), so it goes
11324 // through the SAME parser — an ALTER-only copy would be a
11325 // second place for the two to drift.
11326 if self.peek_mysql_inline_key_start() {
11327 return Ok(match self.parse_mysql_inline_key()? {
11328 Some(c) => {
11329 alloc::vec![crate::ast::AlterTableTarget::AddTableConstraint(c)]
11330 }
11331 // FULLTEXT / SPATIAL parse and are accepted as a
11332 // no-op here exactly as they are inline.
11333 None => Vec::new(),
11334 });
11335 }
11336 // v7.14.0 — ADD CONSTRAINT <name> { FOREIGN KEY |
11337 // PRIMARY KEY | UNIQUE | CHECK }. pg_dump emits
11338 // PRIMARY KEY this way; mysqldump emits both.
11339 // Peek-only dispatch (no advance) — `advance()`
11340 // destructively replaces consumed tokens with Eof,
11341 // so saved-pos restore would land on Eofs.
11342 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint"))
11343 {
11344 // The next-but-one ident is the constraint
11345 // name; the one after THAT is the kind.
11346 let kind_pos = self.pos + 2;
11347 let kind = self.tokens.get(kind_pos).cloned();
11348 if matches!(&kind, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("foreign"))
11349 {
11350 let fk = self.parse_table_level_fk()?;
11351 return Ok(alloc::vec![
11352 crate::ast::AlterTableTarget::AddForeignKey(fk)
11353 ]);
11354 }
11355 if matches!(&kind, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("primary"))
11356 {
11357 self.advance(); // CONSTRAINT
11358 // v7.39 (read01 round 48) — keep the name; the engine
11359 // stores it now instead of dropping it on the floor.
11360 let con_name = self.expect_ident_like()?;
11361 self.advance(); // PRIMARY
11362 self.expect_keyword_ident("key")?;
11363 let cols = self.parse_paren_ident_list("PRIMARY KEY")?;
11364 // v7.39 (round 711) — the ALTER form carries the
11365 // timing too (pg_dump writes it here).
11366 let (deferrable, initially_deferred) =
11367 self.consume_deferrable_clauses_timed()?;
11368 return Ok(alloc::vec![
11369 crate::ast::AlterTableTarget::AddTableConstraint(
11370 crate::ast::TableConstraint::PrimaryKey {
11371 name: Some(con_name),
11372 columns: cols,
11373 deferrable,
11374 initially_deferred,
11375 }
11376 )
11377 ]);
11378 }
11379 if matches!(&kind, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("unique"))
11380 {
11381 self.advance(); // CONSTRAINT
11382 // v7.39 (read01 round 48) — keep the name.
11383 let con_name = self.expect_ident_like()?;
11384 // v7.22 (mailrs round-13 gap 6) — delegate so
11385 // the optional `NULLS [NOT] DISTINCT` modifier
11386 // parses here too (pg_dump emits the ALTER
11387 // form; semantics enforced by the engine
11388 // since v7.13).
11389 let mut uc = self.parse_table_level_unique()?;
11390 if let crate::ast::TableConstraint::Unique { name, .. } = &mut uc {
11391 *name = Some(con_name);
11392 }
11393 return Ok(alloc::vec![
11394 crate::ast::AlterTableTarget::AddTableConstraint(uc)
11395 ]);
11396 }
11397 if matches!(&kind, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("check"))
11398 {
11399 self.advance(); // CONSTRAINT
11400 // v7.39 (read01 round 48) — keep the name.
11401 let con_name = self.expect_ident_like()?;
11402 self.advance(); // CHECK
11403 if !matches!(self.peek(), Token::LParen) {
11404 return Err(self.err(alloc::format!(
11405 "expected '(' after CHECK, got {:?}", self.peek()
11406 )));
11407 }
11408 self.advance();
11409 let expr = self.parse_expr(0)?;
11410 if matches!(self.peek(), Token::RParen) {
11411 self.advance();
11412 }
11413 let not_valid = self.parse_not_valid_suffix();
11414 return Ok(alloc::vec![
11415 crate::ast::AlterTableTarget::AddTableConstraint(
11416 crate::ast::TableConstraint::Check {
11417 name: Some(con_name),
11418 expr,
11419 not_valid,
11420 }
11421 )
11422 ]);
11423 }
11424 // v7.39 (round 211) — ADD CONSTRAINT <name> EXCLUDE
11425 // [USING <am>] (<col> WITH <op>[, …]). pg_dump emits
11426 // exclusion constraints via this ALTER form.
11427 if matches!(&kind, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("exclude"))
11428 {
11429 self.advance(); // CONSTRAINT
11430 let con_name = self.expect_ident_like()?;
11431 let mut ex = self.parse_table_level_exclude()?;
11432 if let crate::ast::TableConstraint::Exclude { name, .. } = &mut ex {
11433 *name = Some(con_name);
11434 }
11435 return Ok(alloc::vec![
11436 crate::ast::AlterTableTarget::AddTableConstraint(ex)
11437 ]);
11438 }
11439 // Unknown kind — fall through to FK path which
11440 // produces a descriptive parse error.
11441 }
11442 let is_fk = matches!(
11443 self.peek(),
11444 Token::Ident(s) if s.eq_ignore_ascii_case("constraint")
11445 || s.eq_ignore_ascii_case("foreign")
11446 );
11447 if is_fk {
11448 let fk = self.parse_table_level_fk()?;
11449 return Ok(alloc::vec![crate::ast::AlterTableTarget::AddForeignKey(fk)]);
11450 }
11451 // v7.14.0 — bare ADD PRIMARY KEY / UNIQUE / CHECK
11452 // (no CONSTRAINT prefix) — same dispatch.
11453 match self.peek().clone() {
11454 Token::Ident(s) if s.eq_ignore_ascii_case("primary") => {
11455 self.advance();
11456 self.expect_keyword_ident("key")?;
11457 let cols = self.parse_paren_ident_list("PRIMARY KEY")?;
11458 let (deferrable, initially_deferred) =
11459 self.consume_deferrable_clauses_timed()?;
11460 return Ok(alloc::vec![
11461 crate::ast::AlterTableTarget::AddTableConstraint(
11462 crate::ast::TableConstraint::PrimaryKey {
11463 name: None,
11464 columns: cols,
11465 deferrable,
11466 initially_deferred,
11467 }
11468 )
11469 ]);
11470 }
11471 Token::Ident(s) if s.eq_ignore_ascii_case("unique") => {
11472 // v7.22 — delegate (NULLS [NOT] DISTINCT).
11473 let uc = self.parse_table_level_unique()?;
11474 return Ok(alloc::vec![
11475 crate::ast::AlterTableTarget::AddTableConstraint(uc)
11476 ]);
11477 }
11478 // v7.39 (round 652) — bare ADD CHECK (no CONSTRAINT
11479 // prefix). The other three bare forms were here and
11480 // this one was not, so it fell through to the column
11481 // path and came back as "unexpected reserved keyword
11482 // 'check' at start of column definition".
11483 _ if self.peek_table_level_check_start() => {
11484 let chk = self.parse_table_level_check()?;
11485 let not_valid = self.parse_not_valid_suffix();
11486 let crate::ast::TableConstraint::Check { expr, .. } = chk else {
11487 unreachable!("parse_table_level_check returns Check")
11488 };
11489 return Ok(alloc::vec![
11490 crate::ast::AlterTableTarget::AddTableConstraint(
11491 crate::ast::TableConstraint::Check {
11492 name: None,
11493 expr,
11494 not_valid,
11495 }
11496 )
11497 ]);
11498 }
11499 // v7.39 (round 211) — bare ADD EXCLUDE (no CONSTRAINT prefix).
11500 Token::Ident(s) if s.eq_ignore_ascii_case("exclude") => {
11501 let ex = self.parse_table_level_exclude()?;
11502 return Ok(alloc::vec![
11503 crate::ast::AlterTableTarget::AddTableConstraint(ex)
11504 ]);
11505 }
11506 _ => {}
11507 }
11508 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("column")) {
11509 self.advance();
11510 }
11511 let mut if_not_exists = false;
11512 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if")) {
11513 self.advance();
11514 if !matches!(self.peek(), Token::Not) {
11515 return Err(self.err(alloc::format!(
11516 "expected NOT after IF in ALTER TABLE ADD COLUMN, got {:?}",
11517 self.peek()
11518 )));
11519 }
11520 self.advance();
11521 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("exists")) {
11522 return Err(self.err(alloc::format!(
11523 "expected EXISTS after IF NOT in ALTER TABLE ADD COLUMN, got {:?}",
11524 self.peek()
11525 )));
11526 }
11527 self.advance();
11528 if_not_exists = true;
11529 }
11530 // v7.13.2 — mailrs round-6 S3: `ADD COLUMN col TYPE
11531 // REFERENCES other(col) [ON DELETE …]`. parse_column_def
11532 // returns ColumnDef + an optional inline FK.
11533 let (column, col_level_fk) = self.parse_column_def_with_fk()?;
11534 let col_name = column.name.clone();
11535 // v7.39.9 — MySQL says where the column goes.
11536 let position = self.parse_column_position();
11537 let mut out = alloc::vec![crate::ast::AlterTableTarget::AddColumn {
11538 column,
11539 if_not_exists,
11540 position,
11541 }];
11542 if let Some(mut fk) = col_level_fk {
11543 if fk.columns.is_empty() {
11544 fk.columns.push(col_name);
11545 }
11546 out.push(crate::ast::AlterTableTarget::AddForeignKey(fk));
11547 }
11548 Ok(out)
11549 }
11550 Token::Drop => {
11551 self.advance();
11552 // v7.13.3 — dispatch on the next token. mailrs round-7
11553 // S8 closed DROP COLUMN; round-6 S7 closed
11554 // DROP CONSTRAINT. Both share IF EXISTS / CASCADE /
11555 // RESTRICT modifiers.
11556 // DROP CONSTRAINT [IF EXISTS] <name> [CASCADE|RESTRICT]
11557 // DROP [COLUMN] [IF EXISTS] <col> [CASCADE|RESTRICT]
11558 let subject = match self.peek() {
11559 Token::Ident(s) if s.eq_ignore_ascii_case("constraint") => {
11560 self.advance();
11561 "constraint"
11562 }
11563 Token::Ident(s) if s.eq_ignore_ascii_case("column") => {
11564 self.advance();
11565 "column"
11566 }
11567 // v7.39 (round 431) — MySQL `DROP {INDEX|KEY} name`.
11568 // `INDEX` lexes as the reserved Token::Index, so it is
11569 // unambiguous. `KEY` is a plain ident, and PG allows a
11570 // column literally named "key", so only read it as the
11571 // keyword when a name follows it.
11572 Token::Index => {
11573 self.advance();
11574 "index"
11575 }
11576 Token::Ident(s)
11577 if s.eq_ignore_ascii_case("key")
11578 && matches!(
11579 self.tokens.get(self.pos + 1),
11580 Some(Token::Ident(_) | Token::QuotedIdent(_))
11581 ) =>
11582 {
11583 self.advance();
11584 "index"
11585 }
11586 // PG-canonical bare `DROP <col>` without COLUMN
11587 // keyword is also valid; treat any other ident
11588 // as the column name.
11589 Token::Ident(_) | Token::QuotedIdent(_) => "column",
11590 other => {
11591 return Err(self.err(alloc::format!(
11592 "expected COLUMN / CONSTRAINT after DROP in ALTER TABLE, got {other:?}"
11593 )));
11594 }
11595 };
11596 let mut if_exists = false;
11597 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if")) {
11598 let n1 = self.tokens.get(self.pos + 1);
11599 if matches!(n1, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("exists")) {
11600 self.advance();
11601 self.advance();
11602 if_exists = true;
11603 }
11604 }
11605 let name = self.expect_ident_like()?;
11606 let mut cascade = false;
11607 if matches!(
11608 self.peek(),
11609 Token::Ident(s) if s.eq_ignore_ascii_case("cascade")
11610 || s.eq_ignore_ascii_case("restrict")
11611 ) {
11612 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("cascade"))
11613 {
11614 cascade = true;
11615 }
11616 self.advance();
11617 }
11618 if subject == "index" {
11619 Ok(alloc::vec![crate::ast::AlterTableTarget::DropIndex {
11620 name,
11621 if_exists,
11622 }])
11623 } else if subject == "constraint" {
11624 Ok(alloc::vec![crate::ast::AlterTableTarget::DropForeignKey {
11625 name,
11626 if_exists,
11627 }])
11628 } else {
11629 Ok(alloc::vec![crate::ast::AlterTableTarget::DropColumn {
11630 column: name,
11631 if_exists,
11632 cascade,
11633 }])
11634 }
11635 }
11636 Token::Ident(s) if s.eq_ignore_ascii_case("alter") => {
11637 self.advance();
11638 // v7.37.18 (18.16) — `ALTER TABLE … ALTER CONSTRAINT
11639 // <name> {DEFERRABLE|NOT DEFERRABLE} [INITIALLY
11640 // {IMMEDIATE|DEFERRED}]`. SPG enforces constraints
11641 // immediately; accept-and-no-op.
11642 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint")) {
11643 self.advance();
11644 self.consume_until_statement_boundary();
11645 return Ok(Vec::new());
11646 }
11647 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("column")) {
11648 self.advance();
11649 }
11650 let col_name = self.expect_ident_like()?;
11651 match self.peek() {
11652 Token::Ident(s) if s.eq_ignore_ascii_case("type") => {
11653 self.advance();
11654 }
11655 // v7.14.0 — pg_dump emits BIGSERIAL via
11656 // `ALTER TABLE … ALTER COLUMN id SET DEFAULT
11657 // nextval('seq')` (the sequence is created
11658 // separately). SPG's BIGSERIAL already uses
11659 // AUTO_INCREMENT; accept SET DEFAULT / DROP
11660 // DEFAULT / SET NOT NULL / DROP NOT NULL as
11661 // engine no-ops by consuming the tail.
11662 Token::Ident(s) if s.eq_ignore_ascii_case("set") => {
11663 // v7.22 (round-13 T2) — `SET DEFAULT
11664 // nextval('…')` is how pg_dump spells a
11665 // SERIAL column (plain integer in CREATE
11666 // TABLE + this ALTER). It used to be
11667 // swallowed as a no-op, which silently
11668 // STRIPPED auto-increment from imported
11669 // schemas — the first post-import INSERT
11670 // without an explicit id then violated NOT
11671 // NULL. Lower it to the auto-increment
11672 // marker instead.
11673 let is_default_nextval =
11674 matches!(self.tokens.get(self.pos + 1), Some(Token::Default))
11675 && matches!(
11676 self.tokens.get(self.pos + 2),
11677 Some(Token::Ident(f)) if f.eq_ignore_ascii_case("nextval")
11678 );
11679 if is_default_nextval {
11680 let seq_name = self.scan_sequence_name_until_boundary();
11681 return Ok(alloc::vec![
11682 crate::ast::AlterTableTarget::SetColumnAutoIncrement {
11683 column: col_name,
11684 seq_name,
11685 }
11686 ]);
11687 }
11688 // v7.37.18 (18.1 + 18.2) — proper lowering.
11689 self.advance(); // consume "set"
11690 match self.peek().clone() {
11691 Token::Default => {
11692 self.advance();
11693 let default_expr = self.parse_expr(0)?;
11694 return Ok(alloc::vec![
11695 crate::ast::AlterTableTarget::AlterColumnSetDefault {
11696 column: col_name,
11697 default_expr,
11698 }
11699 ]);
11700 }
11701 Token::Not => {
11702 self.advance();
11703 if !matches!(self.peek(), Token::Null) {
11704 return Err(self.err(alloc::format!(
11705 "expected NULL after ALTER COLUMN SET NOT, got {:?}",
11706 self.peek()
11707 )));
11708 }
11709 self.advance();
11710 return Ok(alloc::vec![
11711 crate::ast::AlterTableTarget::AlterColumnSetNotNull {
11712 column: col_name,
11713 }
11714 ]);
11715 }
11716 // `SET EXPRESSION AS (expr)` (PG 17) — change a
11717 // stored generated column's expression and
11718 // recompute existing rows.
11719 Token::Ident(s) if s.eq_ignore_ascii_case("expression") => {
11720 self.advance(); // EXPRESSION
11721 if matches!(self.peek(), Token::As) {
11722 self.advance();
11723 }
11724 let expr = self.parse_expr(0)?;
11725 return Ok(alloc::vec![
11726 crate::ast::AlterTableTarget::AlterColumnSetExpression {
11727 column: col_name,
11728 expr,
11729 }
11730 ]);
11731 }
11732 other => {
11733 // Other SET subjects (STATISTICS,
11734 // STORAGE, COMPRESSION, …) stay no-ops —
11735 // storage hints with no SPG semantics.
11736 let _ = other;
11737 self.consume_until_statement_boundary();
11738 return Ok(Vec::new());
11739 }
11740 }
11741 }
11742 Token::Ident(s) if s.eq_ignore_ascii_case("drop") => {
11743 self.advance(); // consume "drop"
11744 return self.parse_alter_column_drop_tail(col_name);
11745 }
11746 Token::Drop => {
11747 self.advance(); // consume Drop token
11748 return self.parse_alter_column_drop_tail(col_name);
11749 }
11750 Token::Ident(s) if s.eq_ignore_ascii_case("add") => {
11751 // v7.22 (round-13 T2) — `ALTER COLUMN c ADD
11752 // GENERATED { ALWAYS | BY DEFAULT } AS
11753 // IDENTITY ( … )`: pg_dump's spelling for
11754 // identity columns. Same auto-increment
11755 // lowering as the nextval default; the
11756 // sequence options inside the parens are
11757 // no-ops under SPG's max+1 semantics.
11758 let is_generated = matches!(
11759 self.tokens.get(self.pos + 1),
11760 Some(Token::Ident(g)) if g.eq_ignore_ascii_case("generated")
11761 );
11762 if !is_generated {
11763 return Err(self.err(alloc::format!(
11764 "expected GENERATED after ALTER COLUMN {col_name} ADD, got {:?}",
11765 self.tokens.get(self.pos + 1)
11766 )));
11767 }
11768 let seq_name = self.scan_sequence_name_until_boundary();
11769 return Ok(alloc::vec![
11770 crate::ast::AlterTableTarget::SetColumnAutoIncrement {
11771 column: col_name,
11772 seq_name,
11773 }
11774 ]);
11775 }
11776 // v7.39 (round 220) — `RESTART [WITH n]` on an identity
11777 // column: floor the next allocated value at n (bare
11778 // RESTART = restart from the start value, 1).
11779 Token::Ident(s) if s.eq_ignore_ascii_case("restart") => {
11780 self.advance();
11781 let with = if matches!(self.peek(), Token::Ident(w) if w.eq_ignore_ascii_case("with"))
11782 {
11783 self.advance();
11784 let neg = if matches!(self.peek(), Token::Minus) {
11785 self.advance();
11786 true
11787 } else {
11788 false
11789 };
11790 match self.advance() {
11791 Token::Integer(v) => Some(if neg { -v } else { v }),
11792 other => {
11793 return Err(self.err(alloc::format!(
11794 "expected integer after RESTART WITH, got {other:?}"
11795 )));
11796 }
11797 }
11798 } else {
11799 None
11800 };
11801 return Ok(alloc::vec![
11802 crate::ast::AlterTableTarget::AlterColumnRestart {
11803 column: col_name,
11804 with,
11805 }
11806 ]);
11807 }
11808 other => {
11809 return Err(self.err(alloc::format!(
11810 "expected TYPE / SET / DROP / ADD after ALTER COLUMN <name>, got {other:?}"
11811 )));
11812 }
11813 }
11814 // v7.39 (round 713) — the type parser has consumed a
11815 // trailing `COLLATE <name>` since Phase 2.5, and
11816 // `parse_column_type_name` discarded it: `ALTER COLUMN t
11817 // TYPE text COLLATE "C"` parsed clean and changed
11818 // nothing. Keep the clause; the engine re-collates.
11819 let (new_type, _, _, _, coll, coll_explicit, coll_name, _, _, _, _, _, _, _) =
11820 self.parse_type_with_implied_flags()?;
11821 let collation = if coll_explicit {
11822 coll_name.map(|n| (coll, n))
11823 } else {
11824 None
11825 };
11826 let using = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using"))
11827 {
11828 self.advance();
11829 Some(self.parse_expr(0)?)
11830 } else {
11831 None
11832 };
11833 Ok(alloc::vec![crate::ast::AlterTableTarget::AlterColumnType {
11834 column: col_name,
11835 new_type,
11836 using,
11837 collation,
11838 }])
11839 }
11840 // v7.15.0 — `ALTER TABLE t RENAME [COLUMN] old TO new`.
11841 // PG also supports `RENAME TO new_table` for table-name
11842 // rename; that surface is deferred (pg_dump never emits
11843 // it). If the first post-RENAME ident is `TO`, the user
11844 // is asking for table rename — error with a clear
11845 // message rather than misparsing `TO` as a column name.
11846 Token::Ident(s) if s.eq_ignore_ascii_case("rename") => {
11847 self.advance();
11848 // v7.16.2 — `ALTER TABLE t RENAME TO new_table`
11849 // table-name rename (mailrs round-10 A.5 — used
11850 // by migrate-042's `RENAME TO email_contacts`).
11851 // `TO` lexes as Token::To.
11852 if matches!(self.peek(), Token::To)
11853 || matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("to"))
11854 {
11855 self.advance();
11856 let new = self.expect_ident_like()?;
11857 return Ok(alloc::vec![crate::ast::AlterTableTarget::RenameTable {
11858 new,
11859 }]);
11860 }
11861 // v7.39 (read01 round 48) — `RENAME CONSTRAINT old TO new`.
11862 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint")) {
11863 self.advance();
11864 let old = self.expect_ident_like()?;
11865 if matches!(self.peek(), Token::To) {
11866 self.advance();
11867 } else {
11868 self.expect_keyword_ident("to")?;
11869 }
11870 let new = self.expect_ident_like()?;
11871 return Ok(alloc::vec![
11872 crate::ast::AlterTableTarget::RenameConstraint { old, new }
11873 ]);
11874 }
11875 // v7.39.9 — MySQL's `RENAME {INDEX|KEY} old TO new`.
11876 // PostgreSQL renames an index with its own top-level
11877 // `ALTER INDEX`, so this spelling had nowhere to go and
11878 // answered 1064; MySQL 9.7.2 parses it and answers 1176
11879 // when the key is not there.
11880 if matches!(self.peek(), Token::Index)
11881 || matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("key"))
11882 {
11883 self.advance();
11884 let old = self.expect_ident_like()?;
11885 if matches!(self.peek(), Token::To) {
11886 self.advance();
11887 } else {
11888 self.expect_keyword_ident("to")?;
11889 }
11890 let new = self.expect_ident_like()?;
11891 return Ok(alloc::vec![crate::ast::AlterTableTarget::RenameIndex {
11892 old,
11893 new,
11894 }]);
11895 }
11896 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("column")) {
11897 self.advance();
11898 }
11899 let old = self.expect_ident_like()?;
11900 // `TO` is a reserved keyword token; accept both
11901 // Token::To and Token::Ident("to") for consistency.
11902 if matches!(self.peek(), Token::To) {
11903 self.advance();
11904 } else {
11905 self.expect_keyword_ident("to")?;
11906 }
11907 let new = self.expect_ident_like()?;
11908 Ok(alloc::vec![crate::ast::AlterTableTarget::RenameColumn {
11909 old,
11910 new,
11911 }])
11912 }
11913 // v7.16.1 — `ALTER TABLE t { ENABLE | DISABLE } TRIGGER
11914 // { ALL | <name> }`. pg_dump --disable-triggers wraps
11915 // every data block with these. Real disable semantics —
11916 // not no-op — because reload correctness assumes the
11917 // triggers don't fire (rows already carry their
11918 // computed values from prod).
11919 Token::Ident(s)
11920 if s.eq_ignore_ascii_case("enable") || s.eq_ignore_ascii_case("disable") =>
11921 {
11922 let enabled = s.eq_ignore_ascii_case("enable");
11923 self.advance();
11924 // PG also accepts ENABLE/DISABLE { REPLICA | ALWAYS }
11925 // TRIGGER … and ENABLE/DISABLE RULE / ROW LEVEL
11926 // SECURITY. v7.16.1 only matches TRIGGER (mailrs's
11927 // pg_dump output) — anything else falls through to
11928 // the catch-all error below.
11929 // v7.22 (round-13 T3) — mysqldump wraps every data
11930 // section in `/*!40000 ALTER TABLE t DISABLE KEYS */`
11931 // + ENABLE KEYS (a MyISAM index-rebuild hint). SPG
11932 // maintains indexes incrementally — engine no-op.
11933 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("keys")) {
11934 self.advance();
11935 return Ok(Vec::new());
11936 }
11937 // v7.37.18 (18.12) — ENABLE/DISABLE ALWAYS TRIGGER
11938 // and ENABLE/DISABLE REPLICA TRIGGER. PG uses these
11939 // to gate triggers on session_replication_role; SPG
11940 // has no replica role, so the prefix is consumed and
11941 // treated identically to the plain ENABLE/DISABLE
11942 // TRIGGER form.
11943 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("always"))
11944 || matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("replica"))
11945 {
11946 self.advance();
11947 }
11948 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("trigger")) {
11949 return Err(self.err(alloc::format!(
11950 "expected TRIGGER after {}, got {:?}",
11951 if enabled { "ENABLE" } else { "DISABLE" },
11952 self.peek()
11953 )));
11954 }
11955 self.advance();
11956 // `ALL` lexes as Token::All (reserved); also
11957 // accept Token::Ident("all") for symmetry.
11958 // v7.37.18 (18.12) — USER / REPLICA / ALWAYS post-
11959 // TRIGGER selectors. USER (= all user triggers) is
11960 // semantically ALL here; REPLICA / ALWAYS gate on
11961 // session_replication_role which SPG doesn't track.
11962 // All map to TriggerSelector::All.
11963 let which = if matches!(self.peek(), Token::All)
11964 || matches!(self.peek(), Token::Ident(s)
11965 if s.eq_ignore_ascii_case("all")
11966 || s.eq_ignore_ascii_case("user")
11967 || s.eq_ignore_ascii_case("replica")
11968 || s.eq_ignore_ascii_case("always"))
11969 {
11970 self.advance();
11971 crate::ast::TriggerSelector::All
11972 } else {
11973 let name = self.expect_ident_like()?;
11974 crate::ast::TriggerSelector::Named(name)
11975 };
11976 Ok(alloc::vec![crate::ast::AlterTableTarget::SetTriggerEnabled {
11977 which,
11978 enabled,
11979 }])
11980 }
11981 // v7.37.16 (16.3) — ATTACH PARTITION child <bounds>
11982 Token::Ident(s) if s.eq_ignore_ascii_case("attach") => {
11983 self.advance();
11984 if !matches!(self.peek(), Token::Partition)
11985 && !matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
11986 if s.eq_ignore_ascii_case("partition"))
11987 {
11988 return Err(self.err(alloc::format!(
11989 "expected PARTITION after ATTACH, got {:?}",
11990 self.peek()
11991 )));
11992 }
11993 self.advance();
11994 let child = self.expect_ident_like()?;
11995 let bounds = self.parse_partition_bounds_tail()?;
11996 Ok(alloc::vec![
11997 crate::ast::AlterTableTarget::AttachPartition { child, bounds }
11998 ])
11999 }
12000 // v7.37.16 (16.4 + 16.5) — DETACH PARTITION child [CONCURRENTLY] [FINALIZE]
12001 Token::Ident(s) if s.eq_ignore_ascii_case("detach") => {
12002 self.advance();
12003 if !matches!(self.peek(), Token::Partition)
12004 && !matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12005 if s.eq_ignore_ascii_case("partition"))
12006 {
12007 return Err(self.err(alloc::format!(
12008 "expected PARTITION after DETACH, got {:?}",
12009 self.peek()
12010 )));
12011 }
12012 self.advance();
12013 let child = self.expect_ident_like()?;
12014 let mut concurrently = false;
12015 let mut finalize = false;
12016 loop {
12017 match self.peek().clone() {
12018 Token::Ident(s) | Token::QuotedIdent(s)
12019 if s.eq_ignore_ascii_case("concurrently") =>
12020 {
12021 self.advance();
12022 concurrently = true;
12023 }
12024 Token::Ident(s) | Token::QuotedIdent(s)
12025 if s.eq_ignore_ascii_case("finalize") =>
12026 {
12027 self.advance();
12028 finalize = true;
12029 }
12030 _ => break,
12031 }
12032 }
12033 Ok(alloc::vec![crate::ast::AlterTableTarget::DetachPartition {
12034 child,
12035 concurrently,
12036 finalize,
12037 }])
12038 }
12039 other => Err(self.err(alloc::format!(
12040 "expected SET / ADD / DROP / ALTER / RENAME / ENABLE / DISABLE / ATTACH / DETACH in ALTER TABLE, got {other:?}"
12041 ))),
12042 }
12043 }
12044
12045 /// v7.37.16 (16.3) — parse the `FOR VALUES …` / `DEFAULT`
12046 /// tail used by both CREATE TABLE … PARTITION OF and ALTER
12047 /// TABLE … ATTACH PARTITION. Shares the same grammar as
12048 /// `parse_partition_of_tail`'s bounds branch.
12049 /// v7.37.18 (18.1 + 18.2) — parse the tail of `ALTER COLUMN
12050 /// col DROP …`. Accepts `DROP DEFAULT` and `DROP NOT NULL`,
12051 /// lowering each to the respective AlterTableTarget. Any
12052 /// other DROP subject (IDENTITY, EXPRESSION, etc.) stays a
12053 /// no-op via consume_until_statement_boundary.
12054 fn parse_alter_column_drop_tail(
12055 &mut self,
12056 col_name: String,
12057 ) -> Result<Vec<crate::ast::AlterTableTarget>, ParseError> {
12058 match self.peek().clone() {
12059 Token::Default => {
12060 self.advance();
12061 Ok(alloc::vec![
12062 crate::ast::AlterTableTarget::AlterColumnDropDefault { column: col_name }
12063 ])
12064 }
12065 Token::Not => {
12066 self.advance();
12067 if !matches!(self.peek(), Token::Null) {
12068 return Err(self.err(alloc::format!(
12069 "expected NULL after ALTER COLUMN DROP NOT, got {:?}",
12070 self.peek()
12071 )));
12072 }
12073 self.advance();
12074 Ok(alloc::vec![
12075 crate::ast::AlterTableTarget::AlterColumnDropNotNull { column: col_name }
12076 ])
12077 }
12078 // `DROP EXPRESSION [IF EXISTS]` — de-generate a stored
12079 // generated column into a plain column.
12080 Token::Ident(s) if s.eq_ignore_ascii_case("expression") => {
12081 self.advance();
12082 // v7.39 (round 187, U10) — IF EXISTS was consumed but
12083 // dropped, so the engine still errored on a plain
12084 // column; PG's semantics are NOTICE + skip.
12085 let mut if_exists = false;
12086 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if")) {
12087 self.advance();
12088 if matches!(self.peek(), Token::Ident(e) if e.eq_ignore_ascii_case("exists")) {
12089 self.advance();
12090 if_exists = true;
12091 }
12092 }
12093 Ok(alloc::vec![
12094 crate::ast::AlterTableTarget::AlterColumnDropExpression {
12095 column: col_name,
12096 if_exists,
12097 }
12098 ])
12099 }
12100 // v7.38 (read01, T28) — `DROP IDENTITY [IF EXISTS]` — de-generate an
12101 // identity column into a plain column.
12102 Token::Ident(s) if s.eq_ignore_ascii_case("identity") => {
12103 self.advance();
12104 let mut if_exists = false;
12105 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if")) {
12106 self.advance();
12107 if matches!(self.peek(), Token::Ident(e) if e.eq_ignore_ascii_case("exists")) {
12108 self.advance();
12109 if_exists = true;
12110 }
12111 }
12112 Ok(alloc::vec![
12113 crate::ast::AlterTableTarget::AlterColumnDropIdentity {
12114 column: col_name,
12115 if_exists,
12116 }
12117 ])
12118 }
12119 _ => {
12120 self.consume_until_statement_boundary();
12121 Ok(Vec::new())
12122 }
12123 }
12124 }
12125
12126 /// Parse the optional trailer of `COPY … TO STDOUT`: nothing (text
12127 /// format, no header), the modern `[WITH] ( opt [, opt]* )` list, or
12128 /// the legacy space-separated `[WITH] CSV|TEXT [HEADER] [DELIMITER
12129 /// 'c'] [NULL 'str'] [QUOTE 'c']` spelling.
12130 fn parse_copy_to_options(&mut self) -> Result<crate::ast::CopyOptions, ParseError> {
12131 let mut opts = crate::ast::CopyOptions::default();
12132 if matches!(self.peek(), Token::Eof | Token::Semicolon) {
12133 return Ok(opts);
12134 }
12135 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with")) {
12136 self.advance();
12137 }
12138 if matches!(self.peek(), Token::LParen) {
12139 self.advance();
12140 loop {
12141 self.parse_one_copy_option(&mut opts)?;
12142 match self.peek() {
12143 Token::Comma => {
12144 self.advance();
12145 }
12146 Token::RParen => {
12147 self.advance();
12148 break;
12149 }
12150 other => {
12151 return Err(self.err(alloc::format!(
12152 "expected ',' or ')' in COPY options, got {other:?}"
12153 )));
12154 }
12155 }
12156 }
12157 } else {
12158 while !matches!(self.peek(), Token::Eof | Token::Semicolon) {
12159 self.parse_one_copy_option(&mut opts)?;
12160 }
12161 }
12162 if !matches!(self.peek(), Token::Eof | Token::Semicolon) {
12163 return Err(self.err(alloc::format!(
12164 "unexpected token after COPY options: {:?}",
12165 self.peek()
12166 )));
12167 }
12168 Ok(opts)
12169 }
12170
12171 fn parse_one_copy_option(
12172 &mut self,
12173 opts: &mut crate::ast::CopyOptions,
12174 ) -> Result<(), ParseError> {
12175 use crate::ast::CopyFormat;
12176 // The option keyword. NULL lexes as its own token; the rest are
12177 // bare identifiers.
12178 let kw = match self.advance() {
12179 Token::Null => alloc::string::String::from("NULL"),
12180 Token::Ident(s) => s.to_uppercase(),
12181 other => {
12182 return Err(self.err(alloc::format!(
12183 "expected a COPY option keyword, got {other:?}"
12184 )));
12185 }
12186 };
12187 match kw.as_str() {
12188 "FORMAT" => {
12189 let fmt = self.expect_ident_like()?;
12190 match fmt.to_ascii_uppercase().as_str() {
12191 "CSV" => opts.format = CopyFormat::Csv,
12192 "TEXT" => opts.format = CopyFormat::Text,
12193 other => {
12194 return Err(self.err(alloc::format!(
12195 "COPY format \"{}\" not recognized",
12196 other.to_ascii_lowercase()
12197 )));
12198 }
12199 }
12200 }
12201 // Legacy bare format keywords.
12202 "CSV" => opts.format = CopyFormat::Csv,
12203 "TEXT" => opts.format = CopyFormat::Text,
12204 "HEADER" => {
12205 opts.header = match self.peek() {
12206 Token::True => {
12207 self.advance();
12208 true
12209 }
12210 Token::False => {
12211 self.advance();
12212 false
12213 }
12214 Token::Ident(s) if s.eq_ignore_ascii_case("on") => {
12215 self.advance();
12216 true
12217 }
12218 Token::Ident(s) if s.eq_ignore_ascii_case("off") => {
12219 self.advance();
12220 false
12221 }
12222 // Bare HEADER (no boolean) means HEADER true.
12223 _ => true,
12224 };
12225 }
12226 // r1066 (7.38 S5.1) — pgbench 14+ loads with
12227 // `COPY … WITH (FREEZE ON)`. The hint's PG effect is
12228 // vacuum bookkeeping on a freshly created/truncated
12229 // table; SPG's per-statement visibility makes it a
12230 // faithful no-op, and rejecting it aborted `pgbench -i`
12231 // against the drop-in. Accept ON/OFF/bare, change nothing.
12232 "FREEZE" => match self.peek() {
12233 Token::True | Token::False => {
12234 self.advance();
12235 }
12236 Token::Ident(s)
12237 if s.eq_ignore_ascii_case("on") || s.eq_ignore_ascii_case("off") =>
12238 {
12239 self.advance();
12240 }
12241 _ => {}
12242 },
12243 "DELIMITER" | "QUOTE" | "ESCAPE" => {
12244 let s = match self.advance() {
12245 Token::String(s) => s,
12246 other => {
12247 return Err(self.err(alloc::format!(
12248 "COPY {kw} expects a single-character string, got {other:?}"
12249 )));
12250 }
12251 };
12252 // v7.39 (round 247) — PG's wording (0A000), keyword in
12253 // lowercase: "COPY delimiter must be a single one-byte
12254 // character".
12255 let one_byte_err = || {
12256 self.err(alloc::format!(
12257 "COPY {} must be a single one-byte character",
12258 kw.to_ascii_lowercase()
12259 ))
12260 };
12261 let mut chars = s.chars();
12262 let c = chars.next().ok_or_else(one_byte_err)?;
12263 if chars.next().is_some() || c.len_utf8() != 1 {
12264 return Err(one_byte_err());
12265 }
12266 match kw.as_str() {
12267 "DELIMITER" => opts.delimiter = Some(c),
12268 "QUOTE" => opts.quote = Some(c),
12269 _ => opts.escape = Some(c),
12270 }
12271 }
12272 // v7.39 (round 247) — `FORCE_QUOTE (col, …)` / `FORCE_QUOTE *`.
12273 "FORCE_QUOTE" => {
12274 if matches!(self.peek(), Token::Star) {
12275 self.advance();
12276 opts.force_quote = Some(Vec::new());
12277 } else {
12278 if !matches!(self.peek(), Token::LParen) {
12279 return Err(self.err(alloc::format!(
12280 "expected '(' or '*' after FORCE_QUOTE, got {:?}",
12281 self.peek()
12282 )));
12283 }
12284 self.advance();
12285 let mut cols = Vec::new();
12286 loop {
12287 cols.push(self.expect_ident_like()?);
12288 match self.peek() {
12289 Token::Comma => {
12290 self.advance();
12291 }
12292 Token::RParen => {
12293 self.advance();
12294 break;
12295 }
12296 other => {
12297 return Err(self.err(alloc::format!(
12298 "expected ',' or ')' in FORCE_QUOTE list, got {other:?}"
12299 )));
12300 }
12301 }
12302 }
12303 opts.force_quote = Some(cols);
12304 }
12305 }
12306 "NULL" => {
12307 opts.null_str = Some(match self.advance() {
12308 Token::String(s) => s,
12309 other => {
12310 return Err(self.err(alloc::format!(
12311 "COPY NULL expects a quoted string, got {other:?}"
12312 )));
12313 }
12314 });
12315 }
12316 // v7.39 (round 265) — the two CSV FROM-side column lists. Same
12317 // grammar as FORCE_QUOTE; PG accepts `*` for FORCE_NOT_NULL /
12318 // FORCE_NULL too.
12319 "FORCE_NOT_NULL" | "FORCE_NULL" => {
12320 let cols = self.parse_copy_column_list(&kw)?;
12321 if kw == "FORCE_NOT_NULL" {
12322 opts.force_not_null = Some(cols);
12323 } else {
12324 opts.force_null = Some(cols);
12325 }
12326 }
12327 other => {
12328 // PG's wording, lowercased option name.
12329 return Err(self.err(alloc::format!(
12330 "option \"{}\" not recognized",
12331 other.to_ascii_lowercase()
12332 )));
12333 }
12334 }
12335 Ok(())
12336 }
12337
12338 /// v7.39 (round 265) — `( col, … )` or `*` after a COPY column-list
12339 /// option (FORCE_QUOTE / FORCE_NOT_NULL / FORCE_NULL). An empty vec
12340 /// is the `*` spelling.
12341 fn parse_copy_column_list(&mut self, kw: &str) -> Result<Vec<String>, ParseError> {
12342 if matches!(self.peek(), Token::Star) {
12343 self.advance();
12344 return Ok(Vec::new());
12345 }
12346 if !matches!(self.peek(), Token::LParen) {
12347 return Err(self.err(alloc::format!(
12348 "expected '(' or '*' after {kw}, got {:?}",
12349 self.peek()
12350 )));
12351 }
12352 self.advance();
12353 let mut cols = Vec::new();
12354 loop {
12355 cols.push(self.expect_ident_like()?);
12356 match self.peek() {
12357 Token::Comma => {
12358 self.advance();
12359 }
12360 Token::RParen => {
12361 self.advance();
12362 break;
12363 }
12364 other => {
12365 return Err(self.err(alloc::format!(
12366 "expected ',' or ')' in {kw} list, got {other:?}"
12367 )));
12368 }
12369 }
12370 }
12371 Ok(cols)
12372 }
12373
12374 fn parse_partition_bounds_tail(
12375 &mut self,
12376 ) -> Result<crate::ast::PartitionOfBoundsAst, ParseError> {
12377 use crate::ast::PartitionOfBoundsAst;
12378 match self.peek() {
12379 Token::Default => {
12380 self.advance();
12381 Ok(PartitionOfBoundsAst::Default)
12382 }
12383 Token::For => {
12384 self.advance();
12385 if !matches!(self.peek(), Token::Values) {
12386 return Err(
12387 self.err(format!("expected VALUES after FOR, got {:?}", self.peek()))
12388 );
12389 }
12390 self.advance();
12391 let want_with = matches!(
12392 self.peek(),
12393 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("with")
12394 );
12395 if want_with {
12396 self.advance();
12397 if !matches!(self.peek(), Token::LParen) {
12398 return Err(self.err(format!(
12399 "expected '(' after FOR VALUES WITH, got {:?}",
12400 self.peek()
12401 )));
12402 }
12403 self.advance();
12404 let (mut modulus, mut remainder): (Option<u32>, Option<u32>) = (None, None);
12405 loop {
12406 let key = self.expect_ident_like()?;
12407 let n = match self.peek().clone() {
12408 Token::Integer(v) if u32::try_from(v).is_ok() => {
12409 self.advance();
12410 v as u32
12411 }
12412 other => {
12413 return Err(self.err(format!(
12414 "FOR VALUES WITH: expected unsigned integer literal, got {other:?}"
12415 )));
12416 }
12417 };
12418 match key.to_ascii_uppercase().as_str() {
12419 "MODULUS" => modulus = Some(n),
12420 "REMAINDER" => remainder = Some(n),
12421 other => {
12422 return Err(self.err(format!(
12423 "FOR VALUES WITH: unknown key {other:?}; \
12424 expected MODULUS or REMAINDER"
12425 )));
12426 }
12427 }
12428 match self.peek() {
12429 Token::Comma => {
12430 self.advance();
12431 }
12432 Token::RParen => {
12433 self.advance();
12434 break;
12435 }
12436 other => {
12437 return Err(self.err(format!(
12438 "expected ',' or ')' in FOR VALUES WITH list, got {other:?}"
12439 )));
12440 }
12441 }
12442 }
12443 let modulus = modulus
12444 .ok_or_else(|| self.err("FOR VALUES WITH: missing MODULUS".to_string()))?;
12445 let remainder = remainder.ok_or_else(|| {
12446 self.err("FOR VALUES WITH: missing REMAINDER".to_string())
12447 })?;
12448 if modulus == 0 {
12449 return Err(self.err("FOR VALUES WITH: MODULUS must be > 0".to_string()));
12450 }
12451 if remainder >= modulus {
12452 return Err(self.err(format!(
12453 "FOR VALUES WITH: REMAINDER ({remainder}) must be < MODULUS ({modulus})"
12454 )));
12455 }
12456 return Ok(PartitionOfBoundsAst::Hash { modulus, remainder });
12457 }
12458 match self.peek() {
12459 Token::From => {
12460 self.advance();
12461 let lower = Box::new(self.parse_partition_bound_expr()?);
12462 if !matches!(self.peek(), Token::To) {
12463 return Err(self.err(format!(
12464 "expected TO after FROM (...), got {:?}",
12465 self.peek()
12466 )));
12467 }
12468 self.advance();
12469 let upper = Box::new(self.parse_partition_bound_expr()?);
12470 Ok(PartitionOfBoundsAst::Range { lower, upper })
12471 }
12472 Token::In => {
12473 self.advance();
12474 if !matches!(self.peek(), Token::LParen) {
12475 return Err(self.err(format!(
12476 "expected '(' after FOR VALUES IN, got {:?}",
12477 self.peek()
12478 )));
12479 }
12480 self.advance();
12481 let mut values = Vec::new();
12482 loop {
12483 values.push(self.parse_expr(0)?);
12484 match self.peek() {
12485 Token::Comma => {
12486 self.advance();
12487 }
12488 Token::RParen => {
12489 self.advance();
12490 break;
12491 }
12492 other => {
12493 return Err(self.err(format!(
12494 "expected ',' or ')' in FOR VALUES IN list, got {other:?}"
12495 )));
12496 }
12497 }
12498 }
12499 if values.is_empty() {
12500 return Err(
12501 self.err("FOR VALUES IN requires at least one literal".to_string())
12502 );
12503 }
12504 Ok(PartitionOfBoundsAst::List { values })
12505 }
12506 other => Err(self.err(format!(
12507 "expected FROM / IN / WITH after FOR VALUES, got {other:?}"
12508 ))),
12509 }
12510 }
12511 other => Err(self.err(format!(
12512 "expected DEFAULT or FOR VALUES after ATTACH PARTITION child, got {other:?}"
12513 ))),
12514 }
12515 }
12516
12517 /// v7.16.2 — peek for `information_schema.<tbl>` /
12518 /// `pg_catalog.<tbl>` triples and, if matched, consume all
12519 /// three tokens + return a synthetic table name the engine's
12520 /// SELECT path recognises as a virtual view. Returns `None`
12521 /// when the head doesn't look like a meta-qualified name.
12522 /// Used by `parse_table_ref` to bypass the
12523 /// `expect_ident_like` schema-strip for these specific PG
12524 /// meta schemas (mailrs round-10 A.3).
12525 fn try_peek_meta_qualified(&mut self) -> Option<(String, String)> {
12526 // Extract the schema name. Must be a plain ident token.
12527 let schema = match self.tokens.get(self.pos) {
12528 Some(Token::Ident(s) | Token::QuotedIdent(s)) => s.clone(),
12529 _ => return None,
12530 };
12531 // Dot.
12532 if !matches!(self.tokens.get(self.pos + 1), Some(Token::Dot)) {
12533 return None;
12534 }
12535 // The table-side ident may lex as a reserved keyword
12536 // (e.g. `Token::Tables`). Tolerate the common ones via a
12537 // helper that reads the trailing token's underlying name.
12538 let tbl = match self.tokens.get(self.pos + 2)? {
12539 Token::Ident(t) | Token::QuotedIdent(t) => t.clone(),
12540 Token::Tables => "tables".to_string(),
12541 // Other PG meta table names that may collide with
12542 // reserved keywords land here as needed.
12543 _ => return None,
12544 };
12545 // Strip the `pg_` prefix from `pg_catalog.pg_class`-style
12546 // names so the synthetic name doesn't double-prefix
12547 // (`__spg_pg_class`, not `__spg_pg_pg_class`).
12548 let (prefix, normalised) = if schema.eq_ignore_ascii_case("information_schema") {
12549 ("__spg_info_", tbl.to_ascii_lowercase())
12550 } else if schema.eq_ignore_ascii_case("pg_catalog") {
12551 // v7.39 (round 541) — only the catalogs SPG actually
12552 // synthesises are rewritten, which is what the BARE path
12553 // has always checked. Anything else keeps its own name and
12554 // takes the ordinary route: `pg_stat_activity` and friends
12555 // resolve through meta_view_result, and a name that is no
12556 // catalog at all gets PG's "relation does not exist"
12557 // instead of a message about a view SPG cannot materialise.
12558 let lowered = tbl.to_ascii_lowercase();
12559 if !SYNTHESISED_PG_CATALOGS.contains(&lowered.as_str()) {
12560 self.advance(); // schema
12561 self.advance(); // dot
12562 self.advance(); // tbl
12563 return Some((lowered.clone(), lowered));
12564 }
12565 let bare = lowered
12566 .strip_prefix("pg_")
12567 .map(alloc::string::String::from)
12568 .unwrap_or(lowered);
12569 ("__spg_pg_", bare)
12570 } else if schema.eq_ignore_ascii_case("mysql") {
12571 // v7.17.0 Phase 3.P0-65 — MySQL system schema
12572 // (`mysql.user`, `mysql.db`). Same synthetic-name
12573 // shape as pg_catalog.
12574 ("__spg_mysql_", tbl.to_ascii_lowercase())
12575 } else {
12576 return None;
12577 };
12578 self.advance(); // schema
12579 self.advance(); // dot
12580 self.advance(); // tbl
12581 Some((
12582 alloc::format!("{prefix}{normalised}"),
12583 tbl.to_ascii_lowercase(),
12584 ))
12585 }
12586
12587 /// Unqualified PG meta-table names (`FROM pg_extension`, `FROM
12588 /// pg_class`) resolve the same way: PG puts `pg_catalog` at the
12589 /// implicit front of every search_path, so a bare reference to a
12590 /// known catalog table always means the catalog table. Only the
12591 /// names the engine actually synthesises are recognised — any
12592 /// other `pg_*` ident stays a user table (mailrs embed round-12).
12593 fn try_peek_meta_bare(&mut self) -> Option<(String, String)> {
12594 // v7.38 (read01 P3.21) — every catalog view SPG synthesises
12595 // (`__spg_pg_*`) is bare-resolvable, matching PG's implicit
12596 // `pg_catalog` at the front of every search_path. (pg_stat_activity
12597 // / pg_stat_statements / pg_locks / pg_statio_user_tables route
12598 // through the meta_view_result path instead, and already resolve
12599 // bare — they must NOT be listed here or the __spg_ rewrite would
12600 // mis-target them.)
12601 const PG_META_TABLES: &[&str] = SYNTHESISED_PG_CATALOGS;
12602 let name = match self.tokens.get(self.pos) {
12603 Some(Token::Ident(s)) => s.to_ascii_lowercase(),
12604 _ => return None,
12605 };
12606 // A following dot means this ident is a schema qualifier,
12607 // not a table name — let the qualified path handle it.
12608 if matches!(self.tokens.get(self.pos + 1), Some(Token::Dot)) {
12609 return None;
12610 }
12611 if !PG_META_TABLES.contains(&name.as_str()) {
12612 return None;
12613 }
12614 self.advance();
12615 let bare = name.strip_prefix("pg_").unwrap_or(&name);
12616 Some((alloc::format!("__spg_pg_{bare}"), name.clone()))
12617 }
12618
12619 /// Consume a bare ident if its lowercase matches `kw`, else err.
12620 /// v7.39 (read01 round 57) — is the next token this bare keyword-ident?
12621 /// Peeks only; the caller advances.
12622 fn peek_keyword_ident(&self, kw: &str) -> bool {
12623 matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case(kw))
12624 }
12625
12626 fn expect_keyword_ident(&mut self, kw: &str) -> Result<(), ParseError> {
12627 match self.advance() {
12628 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case(kw) => Ok(()),
12629 other => Err(ParseError {
12630 message: format!("expected {kw:?}, got {other:?}"),
12631 token_pos: self.consumed_pos(),
12632 }),
12633 }
12634 }
12635
12636 /// Accept either a quoted identifier (`"foo"`) or a quoted string
12637 /// literal (`'foo'`) — same shape used by CREATE USER for the
12638 /// username slot.
12639 fn expect_ident_or_string(&mut self) -> Result<String, ParseError> {
12640 match self.advance() {
12641 Token::Ident(s) | Token::QuotedIdent(s) | Token::String(s) => Ok(s),
12642 other => Err(ParseError {
12643 message: format!("expected identifier or string, got {other:?}"),
12644 token_pos: self.consumed_pos(),
12645 }),
12646 }
12647 }
12648
12649 fn expect_string_literal(&mut self) -> Result<String, ParseError> {
12650 match self.advance() {
12651 Token::String(s) => Ok(s),
12652 other => Err(ParseError {
12653 message: format!("expected quoted string, got {other:?}"),
12654 token_pos: self.consumed_pos(),
12655 }),
12656 }
12657 }
12658
12659 fn parse_select_stmt(&mut self) -> Result<Statement, ParseError> {
12660 // v7.30.2 (mailrs round-25 ask 2) — derived tables /
12661 // subqueries recurse through here without passing
12662 // parse_expr; share the same nesting budget.
12663 self.enter_nested()?;
12664 let r = self.parse_select_stmt_inner();
12665 self.nest_depth -= 1;
12666 r
12667 }
12668
12669 fn parse_select_stmt_inner(&mut self) -> Result<Statement, ParseError> {
12670 // Caller dispatches on Token::Select; the inner helper handles
12671 // the rest. ORDER BY / LIMIT bind at this top level; UNION peers
12672 // get a fresh bare-select parse and may not have their own ORDER
12673 // BY / LIMIT.
12674 let mut head = self.parse_bare_select()?;
12675 let into = self.pending_select_into.take();
12676 self.parse_setop_chain_into(&mut head)?;
12677 self.parse_select_tail_into(&mut head)?;
12678 // v7.38.19 — `SELECT … INTO t` lowers to the SAME node as
12679 // `CREATE TABLE t AS SELECT …`, which is what a comment in
12680 // `ast.rs` has claimed since v7.38 and what only CTAS actually
12681 // did. The tail (ORDER BY / LIMIT) is parsed first so it belongs
12682 // to the body, as it does in PostgreSQL.
12683 if let Some((name, temporary)) = into {
12684 return Ok(Statement::CreateMaterializedView(
12685 crate::ast::CreateMaterializedViewStatement {
12686 temporary,
12687 name,
12688 if_not_exists: false,
12689 columns: Vec::new(),
12690 body: head,
12691 with_data: true,
12692 as_plain_table: true,
12693 },
12694 ));
12695 }
12696 Ok(Statement::Select(head))
12697 }
12698
12699 /// v7.37.17 (17.6 siblings) — the three SQL set operations
12700 /// share the peer chain: UNION [ALL], EXCEPT [ALL] (a reserved
12701 /// token), and INTERSECT [ALL] (a bare ident — it was never
12702 /// reserved in SPG's lexer). PG precedence: INTERSECT binds
12703 /// tighter than UNION / EXCEPT — the executor folds the chain
12704 /// left-to-right, which is already correct for LEADING
12705 /// intersects; an INTERSECT pair that FOLLOWS a union/except
12706 /// pair nests into that previous peer, so A UNION B INTERSECT C
12707 /// = A ∪ (B ∩ C). Shared by the top level and parenthesized
12708 /// groups.
12709 fn parse_setop_chain_into(&mut self, head: &mut SelectStatement) -> Result<(), ParseError> {
12710 // A parenthesized group arrives with its own (already
12711 // regrouped) unions on `head`; only the pairs THIS chain
12712 // appends participate in the precedence regroup below —
12713 // nesting an outer INTERSECT into a group-internal peer
12714 // would dissolve the explicit grouping.
12715 let boundary = head.unions.len();
12716 loop {
12717 let base = match self.peek() {
12718 Token::Union => UnionKind::Distinct,
12719 Token::Except => UnionKind::Except,
12720 Token::Ident(s) if s.eq_ignore_ascii_case("intersect") => UnionKind::Intersect,
12721 _ => break,
12722 };
12723 self.advance();
12724 let kind = if matches!(self.peek(), Token::All) {
12725 self.advance();
12726 match base {
12727 UnionKind::Distinct => UnionKind::All,
12728 UnionKind::Except => UnionKind::ExceptAll,
12729 _ => UnionKind::IntersectAll,
12730 }
12731 } else {
12732 base
12733 };
12734 let peer = self.parse_bare_select()?;
12735 head.unions.push((kind, peer));
12736 }
12737 let mut pairs = core::mem::take(&mut head.unions);
12738 let tail = pairs.split_off(boundary);
12739 let mut regrouped: Vec<(UnionKind, SelectStatement)> = pairs;
12740 for (kind, peer) in tail {
12741 let is_intersect = matches!(kind, UnionKind::Intersect | UnionKind::IntersectAll);
12742 // An intersect nests into the previous element of THIS
12743 // chain only; with no new previous element it stays at
12744 // the outer level (the left fold applies it to the
12745 // whole head, group included).
12746 match (
12747 is_intersect,
12748 regrouped.len() > boundary,
12749 regrouped.last_mut(),
12750 ) {
12751 (true, true, Some((_, prev))) => prev.unions.push((kind, peer)),
12752 _ => regrouped.push((kind, peer)),
12753 }
12754 }
12755 head.unions = regrouped;
12756 Ok(())
12757 }
12758
12759 /// v7.37.17 (17.6 siblings) — the shared SELECT tail: ORDER BY /
12760 /// LIMIT / OFFSET / FETCH FIRST / FOR-lock clauses. Extracted so
12761 /// the top-level bare VALUES statement reuses it verbatim.
12762 /// v6.4.0 — parse an optional `ORDER BY <expr> [ASC|DESC] [NULLS …], …`
12763 /// clause into its key list (empty when no `ORDER BY` follows). Extracted
12764 /// (v7.39 round 135) so the grouping-set path can parse ORDER BY early,
12765 /// where the grouping-set universe is still in scope.
12766 fn parse_order_by_keys(&mut self) -> Result<Vec<OrderBy>, ParseError> {
12767 if !matches!(self.peek(), Token::Order) {
12768 return Ok(Vec::new());
12769 }
12770 self.advance();
12771 if !self.peek_is_by() {
12772 return Err(self.err(format!("expected BY after ORDER, got {:?}", self.peek())));
12773 }
12774 self.advance();
12775 let mut keys = Vec::new();
12776 loop {
12777 // v7.39 (round 691) — save/restore, the discipline this parser
12778 // already uses around `pending_sample_preds`, so a subquery inside
12779 // a key neither inherits nor leaks the channel.
12780 let saved_flag = core::mem::replace(&mut self.in_order_by_key, true);
12781 let saved_coll = self.order_key_collation.take();
12782 let parsed = self.parse_expr(0);
12783 self.in_order_by_key = saved_flag;
12784 let collation = core::mem::replace(&mut self.order_key_collation, saved_coll);
12785 let expr = parsed?;
12786 let desc = if matches!(self.peek(), Token::Desc) {
12787 self.advance();
12788 true
12789 } else if matches!(self.peek(), Token::Asc) {
12790 self.advance();
12791 false
12792 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
12793 // `ORDER BY x USING <op>` — PG's operator-class spelling. SPG has
12794 // one ordering per type, so the btree comparison operators map
12795 // onto it: < / <= are ASC, > / >= are DESC. Any other operator
12796 // would need a custom operator class — honest error.
12797 self.advance();
12798 match self.advance() {
12799 Token::Lt | Token::LtEq => false,
12800 Token::Gt | Token::GtEq => true,
12801 other => {
12802 return Err(self.err(alloc::format!(
12803 "ORDER BY USING supports the btree comparison \
12804 operators (< <= > >=); got {other:?}"
12805 )));
12806 }
12807 }
12808 } else {
12809 false
12810 };
12811 // v7.24 (round-16 A) — explicit NULLS FIRST/LAST.
12812 let nulls_first = self.parse_optional_nulls_placement()?;
12813 keys.push(OrderBy {
12814 expr,
12815 desc,
12816 nulls_first,
12817 collation,
12818 });
12819 if matches!(self.peek(), Token::Comma) {
12820 self.advance();
12821 } else {
12822 break;
12823 }
12824 }
12825 Ok(keys)
12826 }
12827
12828 fn parse_select_tail_into(&mut self, head: &mut SelectStatement) -> Result<(), ParseError> {
12829 // v7.39 (round 135) — a grouping-set query may have already parsed +
12830 // rewritten its ORDER BY (to reference synthetic grouping columns); if
12831 // no ORDER BY token is present, keep that pre-set order_by rather than
12832 // clobbering it with an empty list.
12833 let parsed_keys = self.parse_order_by_keys()?;
12834 head.order_by = if parsed_keys.is_empty() {
12835 core::mem::take(&mut head.order_by)
12836 } else {
12837 parsed_keys
12838 };
12839 // v7.39 (round 314, V39) — the row-count clauses come in EITHER
12840 // order. PG's grammar takes a limit clause and an offset clause
12841 // as an unordered pair, so `OFFSET 2 LIMIT 3` means exactly what
12842 // `LIMIT 3 OFFSET 2` does (measured: same rows). This used to
12843 // parse them in a fixed LIMIT-then-OFFSET sequence, so the other
12844 // spelling died on `expected end of input, got Limit`.
12845 //
12846 // Each may appear at most once, and LIMIT and FETCH FIRST are
12847 // two spellings of the same clause — PG rejects `LIMIT 1 LIMIT 2`,
12848 // `OFFSET 1 OFFSET 2` and `LIMIT 2 FETCH FIRST 3 ROWS ONLY` alike.
12849 // A second one is left unconsumed here, which the caller reports
12850 // as trailing input rather than silently taking the last.
12851 let mut saw_limit = false;
12852 let mut saw_offset = false;
12853 loop {
12854 if !saw_limit && matches!(self.peek(), Token::Limit) {
12855 self.advance();
12856 // v7.17.0 Phase 5.1 — `LIMIT NULL` / `LIMIT ALL` are
12857 // PG synonyms for "no limit". Treat both as None
12858 // (no head.limit set) so the engine's existing
12859 // unlimited-result path takes over. Reject was the
12860 // pre-5.1 behaviour and broke pg_dump-flavoured
12861 // tooling that occasionally emits LIMIT NULL.
12862 if self.consume_limit_unbounded_sentinel() {
12863 head.limit = None;
12864 } else {
12865 let first = self.parse_limit_expr("LIMIT")?;
12866 // MySQL `LIMIT offset, count` — the first number is
12867 // the offset when a comma follows.
12868 if matches!(self.peek(), Token::Comma) {
12869 self.advance();
12870 let count = self.parse_limit_expr("LIMIT")?;
12871 head.offset = Some(first);
12872 saw_offset = true;
12873 head.limit = Some(count);
12874 } else {
12875 head.limit = Some(first);
12876 }
12877 }
12878 saw_limit = true;
12879 continue;
12880 }
12881 if !saw_offset && matches!(self.peek(), Token::Offset) {
12882 self.advance();
12883 // PG also accepts an optional `ROW` / `ROWS` trailer
12884 // after the offset value (`OFFSET 10 ROWS`). The
12885 // FETCH-FIRST branch below relies on the same.
12886 let off = self.parse_limit_expr("OFFSET")?;
12887 self.consume_optional_rows_keyword();
12888 head.offset = Some(off);
12889 saw_offset = true;
12890 continue;
12891 }
12892 // v7.17.0 Phase 5.1 — `FETCH FIRST <int|$N> ROWS ONLY` is
12893 // the SQL-standard alias for LIMIT. PG accepts both
12894 // spellings interchangeably; pg_dump emits FETCH FIRST in
12895 // newer versions. We map it onto `head.limit` so the
12896 // engine path is unified.
12897 if !saw_limit
12898 && matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12899 if s.eq_ignore_ascii_case("fetch"))
12900 {
12901 self.advance(); // FETCH
12902 // `FIRST` or `NEXT` (both legal per SQL standard).
12903 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12904 if s.eq_ignore_ascii_case("first") || s.eq_ignore_ascii_case("next"))
12905 {
12906 self.advance();
12907 }
12908 // Count (optional in the bare `FETCH FIRST ROW ONLY` —
12909 // implicit 1 — but we always consume one if present).
12910 let count = if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12911 if s.eq_ignore_ascii_case("row") || s.eq_ignore_ascii_case("rows"))
12912 {
12913 // Bare `FETCH FIRST ROW ONLY` = LIMIT 1.
12914 crate::ast::LimitExpr::Literal(1)
12915 } else {
12916 self.parse_limit_expr("FETCH FIRST")?
12917 };
12918 // Eat `ROW` / `ROWS` if not already consumed above.
12919 self.consume_optional_rows_keyword();
12920 // Optional `ONLY` (the spec form) — or the SQL:2008
12921 // `WITH TIES` form. v7.17.0 Phase 3.P0-49: the executor
12922 // now honours WITH TIES by extending past the LIMIT
12923 // truncation point through every row that shares the
12924 // last-kept row's ORDER BY key.
12925 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12926 if s.eq_ignore_ascii_case("only"))
12927 {
12928 self.advance();
12929 } else if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12930 if s.eq_ignore_ascii_case("with"))
12931 {
12932 self.advance(); // WITH
12933 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12934 if s.eq_ignore_ascii_case("ties"))
12935 {
12936 self.advance();
12937 head.limit_with_ties = true;
12938 }
12939 }
12940 head.limit = Some(count);
12941 saw_limit = true;
12942 continue;
12943 }
12944 break;
12945 }
12946 // v7.17.0 Phase 3.4 — trailing row-lock clauses:
12947 // FOR { UPDATE | NO KEY UPDATE | SHARE | KEY SHARE }
12948 // [ OF table_name [, …] ]
12949 // [ NOWAIT | SKIP LOCKED ]
12950 // Multiple FOR clauses may stack (PG: `FOR UPDATE OF t1
12951 // FOR SHARE OF t2`). SPG is a single-writer engine — every
12952 // SELECT already returns a consistent snapshot — so these
12953 // are accept-and-discard: the parser absorbs them so
12954 // mailrs / Rails / Django code paths that emit `SELECT
12955 // … FOR UPDATE` for advisory pessimistic locking load
12956 // without a parser error. The on-disk locking model is
12957 // unchanged; callers that rely on FOR UPDATE for read-
12958 // through-write ordering still get the right answer
12959 // because SPG serialises writes anyway.
12960 head.locking = self
12961 .consume_optional_for_lock_clauses()
12962 .map(alloc::boxed::Box::new);
12963 Ok(())
12964 }
12965
12966 /// v7.17.0 Phase 3.4 — eat zero or more `FOR { UPDATE | NO KEY
12967 /// UPDATE | SHARE | KEY SHARE } [ OF tbl[, …] ] [ NOWAIT | SKIP
12968 /// LOCKED ]` trailers. Each clause is fully accepted and
12969 /// discarded — SPG's single-writer model already satisfies the
12970 /// callers' implicit ordering requirement. Stops at the first
12971 /// token that isn't `FOR`.
12972 fn consume_optional_for_lock_clauses(&mut self) -> Option<crate::ast::LockingClause> {
12973 // v7.39 (round 293, E3 Phase 1) — the clause is REPORTED now,
12974 // not discarded. PG keeps the strongest of several clauses; the
12975 // policy of the last one wins, which is what this loop records.
12976 let mut seen: Option<crate::ast::LockingClause> = None;
12977 while matches!(self.peek(), Token::For) {
12978 // v7.37.14 (A2.5-stub) — record that this query asked
12979 // for a row lock the parser is about to silently
12980 // discard. Operators surface the count via
12981 // `spg_sql::silent_for_update_count()` so they can
12982 // gauge how much of the workload depends on advisory
12983 // FOR UPDATE / FOR SHARE / FOR KEY SHARE semantics
12984 // before v7.37.15's per-row tuple locking lands.
12985 crate::record_silent_for_update_clause();
12986 self.advance(); // FOR
12987 // `NO KEY` prefix (PG) — `NO` is reserved-keyword-shaped
12988 // (`Token::Not` isn't it; PG `NO` lexes as Token::Ident).
12989 let mut no_key = false;
12990 let mut key = false;
12991 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12992 if s.eq_ignore_ascii_case("no"))
12993 {
12994 self.advance(); // NO
12995 no_key = true;
12996 // The next ident should be KEY but be generous;
12997 // anything followed by UPDATE/SHARE is accepted.
12998 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
12999 if s.eq_ignore_ascii_case("key"))
13000 {
13001 self.advance(); // KEY
13002 }
13003 }
13004 // `KEY` prefix (PG `FOR KEY SHARE`).
13005 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13006 if s.eq_ignore_ascii_case("key"))
13007 {
13008 self.advance(); // KEY
13009 key = true;
13010 }
13011 // Lock-strength keyword: UPDATE / SHARE. Required, but
13012 // we're lenient — an unexpected token here just bails
13013 // (we already consumed FOR; caller's downstream
13014 // dispatch will error if anything actually depends on
13015 // the trailing tokens).
13016 let is_update = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13017 if s.eq_ignore_ascii_case("update"));
13018 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13019 if s.eq_ignore_ascii_case("update") || s.eq_ignore_ascii_case("share"))
13020 {
13021 self.advance();
13022 use crate::ast::LockStrength as LS;
13023 let strength = match (is_update, no_key, key) {
13024 (true, true, _) => LS::NoKeyUpdate,
13025 (true, _, _) => LS::Update,
13026 (false, _, true) => LS::KeyShare,
13027 (false, _, _) => LS::Share,
13028 };
13029 seen = Some(crate::ast::LockingClause {
13030 strength,
13031 of_tables: alloc::vec::Vec::new(),
13032 policy: crate::ast::LockWait::Wait,
13033 });
13034 } else {
13035 // FOR by itself (or `FOR KEY` with nothing after) —
13036 // give up on the lock-clause path. We've already
13037 // advanced past FOR; further attempts to parse
13038 // here would clobber state.
13039 return seen;
13040 }
13041 // Optional `OF tbl[, tbl …]`. mailrs emits this when
13042 // joining and locking only a subset of tables.
13043 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13044 if s.eq_ignore_ascii_case("of"))
13045 {
13046 self.advance(); // OF
13047 #[allow(clippy::while_let_loop)]
13048 loop {
13049 match self.peek() {
13050 Token::Ident(_) | Token::QuotedIdent(_) => {
13051 // v7.39 (round 294) — the name is CAPTURED now: PG
13052 // validates it against the FROM clause, and an
13053 // uncaptured list silently means "lock everything".
13054 let mut nm = match self.advance() {
13055 Token::Ident(n) | Token::QuotedIdent(n) => n,
13056 _ => alloc::string::String::new(),
13057 };
13058 // Optional schema-qualified `schema.table`.
13059 if matches!(self.peek(), Token::Dot) {
13060 self.advance();
13061 if let Token::Ident(n) | Token::QuotedIdent(n) = self.peek().clone()
13062 {
13063 self.advance();
13064 nm = n;
13065 }
13066 }
13067 if let Some(c) = seen.as_mut() {
13068 c.of_tables.push(nm);
13069 }
13070 }
13071 _ => break,
13072 }
13073 if matches!(self.peek(), Token::Comma) {
13074 self.advance();
13075 } else {
13076 break;
13077 }
13078 }
13079 }
13080 // Optional `NOWAIT` | `SKIP LOCKED`.
13081 match self.peek().clone() {
13082 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("nowait") => {
13083 self.advance();
13084 if let Some(c) = seen.as_mut() {
13085 c.policy = crate::ast::LockWait::NoWait;
13086 }
13087 }
13088 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("skip") => {
13089 self.advance(); // SKIP
13090 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13091 if s.eq_ignore_ascii_case("locked"))
13092 {
13093 self.advance(); // LOCKED
13094 if let Some(c) = seen.as_mut() {
13095 c.policy = crate::ast::LockWait::SkipLocked;
13096 }
13097 }
13098 }
13099 _ => {}
13100 }
13101 // Loop: PG allows multiple FOR clauses chained.
13102 }
13103 seen
13104 }
13105
13106 /// v7.9.24 — accept `LIMIT <int>` or `LIMIT $N`. mailrs H2.
13107 /// Bind value gets resolved during prepared-statement Execute;
13108 /// the Pratt expression parser would over-accept here (e.g.
13109 /// `LIMIT 5 + 5`), so we narrowly accept only the two PG forms.
13110 /// v7.17.0 Phase 5.1 — consume the `LIMIT NULL` / `LIMIT ALL`
13111 /// sentinel tokens (PG synonyms for "no limit"). Returns true
13112 /// when one was consumed; caller skips the regular
13113 /// limit-value parse and leaves `head.limit` at None.
13114 fn consume_limit_unbounded_sentinel(&mut self) -> bool {
13115 if matches!(self.peek(), Token::Null) {
13116 self.advance();
13117 return true;
13118 }
13119 if matches!(self.peek(), Token::All) {
13120 self.advance();
13121 return true;
13122 }
13123 false
13124 }
13125
13126 /// v7.17.0 Phase 5.1 — eat an optional trailing `ROW` / `ROWS`
13127 /// keyword after a LIMIT / OFFSET / FETCH FIRST value, the
13128 /// SQL-standard shape. No-op when missing.
13129 fn consume_optional_rows_keyword(&mut self) {
13130 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13131 if s.eq_ignore_ascii_case("row") || s.eq_ignore_ascii_case("rows"))
13132 {
13133 self.advance();
13134 }
13135 }
13136
13137 /// v7.39 (round 284) — `LIMIT` / `OFFSET` over a general expression.
13138 ///
13139 /// PG's row-count clause takes an `a_expr`, so `LIMIT 1+1` and
13140 /// `OFFSET 2+3` are legal; only `FETCH FIRST` is restricted to a
13141 /// constant, which is why that spelling keeps the token path below.
13142 ///
13143 /// Constants are folded here rather than carried into the tree: the
13144 /// 15+ execution paths that read the row count go through
13145 /// `limit_literal()`, which answers `Option<u32>` — and `None` there
13146 /// means "no limit". A clause the engine could not resolve would
13147 /// therefore return the WHOLE table instead of failing. Folding at
13148 /// parse time keeps that impossible; a non-constant clause is still
13149 /// a clean error (recorded residual — closing it wants a resolution
13150 /// pre-pass on the simple-query path, where `substitute_placeholders`
13151 /// does not run).
13152 fn parse_limit_expr(&mut self, label: &str) -> Result<crate::ast::LimitExpr, ParseError> {
13153 // PG restricts FETCH FIRST to a constant or a PARENTHESISED
13154 // expression: `FETCH FIRST 1+1 ROWS ONLY` is a syntax error, but
13155 // `FETCH FIRST (1+1) ROWS ONLY` and `FETCH FIRST (SELECT 3) ROWS
13156 // ONLY` both work (its grammar takes a c_expr). Both measured
13157 // against PG 18.4 in round 305.
13158 if label == "FETCH FIRST" && !matches!(self.peek(), Token::LParen) {
13159 return self.parse_limit_constant(label);
13160 }
13161 // One pass, no rewind: `advance()` takes each token by
13162 // `mem::replace`, so a consumed token reads back as Eof and this
13163 // parser cannot backtrack. Everything — bare literal included —
13164 // is therefore folded from the parsed expression rather than
13165 // re-read from the token stream.
13166 let start = self.pos;
13167 let e = self.parse_expr(0)?;
13168 if let crate::ast::Expr::Placeholder(n) = e {
13169 return Ok(crate::ast::LimitExpr::Placeholder(n));
13170 }
13171 let neg_label = if label == "OFFSET" { "OFFSET" } else { "LIMIT" };
13172 match fold_limit_constant(&e) {
13173 Some(Ok(v)) if v < 0 => Err(ParseError {
13174 message: alloc::format!("{neg_label} must not be negative"),
13175 token_pos: start,
13176 }),
13177 Some(Ok(v)) => u32::try_from(v)
13178 .map(crate::ast::LimitExpr::Literal)
13179 .map_err(|_| ParseError {
13180 message: alloc::format!("{label} value too large: {v}"),
13181 token_pos: start,
13182 }),
13183 Some(Err(message)) => Err(ParseError {
13184 message: message.replace("{L}", neg_label),
13185 token_pos: start,
13186 }),
13187 // v7.39 (round 305, V23) — not foldable at parse time
13188 // (`LIMIT (SELECT 4)`, `LIMIT greatest(2,3)`). Carry the
13189 // expression; the engine evaluates it once before dispatch.
13190 None => Ok(crate::ast::LimitExpr::Expr(alloc::boxed::Box::new(e))),
13191 }
13192 }
13193
13194 fn parse_limit_constant(&mut self, label: &str) -> Result<crate::ast::LimitExpr, ParseError> {
13195 // v7.39 (round 239) — PG's row-count clause takes a bigint with its
13196 // coercion rules, not just an integer token: a NUMERIC rounds half
13197 // away from zero (`LIMIT 2.5` keeps 3 rows), a negative count is
13198 // refused with PG's wording ("LIMIT must not be negative", 2201W /
13199 // 2201X — FETCH FIRST shares LIMIT's), and a string coerces by its
13200 // content, failing as an input-syntax error on the value. General
13201 // expressions (`LIMIT 1+1`) stay unsupported — a recorded residual;
13202 // they need an Expr-carrying LimitExpr variant.
13203 let neg_label = if label == "OFFSET" { "OFFSET" } else { "LIMIT" };
13204 let err_at = |message: alloc::string::String, pos: usize| ParseError {
13205 message,
13206 token_pos: pos,
13207 };
13208 match self.advance() {
13209 Token::Integer(n) if n >= 0 => u32::try_from(n)
13210 .map(crate::ast::LimitExpr::Literal)
13211 .map_err(|_| ParseError {
13212 message: alloc::format!("{label} value too large: {n}"),
13213 token_pos: self.consumed_pos(),
13214 }),
13215 Token::Integer(_) => Err(err_at(
13216 alloc::format!("{neg_label} must not be negative"),
13217 self.pos.saturating_sub(1),
13218 )),
13219 Token::Numeric(t) => {
13220 let pos = self.pos.saturating_sub(1);
13221 let v: f64 = t.parse().map_err(|_| {
13222 err_at(
13223 alloc::format!("invalid input syntax for type bigint: \"{t}\""),
13224 pos,
13225 )
13226 })?;
13227 if v < 0.0 {
13228 return Err(err_at(
13229 alloc::format!("{neg_label} must not be negative"),
13230 pos,
13231 ));
13232 }
13233 // Round half away from zero — PG's numeric→bigint cast.
13234 // (no_std: no f64::round; v is non-negative, so truncating
13235 // v + 0.5 is the same thing.)
13236 #[allow(clippy::cast_possible_truncation, clippy::cast_sign_loss)]
13237 let rounded = (v + 0.5) as u64;
13238 u32::try_from(rounded)
13239 .map(crate::ast::LimitExpr::Literal)
13240 .map_err(|_| err_at(alloc::format!("{label} value too large: {t}"), pos))
13241 }
13242 Token::Minus => {
13243 let pos = self.pos.saturating_sub(1);
13244 match self.peek() {
13245 Token::Integer(_) | Token::Numeric(_) => {
13246 self.advance();
13247 Err(err_at(
13248 alloc::format!("{neg_label} must not be negative"),
13249 pos,
13250 ))
13251 }
13252 other => Err(err_at(
13253 alloc::format!(
13254 "expected non-negative integer or $N placeholder after {label}, got {other:?}"
13255 ),
13256 pos,
13257 )),
13258 }
13259 }
13260 Token::String(t) => {
13261 let pos = self.pos.saturating_sub(1);
13262 match t.trim().parse::<i64>() {
13263 Ok(n) if n < 0 => Err(err_at(
13264 alloc::format!("{neg_label} must not be negative"),
13265 pos,
13266 )),
13267 Ok(n) => u32::try_from(n)
13268 .map(crate::ast::LimitExpr::Literal)
13269 .map_err(|_| err_at(alloc::format!("{label} value too large: {t}"), pos)),
13270 Err(_) => Err(err_at(
13271 alloc::format!("invalid input syntax for type bigint: \"{t}\""),
13272 pos,
13273 )),
13274 }
13275 }
13276 Token::Placeholder(n) => Ok(crate::ast::LimitExpr::Placeholder(n)),
13277 other => Err(ParseError {
13278 message: alloc::format!(
13279 "expected non-negative integer or $N placeholder after {label}, got {other:?}"
13280 ),
13281 token_pos: self.consumed_pos(),
13282 }),
13283 }
13284 }
13285
13286 /// Parse one SELECT block without ORDER BY / LIMIT / UNION chaining —
13287 /// just `[DISTINCT] items [FROM] [WHERE] [GROUP BY]`. Returned with
13288 /// `unions` empty and `order_by` / `limit` `None`; the top-level
13289 /// `parse_select_stmt` is responsible for filling those in.
13290 /// v7.37.17 (17.6 siblings) — rewrite every `grouping(keys…)`
13291 /// call in the expression tree to the per-set integer bitmask
13292 /// (PG semantics: one bit per argument, MSB first; 1 = the key
13293 /// is dropped in this grouping set). Runs during the ROLLUP /
13294 /// CUBE / GROUPING SETS expansion, where the set is known.
13295 /// v7.39 (round 135) — collect the distinct `grouping(...)` calls appearing
13296 /// anywhere in `expr` (an ORDER BY key), without recursing into their args.
13297 fn collect_grouping_calls(expr: &Expr, out: &mut Vec<Expr>) {
13298 if let Expr::FunctionCall { name, .. } = expr
13299 && name.eq_ignore_ascii_case("grouping")
13300 {
13301 if !out.iter().any(|e| e == expr) {
13302 out.push(expr.clone());
13303 }
13304 return;
13305 }
13306 match expr {
13307 Expr::Binary { lhs, rhs, .. } => {
13308 Self::collect_grouping_calls(lhs, out);
13309 Self::collect_grouping_calls(rhs, out);
13310 }
13311 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } | Expr::IsNull { expr, .. } => {
13312 Self::collect_grouping_calls(expr, out)
13313 }
13314 Expr::FunctionCall { args, .. } => {
13315 for a in args {
13316 Self::collect_grouping_calls(a, out);
13317 }
13318 }
13319 Expr::Case {
13320 operand,
13321 branches,
13322 else_branch,
13323 } => {
13324 if let Some(o) = operand {
13325 Self::collect_grouping_calls(o, out);
13326 }
13327 for (c, v) in branches {
13328 Self::collect_grouping_calls(c, out);
13329 Self::collect_grouping_calls(v, out);
13330 }
13331 if let Some(x) = else_branch {
13332 Self::collect_grouping_calls(x, out);
13333 }
13334 }
13335 _ => {}
13336 }
13337 }
13338
13339 /// v7.39 (round 135) — replace each `grouping(...)` call in `expr` equal to
13340 /// `grp_exprs[k]` with a reference to the synthetic ordering column
13341 /// `__grp_ord_k` (injected per grouping-set branch).
13342 fn rewrite_grouping_to_col(expr: &mut Expr, grp_exprs: &[Expr]) {
13343 if let Expr::FunctionCall { name, .. } = expr
13344 && name.eq_ignore_ascii_case("grouping")
13345 {
13346 if let Some(k) = grp_exprs.iter().position(|e| e == expr) {
13347 *expr = Expr::Column(crate::ast::ColumnName {
13348 qualifier: None,
13349 name: alloc::format!("__grp_ord_{k}"),
13350 });
13351 }
13352 return;
13353 }
13354 match expr {
13355 Expr::Binary { lhs, rhs, .. } => {
13356 Self::rewrite_grouping_to_col(lhs, grp_exprs);
13357 Self::rewrite_grouping_to_col(rhs, grp_exprs);
13358 }
13359 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } | Expr::IsNull { expr, .. } => {
13360 Self::rewrite_grouping_to_col(expr, grp_exprs)
13361 }
13362 Expr::FunctionCall { args, .. } => {
13363 for a in args {
13364 Self::rewrite_grouping_to_col(a, grp_exprs);
13365 }
13366 }
13367 Expr::Case {
13368 operand,
13369 branches,
13370 else_branch,
13371 } => {
13372 if let Some(o) = operand {
13373 Self::rewrite_grouping_to_col(o, grp_exprs);
13374 }
13375 for (c, v) in branches {
13376 Self::rewrite_grouping_to_col(c, grp_exprs);
13377 Self::rewrite_grouping_to_col(v, grp_exprs);
13378 }
13379 if let Some(x) = else_branch {
13380 Self::rewrite_grouping_to_col(x, grp_exprs);
13381 }
13382 }
13383 _ => {}
13384 }
13385 }
13386
13387 /// v7.39 (round 242) — one grouping element of PG's GROUP BY grammar,
13388 /// as the list of key sets it contributes. A bare expression is one
13389 /// single-key set; `ROLLUP (u1, …, un)` the n+1 unit-prefixes (largest
13390 /// first); `CUBE` every unit-subset (largest first); `GROUPING SETS`
13391 /// the concatenation of its items' sets, where an item is itself an
13392 /// element, a parenthesized key list, or the empty set `()`. A
13393 /// ROLLUP/CUBE member in parentheses is a composite UNIT: its keys
13394 /// move together.
13395 fn parse_grouping_element(&mut self) -> Result<Vec<Vec<Expr>>, ParseError> {
13396 let is_kw = |t: &Token, kw: &str| matches!(t, Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case(kw));
13397 // ROLLUP ( … ) / CUBE ( … )
13398 if (is_kw(self.peek(), "rollup") || is_kw(self.peek(), "cube"))
13399 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
13400 {
13401 let is_cube = is_kw(self.peek(), "cube");
13402 self.advance(); // ROLLUP / CUBE
13403 self.advance(); // (
13404 let mut units: Vec<Vec<Expr>> = Vec::new();
13405 loop {
13406 if matches!(self.peek(), Token::LParen) {
13407 // Composite unit: (a, b) rolls up as one.
13408 self.advance();
13409 let mut unit = Vec::new();
13410 if !matches!(self.peek(), Token::RParen) {
13411 loop {
13412 unit.push(self.parse_expr(0)?);
13413 match self.peek() {
13414 Token::Comma => {
13415 self.advance();
13416 }
13417 Token::RParen => break,
13418 other => {
13419 return Err(self.err(format!(
13420 "expected ',' or ')' in grouping unit, got {other:?}"
13421 )));
13422 }
13423 }
13424 }
13425 }
13426 self.advance(); // )
13427 units.push(unit);
13428 } else {
13429 units.push(alloc::vec![self.parse_expr(0)?]);
13430 }
13431 match self.peek() {
13432 Token::Comma => {
13433 self.advance();
13434 }
13435 Token::RParen => break,
13436 other => {
13437 return Err(self.err(format!(
13438 "expected ',' or ')' in grouping list, got {other:?}"
13439 )));
13440 }
13441 }
13442 }
13443 self.advance(); // )
13444 let flatten = |unit_sel: &[bool]| -> Vec<Expr> {
13445 units
13446 .iter()
13447 .zip(unit_sel.iter())
13448 .filter(|(_, keep)| **keep)
13449 .flat_map(|(u, _)| u.iter().cloned())
13450 .collect()
13451 };
13452 let n = units.len();
13453 if is_cube {
13454 let mut subsets: Vec<Vec<bool>> = (0..(1u32 << n))
13455 .map(|mask| (0..n).map(|i| mask & (1 << i) != 0).collect())
13456 .collect();
13457 subsets.sort_by_key(|sel| core::cmp::Reverse(sel.iter().filter(|b| **b).count()));
13458 return Ok(subsets.iter().map(|sel| flatten(sel)).collect());
13459 }
13460 return Ok((0..=n)
13461 .rev()
13462 .map(|keep| {
13463 let sel: Vec<bool> = (0..n).map(|i| i < keep).collect();
13464 flatten(&sel)
13465 })
13466 .collect());
13467 }
13468 // GROUPING SETS ( item [, item]* )
13469 if is_kw(self.peek(), "grouping")
13470 && matches!(self.tokens.get(self.pos + 1), Some(t) if is_kw(t, "sets"))
13471 {
13472 self.advance(); // GROUPING
13473 self.advance(); // SETS
13474 if !matches!(self.peek(), Token::LParen) {
13475 return Err(self.err(format!(
13476 "expected '(' after GROUPING SETS, got {:?}",
13477 self.peek()
13478 )));
13479 }
13480 self.advance(); // outer (
13481 let mut sets: Vec<Vec<Expr>> = Vec::new();
13482 loop {
13483 if matches!(self.peek(), Token::LParen) {
13484 // A parenthesized key list (or the empty set).
13485 self.advance();
13486 let mut set = Vec::new();
13487 if !matches!(self.peek(), Token::RParen) {
13488 loop {
13489 set.push(self.parse_expr(0)?);
13490 match self.peek() {
13491 Token::Comma => {
13492 self.advance();
13493 }
13494 Token::RParen => break,
13495 other => {
13496 return Err(self.err(format!(
13497 "expected ',' or ')' in grouping set, got {other:?}"
13498 )));
13499 }
13500 }
13501 }
13502 }
13503 self.advance(); // )
13504 sets.push(set);
13505 } else {
13506 // A nested element: ROLLUP/CUBE/GROUPING SETS or a
13507 // bare expression.
13508 sets.extend(self.parse_grouping_element()?);
13509 }
13510 match self.peek() {
13511 Token::Comma => {
13512 self.advance();
13513 }
13514 Token::RParen => break,
13515 other => {
13516 return Err(self.err(format!(
13517 "expected ',' or ')' after a grouping set, got {other:?}"
13518 )));
13519 }
13520 }
13521 }
13522 self.advance(); // outer )
13523 return Ok(sets);
13524 }
13525 Ok(alloc::vec![alloc::vec![self.parse_expr(0)?]])
13526 }
13527
13528 fn substitute_grouping_calls(expr: &mut Expr, dropped: &[Expr]) {
13529 // v7.38 (read01) — a reference to a key that is dropped in this grouping
13530 // set evaluates to NULL, at any depth. Previously only a *top-level*
13531 // select item equal to a dropped key was nullified, so a key nested in
13532 // an expression (`COALESCE(g,'TOTAL')`, `g || sum(v)`) survived as a raw
13533 // column and failed to resolve against the set's synthetic schema.
13534 if dropped.iter().any(|d| d == expr) {
13535 *expr = Expr::Literal(Literal::Null);
13536 return;
13537 }
13538 if let Expr::FunctionCall { name, args } = expr
13539 && name.eq_ignore_ascii_case("grouping")
13540 {
13541 let mut mask: i64 = 0;
13542 for a in args.iter() {
13543 mask <<= 1;
13544 if dropped.iter().any(|d| d == a) {
13545 mask |= 1;
13546 }
13547 }
13548 // v7.39 (round 242) — wrapped in a cast, NOT a bare integer
13549 // literal: a bare integer in a select item is indistinguishable
13550 // from a positional reference once `ORDER BY 1` substitutes the
13551 // item back in, and the round-232 position check then read the
13552 // mask value as an out-of-range position. The cast changes
13553 // nothing semantically (grouping() is integer).
13554 *expr = Expr::Cast {
13555 expr: alloc::boxed::Box::new(Expr::Literal(Literal::Integer(mask))),
13556 target: crate::ast::CastTarget::Int,
13557 };
13558 return;
13559 }
13560 // Generic recursion over the common expression shapes the
13561 // SELECT list uses; anything without child expressions is
13562 // left alone.
13563 match expr {
13564 Expr::FunctionCall { args, .. } => {
13565 for a in args {
13566 Self::substitute_grouping_calls(a, dropped);
13567 }
13568 }
13569 Expr::Binary { lhs, rhs, .. } => {
13570 Self::substitute_grouping_calls(lhs, dropped);
13571 Self::substitute_grouping_calls(rhs, dropped);
13572 }
13573 Expr::Unary { expr: inner, .. } => {
13574 Self::substitute_grouping_calls(inner, dropped);
13575 }
13576 Expr::Cast { expr: inner, .. } => {
13577 Self::substitute_grouping_calls(inner, dropped);
13578 }
13579 Expr::Case {
13580 operand,
13581 branches,
13582 else_branch,
13583 } => {
13584 if let Some(op) = operand {
13585 Self::substitute_grouping_calls(op, dropped);
13586 }
13587 for (w, t) in branches {
13588 Self::substitute_grouping_calls(w, dropped);
13589 Self::substitute_grouping_calls(t, dropped);
13590 }
13591 if let Some(e) = else_branch {
13592 Self::substitute_grouping_calls(e, dropped);
13593 }
13594 }
13595 // v7.38 (read01) — recurse into the remaining child-bearing shapes
13596 // so a dropped key nested in `IS NULL` / `LIKE` / `IN (…)` / `EXTRACT`
13597 // / a subscript / `ANY`/`ALL` is nullified too (`CASE WHEN g IS NULL
13598 // …` is the canonical rollup-total label idiom).
13599 Expr::IsNull { expr: inner, .. } => Self::substitute_grouping_calls(inner, dropped),
13600 Expr::Like { expr, pattern, .. } => {
13601 Self::substitute_grouping_calls(expr, dropped);
13602 Self::substitute_grouping_calls(pattern, dropped);
13603 }
13604 Expr::InList { expr, list, .. } => {
13605 Self::substitute_grouping_calls(expr, dropped);
13606 for item in list {
13607 Self::substitute_grouping_calls(item, dropped);
13608 }
13609 }
13610 Expr::Extract { source, .. } => Self::substitute_grouping_calls(source, dropped),
13611 Expr::Array(items) => {
13612 for item in items {
13613 Self::substitute_grouping_calls(item, dropped);
13614 }
13615 }
13616 Expr::ArraySubscript { target, index } => {
13617 Self::substitute_grouping_calls(target, dropped);
13618 Self::substitute_grouping_calls(index, dropped);
13619 }
13620 Expr::ArraySlice { target, lo, hi } => {
13621 Self::substitute_grouping_calls(target, dropped);
13622 if let Some(lo) = lo {
13623 Self::substitute_grouping_calls(lo, dropped);
13624 }
13625 if let Some(hi) = hi {
13626 Self::substitute_grouping_calls(hi, dropped);
13627 }
13628 }
13629 Expr::AnyAll { expr, array, .. } => {
13630 Self::substitute_grouping_calls(expr, dropped);
13631 Self::substitute_grouping_calls(array, dropped);
13632 }
13633 _ => {}
13634 }
13635 }
13636
13637 fn parse_bare_select(&mut self) -> Result<SelectStatement, ParseError> {
13638 // v7.37.17 (17.6 siblings) — parenthesized set-operation
13639 // group: `( <select chain> )` usable anywhere a query block
13640 // is (head or peer of an outer chain). The group's own
13641 // unions ride the returned SelectStatement; the executor's
13642 // nested-peer recursion runs them.
13643 if matches!(self.peek(), Token::LParen)
13644 && matches!(
13645 self.tokens.get(self.pos + 1),
13646 Some(Token::Select | Token::LParen | Token::Values)
13647 )
13648 {
13649 self.advance(); // (
13650 self.enter_nested()?;
13651 // v7.37 D.20 — a group whose head is a VALUES list:
13652 // `(VALUES (1),(2)) UNION (VALUES (3))`. Parse the VALUES body,
13653 // otherwise recurse into a nested SELECT/group head.
13654 let mut head = (if matches!(self.peek(), Token::Values) {
13655 self.advance(); // VALUES
13656 self.parse_values_rows_body()
13657 } else {
13658 self.parse_bare_select()
13659 })
13660 .and_then(|mut h| {
13661 self.parse_setop_chain_into(&mut h)?;
13662 Ok(h)
13663 });
13664 self.nest_depth -= 1;
13665 let mut head = match &mut head {
13666 Ok(h) => core::mem::take(h),
13667 Err(_) => return head,
13668 };
13669 // v7.37.17 (17.6 siblings) — group-internal tail:
13670 // `(A UNION B ORDER BY 1 LIMIT 5)`. Parse it into the
13671 // group head, then wrap the group as a derived table
13672 // (SELECT * FROM (group)) so the outer chain / outer
13673 // tail can't clobber the group's own ordering or limit.
13674 let has_tail = matches!(self.peek(), Token::Order | Token::Limit | Token::Offset)
13675 || matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
13676 if s.eq_ignore_ascii_case("fetch"));
13677 if has_tail {
13678 self.parse_select_tail_into(&mut head)?;
13679 head = SelectStatement {
13680 locking: None,
13681 ctes: Vec::new(),
13682 distinct: false,
13683 distinct_on: Vec::new(),
13684 items: alloc::vec![SelectItem::Wildcard],
13685 from: Some(FromClause {
13686 primary: TableRef {
13687 name: "subquery".to_string(),
13688 alias: None,
13689 only: false,
13690 as_of_segment: None,
13691 unnest_expr: None,
13692 unnest_column_aliases: Vec::new(),
13693 with_ordinality: false,
13694 generate_series_args: None,
13695 lateral_subquery: Some(Box::new(head)),
13696 jsonb_each_text_arg: None,
13697 table_fn_call: None,
13698 rows_from: None,
13699 json_table: None,
13700 scalar_fn_item: false,
13701 },
13702 joins: Vec::new(),
13703 }),
13704 where_: None,
13705 group_by: None,
13706 group_by_all: false,
13707 having: None,
13708 unions: Vec::new(),
13709 order_by: Vec::new(),
13710 limit: None,
13711 offset: None,
13712 limit_with_ties: false,
13713 window_check_exprs: Vec::new(),
13714 };
13715 }
13716 if !matches!(self.peek(), Token::RParen) {
13717 return Err(self.err(format!(
13718 "expected ')' after parenthesized query group, got {:?}",
13719 self.peek()
13720 )));
13721 }
13722 self.advance();
13723 return Ok(head);
13724 }
13725 // `TABLE name` shorthand as a query block — valid anywhere
13726 // a SELECT head is (set-op peers included).
13727 if matches!(self.peek(), Token::Table)
13728 && matches!(
13729 self.tokens.get(self.pos + 1),
13730 Some(Token::Ident(_) | Token::QuotedIdent(_))
13731 )
13732 {
13733 return self.parse_table_shorthand();
13734 }
13735 if !matches!(self.peek(), Token::Select) {
13736 return Err(self.err(format!(
13737 "expected SELECT to start a query block, got {:?}",
13738 self.peek()
13739 )));
13740 }
13741 self.advance();
13742 // v7.39.9 — MySQL's `SELECT STRAIGHT_JOIN …` join-order hint.
13743 //
13744 // It sits where `DISTINCT` sits and tells the optimiser to join
13745 // in the written order. SPG plans its own joins, so the hint is
13746 // accepted and not acted on — but it has to PARSE, because as a
13747 // bare identifier it became a column: measured on the published
13748 // image, `SELECT STRAIGHT_JOIN a FROM t` answered `Unknown
13749 // column 'straight_join' in 'field list'` where MySQL 9.7.2
13750 // returns the rows. Only in this position, which is the only one
13751 // MySQL accepts either — a trailing `STRAIGHT_JOIN` is its 1064.
13752 if self.mysql_dialect
13753 && matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("straight_join"))
13754 {
13755 self.advance();
13756 }
13757 let distinct = if matches!(self.peek(), Token::Distinct) {
13758 self.advance();
13759 true
13760 } else {
13761 false
13762 };
13763 // v7.37.17 (17.6 siblings) — `DISTINCT ON (expr [, …])`:
13764 // keep the first row (per ORDER BY) of each group the
13765 // expressions define. Django's .distinct('field') shape.
13766 let distinct_on: Vec<Expr> = if distinct && matches!(self.peek(), Token::On) {
13767 self.advance(); // ON
13768 if !matches!(self.peek(), Token::LParen) {
13769 return Err(self.err(format!(
13770 "expected '(' after DISTINCT ON, got {:?}",
13771 self.peek()
13772 )));
13773 }
13774 self.advance();
13775 let mut exprs = Vec::new();
13776 loop {
13777 exprs.push(self.parse_expr(0)?);
13778 match self.peek() {
13779 Token::Comma => {
13780 self.advance();
13781 }
13782 Token::RParen => break,
13783 other => {
13784 return Err(self.err(format!(
13785 "expected ',' or ')' in DISTINCT ON list, got {other:?}"
13786 )));
13787 }
13788 }
13789 }
13790 self.advance(); // )
13791 exprs
13792 } else {
13793 Vec::new()
13794 };
13795 let mut items = self.parse_select_list()?;
13796 // v7.38.19 — `SELECT … INTO <table>`, PostgreSQL's other spelling
13797 // of CTAS. It sits exactly here in PG's grammar, right after the
13798 // target list.
13799 //
13800 // A comment in `ast.rs` has said since v7.38 that CTAS and
13801 // `SELECT INTO` lower to the same node. Only CTAS ever did:
13802 // `SELECT i INTO t FROM src` answered `syntax error at or near
13803 // "INTO"`, which the differential found while measuring what
13804 // PostgreSQL tags each of the five materialising forms with. A
13805 // comment describing a capability the code does not have is the
13806 // defect this version has been finding all day, and this is the
13807 // one it found in the parser.
13808 //
13809 // `INTO` is captured rather than consumed here: the name has to
13810 // travel out of a function that returns a `SelectStatement`, and
13811 // the caller lowers the whole thing to the CTAS node.
13812 if matches!(self.peek(), Token::Into) {
13813 self.advance();
13814 // `TEMP` / `TEMPORARY` / `UNLOGGED` / `TABLE` are modifiers on
13815 // the target, not part of its name. SPG has one storage
13816 // class, so `UNLOGGED` is accepted and means nothing, which
13817 // is what it already means on `CREATE TABLE`.
13818 let mut temporary = false;
13819 loop {
13820 match self.peek().clone() {
13821 Token::Ident(w) | Token::QuotedIdent(w)
13822 if w.eq_ignore_ascii_case("temp")
13823 || w.eq_ignore_ascii_case("temporary") =>
13824 {
13825 temporary = true;
13826 self.advance();
13827 }
13828 Token::Ident(w) | Token::QuotedIdent(w)
13829 if w.eq_ignore_ascii_case("unlogged") =>
13830 {
13831 self.advance();
13832 }
13833 Token::Table => {
13834 self.advance();
13835 }
13836 _ => break,
13837 }
13838 }
13839 let name = match self.peek().clone() {
13840 Token::Ident(w) | Token::QuotedIdent(w) => {
13841 self.advance();
13842 w
13843 }
13844 other => {
13845 return Err(self.err(alloc::format!(
13846 "expected a table name after SELECT … INTO, got {other:?}"
13847 )));
13848 }
13849 };
13850 self.pending_select_into = Some((name, temporary));
13851 }
13852 // Scope the TABLESAMPLE lowering channel to this SELECT:
13853 // stash whatever an enclosing select accumulated, collect
13854 // our own FROM's predicates, restore after the combine.
13855 let enclosing_sample_preds = core::mem::take(&mut self.pending_sample_preds);
13856 let mut from = if matches!(self.peek(), Token::From) {
13857 self.advance();
13858 Some(self.parse_from_clause()?)
13859 } else {
13860 None
13861 };
13862 // v7.37 D.22 — a set-returning function in the projection with no FROM
13863 // (`SELECT unnest(arr)`, `SELECT 'x', generate_series(a,b)`) expands to
13864 // rows. Move the first SRF projection item to a FROM-position derived
13865 // table and replace it in the projection with a reference to its output
13866 // column; sibling scalar columns repeat per SRF row. PG names the output
13867 // column after the function (or its AS alias). Reuses the FROM-SRF
13868 // machinery. Only fires with no FROM — mixed SRF over a real FROM already
13869 // works via the targetlist-SRF path.
13870 // v7.39 (read01 round 69) — lower `(f(args)).*`. With no FROM it IS
13871 // `SELECT * FROM f(args)` — the record's fields become the columns, which
13872 // is exactly what the function's own row shape already is. Anywhere else
13873 // (per outer row, or beside other items) it would need a real record-typed
13874 // projection, so it says so rather than answering something else.
13875 if let [
13876 SelectItem::Expr {
13877 expr: Expr::FunctionCall { name, args },
13878 ..
13879 },
13880 ] = items.as_slice()
13881 && name == "__record_expand"
13882 {
13883 let Some(Expr::FunctionCall {
13884 name: inner_name,
13885 args: inner_args,
13886 }) = args.first()
13887 else {
13888 return Err(self.err(
13889 "(<expr>).* expands a function's record — it needs a function call".into(),
13890 ));
13891 };
13892 if from.is_some() {
13893 return Err(self.err(
13894 "(<fn>).* over a FROM clause is not supported — call the function in FROM"
13895 .into(),
13896 ));
13897 }
13898 let fn_ref = TableRef {
13899 name: inner_name.clone(),
13900 alias: None,
13901 only: false,
13902 as_of_segment: None,
13903 unnest_expr: None,
13904 unnest_column_aliases: Vec::new(),
13905 with_ordinality: false,
13906 generate_series_args: None,
13907 lateral_subquery: None,
13908 jsonb_each_text_arg: None,
13909 table_fn_call: Some(Box::new((
13910 inner_name.to_ascii_lowercase(),
13911 inner_args.clone(),
13912 ))),
13913 rows_from: None,
13914 json_table: None,
13915 scalar_fn_item: false,
13916 };
13917 items = alloc::vec![SelectItem::Wildcard];
13918 from = Some(FromClause {
13919 primary: fn_ref,
13920 joins: Vec::new(),
13921 });
13922 }
13923 // v7.39 (read01 round 74) — `(f(args)).*` beside other items, or over a
13924 // FROM, keeps its marker: the ENGINE lowers it, because naming the
13925 // record's fields takes the catalog. It becomes a LATERAL of the same
13926 // function plus one item per declared column — the machinery rounds 65
13927 // and 69 already built.
13928 // v7.39 (read01 round 67) — the lift moves ONE SRF into FROM. With two
13929 // (`SELECT generate_series(1,3), generate_series(10,11)`) PG runs them in
13930 // LOCKSTEP, padding the shorter with NULLs — a shape the lift cannot
13931 // express, since the lifted one becomes a scan and the other would
13932 // expand per its rows (a cross product, not a zip). So when the
13933 // projection holds more than one top-level function call, the lift steps
13934 // aside and the engine's target-list expansion takes the whole list.
13935 let fn_call_items = items
13936 .iter()
13937 .filter(|it| {
13938 matches!(
13939 it,
13940 SelectItem::Expr {
13941 expr: Expr::FunctionCall { .. },
13942 ..
13943 }
13944 )
13945 })
13946 .count();
13947 if from.is_none() && fn_call_items <= 1 {
13948 let mut found: Option<(usize, TableRef, String)> = None;
13949 for (i, item) in items.iter().enumerate() {
13950 if let SelectItem::Expr {
13951 expr: Expr::FunctionCall { name, args },
13952 alias,
13953 } = item
13954 {
13955 let lname = name.to_ascii_lowercase();
13956 let colname = alias.clone().unwrap_or_else(|| lname.clone());
13957 let (unnest, gs) = match lname.as_str() {
13958 "unnest" if args.len() == 1 => (Some(Box::new(args[0].clone())), None),
13959 "generate_series" if (2..=3).contains(&args.len()) => {
13960 (None, Some(args.clone()))
13961 }
13962 // v7.38 (read01) — generate_subscripts(arr, dim) in a
13963 // no-FROM projection yields the 1-based subscripts, i.e.
13964 // generate_series(1, array_length(arr, dim)); an invalid
13965 // dimension makes array_length NULL → 0 rows, as in PG.
13966 "generate_subscripts" if args.len() == 2 => (
13967 None,
13968 Some(alloc::vec![
13969 Expr::Literal(Literal::Integer(1)),
13970 Expr::FunctionCall {
13971 name: "array_length".to_string(),
13972 args: args.clone(),
13973 },
13974 ]),
13975 ),
13976 // v7.38 (read01, T-srf) — string_to_table / regexp_split_to_table
13977 // in a no-FROM projection unnest their *_to_array form.
13978 "string_to_table" | "regexp_split_to_table" => {
13979 let array_fn = if lname == "string_to_table" {
13980 "string_to_array"
13981 } else {
13982 "regexp_split_to_array"
13983 };
13984 (
13985 Some(Box::new(Expr::FunctionCall {
13986 name: array_fn.to_string(),
13987 args: args.clone(),
13988 })),
13989 None,
13990 )
13991 }
13992 // v7.38 (read01, T15) — jsonb/json_array_elements[_text] in
13993 // a no-FROM projection expand per element. The scalar form
13994 // returns the elements as a TEXT array, so unnest over the
13995 // same call materialises one row each (same rewrite the
13996 // FROM-clause form uses).
13997 "jsonb_array_elements"
13998 | "json_array_elements"
13999 | "jsonb_array_elements_text"
14000 | "json_array_elements_text"
14001 if args.len() == 1 =>
14002 {
14003 (
14004 Some(Box::new(Expr::FunctionCall {
14005 name: lname.clone(),
14006 args: args.clone(),
14007 })),
14008 None,
14009 )
14010 }
14011 // v7.38 (read01, T15) — jsonb/json_path_query(doc, path)
14012 // in a no-FROM projection expands per match (scalar form
14013 // returns the matches as a TEXT array → unnest).
14014 "jsonb_path_query" | "json_path_query" if args.len() == 2 => (
14015 Some(Box::new(Expr::FunctionCall {
14016 name: lname.clone(),
14017 args: args.clone(),
14018 })),
14019 None,
14020 ),
14021 _ => continue,
14022 };
14023 found = Some((
14024 i,
14025 TableRef {
14026 name: colname.clone(),
14027 alias: Some(colname.clone()),
14028 only: false,
14029 as_of_segment: None,
14030 unnest_expr: unnest,
14031 unnest_column_aliases: alloc::vec![colname.clone()],
14032 with_ordinality: false,
14033 generate_series_args: gs,
14034 lateral_subquery: None,
14035 jsonb_each_text_arg: None,
14036 table_fn_call: None,
14037 rows_from: None,
14038 json_table: None,
14039 scalar_fn_item: false,
14040 },
14041 colname,
14042 ));
14043 break;
14044 }
14045 }
14046 if let Some((idx, tref, colname)) = found {
14047 from = Some(FromClause {
14048 primary: tref,
14049 joins: Vec::new(),
14050 });
14051 items[idx] = SelectItem::Expr {
14052 expr: Expr::Column(ColumnName {
14053 qualifier: None,
14054 name: colname.clone(),
14055 }),
14056 alias: Some(colname),
14057 };
14058 }
14059 }
14060 let sample_preds = core::mem::take(&mut self.pending_sample_preds);
14061 let where_ = if matches!(self.peek(), Token::Where) {
14062 self.advance();
14063 Some(self.parse_expr(0)?)
14064 } else {
14065 None
14066 };
14067 let where_ = sample_preds.into_iter().fold(where_, |acc, pred| {
14068 Some(match acc {
14069 Some(w) => Expr::Binary {
14070 lhs: Box::new(pred),
14071 op: crate::ast::BinOp::And,
14072 rhs: Box::new(w),
14073 },
14074 None => pred,
14075 })
14076 });
14077 self.pending_sample_preds = enclosing_sample_preds;
14078 let mut group_by_all = false;
14079 // v7.37.17 (17.6 siblings) — ROLLUP / CUBE / GROUPING SETS
14080 // share one expansion: `grouping_sets` lists the key subsets
14081 // (first = primary, assigned to stmt.group_by; the rest
14082 // become UNION ALL peers), `grouping_universe` is the full
14083 // key list used to compute each peer's dropped keys.
14084 let mut grouping_sets: Vec<Vec<Expr>> = Vec::new();
14085 let mut grouping_universe: Vec<Expr> = Vec::new();
14086 // v7.39 (round 472) — did the GROUP BY end in MySQL's `WITH ROLLUP`?
14087 // A BOOL, not the key list: this frame is the statement parser's, and
14088 // round 430 measured that a `Vec` local here is enough on its own to
14089 // tip the 512 KiB nesting guard. The keys are recoverable from
14090 // `grouping_universe`, which a rollup fills with exactly them.
14091 let mut mysql_rollup = false;
14092 let group_by = if matches!(self.peek(), Token::Group) {
14093 self.advance();
14094 if !self.peek_is_by() {
14095 return Err(self.err(format!("expected BY after GROUP, got {:?}", self.peek())));
14096 }
14097 self.advance();
14098 // v6.4.1 — `GROUP BY ALL` shortcut. Planner expands to
14099 // every non-aggregate SELECT-list item later.
14100 if matches!(self.peek(), Token::All) {
14101 self.advance();
14102 group_by_all = true;
14103 None
14104 } else {
14105 // v7.39 (round 242) — PG's general grouping-element grammar:
14106 // GROUP BY [DISTINCT] element [, element]*, where an element
14107 // is a bare expression, ROLLUP (…), CUBE (…) or GROUPING
14108 // SETS (…) — mixed freely. Each element yields a list of
14109 // key sets; the query's grouping sets are the CARTESIAN
14110 // PRODUCT of the elements' lists (so `a, ROLLUP (b)` is
14111 // {(a,b),(a)}), and DISTINCT drops duplicate sets by
14112 // content. ROLLUP/CUBE members may be composite
14113 // (`ROLLUP ((a, b))` moves a and b as one unit), and a
14114 // GROUPING SETS item may itself be a ROLLUP/CUBE. The old
14115 // parser handled only a lone ROLLUP/CUBE/GS as the whole
14116 // clause.
14117 let distinct_sets = if matches!(self.peek(), Token::Distinct) {
14118 self.advance();
14119 true
14120 } else {
14121 false
14122 };
14123 let mut element_sets: Vec<Vec<Vec<Expr>>> = Vec::new();
14124 loop {
14125 element_sets.push(self.parse_grouping_element()?);
14126 if matches!(self.peek(), Token::Comma) {
14127 self.advance();
14128 } else {
14129 break;
14130 }
14131 }
14132 let mut total: Vec<Vec<Expr>> = alloc::vec![Vec::new()];
14133 for el in &element_sets {
14134 let mut next: Vec<Vec<Expr>> = Vec::new();
14135 for base in &total {
14136 for set in el {
14137 let mut merged = base.clone();
14138 for k in set {
14139 if !merged.iter().any(|m| m == k) {
14140 merged.push(k.clone());
14141 }
14142 }
14143 next.push(merged);
14144 }
14145 }
14146 total = next;
14147 }
14148 // v7.39 (round 472) — MySQL spells a rollup as a SUFFIX:
14149 // `GROUP BY a, b WITH ROLLUP` is PG's `GROUP BY ROLLUP(a, b)`.
14150 // The keys and the aggregates come out identical; the ROW
14151 // ORDER does not, and that is the part a report depends on.
14152 // MySQL interleaves each group's subtotal right after its
14153 // own rows (east/a, east/b, east/NULL, west/a, …, NULL/NULL)
14154 // where the union-of-grouping-sets expansion emits every
14155 // leaf first and then every subtotal. MariaDB REFUSES an
14156 // ORDER BY next to ROLLUP (1221), so a client cannot fix the
14157 // order itself — measured on MariaDB 11 and MySQL 9.7, which
14158 // agree on the order and disagree only on whether ORDER BY
14159 // is allowed (MySQL allows it; SPG allows it too, since
14160 // refusing would break the clients that can write it).
14161 if self.mysql_dialect
14162 && matches!(self.peek(), Token::Ident(w) if w.eq_ignore_ascii_case("with"))
14163 && matches!(
14164 self.tokens.get(self.pos + 1),
14165 Some(Token::Ident(r)) if r.eq_ignore_ascii_case("rollup")
14166 )
14167 {
14168 self.advance(); // WITH
14169 self.advance(); // ROLLUP
14170 let keys = total.into_iter().next().unwrap_or_default();
14171 mysql_rollup = true;
14172 // n+1 prefixes, largest first — the same expansion
14173 // `ROLLUP (…)` produces.
14174 total = (0..=keys.len()).rev().map(|n| keys[..n].to_vec()).collect();
14175 }
14176 if distinct_sets {
14177 let mut seen: Vec<Vec<String>> = Vec::new();
14178 total.retain(|set| {
14179 let mut key: Vec<String> =
14180 set.iter().map(|e| alloc::format!("{e}")).collect();
14181 key.sort();
14182 if seen.contains(&key) {
14183 false
14184 } else {
14185 seen.push(key);
14186 true
14187 }
14188 });
14189 }
14190 if total.len() > 1 {
14191 let mut universe: Vec<Expr> = Vec::new();
14192 for set in &total {
14193 for k in set {
14194 if !universe.iter().any(|u| u == k) {
14195 universe.push(k.clone());
14196 }
14197 }
14198 }
14199 grouping_universe = universe;
14200 let primary = total[0].clone();
14201 grouping_sets = total;
14202 Some(primary)
14203 } else {
14204 // One set (a plain GROUP BY list, or a single-set
14205 // spelling like GROUPING SETS ((a, b))). An EMPTY
14206 // single set — GROUPING SETS (()) — stays
14207 // `Some(vec![])`: the grand-total group, which must
14208 // run the aggregate path.
14209 Some(total.into_iter().next().unwrap_or_default())
14210 }
14211 }
14212 } else {
14213 None
14214 };
14215 let having = if matches!(self.peek(), Token::Having) {
14216 self.advance();
14217 Some(self.parse_expr(0)?)
14218 } else {
14219 None
14220 };
14221 // `WINDOW w AS ( <window-def> ) [, ...]` — named windows.
14222 // OVER w parsed to a marker above; inline each definition
14223 // into the referencing WindowFunction nodes.
14224 let mut window_defs: Vec<(String, WindowDef)> = Vec::new();
14225 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("window")) {
14226 self.advance();
14227 loop {
14228 let wname = self.expect_ident_like()?;
14229 if !matches!(self.peek(), Token::As) {
14230 return Err(self.err(format!(
14231 "expected AS after WINDOW {wname}, got {:?}",
14232 self.peek()
14233 )));
14234 }
14235 self.advance();
14236 // v7.39 (round 229) — PG rejects a redefinition outright.
14237 if window_defs
14238 .iter()
14239 .any(|(n, _)| n.eq_ignore_ascii_case(&wname))
14240 {
14241 return Err(self.err(alloc::format!("window \"{wname}\" is already defined")));
14242 }
14243 let def = self.parse_over_clause()?;
14244 // A definition may itself copy an earlier one
14245 // (`WINDOW w1 AS (PARTITION BY g), w2 AS (w1 ORDER BY v)`),
14246 // so resolve it against the defs already in scope. Same
14247 // copy rules as an `OVER (w1 …)` in the select list.
14248 let mut probe = Expr::WindowFunction {
14249 name: String::new(),
14250 args: Vec::new(),
14251 partition_by: def.0,
14252 order_by: def.1,
14253 frame: def.2,
14254 null_treatment: crate::ast::NullTreatment::Respect,
14255 filter: None,
14256 };
14257 Self::substitute_named_windows(&mut probe, &window_defs)
14258 .map_err(|m| self.err(m))?;
14259 let Expr::WindowFunction {
14260 partition_by,
14261 order_by,
14262 frame,
14263 ..
14264 } = probe
14265 else {
14266 unreachable!("probe is a WindowFunction")
14267 };
14268 window_defs.push((wname, (partition_by, order_by, frame)));
14269 if matches!(self.peek(), Token::Comma) {
14270 self.advance();
14271 continue;
14272 }
14273 break;
14274 }
14275 }
14276 // v7.39 (round 705) — which definitions did anything reference?
14277 // The ones nothing did used to be dropped here, unexamined, so
14278 // `WINDOW w AS (ORDER BY nosuch)` succeeded — PG analyses every
14279 // definition whether referenced or not. Their key expressions ride
14280 // out on the statement for the engine to resolve.
14281 let mut window_refs: Vec<String> = Vec::new();
14282 if !window_defs.is_empty() {
14283 for it in &items {
14284 if let SelectItem::Expr { expr, .. } = it {
14285 Self::collect_named_window_refs(expr, &mut window_refs);
14286 }
14287 }
14288 }
14289 let window_check_exprs: Vec<Expr> = window_defs
14290 .iter()
14291 .filter(|(n, _)| !window_refs.iter().any(|r| r.eq_ignore_ascii_case(n)))
14292 .flat_map(|(_, (partition, order, _))| {
14293 partition
14294 .iter()
14295 .cloned()
14296 .chain(order.iter().map(|(e, _, _)| e.clone()))
14297 })
14298 .collect();
14299 if !window_defs.is_empty()
14300 || items
14301 .iter()
14302 .any(|it| matches!(it, SelectItem::Expr { expr, .. } if Self::expr_has_named_window(expr)))
14303 {
14304 for it in &mut items {
14305 if let SelectItem::Expr { expr, .. } = it {
14306 Self::substitute_named_windows(expr, &window_defs)
14307 .map_err(|m| self.err(m))?;
14308 }
14309 }
14310 }
14311 // `GROUP BY 1` — positional keys substitute with the Nth
14312 // select item's expression (same contract ORDER BY has had
14313 // since v6.x). Out-of-range positions error.
14314 let group_by = match group_by {
14315 Some(mut keys) => {
14316 for k in &mut keys {
14317 if let Expr::Literal(Literal::Integer(n)) = k {
14318 let idx = *n;
14319 if idx < 1 || idx as usize > items.len() {
14320 return Err(self.err(alloc::format!(
14321 "GROUP BY position {idx} is not in select list"
14322 )));
14323 }
14324 match &items[(idx - 1) as usize] {
14325 SelectItem::Expr { expr, .. } => *k = expr.clone(),
14326 SelectItem::Wildcard | SelectItem::QualifiedWildcard(_) => {
14327 return Err(self.err(alloc::format!(
14328 "GROUP BY position {idx} references a wildcard item"
14329 )));
14330 }
14331 }
14332 }
14333 }
14334 Some(keys)
14335 }
14336 None => None,
14337 };
14338 let mut stmt = SelectStatement {
14339 locking: None,
14340 ctes: Vec::new(),
14341 distinct,
14342 distinct_on,
14343 items,
14344 from,
14345 where_,
14346 group_by,
14347 group_by_all,
14348 having,
14349 unions: Vec::new(),
14350 order_by: Vec::new(),
14351 limit: None,
14352 offset: None,
14353 limit_with_ties: false,
14354 window_check_exprs,
14355 };
14356 // Grouping expansion (ROLLUP / CUBE / GROUPING SETS): the
14357 // first set is the primary (already on stmt.group_by); each
14358 // further set becomes a UNION ALL peer with its dropped
14359 // keys (universe minus the set) replaced by NULL literals
14360 // in the peer's items and group_by. PG-legal: non-grouped
14361 // select items must be group keys or aggregates, so a
14362 // dropped key's occurrences in the projection are exactly
14363 // the ones to nullify.
14364 // v7.39 (round 242) — grouping() OUTSIDE an expansion: PG allows it
14365 // over a plain GROUP BY (every argument must be a group key; the
14366 // mask is then 0) and rejects anything else with 42803. SPG's
14367 // rewrite only ran during the ROLLUP/CUBE expansion, so a plain
14368 // `SELECT grouping(a) … GROUP BY a` died at eval with "unknown
14369 // function `grouping`".
14370 if grouping_sets.len() <= 1 {
14371 let keys: Vec<Expr> = stmt.group_by.clone().unwrap_or_default();
14372 let mut calls: Vec<Expr> = Vec::new();
14373 for item in &stmt.items {
14374 if let SelectItem::Expr { expr, .. } = item {
14375 Self::collect_grouping_calls(expr, &mut calls);
14376 }
14377 }
14378 if let Some(h) = &stmt.having {
14379 Self::collect_grouping_calls(h, &mut calls);
14380 }
14381 for call in &calls {
14382 let Expr::FunctionCall { args, .. } = call else {
14383 continue;
14384 };
14385 for a in args {
14386 if !keys.iter().any(|k| k == a) {
14387 return Err(self.err(
14388 "arguments to GROUPING must be grouping expressions of the associated query level"
14389 .to_string(),
14390 ));
14391 }
14392 }
14393 }
14394 if !calls.is_empty() {
14395 for item in &mut stmt.items {
14396 if let SelectItem::Expr { expr, .. } = item {
14397 Self::substitute_grouping_calls(expr, &[]);
14398 }
14399 }
14400 if let Some(h) = &mut stmt.having {
14401 Self::substitute_grouping_calls(h, &[]);
14402 }
14403 }
14404 }
14405 if grouping_sets.len() > 1 {
14406 // The primary set's own dropped keys nullify in the
14407 // HEAD's projection too (GROUPING SETS's first set may
14408 // omit keys other sets use).
14409 let primary = grouping_sets[0].clone();
14410 let head_dropped: Vec<Expr> = grouping_universe
14411 .iter()
14412 .filter(|u| !primary.iter().any(|k| k == *u))
14413 .cloned()
14414 .collect();
14415 for set in grouping_sets.iter().skip(1) {
14416 let mut peer = stmt.clone();
14417 peer.unions = Vec::new();
14418 let dropped: Vec<&Expr> = grouping_universe
14419 .iter()
14420 .filter(|u| !set.iter().any(|k| k == *u))
14421 .collect();
14422 // Empty set = grand-total group: `Some(vec![])` forces
14423 // the aggregate path (one collapsed row) instead of a
14424 // per-row passthrough. See the primary-set note above.
14425 peer.group_by = Some(set.clone());
14426 let dropped_owned: Vec<Expr> = dropped.iter().map(|d| (*d).clone()).collect();
14427 for item in &mut peer.items {
14428 if let SelectItem::Expr { expr, alias } = item {
14429 if dropped.iter().any(|d| *d == expr) {
14430 // v7.39 — keep the dropped key's name on the
14431 // NULL literal so the UNION output column
14432 // (and any top-level ORDER BY on it) still
14433 // resolves.
14434 if alias.is_none()
14435 && let Expr::Column(c) = &expr
14436 {
14437 *alias = Some(c.name.clone());
14438 }
14439 *expr = Expr::Literal(Literal::Null);
14440 } else {
14441 Self::substitute_grouping_calls(expr, &dropped_owned);
14442 }
14443 }
14444 }
14445 if let Some(h) = &mut peer.having {
14446 Self::substitute_grouping_calls(h, &dropped_owned);
14447 }
14448 stmt.unions.push((UnionKind::All, peer));
14449 }
14450 for item in &mut stmt.items {
14451 if let SelectItem::Expr { expr, alias } = item {
14452 if head_dropped.iter().any(|d| d == expr) {
14453 if alias.is_none()
14454 && let Expr::Column(c) = &expr
14455 {
14456 *alias = Some(c.name.clone());
14457 }
14458 *expr = Expr::Literal(Literal::Null);
14459 } else {
14460 Self::substitute_grouping_calls(expr, &head_dropped);
14461 }
14462 }
14463 }
14464 if let Some(h) = &mut stmt.having {
14465 Self::substitute_grouping_calls(h, &head_dropped);
14466 }
14467 // v7.39 (round 135) — GROUPING() in ORDER BY. Parse the ORDER BY now
14468 // (while `grouping_universe` / the per-branch sets are in scope). For
14469 // each grouping() call in it, inject a per-branch hidden column
14470 // `__grp_ord_K` carrying that branch's mask into the head + every
14471 // peer, and rewrite the ORDER BY to reference it. `parse_select_tail_into`
14472 // preserves this pre-set order_by; the engine strips `__grp_ord_*`
14473 // from the final output. A standalone grouping-set query has ORDER BY
14474 // (not an explicit set-op) next, so consuming it here is safe.
14475 // v7.39 (round 472) — absent the client's own ORDER BY, a MySQL
14476 // rollup carries the hierarchical order: sort by the grouping
14477 // keys with the rolled-up NULLs last, which is exactly the
14478 // interleaving both oracles emit. A client's own ORDER BY wins,
14479 // which is what MySQL does (MariaDB refuses to let one be
14480 // written at all).
14481 // The synthesised keys have to travel the SAME path a written
14482 // ORDER BY does: the block below is what turns a `grouping()`
14483 // call into the per-branch `__grp_ord_K` column the engine can
14484 // actually sort on. Bypassing it left a bare `grouping(text)`
14485 // for the evaluator to reject.
14486 let synthesised_or_parsed: Vec<OrderBy> = if matches!(self.peek(), Token::Order) {
14487 self.parse_order_by_keys()?
14488 } else if mysql_rollup {
14489 Self::mysql_rollup_order(&grouping_universe)
14490 } else {
14491 Vec::new()
14492 };
14493 if !synthesised_or_parsed.is_empty() {
14494 let mut order_keys = synthesised_or_parsed;
14495 let mut grp_exprs: Vec<Expr> = Vec::new();
14496 for ob in &order_keys {
14497 Self::collect_grouping_calls(&ob.expr, &mut grp_exprs);
14498 }
14499 for (k, gexpr) in grp_exprs.iter().enumerate() {
14500 let colname = alloc::format!("__grp_ord_{k}");
14501 // Head branch (primary set) uses `head_dropped`.
14502 let mut he = gexpr.clone();
14503 Self::substitute_grouping_calls(&mut he, &head_dropped);
14504 stmt.items.push(SelectItem::Expr {
14505 expr: he,
14506 alias: Some(colname.clone()),
14507 });
14508 // Each peer `stmt.unions[i]` corresponds to `grouping_sets[i+1]`.
14509 for (i, (_, peer)) in stmt.unions.iter_mut().enumerate() {
14510 let set = &grouping_sets[i + 1];
14511 let dropped: Vec<Expr> = grouping_universe
14512 .iter()
14513 .filter(|u| !set.iter().any(|k| k == *u))
14514 .cloned()
14515 .collect();
14516 let mut pe = gexpr.clone();
14517 Self::substitute_grouping_calls(&mut pe, &dropped);
14518 peer.items.push(SelectItem::Expr {
14519 expr: pe,
14520 alias: Some(colname.clone()),
14521 });
14522 }
14523 }
14524 for ob in &mut order_keys {
14525 Self::rewrite_grouping_to_col(&mut ob.expr, &grp_exprs);
14526 }
14527 stmt.order_by = order_keys;
14528 }
14529 }
14530 Ok(stmt)
14531 }
14532
14533 /// v7.39 (round 472) — the row order MySQL's `WITH ROLLUP` promises,
14534 /// as ORDER BY keys.
14535 ///
14536 /// Per key: the rollup marker, then the key. Sorting on the key alone
14537 /// is not enough, and a table with a NULL in it says why — MariaDB puts
14538 /// the DATA-NULL group where a plain GROUP BY puts it (first) and only
14539 /// the ROLLUP-introduced NULL last, and both print as NULL.
14540 /// `GROUPING(k)` is the one thing that tells them apart: 0 for every
14541 /// real group including the data-NULL one, 1 only for the row the
14542 /// rollup added. Measured on MariaDB 11 — `('a',1),(NULL,2),('b',3)`
14543 /// rolls up to NULL|2, a|1, b|3, NULL|6.
14544 ///
14545 /// `#[inline(never)]`: its locals must not join the statement parser's
14546 /// frame, which round 430 measured sitting against the nesting guard.
14547 #[inline(never)]
14548 fn mysql_rollup_order(keys: &[Expr]) -> Vec<OrderBy> {
14549 let mut out: Vec<OrderBy> = Vec::with_capacity(keys.len() * 2);
14550 for e in keys {
14551 out.push(OrderBy {
14552 expr: Expr::FunctionCall {
14553 name: "grouping".into(),
14554 args: alloc::vec![e.clone()],
14555 },
14556 desc: false,
14557 nulls_first: None,
14558 collation: None,
14559 });
14560 out.push(OrderBy {
14561 expr: e.clone(),
14562 desc: false,
14563 // MySQL orders NULL first on an ascending key.
14564 nulls_first: Some(true),
14565 collation: None,
14566 });
14567 }
14568 out
14569 }
14570
14571 /// v7.39 (round 535) — `REINDEX [(opts)] { INDEX | TABLE | SCHEMA |
14572 /// DATABASE | SYSTEM } [CONCURRENTLY] [<name>]`.
14573 #[inline(never)]
14574 fn parse_reindex_tail(&mut self) -> Result<Statement, ParseError> {
14575 use crate::ast::MaintainKind;
14576 self.skip_paren_option_list();
14577 let kind = match self.peek() {
14578 // `TABLE` and `INDEX` lex as keywords, not identifiers.
14579 Token::Table | Token::Index => {
14580 self.advance();
14581 MaintainKind::ReindexRelation
14582 }
14583 Token::Ident(s) | Token::QuotedIdent(s) => match s.to_ascii_lowercase().as_str() {
14584 "index" | "table" => {
14585 self.advance();
14586 MaintainKind::ReindexRelation
14587 }
14588 "schema" => {
14589 self.advance();
14590 MaintainKind::ReindexSchema
14591 }
14592 "system" | "database" => {
14593 self.advance();
14594 MaintainKind::Whole
14595 }
14596 // PG requires the object type; anything else is the
14597 // caller's problem, not something to swallow.
14598 _ => MaintainKind::ReindexRelation,
14599 },
14600 _ => MaintainKind::Whole,
14601 };
14602 // PG bars `REINDEX … CONCURRENTLY` inside a transaction block and
14603 // allows the plain form, so the modifier is recorded rather than
14604 // skipped. It still has no effect on how the reindex runs.
14605 let mut concurrently = false;
14606 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("concurrently")) {
14607 self.advance();
14608 concurrently = true;
14609 }
14610 let target = self.take_optional_maintain_name();
14611 self.consume_until_statement_boundary();
14612 Ok(Statement::Maintain {
14613 kind,
14614 concurrently,
14615 target,
14616 })
14617 }
14618
14619 /// v7.39 (round 535) — `CLUSTER [VERBOSE] [<table> [USING <index>]]`
14620 /// and `CLUSTER [VERBOSE] <index> ON <table>`.
14621 #[inline(never)]
14622 fn parse_cluster_tail(&mut self) -> Result<Statement, ParseError> {
14623 use crate::ast::MaintainKind;
14624 self.skip_paren_option_list();
14625 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("verbose")) {
14626 self.advance();
14627 }
14628 let target = self.take_optional_maintain_name();
14629 self.consume_until_statement_boundary();
14630 Ok(Statement::Maintain {
14631 kind: if target.is_some() {
14632 MaintainKind::ClusterRelation
14633 } else {
14634 MaintainKind::Whole
14635 },
14636 // CLUSTER has no CONCURRENTLY form, and PG runs it inside a
14637 // transaction block quite happily (measured).
14638 concurrently: false,
14639 target,
14640 })
14641 }
14642
14643 /// The next token as a relation / schema name, when there is one.
14644 fn take_optional_maintain_name(&mut self) -> Option<alloc::string::String> {
14645 match self.peek() {
14646 Token::Ident(_) | Token::QuotedIdent(_) => match self.advance() {
14647 Token::Ident(n) | Token::QuotedIdent(n) => Some(n),
14648 _ => None,
14649 },
14650 _ => None,
14651 }
14652 }
14653
14654 /// A parenthesised option list, absorbed.
14655 fn skip_paren_option_list(&mut self) {
14656 if !matches!(self.peek(), Token::LParen) {
14657 return;
14658 }
14659 let mut depth = 0usize;
14660 loop {
14661 match self.advance() {
14662 Token::LParen => depth += 1,
14663 Token::RParen => {
14664 depth -= 1;
14665 if depth == 0 {
14666 return;
14667 }
14668 }
14669 Token::Eof => return,
14670 _ => {}
14671 }
14672 }
14673 }
14674
14675 /// v7.39 (round 531) — the `LIKE` clause inside a CREATE TABLE
14676 /// column list.
14677 ///
14678 /// PG's option names are COMMENTS / COMPRESSION / CONSTRAINTS /
14679 /// DEFAULTS / GENERATED / IDENTITY / INDEXES / STATISTICS / STORAGE
14680 /// / ALL. The three that describe physical storage have no meaning
14681 /// here, so they parse and change nothing rather than making a
14682 /// dump that mentions them fail to load.
14683 ///
14684 /// `#[inline(never)]` because the CREATE TABLE frame sits on the
14685 /// parse chain the nesting sentinel is tuned against.
14686 #[inline(never)]
14687 fn parse_create_table_like(&mut self, at: usize) -> Result<crate::ast::LikeSpec, ParseError> {
14688 self.advance(); // LIKE
14689 let source = self.expect_ident_like()?;
14690 let mut options = crate::ast::LikeOptions::default();
14691 loop {
14692 let including = match self.peek() {
14693 Token::Ident(s) if s.eq_ignore_ascii_case("including") => true,
14694 Token::Ident(s) if s.eq_ignore_ascii_case("excluding") => false,
14695 _ => break,
14696 };
14697 self.advance();
14698 // `ALL` lexes as its own keyword, not an identifier.
14699 let opt = if matches!(self.peek(), Token::All) {
14700 self.advance();
14701 alloc::string::String::from("all")
14702 } else {
14703 self.expect_ident_like()?
14704 };
14705 let set = |o: &mut crate::ast::LikeOptions, on: bool| {
14706 o.defaults = on;
14707 o.constraints = on;
14708 o.identity = on;
14709 o.generated = on;
14710 o.indexes = on;
14711 o.comments = on;
14712 };
14713 match opt.to_ascii_lowercase().as_str() {
14714 "all" => set(&mut options, including),
14715 "defaults" => options.defaults = including,
14716 "constraints" => options.constraints = including,
14717 "identity" => options.identity = including,
14718 "generated" => options.generated = including,
14719 "indexes" => options.indexes = including,
14720 "comments" => options.comments = including,
14721 // No storage model to copy into.
14722 "storage" | "statistics" | "compression" => {}
14723 other => {
14724 return Err(self.err(alloc::format!("unrecognized LIKE option {other:?}")));
14725 }
14726 }
14727 }
14728 Ok(crate::ast::LikeSpec {
14729 source,
14730 at,
14731 options,
14732 })
14733 }
14734
14735 fn parse_create_table_stmt_after_create(&mut self) -> Result<Statement, ParseError> {
14736 // Caller already consumed CREATE; we're sitting on TABLE.
14737 debug_assert!(matches!(self.peek(), Token::Table));
14738 self.advance();
14739 let if_not_exists = self.consume_if_not_exists();
14740 let name = self.expect_ident_like()?;
14741 // v7.37.6-B — `CREATE TABLE c PARTITION OF parent <bounds>`
14742 // child shape has no column list; the child inherits its
14743 // columns from the parent at engine-DDL time. Detect it
14744 // before the `(` requirement below.
14745 if matches!(self.peek(), Token::Partition)
14746 && Self::tokens_match_ident_ci(self.tokens.get(self.pos + 1), "of")
14747 {
14748 self.advance(); // PARTITION
14749 self.advance(); // of
14750 let partition_of = self.parse_partition_of_tail()?;
14751 return Ok(Statement::CreateTable(CreateTableStatement {
14752 temporary: false,
14753 name,
14754 engine: None,
14755 columns: Vec::new(),
14756 like_specs: Vec::new(),
14757 inherits: Vec::new(),
14758 if_not_exists,
14759 foreign_keys: Vec::new(),
14760 table_constraints: Vec::new(),
14761 partition_by: None,
14762 partition_of: Some(partition_of),
14763 }));
14764 }
14765 // v7.38 (read01 P6.49) — CTAS: `CREATE TABLE name AS <select>`. Reuses
14766 // the materialized-view materialisation path (run the SELECT, infer the
14767 // column types, create + populate the table) but marks the node so the
14768 // executor creates a plain table without a mat-view registry entry.
14769 if matches!(self.peek(), Token::As) {
14770 self.advance();
14771 let body_stmt = self.parse_select_stmt()?;
14772 let Statement::Select(body) = body_stmt else {
14773 return Err(self.err(format!(
14774 "CREATE TABLE {name:?} AS body must be a SELECT, got {body_stmt:?}"
14775 )));
14776 };
14777 let with_data = self.parse_optional_with_data(true)?;
14778 return Ok(Statement::CreateMaterializedView(
14779 crate::ast::CreateMaterializedViewStatement {
14780 temporary: false,
14781 name,
14782 if_not_exists,
14783 columns: Vec::new(),
14784 body,
14785 with_data,
14786 as_plain_table: true,
14787 },
14788 ));
14789 }
14790 if !matches!(self.peek(), Token::LParen) {
14791 return Err(self.err(format!(
14792 "expected '(' after table name, got {:?}",
14793 self.peek()
14794 )));
14795 }
14796 self.advance();
14797 let mut columns = Vec::new();
14798 let mut foreign_keys: Vec<ForeignKeyConstraint> = Vec::new();
14799 let mut table_constraints: Vec<crate::ast::TableConstraint> = Vec::new();
14800 let mut like_specs: Vec<crate::ast::LikeSpec> = Vec::new();
14801 loop {
14802 // v7.39 (round 621) — `CREATE TABLE c () INHERITS (p)`, the empty
14803 // column list. It is how a child that adds nothing of its own is
14804 // written, and this loop demanded at least one entry: `syntax
14805 // error at or near ")"`. The child takes the parent's columns,
14806 // which the INHERITS clause already arranges.
14807 if columns.is_empty() && matches!(self.peek(), Token::RParen) {
14808 self.advance();
14809 break;
14810 }
14811 // v7.6.0 / v7.9.18 — distinguish table-level constraint
14812 // clauses from column definitions. Constraints start
14813 // with `CONSTRAINT <name> …`, `FOREIGN KEY (…)`,
14814 // `PRIMARY KEY (…)`, or `UNIQUE (…)`. Anything else is
14815 // a column.
14816 if self.peek_table_level_pk_start() {
14817 table_constraints.push(self.parse_table_level_primary_key()?);
14818 } else if matches!(self.peek(), Token::Like) {
14819 // v7.39 (round 531) — `LIKE <table> [ {INCLUDING|EXCLUDING}
14820 // <opt> ]*`. The source table's shape lives in the catalog,
14821 // so this records the clause and the engine expands it.
14822 like_specs.push(self.parse_create_table_like(columns.len())?);
14823 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("exclude")) {
14824 // v7.39 (round 210) — bare `EXCLUDE [USING m] (col WITH op, …)`.
14825 table_constraints.push(self.parse_table_level_exclude()?);
14826 } else if self.peek_table_level_unique_start() {
14827 table_constraints.push(self.parse_table_level_unique()?);
14828 } else if self.peek_table_level_check_start() {
14829 // v7.13.0 — table-level CHECK (mailrs round-5 G3).
14830 table_constraints.push(self.parse_table_level_check()?);
14831 } else if self.peek_mysql_inline_key_start() {
14832 // v7.14.0 — mysqldump emits inline `KEY name (cols)`,
14833 // `INDEX name (cols)`, `UNIQUE KEY name (cols)`,
14834 // `FULLTEXT KEY name (cols)`, `SPATIAL KEY name (cols)`
14835 // inside the column list. Skip name + paren list;
14836 // for UNIQUE KEY, register as a UC.
14837 if let Some(uc) = self.parse_mysql_inline_key()? {
14838 table_constraints.push(uc);
14839 }
14840 } else if let Some(kind) = self.peek_named_table_constraint_kind() {
14841 // v7.22 (mailrs round-13 gap 5) — `CONSTRAINT <name>
14842 // { CHECK | UNIQUE | PRIMARY KEY }`: every pg_dump'd
14843 // CHECK is named, and the named-CONSTRAINT arm used
14844 // to accept FOREIGN KEY only. The name is accepted
14845 // and discarded — same handling as every other SPG
14846 // constraint name.
14847 self.advance(); // CONSTRAINT
14848 // v7.39 (read01 round 48) — the name is kept now: the schema
14849 // stores it, so DROP / RENAME CONSTRAINT can find it.
14850 let con_name = self.expect_ident_like()?;
14851 let mut tc = match kind {
14852 NamedTableConstraintKind::Check => self.parse_table_level_check()?,
14853 NamedTableConstraintKind::Unique => self.parse_table_level_unique()?,
14854 NamedTableConstraintKind::PrimaryKey => self.parse_table_level_primary_key()?,
14855 NamedTableConstraintKind::Exclude => self.parse_table_level_exclude()?,
14856 };
14857 match &mut tc {
14858 crate::ast::TableConstraint::Check { name, .. }
14859 | crate::ast::TableConstraint::Unique { name, .. }
14860 | crate::ast::TableConstraint::PrimaryKey { name, .. }
14861 | crate::ast::TableConstraint::Exclude { name, .. } => {
14862 *name = Some(con_name);
14863 }
14864 _ => {}
14865 }
14866 table_constraints.push(tc);
14867 } else if self.peek_constraint_or_fk_start() {
14868 foreign_keys.push(self.parse_table_level_fk()?);
14869 } else {
14870 let (col, col_level_fk) = self.parse_column_def_with_fk()?;
14871 // v7.13.0 — fold inline UNIQUE / CHECK column
14872 // constraints into table-level entries so the
14873 // engine path stays uniform.
14874 if col.is_unique {
14875 table_constraints.push(crate::ast::TableConstraint::Unique {
14876 name: None,
14877 columns: alloc::vec![col.name.clone()],
14878 nulls_not_distinct: col.unique_nulls_not_distinct,
14879 deferrable: col.constraint_deferrable,
14880 initially_deferred: col.constraint_initially_deferred,
14881 });
14882 }
14883 if let Some(check_expr) = col.check.clone() {
14884 table_constraints.push(crate::ast::TableConstraint::Check {
14885 name: None,
14886 expr: check_expr,
14887 not_valid: false,
14888 });
14889 }
14890 columns.push(col);
14891 if let Some(fk) = col_level_fk {
14892 foreign_keys.push(fk);
14893 }
14894 }
14895 match self.peek() {
14896 Token::Comma => {
14897 self.advance();
14898 }
14899 Token::RParen => {
14900 self.advance();
14901 break;
14902 }
14903 other => {
14904 return Err(
14905 self.err(format!("expected ',' or ')' in column list, got {other:?}"))
14906 );
14907 }
14908 }
14909 }
14910 // v7.39 (round 531) — a `LIKE` clause brings its own columns, so
14911 // `CREATE TABLE k (LIKE t)` is a complete definition even though
14912 // nothing is written between the parentheses.
14913 // v7.39 (round 621) — a table with NO columns is legal: PG creates it
14914 // and `INSERT … DEFAULT VALUES` puts a row in it. This refused, so the
14915 // empty parentheses were a parse error in their own right — quite apart
14916 // from `CREATE TABLE c () INHERITS (p)`, which needs table inheritance
14917 // SPG does not have (filed separately).
14918 let _ = &like_specs;
14919 // v7.39 (round 645) — `INHERITS (p1, p2)`, PG table inheritance.
14920 // It sits between the column list and the MySQL table options,
14921 // and it was a syntax error until this round.
14922 let mut inherits: Vec<String> = Vec::new();
14923 if matches!(self.peek(), Token::Ident(k) | Token::QuotedIdent(k)
14924 if k.eq_ignore_ascii_case("inherits"))
14925 {
14926 self.advance();
14927 if !matches!(self.peek(), Token::LParen) {
14928 return Err(self.err(alloc::format!(
14929 "expected ( after INHERITS, got {:?}",
14930 self.peek()
14931 )));
14932 }
14933 self.advance();
14934 loop {
14935 inherits.push(self.expect_ident_like()?);
14936 if matches!(self.peek(), Token::Comma) {
14937 self.advance();
14938 continue;
14939 }
14940 break;
14941 }
14942 if !matches!(self.peek(), Token::RParen) {
14943 return Err(self.err(alloc::format!(
14944 "expected ) closing INHERITS, got {:?}",
14945 self.peek()
14946 )));
14947 }
14948 self.advance();
14949 }
14950 // v7.14.0 — consume MySQL/MariaDB table options after the
14951 // closing `)`. mysqldump emits things like
14952 // `ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_unicode_ci
14953 // AUTO_INCREMENT=42 ROW_FORMAT=DYNAMIC COMMENT='blog posts'`.
14954 // SPG accepts all forms as no-ops (each option is
14955 // `<ident> [=] <ident-or-string>` separated by whitespace).
14956 let engine = self.consume_mysql_table_options();
14957 // v7.38 (read01 P6.55) — PG storage parameters `WITH (opt=val, …)`.
14958 // SPG has no per-table reloptions, so accept and ignore them so a
14959 // pg_dump `CREATE TABLE … WITH (fillfactor=70, …)` restores cleanly.
14960 self.consume_with_reloptions();
14961 // v7.37.6-B — declarative-partition-parent suffix
14962 // (`PARTITION BY RANGE (key_col)`) sits after the column
14963 // list + MySQL table-options. v7.37.6-B only accepts RANGE
14964 // and locks the key column at one ident; the engine then
14965 // verifies the column type is TIMESTAMPTZ.
14966 let partition_by = if matches!(self.peek(), Token::Partition) {
14967 self.advance(); // PARTITION
14968 if !self.peek_is_by() {
14969 return Err(self.err(format!(
14970 "expected BY after PARTITION, got {:?}",
14971 self.peek()
14972 )));
14973 }
14974 self.advance();
14975 Some(self.parse_partition_by_tail()?)
14976 } else {
14977 None
14978 };
14979 Ok(Statement::CreateTable(CreateTableStatement {
14980 temporary: false,
14981 name,
14982 engine,
14983 columns,
14984 like_specs,
14985 inherits,
14986 if_not_exists,
14987 foreign_keys,
14988 table_constraints,
14989 partition_by,
14990 partition_of: None,
14991 }))
14992 }
14993
14994 /// v7.37.6-B — case-insensitive ident match helper for the
14995 /// `PARTITION OF` / `MINVALUE` / `MAXVALUE` keywords. They lex
14996 /// as `Token::Ident("of"/"minvalue"/"maxvalue")` because we
14997 /// didn't burn a global keyword slot for each (see the
14998 /// `Token::Partition` doc-comment in `lexer.rs`).
14999 fn tokens_match_ident_ci(t: Option<&Token>, want: &str) -> bool {
15000 matches!(t, Some(Token::Ident(s) | Token::QuotedIdent(s)) if s.eq_ignore_ascii_case(want))
15001 }
15002
15003 /// v7.37.6-B — after `PARTITION BY`, expect `RANGE (key_col [, ...])`.
15004 /// v7.37.16 (16.1/16.2) — extended to LIST + HASH.
15005 fn parse_partition_by_tail(&mut self) -> Result<crate::ast::PartitionBySpec, ParseError> {
15006 use crate::ast::{PartitionBySpec, PartitionKindAst};
15007 let kind = match self.peek() {
15008 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("range") => {
15009 self.advance();
15010 PartitionKindAst::Range
15011 }
15012 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("list") => {
15013 self.advance();
15014 PartitionKindAst::List
15015 }
15016 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("hash") => {
15017 self.advance();
15018 PartitionKindAst::Hash
15019 }
15020 other => {
15021 return Err(self.err(format!(
15022 "PARTITION BY: expected RANGE / LIST / HASH, got {other:?}"
15023 )));
15024 }
15025 };
15026 if !matches!(self.peek(), Token::LParen) {
15027 return Err(self.err(format!(
15028 "expected '(' after PARTITION BY <strategy>, got {:?}",
15029 self.peek()
15030 )));
15031 }
15032 self.advance();
15033 let mut key_columns = Vec::new();
15034 loop {
15035 key_columns.push(self.expect_ident_like()?);
15036 match self.peek() {
15037 Token::Comma => {
15038 self.advance();
15039 }
15040 Token::RParen => {
15041 self.advance();
15042 break;
15043 }
15044 other => {
15045 return Err(self.err(format!(
15046 "expected ',' or ')' in PARTITION BY key list, got {other:?}"
15047 )));
15048 }
15049 }
15050 }
15051 if key_columns.is_empty() {
15052 return Err(self.err("PARTITION BY requires at least one key column".to_string()));
15053 }
15054 Ok(PartitionBySpec { kind, key_columns })
15055 }
15056
15057 /// v7.37.6-B — after `PARTITION OF`, expect
15058 /// <parent> FOR VALUES FROM ( <expr> ) TO ( <expr> )
15059 /// or
15060 /// <parent> DEFAULT
15061 fn parse_partition_of_tail(&mut self) -> Result<crate::ast::PartitionOfSpec, ParseError> {
15062 use crate::ast::{PartitionOfBoundsAst, PartitionOfSpec};
15063 let parent_name = self.expect_ident_like()?;
15064 // v7.37.6-B rejects an explicit column list — the child
15065 // inherits from the parent. mailrs round-7 taught us that
15066 // CREATE TABLE-side schema reconciliation hides drift, so
15067 // we surface this as a parse error rather than silently
15068 // ignoring user columns.
15069 if matches!(self.peek(), Token::LParen) {
15070 return Err(self.err(
15071 "CREATE TABLE … PARTITION OF parent: explicit column list not supported \
15072 at v7.37.6-B; the child inherits its columns from the parent"
15073 .to_string(),
15074 ));
15075 }
15076 let bounds = match self.peek() {
15077 Token::Default => {
15078 self.advance();
15079 PartitionOfBoundsAst::Default
15080 }
15081 Token::For => {
15082 self.advance();
15083 if !matches!(self.peek(), Token::Values) {
15084 return Err(
15085 self.err(format!("expected VALUES after FOR, got {:?}", self.peek()))
15086 );
15087 }
15088 self.advance();
15089 // WITH is not a reserved Token in the lexer — it lexes
15090 // as Token::Ident("with"). Disambiguate manually.
15091 let want_with = matches!(
15092 self.peek(),
15093 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("with")
15094 );
15095 if want_with {
15096 self.advance();
15097 if !matches!(self.peek(), Token::LParen) {
15098 return Err(self.err(format!(
15099 "expected '(' after FOR VALUES WITH, got {:?}",
15100 self.peek()
15101 )));
15102 }
15103 self.advance();
15104 let (mut modulus, mut remainder): (Option<u32>, Option<u32>) = (None, None);
15105 loop {
15106 let key = self.expect_ident_like()?;
15107 let n = match self.peek().clone() {
15108 Token::Integer(v) if u32::try_from(v).is_ok() => {
15109 self.advance();
15110 v as u32
15111 }
15112 other => {
15113 return Err(self.err(format!(
15114 "FOR VALUES WITH: expected unsigned integer literal, got {other:?}"
15115 )));
15116 }
15117 };
15118 match key.to_ascii_uppercase().as_str() {
15119 "MODULUS" => modulus = Some(n),
15120 "REMAINDER" => remainder = Some(n),
15121 other => {
15122 return Err(self.err(format!(
15123 "FOR VALUES WITH: unknown key {other:?}; \
15124 expected MODULUS or REMAINDER"
15125 )));
15126 }
15127 }
15128 match self.peek() {
15129 Token::Comma => {
15130 self.advance();
15131 }
15132 Token::RParen => {
15133 self.advance();
15134 break;
15135 }
15136 other => {
15137 return Err(self.err(format!(
15138 "expected ',' or ')' in FOR VALUES WITH list, got {other:?}"
15139 )));
15140 }
15141 }
15142 }
15143 let modulus = modulus
15144 .ok_or_else(|| self.err("FOR VALUES WITH: missing MODULUS".to_string()))?;
15145 let remainder = remainder.ok_or_else(|| {
15146 self.err("FOR VALUES WITH: missing REMAINDER".to_string())
15147 })?;
15148 if modulus == 0 {
15149 return Err(self.err("FOR VALUES WITH: MODULUS must be > 0".to_string()));
15150 }
15151 if remainder >= modulus {
15152 return Err(self.err(format!(
15153 "FOR VALUES WITH: REMAINDER ({remainder}) \
15154 must be < MODULUS ({modulus})"
15155 )));
15156 }
15157 PartitionOfBoundsAst::Hash { modulus, remainder }
15158 } else {
15159 match self.peek() {
15160 Token::From => {
15161 self.advance();
15162 let lower = Box::new(self.parse_partition_bound_expr()?);
15163 if !matches!(self.peek(), Token::To) {
15164 return Err(self.err(format!(
15165 "expected TO after FROM (...), got {:?}",
15166 self.peek()
15167 )));
15168 }
15169 self.advance();
15170 let upper = Box::new(self.parse_partition_bound_expr()?);
15171 PartitionOfBoundsAst::Range { lower, upper }
15172 }
15173 // v7.37.16 (16.1) — FOR VALUES IN (lit [, lit, …])
15174 Token::In => {
15175 self.advance();
15176 if !matches!(self.peek(), Token::LParen) {
15177 return Err(self.err(format!(
15178 "expected '(' after FOR VALUES IN, got {:?}",
15179 self.peek()
15180 )));
15181 }
15182 self.advance();
15183 let mut values = Vec::new();
15184 loop {
15185 values.push(self.parse_expr(0)?);
15186 match self.peek() {
15187 Token::Comma => {
15188 self.advance();
15189 }
15190 Token::RParen => {
15191 self.advance();
15192 break;
15193 }
15194 other => {
15195 return Err(self.err(format!(
15196 "expected ',' or ')' in FOR VALUES IN list, got {other:?}"
15197 )));
15198 }
15199 }
15200 }
15201 if values.is_empty() {
15202 return Err(self.err(
15203 "FOR VALUES IN requires at least one literal".to_string(),
15204 ));
15205 }
15206 PartitionOfBoundsAst::List { values }
15207 }
15208 other => {
15209 return Err(self.err(format!(
15210 "expected FROM / IN / WITH after FOR VALUES, got {other:?}"
15211 )));
15212 }
15213 }
15214 }
15215 }
15216 other => {
15217 return Err(self.err(format!(
15218 "expected FOR VALUES or DEFAULT after PARTITION OF parent, got {other:?}"
15219 )));
15220 }
15221 };
15222 Ok(PartitionOfSpec {
15223 parent_name,
15224 bounds,
15225 })
15226 }
15227
15228 /// v7.37.6-B — a single `( <expr> )` bound. `MINVALUE` /
15229 /// `MAXVALUE` lex as Ident; rewrite them into FunctionCall
15230 /// markers (no-arg builtins) so the engine resolves them
15231 /// against [`spg_storage::PartitionBound::{MinValue, MaxValue}`].
15232 fn parse_partition_bound_expr(&mut self) -> Result<crate::ast::Expr, ParseError> {
15233 if !matches!(self.peek(), Token::LParen) {
15234 return Err(self.err(format!(
15235 "expected '(' before partition bound, got {:?}",
15236 self.peek()
15237 )));
15238 }
15239 self.advance();
15240 let expr = match self.peek() {
15241 Token::Ident(s) | Token::QuotedIdent(s)
15242 if s.eq_ignore_ascii_case("minvalue") || s.eq_ignore_ascii_case("maxvalue") =>
15243 {
15244 let name = s.to_ascii_uppercase();
15245 self.advance();
15246 crate::ast::Expr::FunctionCall {
15247 name,
15248 args: Vec::new(),
15249 }
15250 }
15251 _ => self.parse_expr(0)?,
15252 };
15253 if !matches!(self.peek(), Token::RParen) {
15254 return Err(self.err(format!(
15255 "expected ')' after partition bound, got {:?}",
15256 self.peek()
15257 )));
15258 }
15259 self.advance();
15260 Ok(expr)
15261 }
15262
15263 /// v7.14.0 — true when the next tokens look like an inline
15264 /// MySQL index declaration: KEY / INDEX / UNIQUE KEY /
15265 /// UNIQUE INDEX / FULLTEXT [KEY|INDEX] / SPATIAL [KEY|INDEX]
15266 /// — each followed by an optional name + `(...)`. Critical:
15267 /// a column NAMED `key` / `index` (PG accepts as ident) must
15268 /// NOT be mistaken for the KEY constraint shape. We disambig
15269 /// by requiring the keyword to be followed by either `(` or
15270 /// `<ident> (`.
15271 fn peek_mysql_inline_key_start(&self) -> bool {
15272 let cur = self.peek();
15273 // Shapes:
15274 // KEY (cols)
15275 // KEY name (cols)
15276 // INDEX (cols)
15277 // INDEX name (cols)
15278 // UNIQUE KEY [name] (cols)
15279 // UNIQUE INDEX [name] (cols)
15280 // FULLTEXT [KEY|INDEX] [name] (cols)
15281 // SPATIAL [KEY|INDEX] [name] (cols)
15282 let after_keyword_followed_by_paren_or_ident_paren = |skip: usize| -> bool {
15283 // tokens at skip = the position AFTER the index-form
15284 // keywords (KEY/INDEX) have been consumed.
15285 match self.tokens.get(skip) {
15286 Some(Token::LParen) => true,
15287 Some(Token::Ident(_) | Token::QuotedIdent(_)) => {
15288 matches!(self.tokens.get(skip + 1), Some(Token::LParen))
15289 }
15290 _ => false,
15291 }
15292 };
15293 // `INDEX` lexes as Token::Index (reserved), not as
15294 // Token::Ident("index"). Both shapes count as a KEY/INDEX
15295 // start; the peek helper below handles either.
15296 let is_key_or_index_tok = |t: &Token| -> bool {
15297 matches!(t, Token::Index)
15298 || matches!(t, Token::Ident(s) if s.eq_ignore_ascii_case("key") || s.eq_ignore_ascii_case("index"))
15299 };
15300 match cur {
15301 Token::Index => after_keyword_followed_by_paren_or_ident_paren(self.pos + 1),
15302 Token::Ident(s) if s.eq_ignore_ascii_case("key") || s.eq_ignore_ascii_case("index") => {
15303 after_keyword_followed_by_paren_or_ident_paren(self.pos + 1)
15304 }
15305 Token::Ident(s)
15306 if s.eq_ignore_ascii_case("fulltext") || s.eq_ignore_ascii_case("spatial") =>
15307 {
15308 let nxt = self.tokens.get(self.pos + 1);
15309 let after_after = if nxt.is_some_and(is_key_or_index_tok) {
15310 self.pos + 2
15311 } else {
15312 self.pos + 1
15313 };
15314 after_keyword_followed_by_paren_or_ident_paren(after_after)
15315 }
15316 Token::Ident(s) if s.eq_ignore_ascii_case("unique") => {
15317 let nxt = self.tokens.get(self.pos + 1);
15318 if !nxt.is_some_and(is_key_or_index_tok) {
15319 return false;
15320 }
15321 after_keyword_followed_by_paren_or_ident_paren(self.pos + 2)
15322 }
15323 _ => false,
15324 }
15325 }
15326
15327 /// v7.14.0 — parse the MySQL inline KEY/INDEX form. Returns
15328 /// Some(TableConstraint::Unique) for UNIQUE KEY (so SPG
15329 /// enforces uniqueness on INSERT). v7.15.0: plain KEY/INDEX
15330 /// returns Some(TableConstraint::Index) so the engine builds
15331 /// a real BTree index on the leading column (mysqldump
15332 /// `KEY idx_posts_author (author_id)` shape).
15333 /// FULLTEXT / SPATIAL still return None — accepted-as-no-op
15334 /// (the storage layer has no matching AM).
15335 fn parse_mysql_inline_key(
15336 &mut self,
15337 ) -> Result<Option<crate::ast::TableConstraint>, ParseError> {
15338 // Detect UNIQUE prefix.
15339 let is_unique = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("unique"))
15340 {
15341 self.advance();
15342 true
15343 } else {
15344 false
15345 };
15346 // Consume FULLTEXT / SPATIAL prefix and record which one
15347 // it was. v7.17.0 Phase 2.2 — FULLTEXT routes through a
15348 // dedicated TableConstraint variant so the engine can
15349 // build a tsvector-GIN; SPATIAL still has no matching
15350 // AM, so it falls back to accept-as-no-op.
15351 let mut is_fulltext = false;
15352 let mut is_spatial = false;
15353 if let Token::Ident(s) = self.peek().clone() {
15354 if s.eq_ignore_ascii_case("fulltext") {
15355 self.advance();
15356 is_fulltext = true;
15357 } else if s.eq_ignore_ascii_case("spatial") {
15358 self.advance();
15359 is_spatial = true;
15360 }
15361 }
15362 // KEY / INDEX keyword. `INDEX` lexes as Token::Index
15363 // (reserved); accept either token shape.
15364 match self.peek() {
15365 Token::Index => {
15366 self.advance();
15367 }
15368 Token::Ident(s) if s.eq_ignore_ascii_case("key") || s.eq_ignore_ascii_case("index") => {
15369 self.advance();
15370 }
15371 other => {
15372 return Err(self.err(alloc::format!(
15373 "expected KEY/INDEX in inline index declaration, got {other:?}"
15374 )));
15375 }
15376 }
15377 // Optional index name (an ident before the `(`).
15378 // v7.15.0 — capture the name when present so the engine
15379 // builds the secondary index under the user's chosen
15380 // name (matches mysqldump's `KEY idx_x (col)` shape).
15381 let mut idx_name: Option<String> = None;
15382 if matches!(self.peek(), Token::Ident(_) | Token::QuotedIdent(_))
15383 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
15384 {
15385 if let Token::Ident(s) | Token::QuotedIdent(s) = self.advance() {
15386 idx_name = Some(s);
15387 }
15388 }
15389 // Optional `USING BTREE` / `USING HASH` (MySQL).
15390 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
15391 self.advance();
15392 if matches!(self.peek(), Token::Ident(_) | Token::QuotedIdent(_)) {
15393 self.advance();
15394 }
15395 }
15396 // Required column list `(col [, col]*)`.
15397 if !matches!(self.peek(), Token::LParen) {
15398 return Err(self.err(alloc::format!(
15399 "expected '(' in inline KEY/INDEX, got {:?}",
15400 self.peek()
15401 )));
15402 }
15403 self.advance();
15404 let mut cols: Vec<String> = Vec::new();
15405 while let Token::Ident(s) | Token::QuotedIdent(s) = self.peek().clone() {
15406 self.advance();
15407 cols.push(s);
15408 // Skip optional `(length)` per-column prefix.
15409 if matches!(self.peek(), Token::LParen) {
15410 let mut depth = 1usize;
15411 self.advance();
15412 while depth > 0 {
15413 match self.peek() {
15414 Token::LParen => depth += 1,
15415 Token::RParen => depth -= 1,
15416 Token::Eof => break,
15417 _ => {}
15418 }
15419 self.advance();
15420 }
15421 }
15422 // Skip optional ASC / DESC.
15423 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("asc") || s.eq_ignore_ascii_case("desc"))
15424 || matches!(self.peek(), Token::Asc | Token::Desc)
15425 {
15426 self.advance();
15427 }
15428 if matches!(self.peek(), Token::Comma) {
15429 self.advance();
15430 continue;
15431 }
15432 break;
15433 }
15434 if matches!(self.peek(), Token::RParen) {
15435 self.advance();
15436 }
15437 // Trailing options on the inline index — comment / etc.
15438 // Skip until comma or `)`.
15439 while !matches!(self.peek(), Token::Comma | Token::RParen | Token::Eof) {
15440 self.advance();
15441 }
15442 if cols.is_empty() {
15443 return Ok(None);
15444 }
15445 if is_unique {
15446 // Carry the captured idx_name on UNIQUE too so future
15447 // engine work can name the underlying BTree
15448 // accordingly; today the unique-constraint installer
15449 // synthesises the name itself, but Display round-trip
15450 // benefits from preserving it.
15451 Ok(Some(crate::ast::TableConstraint::Unique {
15452 name: idx_name,
15453 columns: cols,
15454 nulls_not_distinct: false,
15455 // MySQL inline UNIQUE KEY has no deferral vocabulary.
15456 deferrable: false,
15457 initially_deferred: false,
15458 }))
15459 } else if is_fulltext {
15460 // v7.17.0 Phase 2.2 — MySQL `FULLTEXT KEY` now
15461 // routes through `TableConstraint::FulltextIndex`;
15462 // the engine builds a tsvector-GIN over each named
15463 // column so MATCH AGAINST gets a real inverted
15464 // index instead of a silently-dropped declaration.
15465 Ok(Some(crate::ast::TableConstraint::FulltextIndex {
15466 name: idx_name,
15467 columns: cols,
15468 }))
15469 } else if is_spatial {
15470 // SPG has no native SPATIAL AM. Accept-as-no-op
15471 // (declaration is parsed, but no index is built).
15472 Ok(None)
15473 } else {
15474 // v7.15.0 — plain KEY / INDEX builds a real BTree
15475 // secondary index.
15476 Ok(Some(crate::ast::TableConstraint::Index {
15477 name: idx_name,
15478 columns: cols,
15479 }))
15480 }
15481 }
15482
15483 /// v7.14.0 — consume MySQL/MariaDB table-options tail after
15484 /// the closing `)`: ENGINE=..., DEFAULT CHARSET=...,
15485 /// COLLATE=..., AUTO_INCREMENT=N, ROW_FORMAT=..., COMMENT='...'
15486 /// (in any order, separated by whitespace).
15487 /// v7.38 (read01 P6.55) — consume and discard a PG `WITH (opt=val, …)`
15488 /// storage-parameter clause on CREATE TABLE. SPG has no per-table
15489 /// reloptions; accepting them keeps pg_dump restores working. `WITH` is a
15490 /// bare ident here, and only the parenthesised form is reloptions (so this
15491 /// never eats a `WITH DATA` / `WITH CHECK OPTION` trailer).
15492 fn consume_with_reloptions(&mut self) {
15493 let is_with = matches!(
15494 self.peek(),
15495 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("with")
15496 );
15497 if !is_with || !matches!(self.tokens.get(self.pos + 1), Some(Token::LParen)) {
15498 return;
15499 }
15500 self.advance(); // WITH
15501 self.advance(); // (
15502 let mut depth = 1u32;
15503 while depth > 0 && !matches!(self.peek(), Token::Eof) {
15504 match self.peek() {
15505 Token::LParen => depth += 1,
15506 Token::RParen => depth -= 1,
15507 _ => {}
15508 }
15509 self.advance();
15510 }
15511 }
15512
15513 /// v7.39 — returns the `ENGINE=` name, which used to be consumed and
15514 /// dropped with everything else here. The rest of the MySQL table
15515 /// options genuinely have no meaning for SPG's storage; the engine
15516 /// name does, because MySQL REFUSES one it does not know and a dump
15517 /// with a typo in it should not quietly become a table.
15518 fn consume_mysql_table_options(&mut self) -> Option<alloc::string::String> {
15519 let mut engine: Option<alloc::string::String> = None;
15520 loop {
15521 // Heuristic: a table option is an ident (or `DEFAULT`
15522 // reserved keyword) followed by `=` and an
15523 // ident / string / integer.
15524 let name_lc = match self.peek().clone() {
15525 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
15526 Token::Default => alloc::string::String::from("default"),
15527 _ => break,
15528 };
15529 let known = matches!(
15530 name_lc.as_str(),
15531 "engine"
15532 | "default"
15533 | "charset"
15534 | "collate"
15535 | "auto_increment"
15536 | "row_format"
15537 | "comment"
15538 | "pack_keys"
15539 | "stats_persistent"
15540 | "stats_auto_recalc"
15541 | "stats_sample_pages"
15542 | "key_block_size"
15543 | "tablespace"
15544 | "min_rows"
15545 | "max_rows"
15546 | "checksum"
15547 | "delay_key_write"
15548 | "insert_method"
15549 | "data"
15550 | "index"
15551 | "encryption"
15552 | "compression"
15553 );
15554 if !known {
15555 break;
15556 }
15557 self.advance(); // option name
15558 // `DEFAULT` optional prefix is followed by `CHARSET` /
15559 // `COLLATE`; consume the next ident too.
15560 if name_lc == "default" {
15561 if matches!(self.peek(), Token::Ident(_) | Token::QuotedIdent(_)) {
15562 self.advance();
15563 }
15564 }
15565 if matches!(self.peek(), Token::Eq) {
15566 self.advance();
15567 }
15568 match self.peek().clone() {
15569 Token::Ident(v) | Token::QuotedIdent(v) | Token::String(v) => {
15570 if name_lc == "engine" {
15571 // v7.39.3 — as WRITTEN. MySQL 9.7.2 refuses an
15572 // engine it does not know and names it back
15573 // exactly: `Unknown storage engine 'NoSuchEng'`,
15574 // measured. The lexer folds a bare identifier, so
15575 // the message quoted a name the dump did not
15576 // contain, which is the one thing that message is
15577 // for. Guarded the same way the column spelling
15578 // is: the span runs to the next token, so what
15579 // comes back has to be the same word.
15580 let written = self
15581 .source_span(self.pos, self.pos)
15582 .map(|raw| raw.trim().trim_matches('`').trim_matches('\''))
15583 .filter(|raw| raw.eq_ignore_ascii_case(&v))
15584 .map(alloc::string::String::from);
15585 engine = Some(written.unwrap_or(v));
15586 }
15587 self.advance();
15588 }
15589 Token::Integer(_) => {
15590 self.advance();
15591 }
15592 _ => {}
15593 }
15594 }
15595 engine
15596 }
15597
15598 /// v7.9.18 — true when the next tokens are `PRIMARY KEY (…)`.
15599 /// PRIMARY and KEY are bare idents; we look-ahead 2 to be
15600 /// sure (otherwise a column literally named `primary` would
15601 /// be mistaken).
15602 fn peek_table_level_pk_start(&self) -> bool {
15603 let cur = self.peek();
15604 let nxt = self.tokens.get(self.pos + 1);
15605 let nxt2 = self.tokens.get(self.pos + 2);
15606 let is_primary = matches!(cur, Token::Ident(s) if s.eq_ignore_ascii_case("primary"));
15607 let is_key = matches!(nxt, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("key"));
15608 let is_lparen = matches!(nxt2, Some(Token::LParen));
15609 is_primary && is_key && is_lparen
15610 }
15611
15612 /// v7.9.18 — true when the next tokens are `UNIQUE (…)`.
15613 /// v7.13.0 — also matches `UNIQUE NULLS [NOT] DISTINCT (…)`
15614 /// (mailrs round-5 G10).
15615 fn peek_table_level_unique_start(&self) -> bool {
15616 let cur = self.peek();
15617 let is_unique = matches!(cur, Token::Ident(s) if s.eq_ignore_ascii_case("unique"));
15618 if !is_unique {
15619 return false;
15620 }
15621 let n1 = self.tokens.get(self.pos + 1);
15622 // Plain `UNIQUE (…)`.
15623 if matches!(n1, Some(Token::LParen)) {
15624 return true;
15625 }
15626 // `UNIQUE NULLS [NOT] DISTINCT (…)`.
15627 let is_nulls = matches!(n1, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("nulls"));
15628 if !is_nulls {
15629 return false;
15630 }
15631 let n2 = self.tokens.get(self.pos + 2);
15632 let n3 = self.tokens.get(self.pos + 3);
15633 let n4 = self.tokens.get(self.pos + 4);
15634 // `UNIQUE NULLS DISTINCT (…)` — 4 tokens before `(`.
15635 if matches!(n2, Some(Token::Distinct)) && matches!(n3, Some(Token::LParen)) {
15636 return true;
15637 }
15638 // `UNIQUE NULLS NOT DISTINCT (…)` — 5 tokens before `(`.
15639 if matches!(n2, Some(Token::Not))
15640 && matches!(n3, Some(Token::Distinct))
15641 && matches!(n4, Some(Token::LParen))
15642 {
15643 return true;
15644 }
15645 false
15646 }
15647
15648 fn parse_table_level_primary_key(&mut self) -> Result<crate::ast::TableConstraint, ParseError> {
15649 self.advance(); // PRIMARY
15650 self.advance(); // KEY
15651 let columns = self.parse_paren_ident_list("PRIMARY KEY")?;
15652 // v7.39 (round 711) — the trailer's values are CARRIED now; round
15653 // 621 consumed and dropped them (the storing half of F08).
15654 let (deferrable, initially_deferred) = self.consume_deferrable_clauses_timed()?;
15655 Ok(crate::ast::TableConstraint::PrimaryKey {
15656 name: None,
15657 columns,
15658 deferrable,
15659 initially_deferred,
15660 })
15661 }
15662
15663 fn parse_table_level_unique(&mut self) -> Result<crate::ast::TableConstraint, ParseError> {
15664 self.advance(); // UNIQUE
15665 // v7.13.0 — optional `NULLS NOT DISTINCT` modifier
15666 // (mailrs round-5 G10, PG 15+ surface). Default behaviour
15667 // is `NULLS DISTINCT` per the SQL standard.
15668 let mut nulls_not_distinct = false;
15669 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("nulls")) {
15670 let n1 = self.tokens.get(self.pos + 1);
15671 let n2 = self.tokens.get(self.pos + 2);
15672 let is_not = matches!(n1, Some(Token::Not));
15673 let is_distinct = matches!(n2, Some(Token::Distinct));
15674 if is_not && is_distinct {
15675 self.advance(); // NULLS
15676 self.advance(); // NOT
15677 self.advance(); // DISTINCT
15678 nulls_not_distinct = true;
15679 } else if matches!(n1, Some(Token::Distinct)) {
15680 self.advance(); // NULLS
15681 self.advance(); // DISTINCT
15682 }
15683 }
15684 let columns = self.parse_paren_ident_list("UNIQUE")?;
15685 let (deferrable, initially_deferred) = self.consume_deferrable_clauses_timed()?;
15686 Ok(crate::ast::TableConstraint::Unique {
15687 name: None,
15688 columns,
15689 nulls_not_distinct,
15690 deferrable,
15691 initially_deferred,
15692 })
15693 }
15694
15695 /// v7.13.0 — table-level `CHECK (<expr>)` constraint
15696 /// (mailrs round-5 G3). Consumes `CHECK` then a parenthesised
15697 /// expression.
15698 /// v7.39 (round 210) — `EXCLUDE [USING <method>] ( <col> WITH <op>
15699 /// [, <col> WITH <op>]* ) [WHERE (...)]`. The operator is read as a
15700 /// standalone token spelling (`&&`, `=`, `@>`, `<@`, `&<`, `&>`).
15701 /// v7.39 (round 652) — the optional `NOT VALID` suffix on a
15702 /// constraint added by ALTER TABLE. `NOT` alone is not enough to
15703 /// commit: `NOT` starts no other suffix here, but reading both
15704 /// tokens before advancing keeps the caller's error message intact
15705 /// if someone writes `NOT NULL` by mistake.
15706 fn parse_not_valid_suffix(&mut self) -> bool {
15707 if !matches!(self.peek(), Token::Not) {
15708 return false;
15709 }
15710 if !matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("valid"))
15711 {
15712 return false;
15713 }
15714 self.advance();
15715 self.advance();
15716 true
15717 }
15718
15719 fn parse_table_level_exclude(&mut self) -> Result<crate::ast::TableConstraint, ParseError> {
15720 self.advance(); // EXCLUDE
15721 // Optional `USING <method>`.
15722 let mut method = None;
15723 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
15724 self.advance();
15725 method = Some(match self.advance() {
15726 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
15727 other => {
15728 return Err(self.err(alloc::format!(
15729 "expected index method after USING, got {other:?}"
15730 )));
15731 }
15732 });
15733 }
15734 if !matches!(self.peek(), Token::LParen) {
15735 return Err(self.err(alloc::format!(
15736 "expected '(' after EXCLUDE, got {:?}",
15737 self.peek()
15738 )));
15739 }
15740 self.advance();
15741 let mut elements: Vec<(String, String)> = Vec::new();
15742 loop {
15743 let col = match self.advance() {
15744 Token::Ident(s) | Token::QuotedIdent(s) => s,
15745 other => {
15746 return Err(self.err(alloc::format!(
15747 "expected column name in EXCLUDE, got {other:?}"
15748 )));
15749 }
15750 };
15751 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with")) {
15752 return Err(self.err(alloc::format!(
15753 "expected WITH after EXCLUDE column, got {:?}",
15754 self.peek()
15755 )));
15756 }
15757 self.advance();
15758 let op = match self.advance() {
15759 Token::InetOverlap => String::from("&&"),
15760 Token::Intersects => String::from("?#"),
15761 Token::IsBelow => String::from("<^"),
15762 Token::IsAbove => String::from(">^"),
15763 Token::PatternLt => String::from("~<~"),
15764 Token::PatternLtEq => String::from("~<=~"),
15765 Token::PatternGt => String::from("~>~"),
15766 Token::PatternGtEq => String::from("~>=~"),
15767 Token::TsMatchOld => String::from("@@@"),
15768 Token::Eq => String::from("="),
15769 Token::JsonContains => String::from("@>"),
15770 Token::JsonContainedBy => String::from("<@"),
15771 Token::OverLeft => String::from("&<"),
15772 Token::OverRight => String::from("&>"),
15773 other => {
15774 return Err(self.err(alloc::format!(
15775 "unsupported EXCLUDE operator {other:?} (SPG supports &&, =, @>, <@, &<, &>)"
15776 )));
15777 }
15778 };
15779 elements.push((col, op));
15780 if matches!(self.peek(), Token::Comma) {
15781 self.advance();
15782 continue;
15783 }
15784 break;
15785 }
15786 if !matches!(self.peek(), Token::RParen) {
15787 return Err(self.err(alloc::format!(
15788 "expected ')' to close EXCLUDE, got {:?}",
15789 self.peek()
15790 )));
15791 }
15792 self.advance();
15793 Ok(crate::ast::TableConstraint::Exclude {
15794 name: None,
15795 method,
15796 elements,
15797 })
15798 }
15799
15800 fn parse_table_level_check(&mut self) -> Result<crate::ast::TableConstraint, ParseError> {
15801 self.advance(); // CHECK
15802 if !matches!(self.peek(), Token::LParen) {
15803 return Err(self.err(alloc::format!(
15804 "expected '(' after CHECK, got {:?}",
15805 self.peek()
15806 )));
15807 }
15808 self.advance();
15809 let expr = self.parse_expr(0)?;
15810 if !matches!(self.peek(), Token::RParen) {
15811 return Err(self.err(alloc::format!(
15812 "expected ')' to close CHECK predicate, got {:?}",
15813 self.peek()
15814 )));
15815 }
15816 self.advance();
15817 // A CHECK written inside CREATE TABLE cannot be NOT VALID: there
15818 // are no existing rows for PG to skip, so it rejects the suffix.
15819 Ok(crate::ast::TableConstraint::Check {
15820 name: None,
15821 expr,
15822 not_valid: false,
15823 })
15824 }
15825
15826 /// v7.13.0 — `true` when the next token is `CHECK` (a bare ident).
15827 fn peek_table_level_check_start(&self) -> bool {
15828 matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("check"))
15829 }
15830
15831 /// v7.22 (round-13 gap 5) — `Some(kind)` when the next tokens are
15832 /// `CONSTRAINT <name> { CHECK | UNIQUE | PRIMARY }`. FOREIGN stays
15833 /// on the dedicated FK path (`parse_table_level_fk` consumes its
15834 /// own CONSTRAINT prefix).
15835 fn peek_named_table_constraint_kind(&self) -> Option<NamedTableConstraintKind> {
15836 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint")) {
15837 return None;
15838 }
15839 // tokens[pos+1] is the constraint name (any ident-like);
15840 // tokens[pos+2] is the kind keyword.
15841 match self.tokens.get(self.pos + 2) {
15842 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("check") => {
15843 Some(NamedTableConstraintKind::Check)
15844 }
15845 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("unique") => {
15846 Some(NamedTableConstraintKind::Unique)
15847 }
15848 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("primary") => {
15849 Some(NamedTableConstraintKind::PrimaryKey)
15850 }
15851 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("exclude") => {
15852 Some(NamedTableConstraintKind::Exclude)
15853 }
15854 _ => None,
15855 }
15856 }
15857
15858 fn parse_paren_ident_list(&mut self, ctx: &str) -> Result<Vec<String>, ParseError> {
15859 if !matches!(self.peek(), Token::LParen) {
15860 return Err(self.err(alloc::format!(
15861 "expected '(' after {ctx}, got {:?}",
15862 self.peek()
15863 )));
15864 }
15865 self.advance();
15866 let mut out = Vec::new();
15867 loop {
15868 out.push(self.expect_ident_like()?);
15869 match self.peek() {
15870 Token::Comma => {
15871 self.advance();
15872 }
15873 Token::RParen => {
15874 self.advance();
15875 break;
15876 }
15877 other => {
15878 return Err(self.err(alloc::format!(
15879 "expected ',' or ')' in {ctx} list, got {other:?}"
15880 )));
15881 }
15882 }
15883 }
15884 if out.is_empty() {
15885 return Err(self.err(alloc::format!("{ctx} requires at least one column")));
15886 }
15887 Ok(out)
15888 }
15889
15890 /// v7.6.0 — true when the next tokens are `CONSTRAINT <name>
15891 /// FOREIGN KEY` or bare `FOREIGN KEY`. Both introduce a
15892 /// table-level FK; a column def never starts with either keyword
15893 /// (column names are not in this reserved set).
15894 fn peek_constraint_or_fk_start(&self) -> bool {
15895 let is_constraint_kw = matches!(
15896 self.peek(),
15897 Token::Ident(s) if s.eq_ignore_ascii_case("constraint")
15898 );
15899 let is_foreign_kw = matches!(
15900 self.peek(),
15901 Token::Ident(s) if s.eq_ignore_ascii_case("foreign")
15902 );
15903 is_constraint_kw || is_foreign_kw
15904 }
15905
15906 /// v7.6.0 — parse a table-level FK clause:
15907 /// `[CONSTRAINT <name>] FOREIGN KEY (<col>[,<col>]*) REFERENCES
15908 /// <tbl> [(<pcol>[,<pcol>]*)] [ON DELETE <action>] [ON UPDATE <action>]`.
15909 fn parse_table_level_fk(&mut self) -> Result<ForeignKeyConstraint, ParseError> {
15910 let mut name: Option<String> = None;
15911 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint")) {
15912 self.advance();
15913 name = Some(self.expect_ident_like()?);
15914 }
15915 // `FOREIGN`
15916 match self.advance() {
15917 Token::Ident(s) if s.eq_ignore_ascii_case("foreign") => {}
15918 other => return Err(self.err(format!("expected FOREIGN, got {other:?}"))),
15919 }
15920 // `KEY`
15921 match self.advance() {
15922 Token::Ident(s) if s.eq_ignore_ascii_case("key") => {}
15923 other => return Err(self.err(format!("expected KEY after FOREIGN, got {other:?}"))),
15924 }
15925 // `(col, col, ...)`
15926 if !matches!(self.peek(), Token::LParen) {
15927 return Err(self.err(format!(
15928 "expected '(' after FOREIGN KEY, got {:?}",
15929 self.peek()
15930 )));
15931 }
15932 self.advance();
15933 let mut columns = Vec::new();
15934 loop {
15935 columns.push(self.expect_ident_like()?);
15936 match self.peek() {
15937 Token::Comma => {
15938 self.advance();
15939 }
15940 Token::RParen => {
15941 self.advance();
15942 break;
15943 }
15944 other => {
15945 return Err(self.err(format!(
15946 "expected ',' or ')' in FK column list, got {other:?}"
15947 )));
15948 }
15949 }
15950 }
15951 if columns.is_empty() {
15952 return Err(self.err("FOREIGN KEY requires at least one column".into()));
15953 }
15954 let (
15955 parent_table,
15956 parent_columns,
15957 on_delete,
15958 on_update,
15959 match_type,
15960 deferrable,
15961 initially_deferred,
15962 ) = self.parse_references_tail(columns.len())?;
15963 Ok(ForeignKeyConstraint {
15964 name,
15965 columns,
15966 parent_table,
15967 parent_columns,
15968 on_delete,
15969 on_update,
15970 match_type,
15971 deferrable,
15972 initially_deferred,
15973 })
15974 }
15975
15976 /// v7.6.0 — parse the tail `REFERENCES <tbl> [(<pcol>...)] [ON
15977 /// DELETE <action>] [ON UPDATE <action>]`. `expected_arity` is
15978 /// the local column count, used to default the parent column
15979 /// list when omitted (SQL spec: parent's PK is implied).
15980 fn parse_references_tail(
15981 &mut self,
15982 expected_arity: usize,
15983 ) -> Result<
15984 (
15985 String,
15986 Vec<String>,
15987 FkAction,
15988 FkAction,
15989 crate::ast::MatchType,
15990 // v7.39 (round 288) — deferrable, initially_deferred.
15991 bool,
15992 bool,
15993 ),
15994 ParseError,
15995 > {
15996 match self.advance() {
15997 Token::Ident(s) if s.eq_ignore_ascii_case("references") => {}
15998 other => return Err(self.err(format!("expected REFERENCES, got {other:?}"))),
15999 }
16000 let parent_table = self.expect_ident_like()?;
16001 let mut parent_columns: Vec<String> = Vec::new();
16002 if matches!(self.peek(), Token::LParen) {
16003 self.advance();
16004 loop {
16005 parent_columns.push(self.expect_ident_like()?);
16006 match self.peek() {
16007 Token::Comma => {
16008 self.advance();
16009 }
16010 Token::RParen => {
16011 self.advance();
16012 break;
16013 }
16014 other => {
16015 return Err(self.err(format!(
16016 "expected ',' or ')' in REFERENCES column list, got {other:?}"
16017 )));
16018 }
16019 }
16020 }
16021 }
16022 if !parent_columns.is_empty() && parent_columns.len() != expected_arity {
16023 return Err(self.err(format!(
16024 "FK arity mismatch: {} local column(s) vs {} parent column(s)",
16025 expected_arity,
16026 parent_columns.len()
16027 )));
16028 }
16029 // Optional `MATCH {SIMPLE | FULL | PARTIAL}`. PG's grammar puts
16030 // it between the referenced column list and the ON / DEFERRABLE
16031 // trailers. SPG implements MATCH SIMPLE semantics (the FK check
16032 // is skipped when any referencing column is NULL), so SIMPLE —
16033 // the default, and the only spelling pg_dump emits — is accepted
16034 // as a no-op. MATCH FULL / MATCH PARTIAL need the per-FK
16035 // mixed-NULL rule, which is not wired yet; reject them honestly
16036 // rather than silently applying SIMPLE (PG itself errors on
16037 // MATCH PARTIAL as "not yet implemented").
16038 let mut match_type = crate::ast::MatchType::Simple;
16039 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("match")) {
16040 self.advance();
16041 // `FULL` is a reserved keyword token (FULL OUTER JOIN);
16042 // SIMPLE / PARTIAL arrive as bare identifiers.
16043 let kind = match self.advance() {
16044 Token::Full => "FULL".to_string(),
16045 Token::Ident(s) => s.to_uppercase(),
16046 other => {
16047 return Err(self.err(format!(
16048 "expected FULL, PARTIAL or SIMPLE after MATCH, got {other:?}"
16049 )));
16050 }
16051 };
16052 match kind.as_str() {
16053 "SIMPLE" => {} // Default — match_type stays Simple.
16054 // v7.38 (read01, T29) — MATCH FULL: the check is skipped only
16055 // when ALL referencing columns are NULL; a mixed-NULL key errors.
16056 "FULL" => match_type = crate::ast::MatchType::Full,
16057 "PARTIAL" => {
16058 return Err(self.err("MATCH PARTIAL not yet implemented".to_string()));
16059 }
16060 _ => {
16061 return Err(self.err(format!(
16062 "expected FULL, PARTIAL or SIMPLE after MATCH, got {kind}"
16063 )));
16064 }
16065 }
16066 }
16067 // v7.6.7 / v7.17.0 Phase 3.1 — interleave `[NOT] DEFERRABLE
16068 // [INITIALLY {DEFERRED | IMMEDIATE}]` and `ON DELETE
16069 // <action>` / `ON UPDATE <action>` in either order. PG /
16070 // pg_dump emits the timing clause AFTER the ON clauses
16071 // (`ON DELETE CASCADE DEFERRABLE INITIALLY DEFERRED`),
16072 // but the SQL spec allows either order. We loop over
16073 // every possible trailer and dispatch on the next token,
16074 // stopping when nothing matches. Phase 3.1 changes the
16075 // bare DEFERRABLE form from hard-error to accept-as-
16076 // immediate; SPG is single-writer with no deferred-
16077 // constraint window so the runtime semantics are always
16078 // immediate even when INITIALLY DEFERRED is requested.
16079 // PG's default referential action (no ON DELETE / ON UPDATE
16080 // clause) is NO ACTION, not RESTRICT — the two enforce
16081 // identically in SPG (single-writer, no deferred window; see the
16082 // shared match arm in constraints.rs) but information_schema.
16083 // referential_constraints must report NO ACTION to match PG.
16084 let mut on_delete = FkAction::NoAction;
16085 let mut on_update = FkAction::NoAction;
16086 let mut seen_on_delete = false;
16087 let mut seen_on_update = false;
16088 let mut deferrable = false;
16089 let mut initially_deferred = false;
16090 loop {
16091 // DEFERRABLE / NOT DEFERRABLE / INITIALLY shapes.
16092 let before = self.pos;
16093 let (d, idef) = self.consume_deferrable_clauses_timed()?;
16094 if self.pos != before {
16095 deferrable = d;
16096 initially_deferred = idef;
16097 continue;
16098 }
16099 // ON DELETE / ON UPDATE.
16100 if !matches!(self.peek(), Token::On) {
16101 break;
16102 }
16103 self.advance();
16104 let which = self.advance();
16105 let action = self.parse_fk_action()?;
16106 match which {
16107 Token::Ident(ref s) if s.eq_ignore_ascii_case("delete") => {
16108 if seen_on_delete {
16109 return Err(self.err("ON DELETE specified twice".into()));
16110 }
16111 seen_on_delete = true;
16112 on_delete = action;
16113 }
16114 Token::Ident(ref s) if s.eq_ignore_ascii_case("update") => {
16115 if seen_on_update {
16116 return Err(self.err("ON UPDATE specified twice".into()));
16117 }
16118 seen_on_update = true;
16119 on_update = action;
16120 }
16121 other => {
16122 return Err(
16123 self.err(format!("expected DELETE or UPDATE after ON, got {other:?}"))
16124 );
16125 }
16126 }
16127 }
16128 Ok((
16129 parent_table,
16130 parent_columns,
16131 on_delete,
16132 on_update,
16133 match_type,
16134 deferrable,
16135 initially_deferred,
16136 ))
16137 }
16138
16139 /// v7.6.0 — parse `CASCADE | RESTRICT | SET NULL | SET DEFAULT |
16140 /// NO ACTION`.
16141 fn parse_fk_action(&mut self) -> Result<FkAction, ParseError> {
16142 match self.advance() {
16143 Token::Ident(s) if s.eq_ignore_ascii_case("cascade") => Ok(FkAction::Cascade),
16144 Token::Ident(s) if s.eq_ignore_ascii_case("restrict") => Ok(FkAction::Restrict),
16145 Token::Ident(s) if s.eq_ignore_ascii_case("set") => match self.advance() {
16146 Token::Null => Ok(FkAction::SetNull),
16147 Token::Default => Ok(FkAction::SetDefault),
16148 other => Err(self.err(format!(
16149 "expected NULL or DEFAULT after SET in FK action, got {other:?}"
16150 ))),
16151 },
16152 Token::Ident(s) if s.eq_ignore_ascii_case("no") => match self.advance() {
16153 Token::Ident(s) if s.eq_ignore_ascii_case("action") => Ok(FkAction::NoAction),
16154 other => Err(self.err(format!(
16155 "expected ACTION after NO in FK action, got {other:?}"
16156 ))),
16157 },
16158 other => Err(self.err(format!(
16159 "expected CASCADE | RESTRICT | SET NULL | SET DEFAULT | NO ACTION, got {other:?}"
16160 ))),
16161 }
16162 }
16163
16164 /// Recognise the optional `IF NOT EXISTS` prefix shared by `CREATE
16165 /// TABLE` and `CREATE INDEX`. Returns `true` if consumed.
16166 fn consume_if_not_exists(&mut self) -> bool {
16167 // `IF` arrives as a bare Ident (we don't reserve it because it
16168 // also appears mid-expression in PG, though we don't support
16169 // those forms yet).
16170 let looks_like_if = matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if"));
16171 if !looks_like_if {
16172 return false;
16173 }
16174 // Peek one ahead before committing: only consume IF when it's
16175 // actually `IF NOT EXISTS`.
16176 if !matches!(self.tokens.get(self.pos + 1), Some(Token::Not)) {
16177 return false;
16178 }
16179 if !matches!(
16180 self.tokens.get(self.pos + 2),
16181 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("exists")
16182 ) {
16183 return false;
16184 }
16185 self.advance(); // IF
16186 self.advance(); // NOT
16187 self.advance(); // EXISTS
16188 true
16189 }
16190
16191 /// v7.12.4 — `IF EXISTS` modifier for DROP statements.
16192 /// Consumes IF EXISTS as a pair; returns false otherwise
16193 /// without consuming any tokens.
16194 /// v7.39 (RLS) — consume the `ROW LEVEL SECURITY` keyword triple after
16195 /// ENABLE/DISABLE/FORCE/NO FORCE.
16196 fn expect_row_level_security(&mut self) -> Result<(), ParseError> {
16197 for kw in ["row", "level", "security"] {
16198 if !matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case(kw))
16199 {
16200 return Err(self.err(alloc::format!(
16201 "expected {} in ROW LEVEL SECURITY, got {:?}",
16202 kw.to_ascii_uppercase(),
16203 self.peek()
16204 )));
16205 }
16206 self.advance();
16207 }
16208 Ok(())
16209 }
16210
16211 fn consume_if_exists(&mut self) -> bool {
16212 let looks_like_if = matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("if"));
16213 if !looks_like_if {
16214 return false;
16215 }
16216 if !matches!(
16217 self.tokens.get(self.pos + 1),
16218 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("exists")
16219 ) {
16220 return false;
16221 }
16222 self.advance(); // IF
16223 self.advance(); // EXISTS
16224 true
16225 }
16226
16227 /// v7.9.14 — consume `ASC | DESC | NULLS FIRST | NULLS LAST`
16228 /// qualifiers after an index column ref. ASC / DESC are
16229 /// reserved tokens; NULLS / FIRST / LAST are bare idents.
16230 /// We accept and discard them since single-column BTree
16231 /// stores rows in natural key order today.
16232 /// v7.24 (round-16 A) — `NULLS FIRST` / `NULLS LAST` after an
16233 /// ORDER BY key. Returns None when absent.
16234 fn parse_optional_nulls_placement(&mut self) -> Result<Option<bool>, ParseError> {
16235 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("nulls")) {
16236 return Ok(None);
16237 }
16238 self.advance();
16239 match self.advance() {
16240 Token::Ident(s) if s.eq_ignore_ascii_case("first") => Ok(Some(true)),
16241 Token::Ident(s) if s.eq_ignore_ascii_case("last") => Ok(Some(false)),
16242 other => Err(self.err(alloc::format!(
16243 "expected FIRST or LAST after NULLS, got {other:?}"
16244 ))),
16245 }
16246 }
16247
16248 /// v7.39 (round 537) — the per-column ordering clause, REPORTED now
16249 /// rather than discarded.
16250 ///
16251 /// SPG's index does not scan in a direction — column ordering is
16252 /// intrinsic to the storage — but `pg_indexes.indexdef` is a
16253 /// reproduction of the DDL, and dropping the clause meant
16254 /// `CREATE INDEX i ON t (a DESC NULLS LAST)` read back as `(a)`. A
16255 /// dump lost it, and a schema diff saw drift on every run.
16256 fn consume_optional_index_column_qualifiers(&mut self) -> crate::ast::IndexColumnOrder {
16257 let mut order = crate::ast::IndexColumnOrder::default();
16258 loop {
16259 match self.peek() {
16260 Token::Asc => {
16261 self.advance();
16262 }
16263 Token::Desc => {
16264 order.descending = true;
16265 self.advance();
16266 }
16267 Token::Ident(s) if s.eq_ignore_ascii_case("nulls") => {
16268 let look = self.tokens.get(self.pos + 1);
16269 if matches!(
16270 look,
16271 Some(Token::Ident(k)) if k.eq_ignore_ascii_case("first")
16272 || k.eq_ignore_ascii_case("last")
16273 ) {
16274 self.advance();
16275 order.nulls_first = Some(matches!(
16276 self.advance(),
16277 Token::Ident(k) if k.eq_ignore_ascii_case("first")
16278 ));
16279 } else {
16280 break;
16281 }
16282 }
16283 _ => break,
16284 }
16285 }
16286 order
16287 }
16288
16289 fn parse_create_index_stmt_after_create(
16290 &mut self,
16291 is_unique: bool,
16292 ) -> Result<Statement, ParseError> {
16293 // Caller consumed CREATE (and the optional UNIQUE); we're on INDEX.
16294 debug_assert!(matches!(self.peek(), Token::Index));
16295 self.advance();
16296 // v7.37.17 (17.6 partial) — CONCURRENTLY noise word (PG 8.2+).
16297 // SPG's CREATE INDEX is synchronous end-to-end today (real
16298 // CONCURRENTLY variant with restartable scans queues with
16299 // v7.39 indexes epic), so the modifier has no runtime effect
16300 // — same accept-and-no-op shape as v7.37.16.5 DETACH
16301 // PARTITION CONCURRENTLY and v7.37.19.8 REFRESH MATERIALIZED
16302 // VIEW CONCURRENTLY.
16303 let mut concurrently = false;
16304 if matches!(
16305 self.peek(),
16306 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("concurrently")
16307 ) {
16308 self.advance();
16309 concurrently = true;
16310 }
16311 let if_not_exists = self.consume_if_not_exists();
16312 // v7.39 (read01 round 93) — the index name is optional (PG since
16313 // forever): `CREATE INDEX ON t (a)` lets the server pick a name.
16314 // When the token after `[IF NOT EXISTS]` is already `ON`, no name
16315 // was given; leave it empty and the engine derives a PG-style
16316 // `<table>_<cols>_idx` name at CREATE time (with collision counter).
16317 let name = if matches!(self.peek(), Token::On) {
16318 String::new()
16319 } else {
16320 self.expect_ident_like()?
16321 };
16322 if !matches!(self.peek(), Token::On) {
16323 return Err(self.err(format!(
16324 "expected ON after CREATE INDEX <name>, got {:?}",
16325 self.peek()
16326 )));
16327 }
16328 self.advance();
16329 let table = self.expect_ident_like()?;
16330 // Optional `USING <method>` — only recognised method in v2.0 is
16331 // `hnsw` (a single-layer NSW graph for kNN). `USING` is the bare
16332 // ident `using` (we don't promote it to a reserved keyword
16333 // because it isn't reserved anywhere else in our SQL surface).
16334 let mut method_name: Option<String> = None;
16335 let method = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
16336 self.advance();
16337 let m = self.expect_ident_like()?;
16338 method_name = Some(m.to_ascii_lowercase());
16339 match m.to_ascii_lowercase().as_str() {
16340 "hnsw" => IndexMethod::Hnsw,
16341 "btree" => IndexMethod::BTree,
16342 "brin" => IndexMethod::Brin,
16343 // v7.12.3 — real GIN inverted index over `tsvector`.
16344 // v7.9.26b's `USING gin` → BTree silent fallback is
16345 // gone; the engine validates that the indexed column
16346 // is `tsvector` at CREATE INDEX time.
16347 "gin" => IndexMethod::Gin,
16348 // v7.9.26b — PG `pg_dump` emits `USING gist` /
16349 // `USING spgist` / `USING hash` for their built-in
16350 // AMs that SPG doesn't have a matching
16351 // implementation for; degrade to BTree on the
16352 // leading column so the schema loads + the index
16353 // catalogue stays consistent. Operator pays the
16354 // planner cost only for the queries that would have
16355 // used the specialised AM.
16356 "gist" | "spgist" | "hash" => IndexMethod::BTree,
16357 // v7.11.3 — pgvector ships both `ivfflat` and
16358 // `hnsw`. Customers shouldn't have to choose
16359 // their on-disk index method based on what SPG
16360 // implements; accept `ivfflat` as a synonym for
16361 // `hnsw` so PG schemas using either method drop
16362 // in. The vector distance op (`<->` / `<#>` /
16363 // `<=>`) at query time still picks the metric.
16364 "ivfflat" => IndexMethod::Hnsw,
16365 other => {
16366 return Err(self.err(alloc::format!(
16367 "unknown index method {other:?}; supported: hnsw, btree, brin, gin (gist/spgist/hash accepted as BTree fallback)"
16368 )));
16369 }
16370 }
16371 } else {
16372 IndexMethod::BTree
16373 };
16374 if !matches!(self.peek(), Token::LParen) {
16375 return Err(self.err(format!(
16376 "expected '(' before indexed column, got {:?}",
16377 self.peek()
16378 )));
16379 }
16380 self.advance();
16381 // v6.8.2 — accept either a bare column ident (legacy) or
16382 // an expression `fn(col, …)` for expression indexes.
16383 // Distinguish by peeking the token *after* the current
16384 // ident: `ident )` is the legacy column-only path;
16385 // anything else triggers the Pratt expression parser.
16386 // (`advance()` uses `mem::replace` to nil out the current
16387 // slot, so we can't save+rewind cleanly — peek-ahead via
16388 // direct index avoids the mutation.)
16389 let mut opclass: Option<String> = None;
16390 let mut key_collation: Option<String> = None;
16391 let (column, expression): (String, Option<Expr>) = match self.peek().clone() {
16392 // Single column with `)` immediately after — fast path.
16393 // v7.9.29 — also: bare column followed by `,` (the
16394 // multi-column form `(a, b, c)`). Without this branch
16395 // the leading ident gets pulled into `parse_expr`
16396 // which then sets `expression = Some(Column(a))` and
16397 // breaks Display round-trip on the multi-column shape.
16398 Token::Ident(s) | Token::QuotedIdent(s)
16399 if matches!(
16400 self.tokens.get(self.pos + 1),
16401 Some(Token::RParen | Token::Comma)
16402 ) =>
16403 {
16404 self.advance();
16405 (s, None)
16406 }
16407 // v7.9.22 — single column followed by a pgvector
16408 // opclass ident: `(col vector_cosine_ops)`. mailrs G5.
16409 // v7.15.0 — capture the opclass instead of discarding
16410 // it so the engine can dispatch (e.g. `gin_trgm_ops`
16411 // → real trigram-shingle GIN over a TEXT column).
16412 // Vector/HNSW opclasses still take their distance
16413 // metric from the query operator (`<->` / `<#>` /
16414 // `<=>`), so for those callers the opclass stays
16415 // informational.
16416 // v7.22 (mailrs round-13 gap 7) — pg_dump qualifies the
16417 // opclass: `(embedding public.vector_cosine_ops)`. Strip
16418 // the schema and dispatch on the bare opclass, the same
16419 // treatment table/type names get.
16420 Token::Ident(s) | Token::QuotedIdent(s)
16421 if matches!(
16422 self.tokens.get(self.pos + 1),
16423 Some(Token::Ident(_) | Token::QuotedIdent(_))
16424 ) && matches!(self.tokens.get(self.pos + 2), Some(Token::Dot))
16425 && matches!(
16426 self.tokens.get(self.pos + 3),
16427 Some(Token::Ident(op) | Token::QuotedIdent(op))
16428 if is_vector_opclass_name(op)
16429 ) =>
16430 {
16431 self.advance(); // column name
16432 self.advance(); // schema qualifier
16433 self.advance(); // dot
16434 let op_tok = self.advance();
16435 if let Token::Ident(op) | Token::QuotedIdent(op) = op_tok {
16436 opclass = Some(op.to_ascii_lowercase());
16437 }
16438 (s, None)
16439 }
16440 // r1038 — an operator class is recognised by its POSITION, not
16441 // by a list of names. It used to be `is_vector_opclass_name`,
16442 // so `USING gin (doc jsonb_path_ops)` — ordinary PG, and what
16443 // sentori's migration wrote — was a syntax error while
16444 // `USING gin (doc)` parsed. Anything sitting between a column
16445 // name and a `,` `)` ASC DESC NULLS COLLATE is an opclass;
16446 // two bare identifiers in a row are not valid there otherwise.
16447 Token::Ident(s) | Token::QuotedIdent(s)
16448 if matches!(
16449 self.tokens.get(self.pos + 1),
16450 Some(Token::Ident(op) | Token::QuotedIdent(op))
16451 if is_vector_opclass_name(op) || Self::opclass_position_follows(
16452 self.tokens.get(self.pos + 2)
16453 )
16454 ) =>
16455 {
16456 self.advance(); // column name
16457 // Capture the opclass token, lower-cased for
16458 // case-insensitive engine dispatch.
16459 let op_tok = self.advance();
16460 if let Token::Ident(op) | Token::QuotedIdent(op) = op_tok {
16461 opclass = Some(op.to_ascii_lowercase());
16462 }
16463 (s, None)
16464 }
16465 Token::Ident(_) | Token::QuotedIdent(_) => {
16466 // v7.39 (round 538) — an explicit COLLATE on the key,
16467 // read by LOOKAHEAD because `parse_expr` absorbs the
16468 // clause as a no-op (SPG orders text by bytes, which is
16469 // the C collation, so it changes nothing to honour). PG
16470 // still PRINTS it: an explicitly written `"C"` and the
16471 // collation a column inherits are different collation
16472 // OBJECTS even where they sort identically, which is why
16473 // `(a COLLATE "C")` shows on a C-collation database too.
16474 if matches!(
16475 self.tokens.get(self.pos + 1),
16476 Some(Token::Ident(w)) if w.eq_ignore_ascii_case("collate")
16477 ) {
16478 key_collation = match self.tokens.get(self.pos + 2) {
16479 Some(Token::Ident(n) | Token::QuotedIdent(n) | Token::String(n)) => {
16480 Some(n.clone())
16481 }
16482 _ => None,
16483 };
16484 }
16485 // v7.39.2 — the clause is read by the LOOKAHEAD above and
16486 // belongs to the KEY, not to the expression. Since
16487 // `COLLATE` became a node, letting `parse_expr` build one
16488 // here put the collation in twice and the key deparsed as
16489 // `(c COLLATE "C" COLLATE "C")`. The ORDER-BY-key channel
16490 // is the same idea and already exists, so this borrows it:
16491 // absorb into the side channel, and the key's own
16492 // lookahead is what carries it.
16493 // v7.39.2 — and the key can only CARRY the byte-order
16494 // spellings. Absorbing into the side channel accepts any
16495 // name, so suppressing the node here without this check
16496 // silently accepted `(name COLLATE "en_US")`, which SPG's
16497 // index cannot honour — a refusal that was doing real
16498 // work, removed by the suppression and put back here.
16499 if let Some(name) = &key_collation {
16500 let lc = name.to_ascii_lowercase();
16501 let byte_order = matches!(
16502 lc.as_str(),
16503 "c" | "posix" | "default" | "ucs_basic" | "pg_c_utf8"
16504 );
16505 let mysql_ok = self.mysql_dialect
16506 && (lc.ends_with("_ci")
16507 || lc.ends_with("_bin")
16508 || lc == "binary"
16509 || matches!(lc.as_str(), "case_insensitive" | "nocase"));
16510 if !byte_order && !mysql_ok {
16511 return Err(self.err(alloc::format!(
16512 "COLLATE {name:?} is not supported in this position: an index \
16513 key carries the byte-order spellings only. Declare it on the \
16514 column (`x text COLLATE {name:?}`) instead"
16515 )));
16516 }
16517 }
16518 let saved_key_ctx = self.in_order_by_key;
16519 self.in_order_by_key = true;
16520 let key_expr = self.parse_expr(0);
16521 self.in_order_by_key = saved_key_ctx;
16522 let key_expr = key_expr?;
16523 let primary = extract_first_column(&key_expr).ok_or_else(|| {
16524 self.err("expression index key must reference at least one column".into())
16525 })?;
16526 (primary, Some(key_expr))
16527 }
16528 // v7.37.43-T4 — parenthesised expression index key
16529 // `CREATE INDEX … ON t ((payload->'bundle'->>'id'))`.
16530 // PG's CREATE INDEX requires the expression to be in
16531 // its own parens to disambiguate function calls from
16532 // column lists, so this `LParen` is the inner open-paren
16533 // of an expression key. parse_expr handles the recursive
16534 // descent and consumes the matching `RParen`.
16535 Token::LParen => {
16536 let key_expr = self.parse_expr(0)?;
16537 let primary = extract_first_column(&key_expr).ok_or_else(|| {
16538 self.err("expression index key must reference at least one column".into())
16539 })?;
16540 (primary, Some(key_expr))
16541 }
16542 other => {
16543 return Err(self.err(format!(
16544 "expected column ident or expression, got {other:?}"
16545 )));
16546 }
16547 };
16548 // v7.9.14 — accept extra comma-separated columns inside
16549 // the index key parens (`CREATE INDEX … (a, b, c)`).
16550 // mailrs F2. Each extra column may carry an optional
16551 // `ASC` / `DESC` / `NULLS FIRST` / `NULLS LAST` clause
16552 // — parsed and discarded; SPG doesn't honour direction
16553 // on a BTree index today (column ordering is intrinsic
16554 // to the storage). v7.10 will widen to genuine composite
16555 // index keys.
16556 let mut extra_columns: Vec<String> = Vec::new();
16557 // The leading column may also have ASC/DESC after it — and that
16558 // one is the column SPG indexes, so its clause is kept.
16559 let key_order = self.consume_optional_index_column_qualifiers();
16560 while matches!(self.peek(), Token::Comma) {
16561 self.advance();
16562 let extra = self.expect_ident_like()?;
16563 let _ = self.consume_optional_index_column_qualifiers();
16564 extra_columns.push(extra);
16565 }
16566 if !matches!(self.peek(), Token::RParen) {
16567 return Err(self.err(format!(
16568 "expected ')' after indexed column / expression, got {:?}",
16569 self.peek()
16570 )));
16571 }
16572 self.advance();
16573 // v6.8.0 — optional `INCLUDE (col1, col2, …)` clause for
16574 // index-only-scan annotation. Bare ident (not a reserved
16575 // keyword) so we test by case-insensitive string match.
16576 let included_columns = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("include"))
16577 {
16578 self.advance();
16579 if !matches!(self.peek(), Token::LParen) {
16580 return Err(self.err(format!("expected '(' after INCLUDE, got {:?}", self.peek())));
16581 }
16582 self.advance();
16583 let mut cols = Vec::new();
16584 loop {
16585 cols.push(self.expect_ident_like()?);
16586 match self.peek() {
16587 Token::Comma => {
16588 self.advance();
16589 }
16590 Token::RParen => {
16591 self.advance();
16592 break;
16593 }
16594 other => {
16595 return Err(self.err(format!(
16596 "expected ',' or ')' in INCLUDE list, got {other:?}"
16597 )));
16598 }
16599 }
16600 }
16601 cols
16602 } else {
16603 Vec::new()
16604 };
16605 // v7.11.3 — accept and discard PG `WITH (k = v, ...)` index
16606 // storage parameters. pgvector emits `WITH (lists = N)` for
16607 // ivfflat and `WITH (m = N, ef_construction = M)` for hnsw;
16608 // SPG's HNSW picks its own parameters today (tunable via
16609 // env vars), so the WITH clause is informational and dropped.
16610 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with")) {
16611 self.advance();
16612 if !matches!(self.peek(), Token::LParen) {
16613 return Err(self.err(format!(
16614 "expected '(' after WITH in CREATE INDEX, got {:?}",
16615 self.peek()
16616 )));
16617 }
16618 self.advance();
16619 loop {
16620 if matches!(self.peek(), Token::RParen) {
16621 self.advance();
16622 break;
16623 }
16624 // Drain `key = value` or bare `key` tokens.
16625 let _ = self.advance(); // key
16626 if matches!(self.peek(), Token::Eq) {
16627 self.advance();
16628 let _ = self.advance(); // value (int / string / ident)
16629 }
16630 match self.peek() {
16631 Token::Comma => {
16632 self.advance();
16633 }
16634 Token::RParen => {
16635 self.advance();
16636 break;
16637 }
16638 other => {
16639 return Err(self.err(format!(
16640 "expected ',' or ')' in WITH (…) clause, got {other:?}"
16641 )));
16642 }
16643 }
16644 }
16645 }
16646 // v7.39 (read01 round 52) — optional `NULLS [NOT] DISTINCT` (PG 15+),
16647 // which sits between the key list and the WHERE clause.
16648 let mut nulls_not_distinct = false;
16649 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("nulls")) {
16650 let n1 = self.tokens.get(self.pos + 1);
16651 let n2 = self.tokens.get(self.pos + 2);
16652 if matches!(n1, Some(Token::Not)) && matches!(n2, Some(Token::Distinct)) {
16653 self.advance(); // NULLS
16654 self.advance(); // NOT
16655 self.advance(); // DISTINCT
16656 nulls_not_distinct = true;
16657 } else if matches!(n1, Some(Token::Distinct)) {
16658 self.advance(); // NULLS
16659 self.advance(); // DISTINCT
16660 }
16661 }
16662 // v6.8.1 — optional `WHERE <expr>` partial-index predicate.
16663 let partial_predicate = if matches!(self.peek(), Token::Where) {
16664 self.advance();
16665 Some(self.parse_expr(0)?)
16666 } else {
16667 None
16668 };
16669 // v7.9.29 — UNIQUE on a vector index (HNSW) makes no
16670 // sense: uniqueness over an ANN structure has no clean
16671 // semantics. Reject early. (BRIN UNIQUE is similarly
16672 // meaningless — block both.)
16673 if is_unique && !matches!(method, IndexMethod::BTree) {
16674 return Err(self.err(alloc::format!(
16675 "UNIQUE is only supported on BTree indexes, got USING {:?}",
16676 method
16677 )));
16678 }
16679 Ok(Statement::CreateIndex(CreateIndexStatement {
16680 concurrently,
16681 name,
16682 key_order,
16683 key_collation,
16684 table,
16685 column,
16686 nulls_not_distinct,
16687 method,
16688 if_not_exists,
16689 included_columns,
16690 partial_predicate,
16691 extra_columns: extra_columns.clone(),
16692 expression,
16693 is_unique,
16694 opclass,
16695 method_name,
16696 }))
16697 }
16698
16699 /// v7.6.0 — wraps `parse_column_def` and consumes an optional
16700 /// column-level `REFERENCES ...` clause. The trailing FK is
16701 /// normalised into table-level shape (single-element columns +
16702 /// parent_columns) so the engine sees one uniform constraint list.
16703 fn parse_column_def_with_fk(
16704 &mut self,
16705 ) -> Result<(ColumnDef, Option<ForeignKeyConstraint>), ParseError> {
16706 let col = self.parse_column_def()?;
16707 // v7.39 (round 308, V29) — an explicitly named inline FK:
16708 // `col INT CONSTRAINT fk_a REFERENCES tbl(pcol)`. The column-def
16709 // loop leaves this spelling intact precisely so the name can be
16710 // kept here; PG reports it in violation messages and matches it
16711 // in `SET CONSTRAINTS`.
16712 let declared_name = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint"))
16713 {
16714 self.advance();
16715 Some(self.expect_ident_like()?)
16716 } else {
16717 None
16718 };
16719 // Inline form: `col INT REFERENCES tbl(pcol) [ON DELETE ...] [ON UPDATE ...]`.
16720 let inline_references = matches!(
16721 self.peek(),
16722 Token::Ident(s) if s.eq_ignore_ascii_case("references")
16723 );
16724 if !inline_references {
16725 return Ok((col, None));
16726 }
16727 let (
16728 parent_table,
16729 parent_columns,
16730 on_delete,
16731 on_update,
16732 match_type,
16733 deferrable,
16734 initially_deferred,
16735 ) = self.parse_references_tail(1)?;
16736 let fk = ForeignKeyConstraint {
16737 name: declared_name,
16738 columns: vec![col.name.clone()],
16739 parent_table,
16740 parent_columns,
16741 on_delete,
16742 on_update,
16743 match_type,
16744 deferrable,
16745 initially_deferred,
16746 };
16747 Ok((col, Some(fk)))
16748 }
16749
16750 /// v7.13.0 — parse a column type (consuming the type ident and
16751 /// any trailing parameters / `[]`), without surrounding column
16752 /// constraints. Used by ALTER COLUMN TYPE (mailrs round-5 G8).
16753 /// Returns the resolved `ColumnTypeName` plus implied
16754 /// `(auto_increment, not_null)` flags from PG SERIAL family
16755 /// shorthands — callers that don't expect those (ALTER COLUMN
16756 /// TYPE) can discard them.
16757 fn parse_column_type_name(&mut self) -> Result<ColumnTypeName, ParseError> {
16758 let (ty, _, _, _, _, _, _, _, _, _, _, _, _, _) = self.parse_type_with_implied_flags()?;
16759 Ok(ty)
16760 }
16761
16762 #[allow(clippy::type_complexity)]
16763 fn parse_type_with_implied_flags(
16764 &mut self,
16765 ) -> Result<
16766 (
16767 ColumnTypeName,
16768 bool,
16769 bool,
16770 Option<String>,
16771 Collation,
16772 // v7.39 (round 370, M4 P4a) — was `COLLATE` written explicitly?
16773 bool,
16774 // v7.39 (round 676) — the collation NAME as written, which the
16775 // `Collation` enum above cannot carry.
16776 Option<String>,
16777 bool,
16778 // v7.17.0 Phase 3.P0-36 — MySQL inline ENUM variant
16779 // list captured at type-parse time. None for all
16780 // non-ENUM types.
16781 Option<Vec<String>>,
16782 // v7.17.0 Phase 3.P0-37 — MySQL inline SET variant
16783 // list. Distinct from ENUM (subset semantics).
16784 Option<Vec<String>>,
16785 // v7.39 (round 386, epic P1) — declared TINYINT / MEDIUMINT
16786 // width, lost when the type collapses to SmallInt / Int.
16787 Option<MysqlIntWidth>,
16788 // v7.39 (round 424) — declared fractional-seconds precision of a
16789 // MySQL temporal column (bare spelling = 0). None under PG.
16790 Option<u8>,
16791 // v7.39.2 — written `TIMESTAMP` rather than `DATETIME`. The
16792 // two are different types on MySQL and SPG stores both as
16793 // `Timestamp`, so the spelling has to travel separately or
16794 // a dump silently rewrites the column.
16795 bool,
16796 // v7.39.3 — a MySQL `FLOAT(m,d)` / `DOUBLE(m,d)` pair. Not a
16797 // display hint: it rounds on write.
16798 Option<(u8, u8)>,
16799 ),
16800 ParseError,
16801 > {
16802 let mut ty_ident = match self.advance() {
16803 Token::Ident(s) => s,
16804 // v7.37.5 β-P2 — `INTERVAL` lexes as a reserved keyword
16805 // (Token::Interval) since v7.9.25 to drive the `INTERVAL
16806 // '<span>'` literal grammar. As a column type it lands
16807 // here directly; downstream resolution still uses the
16808 // canonical lowercase string.
16809 Token::Interval => "interval".to_string(),
16810 other => {
16811 return Err(ParseError {
16812 message: format!("expected column type, got {other:?}"),
16813 token_pos: self.consumed_pos(),
16814 });
16815 }
16816 };
16817 // v7.22 (mailrs round-13 gap 4) — schema-qualified type names:
16818 // pg_dump qualifies extension types (`public.vector(1024)`).
16819 // SPG is single-namespace; drop the schema and resolve the
16820 // bare type — same treatment table names already get.
16821 while matches!(self.peek(), Token::Dot) {
16822 self.advance();
16823 ty_ident = self.expect_ident_like()?;
16824 }
16825 let mut implied_auto_increment = false;
16826 let mut implied_not_null = false;
16827 let mut user_type_ref: Option<String> = None;
16828 // v7.17.0 Phase 3.P0-36 — MySQL inline ENUM('a','b','c')
16829 // value list, captured here and bubbled up through the
16830 // ColumnDef so the engine can attach it to the column
16831 // schema (and validate INSERT cells against it).
16832 let mut inline_enum_variants: Option<Vec<String>> = None;
16833 // v7.17.0 Phase 3.P0-37 — MySQL inline SET variant list.
16834 let mut inline_set_variants: Option<Vec<String>> = None;
16835 // v7.39 (round 386, type-fidelity epic P1) — the declared MySQL
16836 // narrow-int width (TINYINT / MEDIUMINT), captured before the type
16837 // collapses to SmallInt / Int. Only under the MySQL dialect.
16838 let mut mysql_int_width: Option<MysqlIntWidth> = None;
16839 // v7.39 (round 424) — the declared fractional-seconds precision of a
16840 // MySQL temporal column. Set by the temporal arms below; stays None
16841 // for PG (whose temporal columns keep full microseconds).
16842 let mut mysql_fsp: Option<u8> = None;
16843 let mut mysql_declared_timestamp = false;
16844 let mut mysql_float_md: Option<(u8, u8)> = None;
16845 let mut ty = match ty_ident.as_str() {
16846 // PG SERIAL family. Implies NOT NULL + AUTO_INCREMENT.
16847 "smallserial" | "serial2" => {
16848 implied_auto_increment = true;
16849 implied_not_null = true;
16850 ColumnTypeName::SmallInt
16851 }
16852 "serial" | "serial4" => {
16853 implied_auto_increment = true;
16854 implied_not_null = true;
16855 ColumnTypeName::Int
16856 }
16857 "bigserial" | "serial8" => {
16858 implied_auto_increment = true;
16859 implied_not_null = true;
16860 ColumnTypeName::BigInt
16861 }
16862 // MySQL flavours we accept by aliasing to the closest SPG
16863 // type. TINYINT covers MySQL's i8 — held inside SMALLINT
16864 // since SPG doesn't have a dedicated i8. MEDIUMINT (MySQL
16865 // 24-bit) → INT. UNSIGNED modifiers are consumed below
16866 // without semantic effect.
16867 // v7.38 (read01 P4.19-sibling) — `int2` / `int4` / `int8` are
16868 // PG's internal type names; pg_dump and hand-written PG schemas
16869 // use them interchangeably with smallint / int / bigint (the cast
16870 // path already accepted them, only the column grammar didn't).
16871 "smallint" | "int2" => {
16872 // v7.14.0 — MySQL display-width on integers
16873 // (`SMALLINT(5)`, `INT(11)`, `BIGINT(20)`). The
16874 // parenthesised number is purely cosmetic — it
16875 // doesn't change storage. Accept + discard.
16876 self.consume_optional_paren_size();
16877 ColumnTypeName::SmallInt
16878 }
16879 // v7.17.0 Phase 4.3 — MySQL `TINYINT(1)` is the
16880 // canonical encoding for BOOLEAN. Every MySQL driver
16881 // (JDBC `tinyInt1isBit=true`, PHP `mysql_field_type`,
16882 // .NET `MySqlConnection`, sqlx) maps it to bit. Pre-
16883 // 4.3 SPG classified TINYINT(1) as SmallInt, which
16884 // gave the customer i16-shaped values where the app
16885 // expected bool — a Tier-A silent type drift on
16886 // mysqldump restores. Now: `TINYINT(1)` → Bool;
16887 // `TINYINT` (no width) and `TINYINT(N)` for N ≠ 1
16888 // stay SmallInt (the legacy width-agnostic path).
16889 "tinyint" => {
16890 let width = self.peek_optional_paren_size_value();
16891 self.consume_optional_paren_size();
16892 if width == Some(1) {
16893 ColumnTypeName::Bool
16894 } else {
16895 // v7.39 (round 386, epic P1) — TINYINT is i8; record the
16896 // lost width so the write path can enforce -128..127.
16897 if self.mysql_dialect {
16898 mysql_int_width = Some(MysqlIntWidth::Tiny);
16899 }
16900 ColumnTypeName::SmallInt
16901 }
16902 }
16903 "mediumint" => {
16904 self.consume_optional_paren_size();
16905 // v7.39 (round 386, epic P1) — MEDIUMINT is 24-bit; record it.
16906 if self.mysql_dialect {
16907 mysql_int_width = Some(MysqlIntWidth::Medium);
16908 }
16909 ColumnTypeName::Int
16910 }
16911 "int" | "integer" | "int4" => {
16912 self.consume_optional_paren_size();
16913 ColumnTypeName::Int
16914 }
16915 "bigint" | "int8" => {
16916 self.consume_optional_paren_size();
16917 ColumnTypeName::BigInt
16918 }
16919 // v7.13.0 — `DOUBLE PRECISION` (PG canonical spelling)
16920 // (mailrs round-5 G6). Consume the optional `PRECISION`
16921 // tail when the type keyword was `double` / `DOUBLE`.
16922 //
16923 // v7.39 (round 269) — REAL is 32-bit, not "the same as our
16924 // FLOAT". `FLOAT(p)` picks the width the way PG does:
16925 // p in 1..=24 is real, 25..=53 is double precision, and
16926 // anything else is an error.
16927 "float" | "double" | "real" => {
16928 if ty_ident.eq_ignore_ascii_case("double")
16929 && matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("precision"))
16930 {
16931 self.advance();
16932 }
16933 if ty_ident.eq_ignore_ascii_case("real") {
16934 // v7.39 (round 274) — the two dialects genuinely
16935 // disagree: PG's REAL is 4-byte, MySQL's REAL is a
16936 // synonym for DOUBLE (8-byte). Round 269 made REAL
16937 // 32-bit globally and thereby narrowed the stored
16938 // precision of every MySQL REAL column.
16939 if self.mysql_dialect {
16940 ColumnTypeName::Float
16941 } else {
16942 ColumnTypeName::Real
16943 }
16944 } else if self.mysql_dialect
16945 && matches!(self.peek(), Token::LParen)
16946 && self.peek_paren_has_comma()
16947 {
16948 // v7.39 (round 360) — MySQL's `FLOAT(m,d)` / `DOUBLE(m,d)`
16949 // display form (`FLOAT(10,2)`), which PG has no
16950 // equivalent of. It was `syntax error at or near ","`,
16951 // so the whole CREATE failed.
16952 //
16953 // v7.39.2 — the guard said `float` while the comment
16954 // said both, so `DOUBLE(10,2)` — which every legacy
16955 // MySQL schema uses for money — still failed the
16956 // whole CREATE with `syntax error at or near "("`.
16957 // Measured on 9.7.2: both forms are accepted, and the
16958 // digits are NOT a display hint, they round on write
16959 // (3.14159265358979 into either stores 3.14). The
16960 // rounding is recorded as a residual; accepting the
16961 // syntax and keeping the width is the half this
16962 // change makes.
16963 // v7.39.3 — keep the pair. The digits are not a
16964 // display hint: MySQL 9.7.2 ROUNDS on write and
16965 // refuses a value wider than `m` (errno 1264), so a
16966 // column declared for money held more precision here
16967 // than its schema said.
16968 let (m, d) = self.parse_optional_numeric_params()?;
16969 mysql_float_md = Some((
16970 u8::try_from(m).unwrap_or(u8::MAX),
16971 u8::try_from(d.max(0)).unwrap_or(u8::MAX),
16972 ));
16973 if ty_ident.eq_ignore_ascii_case("float") {
16974 ColumnTypeName::Real
16975 } else {
16976 ColumnTypeName::Float
16977 }
16978 } else if ty_ident.eq_ignore_ascii_case("float")
16979 && matches!(self.peek(), Token::LParen)
16980 {
16981 // PG words the two bounds differently, and
16982 // parse_paren_size already rejects a zero.
16983 let p = self.parse_paren_size("FLOAT")?;
16984 if p > 53 {
16985 return Err(self.err(String::from(
16986 "precision for type float must be less than 54 bits",
16987 )));
16988 }
16989 if p <= 24 {
16990 ColumnTypeName::Real
16991 } else {
16992 ColumnTypeName::Float
16993 }
16994 } else if ty_ident.eq_ignore_ascii_case("float") && self.mysql_dialect {
16995 // v7.39.2 — MySQL's bare FLOAT is FOUR bytes; PG's is
16996 // eight (it is `float8`'s spelling there). SPG used
16997 // PG's for both, so a MySQL FLOAT column silently
16998 // kept more precision than MySQL does — measured,
16999 // 3.14159265358979 comes back as 3.14159 there and
17000 // came back whole here — and reported itself as
17001 // `double` to every reflection.
17002 //
17003 // This is the mirror of the REAL split above: the
17004 // two dialects disagree about which spelling means
17005 // which width, and one of them was already honoured.
17006 ColumnTypeName::Real
17007 } else {
17008 ColumnTypeName::Float
17009 }
17010 }
17011 // v7.13.0 — `FLOAT8` (PG short form) maps the same as FLOAT.
17012 "float4" => ColumnTypeName::Real,
17013 "float8" => ColumnTypeName::Float,
17014 "text" => ColumnTypeName::Text,
17015 // v7.39 (round 360) — MySQL's sized TEXT and BLOB families.
17016 // `LONGTEXT`, `BLOB` and `VARBINARY` appear in nearly every
17017 // real MySQL schema and NONE of them existed: the CREATE
17018 // failed outright with `type "blob" does not exist`, so the
17019 // table was never made. The sizes differ only in MySQL's
17020 // maximum length, which SPG does not cap, so they collapse
17021 // onto TEXT and BYTEA the way the unsized spellings do.
17022 "tinytext" | "mediumtext" | "longtext" => ColumnTypeName::Text,
17023 "blob" | "tinyblob" | "mediumblob" | "longblob" => ColumnTypeName::Bytes,
17024 // `VARBINARY(n)` / `BINARY(n)` — a length that SPG does not
17025 // enforce, consumed so the declaration parses.
17026 "varbinary" | "binary" => {
17027 self.consume_optional_paren_size();
17028 ColumnTypeName::Bytes
17029 }
17030 "name" => ColumnTypeName::Name,
17031 "xid" => ColumnTypeName::Xid,
17032 "oid" => ColumnTypeName::Oid,
17033 "xid8" => ColumnTypeName::Xid8,
17034 "bool" | "boolean" => ColumnTypeName::Bool,
17035 // v7.39 (round 620) — an UNBOUNDED `varchar` is the same type as
17036 // an unbounded `character varying`, which the arm below has always
17037 // read as text. Only the short spelling demanded a length, so
17038 // `CREATE TABLE t (x VARCHAR)` — as ordinary a line of DDL as
17039 // there is — failed on `VARCHAR type requires (N)` while the long
17040 // spelling of the same thing was accepted. The same asymmetry
17041 // round 613 closed on the CAST side, here on the DDL side.
17042 "varchar" => {
17043 if matches!(self.peek(), Token::LParen) {
17044 ColumnTypeName::Varchar(self.parse_paren_size("VARCHAR")?)
17045 } else {
17046 ColumnTypeName::Text
17047 }
17048 }
17049 // v7.39 (bpchar epic) — bare `char` = char(1), same as bare
17050 // `character` below (SQL standard).
17051 "char" => {
17052 if matches!(self.peek(), Token::LParen) {
17053 ColumnTypeName::Char(self.parse_paren_size("CHAR")?)
17054 } else {
17055 ColumnTypeName::Char(1)
17056 }
17057 }
17058 // pg_dump's canonical spellings: `character varying(n)` = varchar,
17059 // `character(n)` = char, bare `character` = char(1). Unbounded
17060 // `character varying` maps to text.
17061 "character" => {
17062 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("varying")) {
17063 self.advance();
17064 if matches!(self.peek(), Token::LParen) {
17065 ColumnTypeName::Varchar(self.parse_paren_size("VARCHAR")?)
17066 } else {
17067 ColumnTypeName::Text
17068 }
17069 } else if matches!(self.peek(), Token::LParen) {
17070 ColumnTypeName::Char(self.parse_paren_size("CHAR")?)
17071 } else {
17072 ColumnTypeName::Char(1)
17073 }
17074 }
17075 "vector" => {
17076 let dim = self.parse_paren_size("VECTOR")?;
17077 let encoding = self.parse_optional_vector_encoding()?;
17078 ColumnTypeName::Vector { dim, encoding }
17079 }
17080 // v7.39 (round 345, M5) — `DECIMAL` and `DEC` are the SQL
17081 // standard's own spellings of NUMERIC, and PG 18.4 accepts both
17082 // (measured: `DECIMAL(10,2)` and `DEC(5,1)` both report as
17083 // `numeric`). Only `NUMERIC` parsed, so `CREATE TABLE t (a
17084 // DECIMAL(10,2))` — how nearly every money column is written,
17085 // in either dialect — was a syntax error and the table was
17086 // never created. `FIXED` is MySQL's alias alone, so it is
17087 // taken only in that dialect.
17088 "numeric" | "decimal" | "dec" => {
17089 let (precision, scale) = self.parse_optional_numeric_params()?;
17090 ColumnTypeName::Numeric(precision, scale)
17091 }
17092 "fixed" if self.mysql_dialect => {
17093 let (precision, scale) = self.parse_optional_numeric_params()?;
17094 ColumnTypeName::Numeric(precision, scale)
17095 }
17096 "date" => ColumnTypeName::Date,
17097 // MySQL's `DATETIME` is the same domain as standard
17098 // `TIMESTAMP` — accept both spellings.
17099 "timestamp" | "datetime" => {
17100 // pg_dump emits `TIMESTAMP(6) WITH TIME ZONE` — the optional
17101 // fractional-seconds precision comes BEFORE the `WITH/WITHOUT
17102 // TIME ZONE` clause, so consume it first.
17103 // v7.39 (round 424) — under MySQL the precision is SEMANTIC
17104 // (it truncates on write and pads on render), so capture it;
17105 // a bare spelling means precision 0 there. PG stores µs always
17106 // and keeps `None`.
17107 let n = self.take_optional_paren_size();
17108 if self.mysql_dialect {
17109 mysql_fsp = Some(n.unwrap_or(0).min(6));
17110 // v7.39.2 — remember WHICH spelling was written.
17111 // MySQL and MariaDB keep `timestamp` and `datetime`
17112 // apart everywhere a client can read the type back,
17113 // and SPG reported `datetime` for both — so a dump
17114 // and reload silently changed the column's declared
17115 // type, and MySQL's TIMESTAMP is not DATETIME (a
17116 // different range, and UTC conversion on the way in
17117 // and out).
17118 mysql_declared_timestamp = ty_ident.eq_ignore_ascii_case("timestamp");
17119 }
17120 // v7.14.0 — PG canonical `TIMESTAMP WITH TIME ZONE`
17121 // / `TIMESTAMP WITHOUT TIME ZONE`. pg_dump emits
17122 // the full form. SPG canonicalises:
17123 // - WITH TIME ZONE → Timestamptz
17124 // - WITHOUT TIME ZONE → Timestamp
17125 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with"))
17126 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("time"))
17127 && matches!(self.tokens.get(self.pos + 2), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("zone"))
17128 {
17129 self.advance(); // WITH
17130 self.advance(); // TIME
17131 self.advance(); // ZONE
17132 ColumnTypeName::Timestamptz
17133 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("without"))
17134 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("time"))
17135 && matches!(self.tokens.get(self.pos + 2), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("zone"))
17136 {
17137 self.advance(); // WITHOUT
17138 self.advance(); // TIME
17139 self.advance(); // ZONE
17140 ColumnTypeName::Timestamp
17141 } else {
17142 // A second `(precision)` cannot legally follow, but the
17143 // old grammar tolerated it; keep that tolerance.
17144 self.consume_optional_paren_size();
17145 ColumnTypeName::Timestamp
17146 }
17147 }
17148 // v7.9.2 — `TIMESTAMPTZ` and full PG spelling
17149 // `TIMESTAMP WITH TIME ZONE`. Same storage as TIMESTAMP;
17150 // only PG-wire OID differs.
17151 "timestamptz" => {
17152 self.consume_optional_paren_size();
17153 ColumnTypeName::Timestamptz
17154 }
17155 // v4.9: JSON / JSONB. Stored as raw text — no parse-time
17156 // validation. We accept the JSONB spelling too because
17157 // most PG clients default to it; SPG doesn't distinguish
17158 // the two (no path-operator perf advantage to model).
17159 "json" => ColumnTypeName::Json,
17160 "jsonb" => ColumnTypeName::Jsonb,
17161 // v7.10.4 — PG `BYTEA` and the SPG `BYTES` alias both
17162 // surface here. Same storage shape; mapping happens at
17163 // the engine side via the ColumnTypeName → DataType
17164 // resolver. Literal forms are handled at coerce_value
17165 // time so the lexer stays untouched.
17166 "bytea" | "bytes" => ColumnTypeName::Bytes,
17167 // v7.17.0 Phase 7 — PG network address types
17168 // v7.17.0 had a Text-backed fallback here for
17169 // `inet` / `cidr` / `macaddr`. v7.37.5 ζ-A promoted
17170 // each to a first-class type; the keywords are
17171 // bound below in the ζ-A block.
17172 // v7.12.0 — PG full-text search types. mailrs G-CRIT-3.
17173 // The actual `to_tsvector` / `@@` / `ts_rank` surface
17174 // arrives in v7.12.1+; the type itself loads here so
17175 // mailrs's `scripts/init-schema.sql` runs unmodified.
17176 "tsvector" => ColumnTypeName::TsVector,
17177 "tsquery" => ColumnTypeName::TsQuery,
17178 // v7.17.0 — PG `UUID`. Wire OID 2950. The drop-in PG
17179 // surface for Django / Rails / Hibernate's default
17180 // PK pattern.
17181 "uuid" => ColumnTypeName::Uuid,
17182 // v7.37.5 β-P2 — PG `INTERVAL` as a column type.
17183 // Storage = three-field {months, days, micros}, catalog
17184 // tag 34, FILE_VERSION 48+, wire OID 1186. Prior to this
17185 // line `INTERVAL` was parser-rejected at CREATE TABLE.
17186 "interval" => {
17187 // pg_dump emits field-qualified forms like `INTERVAL DAY TO
17188 // SECOND` and an optional `(p)` precision. SPG stores the full
17189 // {months,days,micros}; consume + ignore the qualifier/precision.
17190 while matches!(self.peek(), Token::To)
17191 || matches!(self.peek(), Token::Ident(s) if matches!(
17192 s.to_ascii_lowercase().as_str(),
17193 "year" | "month" | "day" | "hour" | "minute" | "second"
17194 ))
17195 {
17196 self.advance();
17197 }
17198 self.consume_optional_paren_size();
17199 ColumnTypeName::Interval
17200 }
17201 // v7.17.0 Phase 3.P0-32 — PG `TIME` (without time zone).
17202 // i64 microseconds since 00:00:00. Wire OID 1083.
17203 // pg_dump emits `TIME(6)` and `TIME(6) WITH TIME ZONE`.
17204 "time" => {
17205 // v7.39 (round 424) — MySQL TIME carries a semantic
17206 // fractional-seconds precision, bare meaning 0.
17207 let n = self.take_optional_paren_size();
17208 if self.mysql_dialect {
17209 mysql_fsp = Some(n.unwrap_or(0).min(6));
17210 }
17211 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with"))
17212 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("time"))
17213 && matches!(self.tokens.get(self.pos + 2), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("zone"))
17214 {
17215 self.advance();
17216 self.advance();
17217 self.advance();
17218 ColumnTypeName::TimeTz
17219 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("without"))
17220 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("time"))
17221 && matches!(self.tokens.get(self.pos + 2), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("zone"))
17222 {
17223 self.advance();
17224 self.advance();
17225 self.advance();
17226 ColumnTypeName::Time
17227 } else {
17228 ColumnTypeName::Time
17229 }
17230 }
17231 // v7.17.0 Phase 3.P0-33 — MySQL `YEAR`. u16 in
17232 // 1901..=2155 + zero-year sentinel 0. Wire = INT4.
17233 "year" => ColumnTypeName::Year,
17234 // v7.17.0 Phase 3.P0-34 — PG `TIMETZ` / `TIME WITH
17235 // TIME ZONE`. i64 us + i32 offset_secs. Wire OID 1266.
17236 "timetz" => ColumnTypeName::TimeTz,
17237 // v7.17.0 Phase 3.P0-35 — PG `MONEY` — i64 cents.
17238 // Wire OID 790.
17239 "money" => ColumnTypeName::Money,
17240 // v7.17.0 Phase 3.P0-38 — PG range types.
17241 "int4range" => ColumnTypeName::Range(RangeKindAst::Int4),
17242 "int8range" => ColumnTypeName::Range(RangeKindAst::Int8),
17243 "numrange" => ColumnTypeName::Range(RangeKindAst::Num),
17244 "tsrange" => ColumnTypeName::Range(RangeKindAst::Ts),
17245 "tstzrange" => ColumnTypeName::Range(RangeKindAst::TsTz),
17246 "daterange" => ColumnTypeName::Range(RangeKindAst::Date),
17247 // v7.37.5 δ — PG 14+ multirange keywords.
17248 "int4multirange" => ColumnTypeName::Multirange(RangeKindAst::Int4),
17249 "int8multirange" => ColumnTypeName::Multirange(RangeKindAst::Int8),
17250 "nummultirange" => ColumnTypeName::Multirange(RangeKindAst::Num),
17251 "tsmultirange" => ColumnTypeName::Multirange(RangeKindAst::Ts),
17252 "tstzmultirange" => ColumnTypeName::Multirange(RangeKindAst::TsTz),
17253 "datemultirange" => ColumnTypeName::Multirange(RangeKindAst::Date),
17254 // v7.37.5 ε — PG geometry scalar keywords.
17255 "point" => ColumnTypeName::Point,
17256 "lseg" => ColumnTypeName::Lseg,
17257 "path" => ColumnTypeName::Path,
17258 "box" => ColumnTypeName::PgBox,
17259 "polygon" => ColumnTypeName::Polygon,
17260 "line" => ColumnTypeName::Line,
17261 "circle" => ColumnTypeName::Circle,
17262 // v7.37.5 ζ-A — network / bit / xml / "char" keywords.
17263 "inet" => ColumnTypeName::Inet,
17264 "cidr" => ColumnTypeName::Cidr,
17265 "macaddr" => ColumnTypeName::Macaddr,
17266 "macaddr8" => ColumnTypeName::Macaddr8,
17267 // `bit`, `bit(N)`, `bit varying`, `bit varying(N)`. SPG carries the
17268 // width in the value, so the optional `(N)` typmod is accepted and
17269 // ignored (the column stores whatever width it's given).
17270 "bit" => {
17271 let varying = matches!(
17272 self.peek(),
17273 Token::Ident(k) if k.eq_ignore_ascii_case("varying")
17274 );
17275 if varying {
17276 self.advance();
17277 }
17278 // v7.39 (round 281) — the length used to be parsed and
17279 // dropped, so `bit(3)` accepted a five-bit string.
17280 let n = if matches!(self.peek(), Token::LParen) {
17281 self.parse_paren_size("BIT")?
17282 } else {
17283 0
17284 };
17285 if varying {
17286 ColumnTypeName::BitVarying(n)
17287 } else {
17288 ColumnTypeName::Bit(n)
17289 }
17290 }
17291 "varbit" => {
17292 let n = if matches!(self.peek(), Token::LParen) {
17293 self.parse_paren_size("VARBIT")?
17294 } else {
17295 0
17296 };
17297 ColumnTypeName::BitVarying(n)
17298 }
17299 "xml" => ColumnTypeName::Xml,
17300 // v7.17.0 Phase 3.P0-39 — PG hstore extension type.
17301 "hstore" => ColumnTypeName::Hstore,
17302 // v7.17.0 Phase 3.P0-36 — MySQL inline ENUM
17303 // `ENUM('a','b','c')`. Storage is TEXT; the value
17304 // list lands on `inline_enum_variants` for the
17305 // engine to validate INSERT cells against. Empty
17306 // value list is a parse error (matches MySQL).
17307 "enum" => {
17308 // Expect the opening `(`.
17309 if !matches!(self.peek(), Token::LParen) {
17310 return Err(self.err(alloc::format!(
17311 "expected '(' after ENUM, got {:?}",
17312 self.peek()
17313 )));
17314 }
17315 self.advance();
17316 let mut variants: Vec<String> = Vec::new();
17317 loop {
17318 match self.advance() {
17319 Token::String(s) => variants.push(s),
17320 other => {
17321 return Err(self.err(alloc::format!(
17322 "ENUM(...) expects string literal variants, got {other:?}"
17323 )));
17324 }
17325 }
17326 match self.peek() {
17327 Token::Comma => {
17328 self.advance();
17329 continue;
17330 }
17331 Token::RParen => {
17332 self.advance();
17333 break;
17334 }
17335 other => {
17336 return Err(self.err(alloc::format!(
17337 "expected ',' or ')' in ENUM(...), got {other:?}"
17338 )));
17339 }
17340 }
17341 }
17342 if variants.is_empty() {
17343 return Err(self.err("ENUM(...) must declare at least one variant".into()));
17344 }
17345 inline_enum_variants = Some(variants);
17346 // Storage is plain TEXT; the variant list lives on
17347 // the ColumnSchema side.
17348 ColumnTypeName::Text
17349 }
17350 // v7.17.0 Phase 3.P0-37 — MySQL inline SET
17351 // `SET('a','b','c')`. Same parse shape as ENUM;
17352 // semantics differ (subset rather than pick-one).
17353 "set" => {
17354 if !matches!(self.peek(), Token::LParen) {
17355 return Err(self.err(alloc::format!(
17356 "expected '(' after SET, got {:?}",
17357 self.peek()
17358 )));
17359 }
17360 self.advance();
17361 let mut variants: Vec<String> = Vec::new();
17362 loop {
17363 match self.advance() {
17364 Token::String(s) => variants.push(s),
17365 other => {
17366 return Err(self.err(alloc::format!(
17367 "SET(...) expects string literal variants, got {other:?}"
17368 )));
17369 }
17370 }
17371 match self.peek() {
17372 Token::Comma => {
17373 self.advance();
17374 continue;
17375 }
17376 Token::RParen => {
17377 self.advance();
17378 break;
17379 }
17380 other => {
17381 return Err(self.err(alloc::format!(
17382 "expected ',' or ')' in SET(...), got {other:?}"
17383 )));
17384 }
17385 }
17386 }
17387 if variants.is_empty() {
17388 return Err(self.err("SET(...) must declare at least one variant".into()));
17389 }
17390 inline_set_variants = Some(variants);
17391 ColumnTypeName::Text
17392 }
17393 _other => {
17394 // v7.17.0 Phase 1.4 — unknown ident → defer
17395 // resolution to the engine. Stored as Text in
17396 // ColumnTypeName + the original name carried as
17397 // `user_type_ref` so CREATE TABLE can look up
17398 // user-defined enum / domain types.
17399 user_type_ref = Some(ty_ident.clone());
17400 ColumnTypeName::Text
17401 }
17402 };
17403 // v7.17.0 Phase 4.4 — MySQL's `UNSIGNED` modifier sits
17404 // right after the type keyword. Pre-4.4 SPG consumed +
17405 // discarded the keyword, leaving a customer column
17406 // declared `id INT UNSIGNED NOT NULL` silently accepting
17407 // negative values — a Tier-A correctness drift where
17408 // application invariants (auto-increment-IDs never
17409 // negative) silently broke on cutover. Now: capture as
17410 // a column flag, persist on the schema, enforce at
17411 // INSERT / UPDATE time.
17412 let is_unsigned = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("unsigned"))
17413 {
17414 self.advance();
17415 true
17416 } else {
17417 false
17418 };
17419 // v7.14.0 — mysqldump emits `<type> CHARACTER SET <name>` and
17420 // `<type> COLLATE <name>` post-fixes on text columns. SPG
17421 // stores text as UTF-8 always so CHARACTER SET is still a
17422 // no-op. v7.17.0 Phase 2.5 — COLLATE no longer drops the
17423 // name: it gets classified into a `Collation` variant the
17424 // engine consults at WHERE-eval time. PG `default` /
17425 // `pg_catalog.default` / `C` / `POSIX` collations all
17426 // resolve to `Binary` (the prior behaviour); `_ci` /
17427 // `case_insensitive` / `nocase` shift to CaseInsensitive.
17428 // The schema-qualifier form (`pg_catalog.default`) lexes
17429 // as `Ident '.' Ident` — peek for the `.` and consume both
17430 // halves so it's treated as one collation name. PG's
17431 // `IDENT.IDENT` collation form (which can appear here) is
17432 // resolved by Collation::from_collation_name on the bare
17433 // identifier after the dot.
17434 let mut collation = Collation::Binary;
17435 // v7.39 (round 370, M4 P4a) — whether an explicit `COLLATE <name>`
17436 // clause was written. The engine needs this to tell an explicit
17437 // `COLLATE utf8mb4_bin` (byte-wise) apart from a column with no
17438 // clause at all: both resolve to `Collation::Binary`, but under the
17439 // MySQL dialect the latter takes the folding default collation.
17440 let mut collation_explicit = false;
17441 let mut collation_name: Option<alloc::string::String> = None;
17442 loop {
17443 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("character"))
17444 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("set"))
17445 {
17446 self.advance(); // CHARACTER
17447 self.advance(); // SET
17448 if matches!(
17449 self.peek(),
17450 Token::Ident(_) | Token::QuotedIdent(_) | Token::String(_)
17451 ) {
17452 self.advance();
17453 }
17454 continue;
17455 }
17456 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("collate")) {
17457 self.advance(); // COLLATE
17458 // Accept Ident / QuotedIdent / String AND the
17459 // keyword-tokenised `Default` (PG `pg_catalog.default`
17460 // and bare `DEFAULT` collation names — `default` is a
17461 // reserved word so the lexer hands back Token::Default
17462 // not Token::Ident).
17463 let read_collation_atom = |this: &mut Self| -> Option<alloc::string::String> {
17464 match this.peek().clone() {
17465 Token::Ident(s) | Token::QuotedIdent(s) | Token::String(s) => {
17466 this.advance();
17467 Some(s)
17468 }
17469 Token::Default => {
17470 this.advance();
17471 Some(alloc::string::String::from("default"))
17472 }
17473 _ => None,
17474 }
17475 };
17476 let raw = if let Some(head) = read_collation_atom(self) {
17477 // Schema-qualified PG form: `pg_catalog.default`.
17478 if matches!(self.peek(), Token::Dot) {
17479 self.advance();
17480 let tail = read_collation_atom(self).unwrap_or_default();
17481 alloc::format!("{head}.{tail}")
17482 } else {
17483 head
17484 }
17485 } else {
17486 alloc::string::String::new()
17487 };
17488 if !raw.is_empty() {
17489 collation_explicit = true;
17490 // v7.39 (round 676) — keep the name too. The enum below
17491 // folds C / POSIX / en_US / default into one value, and
17492 // `pg_attribute.attcollation` has to tell them apart.
17493 // The schema qualifier goes: PG's `pg_catalog.default`
17494 // and a bare `default` name the same collation.
17495 // v7.39 (round 679) — strip a SCHEMA qualifier, not an
17496 // encoding suffix. Round 676 used `rsplit('.')` for
17497 // both, and `COLLATE "en_US.utf8"` came out as `utf8`:
17498 // PG writes `pg_catalog.default` (qualifier) and
17499 // `en_US.utf8` (locale + encoding) with the same
17500 // separator. Only `pg_catalog.` is a qualifier, and it
17501 // is the only one PG's own dumps emit.
17502 let bare = raw.trim_matches(|c: char| c == '"' || c == '\'');
17503 let bare = bare.strip_prefix("pg_catalog.").unwrap_or(bare);
17504 collation_name = Some(alloc::string::String::from(bare));
17505 let parsed = Collation::from_collation_name(&raw);
17506 // Last COLLATE clause wins, but `Binary` from a
17507 // bare keyword like `default` should not
17508 // silently downgrade a stronger one set earlier
17509 // on the same column. v7.17 only ships one
17510 // non-Binary variant so a simple OR is enough.
17511 if parsed != Collation::Binary {
17512 collation = parsed;
17513 }
17514 }
17515 continue;
17516 }
17517 break;
17518 }
17519 // v7.10.10 — postfix `[]` widens the base type to its array
17520 // type. PG accepts `TYPE[]` after any base type and so does
17521 // SPG now (round-753 probe: INT[] / NUMERIC[] / TIMESTAMP[]
17522 // all through; the old "only TEXT[]" note was stale).
17523 if matches!(self.peek(), Token::LBracket) {
17524 self.advance();
17525 if !matches!(self.peek(), Token::RBracket) {
17526 return Err(self.err(alloc::format!(
17527 "TEXT[] takes no dimension; got {:?}",
17528 self.peek()
17529 )));
17530 }
17531 self.advance();
17532 // v7.11.13 — widened to INT[] and BIGINT[] in addition
17533 // to TEXT[]. Other base types (BOOL[], NUMERIC[], etc.)
17534 // still error here.
17535 ty = match ty {
17536 ColumnTypeName::Text => ColumnTypeName::TextArray,
17537 ColumnTypeName::Int => ColumnTypeName::IntArray,
17538 ColumnTypeName::BigInt => ColumnTypeName::BigIntArray,
17539 // v7.37.5 β-P4 — INTERVAL[] via the same postfix
17540 // `[]` grammar. Wire OID 1187.
17541 ColumnTypeName::Interval => ColumnTypeName::IntervalArray,
17542 // v7.37.5 γ — full PG array-of-scalar family.
17543 ColumnTypeName::Bool => ColumnTypeName::BoolArray,
17544 ColumnTypeName::SmallInt => ColumnTypeName::SmallIntArray,
17545 ColumnTypeName::Float => ColumnTypeName::FloatArray,
17546 // NUMERIC(p, s) loses its precision params at the
17547 // array level (matches PG: `NUMERIC[]` is untyped,
17548 // per-element precision flows through values).
17549 ColumnTypeName::Numeric(_, _) => ColumnTypeName::NumericArray,
17550 ColumnTypeName::Date => ColumnTypeName::DateArray,
17551 ColumnTypeName::Timestamp => ColumnTypeName::TimestampArray,
17552 ColumnTypeName::Timestamptz => ColumnTypeName::TimestamptzArray,
17553 ColumnTypeName::Uuid => ColumnTypeName::UuidArray,
17554 ColumnTypeName::Json => ColumnTypeName::JsonArray,
17555 ColumnTypeName::Jsonb => ColumnTypeName::JsonbArray,
17556 ColumnTypeName::Bytes => ColumnTypeName::BytesArray,
17557 // VARCHAR(n)[] / CHAR(n)[] drop the length cap at
17558 // the array level (matches PG semantics where the
17559 // element precision is per-row, not column-wide).
17560 ColumnTypeName::Varchar(_) => ColumnTypeName::VarcharArray,
17561 ColumnTypeName::Char(_) => ColumnTypeName::CharArray,
17562 // v7.37.5 ζ-A — MONEY[] (OID 791) ship-triage
17563 // follow-up.
17564 ColumnTypeName::Money => ColumnTypeName::MoneyArray,
17565 other => {
17566 return Err(self.err(alloc::format!("{other:?}[] not yet supported")));
17567 }
17568 };
17569 // v7.17.0 Phase 3.P0-40 — second `[]` widens 1D → 2D
17570 // for INT/TEXT/BIGINT. Anything else is an error.
17571 if matches!(self.peek(), Token::LBracket) {
17572 self.advance();
17573 if !matches!(self.peek(), Token::RBracket) {
17574 return Err(self.err(alloc::format!(
17575 "TYPE[][] second dimension takes no size; got {:?}",
17576 self.peek()
17577 )));
17578 }
17579 self.advance();
17580 ty = match ty {
17581 ColumnTypeName::IntArray => ColumnTypeName::IntArray2D,
17582 ColumnTypeName::BigIntArray => ColumnTypeName::BigIntArray2D,
17583 ColumnTypeName::TextArray => ColumnTypeName::TextArray2D,
17584 // v7.39 (read01 round 75) — bool[][].
17585 ColumnTypeName::BoolArray => ColumnTypeName::BoolArray2D,
17586 other => {
17587 return Err(self.err(alloc::format!(
17588 "v7.17 2D arrays support INT[][] / BIGINT[][] / \
17589 TEXT[][] only; got {other:?}"
17590 )));
17591 }
17592 };
17593 }
17594 }
17595 Ok((
17596 ty,
17597 implied_auto_increment,
17598 implied_not_null,
17599 user_type_ref,
17600 collation,
17601 collation_explicit,
17602 collation_name,
17603 is_unsigned,
17604 inline_enum_variants,
17605 inline_set_variants,
17606 mysql_int_width,
17607 mysql_fsp,
17608 mysql_declared_timestamp,
17609 mysql_float_md,
17610 ))
17611 }
17612
17613 fn parse_column_def(&mut self) -> Result<ColumnDef, ParseError> {
17614 // v7.20 — PG reserves the table-constraint keywords, so a
17615 // BARE `UNIQUE` / `PRIMARY` / … in column position is a
17616 // malformed constraint clause (e.g. `UNIQUE a` missing its
17617 // parens), not a column named "unique". Since v7.17's
17618 // unknown-type leniency (`user_type_ref`) such a clause
17619 // would otherwise parse as a column with a user-defined
17620 // type — silently accepting invalid DDL. Quoted
17621 // identifiers ("unique" / `unique`) remain valid names.
17622 if let Token::Ident(s) = self.peek()
17623 && [
17624 "unique",
17625 "primary",
17626 "foreign",
17627 "constraint",
17628 "check",
17629 "references",
17630 "exclude",
17631 ]
17632 .iter()
17633 .any(|kw| s.eq_ignore_ascii_case(kw))
17634 {
17635 return Err(self.err(alloc::format!(
17636 "unexpected reserved keyword '{s}' at start of column definition \
17637 (malformed table constraint?)"
17638 )));
17639 }
17640 let name_tok = self.pos;
17641 let name = self.expect_ident_like()?;
17642 // v7.39.3 — MySQL 9.7.2 reports a column by the SPELLING it was
17643 // declared with: `MyCol` stays `MyCol` in SHOW COLUMNS, in
17644 // information_schema, and in SHOW CREATE (measured). SPG folded
17645 // an unquoted name, so a table restored from a dump reported
17646 // names the application had never written.
17647 //
17648 // The written form comes back from the source span, which only
17649 // the MySQL dialect keeps. The span runs to the START of the
17650 // next token, so a comment or unusual spacing between them
17651 // arrives with it — hence the check that what came back is the
17652 // same identifier. It is not decoration: without it,
17653 // `CREATE TABLE t (MyCol /* c */ INT)` names the column
17654 // `MyCol /* c */`.
17655 let name = self
17656 .source_span(name_tok, name_tok)
17657 .map(|raw| raw.trim().trim_matches('`').trim_matches('"'))
17658 .filter(|raw| raw.eq_ignore_ascii_case(&name))
17659 .map_or(name, alloc::string::String::from);
17660 let (
17661 ty,
17662 implied_auto_increment,
17663 implied_not_null,
17664 user_type_ref,
17665 collation,
17666 collation_explicit,
17667 collation_name,
17668 is_unsigned,
17669 inline_enum_variants,
17670 inline_set_variants,
17671 mysql_int_width,
17672 mysql_fsp,
17673 mysql_declared_timestamp,
17674 mysql_float_md,
17675 ) = self.parse_type_with_implied_flags()?;
17676 // Column constraints: `DEFAULT <expr>`, `NOT NULL`, and the
17677 // MySQL-flavoured `AUTO_INCREMENT` may appear in any order;
17678 // each at most once.
17679 let mut default: Option<Expr> = None;
17680 let mut nullable = !implied_not_null;
17681 let mut nullability_seen = implied_not_null;
17682 let mut auto_increment = implied_auto_increment;
17683 let mut is_primary_key = false;
17684 let mut is_unique = false;
17685 let mut unique_nulls_not_distinct = false;
17686 let mut constraint_deferrable = false;
17687 let mut constraint_initially_deferred = false;
17688 let mut check: Option<Expr> = None;
17689 let mut on_update_runtime: Option<Expr> = None;
17690 let mut generated_stored_expr: Option<Box<Expr>> = None;
17691 let mut identity_always = false;
17692 loop {
17693 // v7.22 (mailrs round-13 gap 3) — PG 18 catalogs
17694 // not-null constraints by name and pg_dump emits them
17695 // inline: `id bigint CONSTRAINT contacts_id_not_null1
17696 // NOT NULL`. Accept and discard the name; whatever
17697 // constraint follows is parsed by the arms below.
17698 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("constraint")) {
17699 // v7.39 (round 308, V29) — a name on an inline
17700 // REFERENCES belongs to the FOREIGN KEY, and the caller
17701 // (`parse_column_def_with_fk`) is what builds it, so
17702 // leave the whole clause for it. Dropping the name here
17703 // is what made `CONSTRAINT fk_a REFERENCES …` come back
17704 // as the synthesised `c_pid_fkey` — which then could
17705 // not be matched by `SET CONSTRAINTS fk_a`. Peek only:
17706 // `advance()` takes tokens by `mem::replace`, so there
17707 // is no rewinding once consumed.
17708 if matches!(
17709 self.tokens.get(self.pos + 2),
17710 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("references")
17711 ) {
17712 break;
17713 }
17714 self.advance();
17715 let _name = self.expect_ident_like()?;
17716 continue;
17717 }
17718 // v7.39 (round 379) — MySQL's SHORT generated-column form
17719 // omits `GENERATED ALWAYS`: `<col> <type> AS (<expr>)
17720 // [STORED | VIRTUAL]`. mysqldump emits the long form (handled
17721 // below), but hand-written schemas and app migrations use this.
17722 // STORED / VIRTUAL is optional (MySQL defaults to VIRTUAL);
17723 // SPG computes-and-stores either way, like the long form.
17724 if matches!(self.peek(), Token::As) {
17725 self.advance();
17726 if !matches!(self.peek(), Token::LParen) {
17727 return Err(self.err(alloc::format!(
17728 "expected '(' after AS in a generated column, got {:?}",
17729 self.peek()
17730 )));
17731 }
17732 self.advance();
17733 let expr = self.parse_expr(0)?;
17734 if !matches!(self.peek(), Token::RParen) {
17735 return Err(self.err(alloc::format!(
17736 "expected ')' after AS (<expr>), got {:?}",
17737 self.peek()
17738 )));
17739 }
17740 self.advance();
17741 if matches!(self.peek(), Token::Ident(s)
17742 if s.eq_ignore_ascii_case("stored") || s.eq_ignore_ascii_case("virtual"))
17743 {
17744 self.advance();
17745 }
17746 generated_stored_expr = Some(alloc::boxed::Box::new(expr));
17747 continue;
17748 }
17749 // v7.22 (round-13 T2) — inline `GENERATED { ALWAYS |
17750 // BY DEFAULT } AS IDENTITY [(seq options)]` (PG 10+;
17751 // the modern replacement for SERIAL in hand-written
17752 // schemas). Both flavours map onto the auto-increment
17753 // machinery — SPG's serial semantics ≈ BY DEFAULT;
17754 // ALWAYS's reject-explicit-values nuance is documented
17755 // leniency. Generated EXPRESSION columns
17756 // (`AS (expr) STORED`) are not supported: error loudly
17757 // instead of silently storing NULLs.
17758 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("generated")) {
17759 self.advance();
17760 let mut saw_generated_always = false;
17761 match self.peek().clone() {
17762 Token::Ident(s) if s.eq_ignore_ascii_case("always") => {
17763 self.advance();
17764 saw_generated_always = true;
17765 }
17766 Token::Ident(b) | Token::QuotedIdent(b) if b.eq_ignore_ascii_case("by") => {
17767 self.advance();
17768 if !matches!(self.peek(), Token::Default) {
17769 return Err(self.err(alloc::format!(
17770 "expected DEFAULT after GENERATED BY, got {:?}",
17771 self.peek()
17772 )));
17773 }
17774 self.advance();
17775 }
17776 other => {
17777 return Err(self.err(alloc::format!(
17778 "expected ALWAYS or BY DEFAULT after GENERATED, got {other:?}"
17779 )));
17780 }
17781 }
17782 if !matches!(self.peek(), Token::As) {
17783 return Err(self.err(alloc::format!(
17784 "expected AS after GENERATED ALWAYS/BY DEFAULT, got {:?}",
17785 self.peek()
17786 )));
17787 }
17788 self.advance();
17789 // v7.37.7(sentori Epic 3 P1)— `GENERATED ALWAYS AS
17790 // ( <expr> ) STORED` stored computed-column. The
17791 // expression is captured for the engine to recompute
17792 // on every INSERT / UPDATE. v7.37.7 accepts the
17793 // STORED keyword only; PG also has VIRTUAL, which
17794 // v7.37.7 carves out (sentori only uses STORED).
17795 if matches!(self.peek(), Token::LParen) {
17796 self.advance();
17797 let expr = self.parse_expr(0)?;
17798 if !matches!(self.peek(), Token::RParen) {
17799 return Err(self.err(alloc::format!(
17800 "expected ')' after GENERATED ALWAYS AS (<expr>), got {:?}",
17801 self.peek()
17802 )));
17803 }
17804 self.advance();
17805 let stored = match self.peek() {
17806 Token::Ident(s) | Token::QuotedIdent(s)
17807 if s.eq_ignore_ascii_case("stored") =>
17808 {
17809 self.advance();
17810 true
17811 }
17812 // v7.38 (read01 P4.14) — accept PG 18's VIRTUAL
17813 // generated columns. SPG computes them on write and
17814 // persists like STORED; the two are observably
17815 // identical for query results (the value, recompute
17816 // on base-column change, and NOT NULL enforcement all
17817 // match), so a PG 18 schema/dump using VIRTUAL loads
17818 // and behaves correctly. The compute-on-read storage
17819 // saving is an invisible internal difference.
17820 Token::Ident(s) | Token::QuotedIdent(s)
17821 if s.eq_ignore_ascii_case("virtual") =>
17822 {
17823 self.advance();
17824 false
17825 }
17826 other => {
17827 return Err(self.err(alloc::format!(
17828 "expected STORED or VIRTUAL after GENERATED ALWAYS AS (<expr>), \
17829 got {other:?}"
17830 )));
17831 }
17832 };
17833 let _ = stored; // STORED / VIRTUAL both compute-and-store.
17834 generated_stored_expr = Some(Box::new(expr));
17835 continue;
17836 }
17837 self.expect_keyword_ident("identity")?;
17838 // Optional `(START WITH 1 INCREMENT BY 1 …)` —
17839 // consume the balanced parens and discard (SPG's
17840 // auto-increment is max+1-scan based).
17841 if matches!(self.peek(), Token::LParen) {
17842 let mut depth = 0usize;
17843 loop {
17844 match self.advance() {
17845 Token::LParen => depth += 1,
17846 Token::RParen => {
17847 depth -= 1;
17848 if depth == 0 {
17849 break;
17850 }
17851 }
17852 Token::Eof => {
17853 return Err(self.err(
17854 "unterminated sequence-options parens after IDENTITY".into(),
17855 ));
17856 }
17857 _ => {}
17858 }
17859 }
17860 }
17861 auto_increment = true;
17862 // v7.38 (read01) — remember the ALWAYS flavour so the engine
17863 // can reject explicit non-DEFAULT INSERT values (unless
17864 // OVERRIDING SYSTEM VALUE) the way PG does.
17865 identity_always = saw_generated_always;
17866 // PG identity columns are implicitly NOT NULL.
17867 nullable = false;
17868 continue;
17869 }
17870 // v7.17.0 Phase 2.1 — MySQL `ON UPDATE
17871 // CURRENT_TIMESTAMP[(N)]`. Only CURRENT_TIMESTAMP
17872 // is accepted today. The "ON" token is an Ident
17873 // (not reserved) — peek before consuming.
17874 if matches!(self.peek(), Token::On)
17875 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("update"))
17876 {
17877 self.advance(); // ON
17878 self.advance(); // update
17879 // Accept CURRENT_TIMESTAMP / CURRENT_TIMESTAMP(N).
17880 let next = self.peek().clone();
17881 match next {
17882 Token::Ident(s) | Token::QuotedIdent(s)
17883 if s.eq_ignore_ascii_case("current_timestamp") =>
17884 {
17885 self.advance();
17886 // Optional `(N)` precision.
17887 if matches!(self.peek(), Token::LParen) {
17888 self.advance();
17889 if !matches!(self.peek(), Token::Integer(_)) {
17890 return Err(self.err(alloc::format!(
17891 "expected integer precision inside CURRENT_TIMESTAMP(…), got {:?}",
17892 self.peek()
17893 )));
17894 }
17895 self.advance();
17896 if !matches!(self.peek(), Token::RParen) {
17897 return Err(self.err(alloc::format!(
17898 "expected ')' after CURRENT_TIMESTAMP precision, got {:?}",
17899 self.peek()
17900 )));
17901 }
17902 self.advance();
17903 }
17904 on_update_runtime = Some(Expr::FunctionCall {
17905 name: "now".into(),
17906 args: Vec::new(),
17907 });
17908 continue;
17909 }
17910 other => {
17911 return Err(self.err(alloc::format!(
17912 "v7.17 only supports ON UPDATE CURRENT_TIMESTAMP, got {other:?}"
17913 )));
17914 }
17915 }
17916 }
17917 if matches!(self.peek(), Token::Default) {
17918 if default.is_some() {
17919 return Err(self.err("DEFAULT specified twice".into()));
17920 }
17921 self.advance();
17922 default = Some(self.parse_expr(0)?);
17923 continue;
17924 }
17925 // v7.39 (round 621) — `NOT DEFERRABLE` shares this arm's leading
17926 // token with NOT NULL and sits EARLIER in the loop than the
17927 // deferrability arm, so without the lookahead it was reported as
17928 // "NOT NULL specified twice" (or "expected NULL after NOT").
17929 if matches!(self.peek(), Token::Not)
17930 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("deferrable"))
17931 {
17932 // NOT DEFERRABLE — explicit immediate; nothing to carry.
17933 self.consume_optional_deferrable_clauses()?;
17934 continue;
17935 }
17936 if matches!(self.peek(), Token::Not) {
17937 if nullability_seen {
17938 return Err(self.err("NOT NULL specified twice".into()));
17939 }
17940 self.advance();
17941 if !matches!(self.peek(), Token::Null) {
17942 return Err(self.err(format!(
17943 "expected NULL after NOT in column def, got {:?}",
17944 self.peek()
17945 )));
17946 }
17947 self.advance();
17948 nullable = false;
17949 nullability_seen = true;
17950 continue;
17951 }
17952 // v7.14.0 — MySQL accepts a bare `NULL` as an explicit
17953 // "this column is nullable" marker (the default in
17954 // standard SQL anyway). mysqldump emits it routinely
17955 // (`col TYPE NULL DEFAULT NULL` for nullable
17956 // timestamps etc). Accept + no-op.
17957 if matches!(self.peek(), Token::Null) {
17958 if nullability_seen && !nullable {
17959 // v7.39 (round 761, F31 tranche 2 #31) — PG's
17960 // sentence, PG18-measured (the table name is the
17961 // caller's; the column half is exact).
17962 return Err(self.err(alloc::format!(
17963 "conflicting NULL/NOT NULL declarations for column \"{name}\""
17964 )));
17965 }
17966 self.advance();
17967 nullable = true;
17968 nullability_seen = true;
17969 continue;
17970 }
17971 // `AUTO_INCREMENT` or its abbreviated form `AUTOINCREMENT`
17972 // arrives as a bare Ident. Match either, case-insensitive.
17973 if let Token::Ident(s) = self.peek()
17974 && (s.eq_ignore_ascii_case("auto_increment")
17975 || s.eq_ignore_ascii_case("autoincrement"))
17976 {
17977 if auto_increment {
17978 return Err(self.err("AUTO_INCREMENT specified twice".into()));
17979 }
17980 self.advance();
17981 auto_increment = true;
17982 continue;
17983 }
17984 // v7.9.13 — inline `PRIMARY KEY` column constraint
17985 // (mailrs F1). Implies `NOT NULL`. The engine creates
17986 // a BTree index for the PK column at CREATE TABLE time
17987 // so FK parent-side index lookups resolve.
17988 // v7.39 (round 621) — `[NOT] DEFERRABLE [INITIALLY {DEFERRED |
17989 // IMMEDIATE}]` after an inline PK / UNIQUE / REFERENCES. Every
17990 // spelling was a parse error, so a pg_dump carrying one stopped
17991 // mid-restore. The clauses are consumed by the same helper the FK
17992 // path has used since round 288 and recorded nowhere: SPG enforces
17993 // the constraint IMMEDIATELY either way, which fails earlier than
17994 // PG inside a transaction that violates-then-repairs — a refusal,
17995 // not a wrong answer. True deferral is the open remainder of F08.
17996 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("deferrable") || s.eq_ignore_ascii_case("initially"))
17997 || (matches!(self.peek(), Token::Not)
17998 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s)) if s.eq_ignore_ascii_case("deferrable")))
17999 {
18000 // v7.39 (round 711) — CARRIED now (the storing half of
18001 // F08); round 621 only consumed.
18002 let (d, idef) = self.consume_deferrable_clauses_timed()?;
18003 constraint_deferrable |= d;
18004 constraint_initially_deferred |= idef;
18005 continue;
18006 }
18007 if let Token::Ident(s) = self.peek()
18008 && s.eq_ignore_ascii_case("primary")
18009 {
18010 if is_primary_key {
18011 return Err(self.err("PRIMARY KEY specified twice".into()));
18012 }
18013 // Peek-ahead for the required `KEY` token.
18014 let next = self.tokens.get(self.pos + 1);
18015 let next_is_key = matches!(
18016 next,
18017 Some(Token::Ident(k)) if k.eq_ignore_ascii_case("key")
18018 );
18019 if !next_is_key {
18020 return Err(self.err(format!(
18021 "expected KEY after PRIMARY in column def, got {:?}",
18022 next
18023 )));
18024 }
18025 self.advance(); // PRIMARY
18026 self.advance(); // KEY
18027 is_primary_key = true;
18028 if nullability_seen && nullable {
18029 return Err(self.err(
18030 "column declared NULL but inline PRIMARY KEY implies NOT NULL".into(),
18031 ));
18032 }
18033 nullable = false;
18034 nullability_seen = true;
18035 continue;
18036 }
18037 // v7.13.0 — inline `UNIQUE` column constraint
18038 // (mailrs round-5 G2). Fold into a single-column
18039 // table-level UNIQUE at CREATE TABLE post-process time.
18040 if let Token::Ident(s) = self.peek()
18041 && s.eq_ignore_ascii_case("unique")
18042 {
18043 if is_unique {
18044 return Err(self.err("UNIQUE specified twice".into()));
18045 }
18046 self.advance();
18047 is_unique = true;
18048 // v7.38 (read01 P4.19) — optional `NULLS [NOT] DISTINCT`
18049 // (PG 15+); default is NULLS DISTINCT per the SQL standard.
18050 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("nulls")) {
18051 let n1 = self.tokens.get(self.pos + 1);
18052 let n2 = self.tokens.get(self.pos + 2);
18053 if matches!(n1, Some(Token::Not)) && matches!(n2, Some(Token::Distinct)) {
18054 self.advance(); // NULLS
18055 self.advance(); // NOT
18056 self.advance(); // DISTINCT
18057 unique_nulls_not_distinct = true;
18058 } else if matches!(n1, Some(Token::Distinct)) {
18059 self.advance(); // NULLS
18060 self.advance(); // DISTINCT
18061 }
18062 }
18063 continue;
18064 }
18065 // v7.13.0 — inline `CHECK (<expr>)` column constraint
18066 // (mailrs round-5 G3). PG semantics: column-level
18067 // CHECK is equivalent to a table-level CHECK. Multiple
18068 // inline CHECKs on the same column AND together.
18069 if let Token::Ident(s) = self.peek()
18070 && s.eq_ignore_ascii_case("check")
18071 {
18072 self.advance();
18073 if !matches!(self.peek(), Token::LParen) {
18074 return Err(self.err(alloc::format!(
18075 "expected '(' after CHECK in column def, got {:?}",
18076 self.peek()
18077 )));
18078 }
18079 self.advance();
18080 let pred = self.parse_expr(0)?;
18081 if !matches!(self.peek(), Token::RParen) {
18082 return Err(self.err(alloc::format!(
18083 "expected ')' to close CHECK predicate, got {:?}",
18084 self.peek()
18085 )));
18086 }
18087 self.advance();
18088 check = Some(match check.take() {
18089 Some(prev) => Expr::Binary {
18090 op: BinOp::And,
18091 lhs: Box::new(prev),
18092 rhs: Box::new(pred),
18093 },
18094 None => pred,
18095 });
18096 continue;
18097 }
18098 break;
18099 }
18100 Ok(ColumnDef {
18101 name,
18102 ty,
18103 nullable,
18104 default,
18105 auto_increment,
18106 is_primary_key,
18107 is_unique,
18108 unique_nulls_not_distinct,
18109 constraint_deferrable,
18110 constraint_initially_deferred,
18111 check,
18112 user_type_ref,
18113 on_update_runtime,
18114 collation,
18115 collation_explicit,
18116 collation_name,
18117 is_unsigned,
18118 inline_enum_variants,
18119 inline_set_variants,
18120 generated_stored_expr,
18121 identity_always,
18122 mysql_int_width,
18123 mysql_fsp,
18124 mysql_declared_timestamp,
18125 mysql_float_md,
18126 })
18127 }
18128
18129 /// `NUMERIC` may appear without parameters, with one (precision
18130 /// only, scale=0), or with both. Returns `(precision, scale)` with
18131 /// 0 = unspecified for the bare form.
18132 fn parse_optional_numeric_params(&mut self) -> Result<(u16, i16), ParseError> {
18133 if !matches!(self.peek(), Token::LParen) {
18134 // Bare `NUMERIC` — PG treats this as "unlimited precision";
18135 // we surface it as precision=0 to mean "unconstrained" so
18136 // the engine doesn't need a separate variant.
18137 return Ok((0, 0));
18138 }
18139 self.advance();
18140 // v7.39 (round 272) — PG's declared precision runs to 1000, and
18141 // it words the out-of-range case with the value it saw. SPG
18142 // capped at 38 (i128's width), so a `numeric(50,10)` column PG
18143 // accepts failed to parse at all; values wider than i128 are
18144 // carried by the arbitrary-precision form.
18145 let precision = match self.advance() {
18146 Token::Integer(n) if (1..=1000).contains(&n) => {
18147 u16::try_from(n).expect("range-checked")
18148 }
18149 Token::Integer(n) => {
18150 return Err(ParseError {
18151 message: format!("NUMERIC precision {n} must be between 1 and 1000"),
18152 token_pos: self.consumed_pos(),
18153 });
18154 }
18155 other => {
18156 return Err(ParseError {
18157 message: format!(
18158 "NUMERIC precision must be an integer in 1..=1000, got {other:?}"
18159 ),
18160 token_pos: self.consumed_pos(),
18161 });
18162 }
18163 };
18164 // v7.39 (round 273) — PG's declared scale runs -1000..=1000 and is
18165 // NOT bounded by the precision (`numeric(10,11)` is legal; a value
18166 // then overflows). A negative scale rounds to tens / hundreds / …
18167 let scale = if matches!(self.peek(), Token::Comma) {
18168 self.advance();
18169 let neg = if matches!(self.peek(), Token::Minus) {
18170 self.advance();
18171 true
18172 } else {
18173 false
18174 };
18175 match self.advance() {
18176 Token::Integer(n) => {
18177 let signed = if neg { -n } else { n };
18178 if !(-1000..=1000).contains(&signed) {
18179 return Err(ParseError {
18180 message: format!(
18181 "NUMERIC scale {signed} must be between -1000 and 1000"
18182 ),
18183 token_pos: self.consumed_pos(),
18184 });
18185 }
18186 i16::try_from(signed).expect("range-checked")
18187 }
18188 other => {
18189 return Err(ParseError {
18190 message: format!("NUMERIC scale must be an integer, got {other:?}"),
18191 token_pos: self.consumed_pos(),
18192 });
18193 }
18194 }
18195 } else {
18196 0
18197 };
18198 if !matches!(self.peek(), Token::RParen) {
18199 return Err(self.err(format!(
18200 "expected ')' to close NUMERIC params, got {:?}",
18201 self.peek()
18202 )));
18203 }
18204 self.advance();
18205 Ok((precision, scale))
18206 }
18207
18208 /// Parse `(N)` where `N` is a positive integer literal — used by the
18209 /// `VARCHAR`/`CHAR`/`VECTOR` column types. `label` is the type name
18210 /// for the error message.
18211 /// v6.0.1: parse the optional `USING <encoding>` clause that
18212 /// follows `VECTOR(N)` in a column definition. Missing clause
18213 /// → `VecEncoding::F32` (pre-v6 default). Unknown encoding
18214 /// ident → `ParseError` listing the encodings recognised today.
18215 fn parse_optional_vector_encoding(&mut self) -> Result<VecEncoding, ParseError> {
18216 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
18217 return Ok(VecEncoding::F32);
18218 }
18219 // v7.13.2 — mailrs round-6 S6: `USING` after a vector type
18220 // overlaps with `ALTER COLUMN TYPE … USING <expr>`. Only
18221 // consume the token when the very next token is a known
18222 // vector-encoding keyword (SQ8 / HALF). Otherwise leave
18223 // `USING` for the caller — it's the rewrite-expression form.
18224 let n1 = self.tokens.get(self.pos + 1);
18225 let next_is_encoding = matches!(
18226 n1,
18227 Some(Token::Ident(s))
18228 if s.eq_ignore_ascii_case("sq8") || s.eq_ignore_ascii_case("half")
18229 );
18230 if !next_is_encoding {
18231 return Ok(VecEncoding::F32);
18232 }
18233 self.advance();
18234 let enc_ident = match self.advance() {
18235 Token::Ident(s) => s,
18236 other => {
18237 return Err(self.err(format!(
18238 "expected vector encoding after USING, got {other:?}"
18239 )));
18240 }
18241 };
18242 match enc_ident.to_ascii_lowercase().as_str() {
18243 "sq8" => Ok(VecEncoding::Sq8),
18244 // v6.0.3: `HALF` (pgvector convention) selects IEEE-754
18245 // binary16 per-element storage.
18246 "half" => Ok(VecEncoding::F16),
18247 other => Err(self.err(format!(
18248 "unknown vector encoding {other:?}; supported: SQ8, HALF"
18249 ))),
18250 }
18251 }
18252
18253 /// v7.17.0 Phase 4.3 — peek at the MySQL display-width
18254 /// without consuming it. Returns `Some(N)` when the next
18255 /// tokens are `( <int> )`; None otherwise. Used by the
18256 /// TINYINT classifier to decide whether to map to Bool or
18257 /// SmallInt.
18258 fn peek_optional_paren_size_value(&self) -> Option<i64> {
18259 if !matches!(self.peek(), Token::LParen) {
18260 return None;
18261 }
18262 let next = self.tokens.get(self.pos + 1)?;
18263 let n = match next {
18264 Token::Integer(n) => *n,
18265 _ => return None,
18266 };
18267 if !matches!(self.tokens.get(self.pos + 2), Some(Token::RParen)) {
18268 return None;
18269 }
18270 Some(n)
18271 }
18272
18273 /// v7.14.0 — consume an optional MySQL display-width
18274 /// parenthesised number after an integer type, returning
18275 /// nothing. `TINYINT(1)` etc.
18276 /// v7.39 (round 360) — does the parenthesised group ahead contain a
18277 /// comma, i.e. is it MySQL's `(m,d)` rather than PG's `(p)`?
18278 fn peek_paren_has_comma(&self) -> bool {
18279 let mut i = self.pos + 1;
18280 let mut depth = 1usize;
18281 while depth > 0 {
18282 match self.tokens.get(i) {
18283 Some(Token::LParen) => depth += 1,
18284 Some(Token::RParen) => depth -= 1,
18285 Some(Token::Comma) if depth == 1 => return true,
18286 None | Some(Token::Eof) => return false,
18287 _ => {}
18288 }
18289 i += 1;
18290 }
18291 false
18292 }
18293
18294 /// v7.39 (round 424) — the same optional `(N)` modifier, but RETURNING
18295 /// the number. Temporal columns need it: MySQL's `DATETIME(3)` declares a
18296 /// fractional-seconds precision that drives write truncation and render
18297 /// padding, where `consume_optional_paren_size` throws it away.
18298 /// `Some(0)` for an explicit `(0)`, `None` when no modifier is written.
18299 fn take_optional_paren_size(&mut self) -> Option<u8> {
18300 let Some(Token::Integer(n)) = self
18301 .tokens
18302 .get(self.pos + 1)
18303 .filter(|_| matches!(self.peek(), Token::LParen))
18304 .cloned()
18305 else {
18306 self.consume_optional_paren_size();
18307 return None;
18308 };
18309 if !matches!(self.tokens.get(self.pos + 2), Some(Token::RParen)) {
18310 self.consume_optional_paren_size();
18311 return None;
18312 }
18313 self.consume_optional_paren_size();
18314 u8::try_from(n).ok()
18315 }
18316
18317 fn consume_optional_paren_size(&mut self) {
18318 if !matches!(self.peek(), Token::LParen) {
18319 return;
18320 }
18321 self.advance();
18322 // Skip until matching RParen (allow nested or any tokens).
18323 let mut depth = 1usize;
18324 while depth > 0 {
18325 match self.peek() {
18326 Token::LParen => depth += 1,
18327 Token::RParen => depth -= 1,
18328 Token::Eof => return,
18329 _ => {}
18330 }
18331 self.advance();
18332 }
18333 }
18334
18335 fn parse_paren_size(&mut self, label: &str) -> Result<u32, ParseError> {
18336 if !matches!(self.peek(), Token::LParen) {
18337 return Err(self.err(format!("{label} type requires (N), got {:?}", self.peek())));
18338 }
18339 self.advance();
18340 let n = match self.advance() {
18341 Token::Integer(n) if n > 0 => u32::try_from(n).map_err(|_| ParseError {
18342 message: format!("{label} size too large: {n}"),
18343 token_pos: self.consumed_pos(),
18344 })?,
18345 other => {
18346 return Err(ParseError {
18347 message: format!("expected positive integer {label} size, got {other:?}"),
18348 token_pos: self.consumed_pos(),
18349 });
18350 }
18351 };
18352 if !matches!(self.peek(), Token::RParen) {
18353 return Err(self.err(format!(
18354 "expected ')' after {label} size, got {:?}",
18355 self.peek()
18356 )));
18357 }
18358 self.advance();
18359 Ok(n)
18360 }
18361
18362 /// v7.39 (round 406) — the `ON CONFLICT DO NOTHING` clause that MySQL's
18363 /// `INSERT IGNORE` lowers to: a bare target (arbitrate on every unique
18364 /// key, like MySQL) whose action skips conflicting rows.
18365 /// v7.39 (round 419) — resolve the conflict clause for ANY of the four
18366 /// INSERT source forms (VALUES / SELECT / parenthesized source / WITH).
18367 /// Before this the MySQL upsert lowerings (`ON DUPLICATE KEY UPDATE`,
18368 /// `REPLACE INTO`) were wired into the VALUES branch ONLY, so the very
18369 /// common bulk-upsert spellings —
18370 /// INSERT INTO t SELECT … ON DUPLICATE KEY UPDATE c = VALUES(c)
18371 /// REPLACE INTO t SELECT …
18372 /// — were a parse error / a duplicate-key failure respectively.
18373 ///
18374 /// Precedence: an explicitly written clause beats a statement-level flag.
18375 /// `ON DUPLICATE KEY UPDATE` first, then PG's own `ON CONFLICT`, then the
18376 /// implicit `REPLACE` and `IGNORE` lowerings.
18377 fn parse_insert_conflict_clause(
18378 &mut self,
18379 replace: bool,
18380 ignore: bool,
18381 ) -> Result<Option<crate::ast::OnConflictClause>, ParseError> {
18382 if let Some(c) = self.parse_optional_on_duplicate_key()? {
18383 return Ok(Some(c));
18384 }
18385 if let Some(c) = self.parse_optional_on_conflict()? {
18386 return Ok(Some(c));
18387 }
18388 if replace {
18389 // REPLACE INTO = delete-then-insert, which PG spells as
18390 // `ON CONFLICT DO UPDATE SET` over every column; the engine
18391 // reads an empty assignment list as "take the incoming row".
18392 return Ok(Some(crate::ast::OnConflictClause {
18393 target_columns: Vec::new(),
18394 index_where: None,
18395 constraint_name: None,
18396 mysql_lowered: true,
18397 action: crate::ast::OnConflictAction::Update {
18398 assignments: Vec::new(),
18399 where_: None,
18400 },
18401 }));
18402 }
18403 if ignore {
18404 return Ok(Some(Self::insert_ignore_clause()));
18405 }
18406 Ok(None)
18407 }
18408
18409 /// v7.39 (round 419, extracted from the VALUES branch) — MySQL's
18410 /// `ON DUPLICATE KEY UPDATE col = expr [, …]`. Bare target (MySQL
18411 /// watches every unique key, which `mysql_lowered` records); `VALUES(col)`
18412 /// in an assignment is MySQL's spelling of `EXCLUDED.col`.
18413 fn parse_optional_on_duplicate_key(
18414 &mut self,
18415 ) -> Result<Option<crate::ast::OnConflictClause>, ParseError> {
18416 if !(matches!(self.peek(), Token::On)
18417 && matches!(self.tokens.get(self.pos + 1),
18418 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("duplicate")))
18419 {
18420 return Ok(None);
18421 }
18422 self.advance(); // ON
18423 self.advance(); // DUPLICATE
18424 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("key")) {
18425 return Err(self.err(format!(
18426 "expected KEY after ON DUPLICATE, got {:?}",
18427 self.peek()
18428 )));
18429 }
18430 self.advance();
18431 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("update")) {
18432 return Err(self.err(format!(
18433 "expected UPDATE after ON DUPLICATE KEY, got {:?}",
18434 self.peek()
18435 )));
18436 }
18437 self.advance();
18438 let mut assignments: Vec<(String, Expr)> = Vec::new();
18439 loop {
18440 let col = self.expect_ident_like()?;
18441 if !matches!(self.peek(), Token::Eq) {
18442 return Err(self.err(format!(
18443 "expected '=' in ON DUPLICATE KEY UPDATE, got {:?}",
18444 self.peek()
18445 )));
18446 }
18447 self.advance();
18448 let mut expr = self.parse_expr(0)?;
18449 Self::rewrite_mysql_values_refs(&mut expr);
18450 assignments.push((col, expr));
18451 if matches!(self.peek(), Token::Comma) {
18452 self.advance();
18453 continue;
18454 }
18455 break;
18456 }
18457 Ok(Some(crate::ast::OnConflictClause {
18458 target_columns: Vec::new(),
18459 index_where: None,
18460 constraint_name: None,
18461 mysql_lowered: true,
18462 action: crate::ast::OnConflictAction::Update {
18463 assignments,
18464 where_: None,
18465 },
18466 }))
18467 }
18468
18469 fn insert_ignore_clause() -> crate::ast::OnConflictClause {
18470 crate::ast::OnConflictClause {
18471 target_columns: Vec::new(),
18472 index_where: None,
18473 constraint_name: None,
18474 mysql_lowered: true,
18475 action: crate::ast::OnConflictAction::Nothing,
18476 }
18477 }
18478
18479 fn parse_insert_stmt(&mut self, replace: bool) -> Result<Statement, ParseError> {
18480 debug_assert!(
18481 matches!(self.peek(), Token::Insert)
18482 || (replace
18483 && matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("replace")))
18484 );
18485 self.advance();
18486 // v7.39 (round 406) — MySQL `INSERT IGNORE INTO t …` skips a row that
18487 // would raise a duplicate-key error instead of failing the statement,
18488 // i.e. `ON CONFLICT DO NOTHING` over every unique key. IGNORE is a
18489 // plain ident to the lexer; only the MySQL dialect accepts it here.
18490 let ignore = self.mysql_dialect
18491 && matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("ignore"));
18492 if ignore {
18493 self.advance();
18494 }
18495 if !matches!(self.peek(), Token::Into) {
18496 return Err(self.err(format!("expected INTO after INSERT, got {:?}", self.peek())));
18497 }
18498 self.advance();
18499 let table = self.expect_ident_like()?;
18500 // v7.39 (round 240) — `INSERT INTO t AS alias`: PG's insert_target
18501 // grammar requires the AS keyword here (a bare identifier would be
18502 // ambiguous with a column list). The alias is what the ON CONFLICT
18503 // DO UPDATE expressions refer to the target row by.
18504 let alias = if matches!(self.peek(), Token::As) {
18505 self.advance();
18506 Some(self.expect_ident_like()?)
18507 } else {
18508 None
18509 };
18510 // v7.39 (round 428) — MySQL's SET-form INSERT:
18511 // INSERT INTO t SET a = 1, b = 'x'
18512 // It is exactly `INSERT INTO t (a, b) VALUES (1, 'x')` — omitted
18513 // columns take their DEFAULT, `SET a = DEFAULT` is legal, and it
18514 // composes with IGNORE / ON DUPLICATE KEY UPDATE / REPLACE (all
18515 // measured). So it lowers to the column list + one VALUES row and
18516 // rejoins the ordinary path, which already handles every one of
18517 // those. PG has no such spelling, hence the dialect gate.
18518 if self.mysql_dialect
18519 && matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("set"))
18520 {
18521 self.advance(); // SET
18522 let mut names = Vec::new();
18523 let mut values = Vec::new();
18524 loop {
18525 names.push(self.expect_ident_like()?);
18526 if !matches!(self.peek(), Token::Eq) {
18527 return Err(self.err(alloc::format!(
18528 "expected '=' in INSERT … SET, got {:?}",
18529 self.peek()
18530 )));
18531 }
18532 self.advance();
18533 // `SET a = DEFAULT` rides the same `__column_default` marker
18534 // the VALUES-row and UPDATE-SET paths use; the INSERT
18535 // executor resolves it against the target column.
18536 if matches!(self.peek(), Token::Default) {
18537 self.advance();
18538 values.push(Expr::FunctionCall {
18539 name: "__column_default".to_string(),
18540 args: Vec::new(),
18541 });
18542 } else {
18543 values.push(self.parse_expr(0)?);
18544 }
18545 if matches!(self.peek(), Token::Comma) {
18546 self.advance();
18547 continue;
18548 }
18549 break;
18550 }
18551 let on_conflict = self.parse_insert_conflict_clause(replace, ignore)?;
18552 let returning = self.parse_optional_returning()?;
18553 return Ok(Statement::Insert(InsertStatement {
18554 ctes: Vec::new(),
18555 table,
18556 alias,
18557 columns: Some(names),
18558 rows: alloc::vec![values],
18559 select_source: None,
18560 // MySQL's SET form has no `OVERRIDING …` clause (that is
18561 // PG's identity-column spelling).
18562 overriding: Overriding::None,
18563 mysql_ignore: ignore,
18564 on_conflict,
18565 returning,
18566 }));
18567 }
18568 // Optional column list — `INSERT INTO t (a, b) VALUES ...`.
18569 // v7.39 (round 151) — a SELECT or WITH right after the paren is
18570 // a parenthesized query source instead (PG select_with_parens:
18571 // `INSERT INTO t (SELECT …)` / `INSERT INTO t (WITH … SELECT …)`);
18572 // both keywords are reserved in PG, so no column list can start
18573 // with them.
18574 let columns = if matches!(self.peek(), Token::LParen) {
18575 self.advance();
18576 if matches!(self.peek(), Token::Select) || self.peek_is_with_kw() {
18577 let select_stmt = if self.peek_is_with_kw() {
18578 self.advance();
18579 self.parse_nested_with_select()?
18580 } else {
18581 match self.parse_select_stmt()? {
18582 Statement::Select(s) => s,
18583 other => {
18584 return Err(self.err(alloc::format!(
18585 "expected SELECT in parenthesized INSERT source, got {other:?}"
18586 )));
18587 }
18588 }
18589 };
18590 if !matches!(self.peek(), Token::RParen) {
18591 return Err(self.err(format!(
18592 "expected ')' after parenthesized INSERT source, got {:?}",
18593 self.peek()
18594 )));
18595 }
18596 self.advance();
18597 let on_conflict = self.parse_insert_conflict_clause(replace, ignore)?;
18598 let returning = self.parse_optional_returning()?;
18599 return Ok(Statement::Insert(InsertStatement {
18600 ctes: Vec::new(),
18601 table,
18602 alias: alias.clone(),
18603 columns: None,
18604 rows: Vec::new(),
18605 select_source: Some(Box::new(select_stmt)),
18606 on_conflict,
18607 returning,
18608 overriding: Overriding::None,
18609 mysql_ignore: ignore,
18610 }));
18611 }
18612 let mut names = Vec::new();
18613 loop {
18614 names.push(self.expect_ident_like()?);
18615 match self.peek() {
18616 Token::Comma => {
18617 self.advance();
18618 }
18619 Token::RParen => {
18620 self.advance();
18621 break;
18622 }
18623 other => {
18624 return Err(self.err(format!(
18625 "expected ',' or ')' in INSERT column list, got {other:?}"
18626 )));
18627 }
18628 }
18629 }
18630 Some(names)
18631 } else {
18632 None
18633 };
18634 // PG 10+ `OVERRIDING {SYSTEM | USER} VALUE` — pg_dump emits
18635 // OVERRIDING SYSTEM VALUE for its identity columns. The clause
18636 // is captured on the statement so the engine can apply PG's
18637 // GENERATED ALWAYS / BY DEFAULT interaction (v7.38, read01).
18638 let overriding = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("overriding"))
18639 {
18640 self.advance();
18641 let which = self.expect_ident_like()?;
18642 let ov = if which.eq_ignore_ascii_case("system") {
18643 Overriding::System
18644 } else if which.eq_ignore_ascii_case("user") {
18645 Overriding::User
18646 } else {
18647 return Err(self.err(format!(
18648 "expected SYSTEM or USER after OVERRIDING, got {which:?}"
18649 )));
18650 };
18651 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("value")) {
18652 return Err(self.err(format!(
18653 "expected VALUE after OVERRIDING {}, got {:?}",
18654 which.to_ascii_uppercase(),
18655 self.peek()
18656 )));
18657 }
18658 self.advance();
18659 ov
18660 } else {
18661 Overriding::None
18662 };
18663 // `INSERT INTO t DEFAULT VALUES` — a single row made
18664 // entirely of column defaults. Lower to the permuted
18665 // column-list path with an empty list: every schema column
18666 // is unmapped, so the engine fills each from its default
18667 // (serials advance, plain defaults evaluate, the rest NULL).
18668 if matches!(self.peek(), Token::Default) {
18669 self.advance();
18670 if !matches!(self.peek(), Token::Values) {
18671 return Err(self.err(format!(
18672 "expected VALUES after DEFAULT in INSERT, got {:?}",
18673 self.peek()
18674 )));
18675 }
18676 self.advance();
18677 if columns.is_some() {
18678 return Err(self.err("DEFAULT VALUES cannot follow an INSERT column list".into()));
18679 }
18680 let on_conflict = self.parse_insert_conflict_clause(replace, ignore)?;
18681 let returning = self.parse_optional_returning()?;
18682 return Ok(Statement::Insert(InsertStatement {
18683 ctes: Vec::new(),
18684 table,
18685 alias: alias.clone(),
18686 columns: Some(Vec::new()),
18687 rows: alloc::vec![Vec::new()],
18688 select_source: None,
18689 on_conflict,
18690 returning,
18691 overriding,
18692 mysql_ignore: ignore,
18693 }));
18694 }
18695 // v7.13.0 — `INSERT INTO t [(cols)] SELECT …` (mailrs
18696 // round-5 G4). Dispatch on VALUES vs SELECT. v7.39 (round 151)
18697 // — a WITH-headed source query (`INSERT INTO t WITH c AS (…)
18698 // SELECT …`) heads the SOURCE select, as in PG (the statement's
18699 // own WITH comes before INSERT).
18700 if matches!(self.peek(), Token::Select) || self.peek_is_with_kw() {
18701 let select_stmt = if self.peek_is_with_kw() {
18702 self.advance();
18703 self.parse_nested_with_select()?
18704 } else {
18705 match self.parse_select_stmt()? {
18706 Statement::Select(s) => s,
18707 other => {
18708 return Err(self.err(alloc::format!(
18709 "expected SELECT after INSERT INTO ... target, got {other:?}"
18710 )));
18711 }
18712 }
18713 };
18714 let on_conflict = self.parse_insert_conflict_clause(replace, ignore)?;
18715 let returning = self.parse_optional_returning()?;
18716 return Ok(Statement::Insert(InsertStatement {
18717 ctes: Vec::new(),
18718 table,
18719 alias: alias.clone(),
18720 columns,
18721 rows: Vec::new(),
18722 select_source: Some(Box::new(select_stmt)),
18723 on_conflict,
18724 returning,
18725 overriding,
18726 mysql_ignore: ignore,
18727 }));
18728 }
18729 if !matches!(self.peek(), Token::Values) {
18730 return Err(self.err(format!(
18731 "expected VALUES or SELECT after table name, got {:?}",
18732 self.peek()
18733 )));
18734 }
18735 self.advance();
18736 if !matches!(self.peek(), Token::LParen) {
18737 return Err(self.err(format!("expected '(' after VALUES, got {:?}", self.peek())));
18738 }
18739 let mut rows = Vec::new();
18740 loop {
18741 // Each iteration consumes one `(expr, expr, …)` tuple.
18742 if !matches!(self.peek(), Token::LParen) {
18743 return Err(self.err(format!(
18744 "expected '(' for next VALUES tuple, got {:?}",
18745 self.peek()
18746 )));
18747 }
18748 self.advance();
18749 let mut tuple = Vec::new();
18750 loop {
18751 // v7.38 (read01) — `INSERT INTO t VALUES (…, DEFAULT, …)` uses
18752 // the column's declared default for that slot. Rides out as the
18753 // same `__column_default` marker call the UPDATE `SET c = DEFAULT`
18754 // path uses; the INSERT executor resolves it per target column.
18755 if matches!(self.peek(), Token::Default) {
18756 self.advance();
18757 tuple.push(Expr::FunctionCall {
18758 name: "__column_default".to_string(),
18759 args: Vec::new(),
18760 });
18761 } else {
18762 tuple.push(self.parse_expr(0)?);
18763 }
18764 match self.peek() {
18765 Token::Comma => {
18766 self.advance();
18767 }
18768 Token::RParen => {
18769 self.advance();
18770 break;
18771 }
18772 other => {
18773 return Err(self.err(format!(
18774 "expected ',' or ')' in VALUES tuple, got {other:?}"
18775 )));
18776 }
18777 }
18778 }
18779 if tuple.is_empty() {
18780 return Err(self.err("INSERT VALUES tuple requires at least one value".into()));
18781 }
18782 rows.push(tuple);
18783 // Continue with comma-separated tuples.
18784 if matches!(self.peek(), Token::Comma) {
18785 self.advance();
18786 } else {
18787 break;
18788 }
18789 }
18790 // MySQL `ON DUPLICATE KEY UPDATE col = expr [, …]` — lowers
18791 // to ON CONFLICT DO UPDATE with an empty conflict target
18792 // (the engine picks the table's first unique index, which
18793 // matches MySQL's any-unique-key behaviour for the common
18794 // single-key case). `VALUES(col)` in the assignments is
18795 // MySQL's spelling of EXCLUDED.col.
18796 let on_conflict = self.parse_insert_conflict_clause(replace, ignore)?;
18797 let returning = self.parse_optional_returning()?;
18798 Ok(Statement::Insert(InsertStatement {
18799 ctes: Vec::new(),
18800 table,
18801 alias,
18802 columns,
18803 rows,
18804 select_source: None,
18805 on_conflict,
18806 returning,
18807 overriding,
18808 mysql_ignore: ignore,
18809 }))
18810 }
18811
18812 /// MySQL's `VALUES(col)` inside ON DUPLICATE KEY UPDATE reads
18813 /// the incoming row's value — exactly PG's EXCLUDED.col.
18814 fn rewrite_mysql_values_refs(e: &mut Expr) {
18815 match e {
18816 Expr::FunctionCall { name, args }
18817 if name.eq_ignore_ascii_case("values")
18818 && args.len() == 1
18819 && matches!(&args[0], Expr::Column(c) if c.qualifier.is_none()) =>
18820 {
18821 let Expr::Column(c) = &args[0] else {
18822 unreachable!("guarded above");
18823 };
18824 *e = Expr::Column(crate::ast::ColumnName {
18825 qualifier: Some("EXCLUDED".to_string()),
18826 name: c.name.clone(),
18827 });
18828 }
18829 Expr::FunctionCall { args, .. } => {
18830 for a in args {
18831 Self::rewrite_mysql_values_refs(a);
18832 }
18833 }
18834 Expr::Binary { lhs, rhs, .. } => {
18835 Self::rewrite_mysql_values_refs(lhs);
18836 Self::rewrite_mysql_values_refs(rhs);
18837 }
18838 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } => {
18839 Self::rewrite_mysql_values_refs(expr);
18840 }
18841 Expr::Case {
18842 operand,
18843 branches,
18844 else_branch,
18845 } => {
18846 if let Some(op) = operand {
18847 Self::rewrite_mysql_values_refs(op);
18848 }
18849 for (w, t) in branches {
18850 Self::rewrite_mysql_values_refs(w);
18851 Self::rewrite_mysql_values_refs(t);
18852 }
18853 if let Some(el) = else_branch {
18854 Self::rewrite_mysql_values_refs(el);
18855 }
18856 }
18857 _ => {}
18858 }
18859 }
18860
18861 /// v7.9.7 — parse the optional `ON CONFLICT (cols) DO …`
18862 /// clause sitting between the INSERT body and the trailing
18863 /// RETURNING. All keywords come in as bare idents; `ON` is
18864 /// a reserved Token though.
18865 fn parse_optional_on_conflict(
18866 &mut self,
18867 ) -> Result<Option<crate::ast::OnConflictClause>, ParseError> {
18868 if !matches!(self.peek(), Token::On) {
18869 return Ok(None);
18870 }
18871 // Peek further: we want exactly "ON CONFLICT ...". If the
18872 // next ident isn't "conflict", let some other parser handle.
18873 let next_is_conflict = matches!(
18874 self.tokens.get(self.pos + 1),
18875 Some(Token::Ident(s) | Token::QuotedIdent(s)) if s.eq_ignore_ascii_case("conflict")
18876 );
18877 if !next_is_conflict {
18878 return Ok(None);
18879 }
18880 self.advance(); // ON
18881 self.advance(); // CONFLICT
18882 // v7.37.17 (17.6 siblings) — `ON CONSTRAINT <name>` names
18883 // the constraint instead of listing columns (the pg_dump
18884 // form); the engine resolves it.
18885 let mut constraint_name: Option<String> = None;
18886 if matches!(self.peek(), Token::On) {
18887 self.advance(); // ON
18888 match self.advance() {
18889 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("constraint") => {
18890 }
18891 other => {
18892 return Err(self.err(alloc::format!(
18893 "expected CONSTRAINT after ON CONFLICT ON, got {other:?}"
18894 )));
18895 }
18896 }
18897 constraint_name = Some(self.expect_ident_like()?);
18898 }
18899 // Optional `(col [, col]*)` target list.
18900 let mut target_columns: Vec<String> = Vec::new();
18901 if matches!(self.peek(), Token::LParen) {
18902 self.advance();
18903 loop {
18904 target_columns.push(self.expect_ident_like()?);
18905 match self.peek() {
18906 Token::Comma => {
18907 self.advance();
18908 }
18909 Token::RParen => {
18910 self.advance();
18911 break;
18912 }
18913 other => {
18914 return Err(self.err(alloc::format!(
18915 "expected ',' or ')' in ON CONFLICT target list, got {other:?}"
18916 )));
18917 }
18918 }
18919 }
18920 }
18921 // v7.39 (round 240) — optional index predicate after the target
18922 // list: `ON CONFLICT (col) WHERE pred DO …`. PG uses it to infer a
18923 // PARTIAL unique index; SPG's arbiters are full indexes, which
18924 // satisfy any predicate, so it is parsed and carried but not
18925 // consulted (recorded residual: partial-unique-index arbiters).
18926 let index_where = if !target_columns.is_empty() && matches!(self.peek(), Token::Where) {
18927 self.advance();
18928 Some(self.parse_expr(0)?)
18929 } else {
18930 None
18931 };
18932 // Required `DO`.
18933 match self.advance() {
18934 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("do") => {}
18935 other => {
18936 return Err(self.err(alloc::format!(
18937 "expected DO after ON CONFLICT [(…)], got {other:?}"
18938 )));
18939 }
18940 }
18941 // Action: NOTHING | UPDATE SET …
18942 let action = match self.advance() {
18943 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("nothing") => {
18944 crate::ast::OnConflictAction::Nothing
18945 }
18946 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("update") => {
18947 self.parse_on_conflict_update_action()?
18948 }
18949 other => {
18950 return Err(self.err(alloc::format!(
18951 "expected NOTHING or UPDATE after ON CONFLICT DO, got {other:?}"
18952 )));
18953 }
18954 };
18955 Ok(Some(crate::ast::OnConflictClause {
18956 target_columns,
18957 index_where,
18958 constraint_name,
18959 mysql_lowered: false,
18960 action,
18961 }))
18962 }
18963
18964 /// v7.9.7 — tail of `ON CONFLICT … DO UPDATE`: parse
18965 /// `SET col = expr [, …] [WHERE cond]`. Caller already
18966 /// consumed `UPDATE`.
18967 fn parse_on_conflict_update_action(
18968 &mut self,
18969 ) -> Result<crate::ast::OnConflictAction, ParseError> {
18970 // `SET`
18971 match self.advance() {
18972 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("set") => {}
18973 other => {
18974 return Err(self.err(alloc::format!(
18975 "expected SET after ON CONFLICT DO UPDATE, got {other:?}"
18976 )));
18977 }
18978 }
18979 let mut assignments: Vec<(String, Expr)> = Vec::new();
18980 loop {
18981 let col = self.expect_ident_like()?;
18982 if !matches!(self.peek(), Token::Eq) {
18983 return Err(self.err(alloc::format!(
18984 "expected `=` after column in ON CONFLICT DO UPDATE SET, got {:?}",
18985 self.peek()
18986 )));
18987 }
18988 self.advance();
18989 let value = self.parse_expr(0)?;
18990 assignments.push((col, value));
18991 if matches!(self.peek(), Token::Comma) {
18992 self.advance();
18993 continue;
18994 }
18995 break;
18996 }
18997 let where_ = if matches!(self.peek(), Token::Where) {
18998 self.advance();
18999 Some(self.parse_expr(0)?)
19000 } else {
19001 None
19002 };
19003 Ok(crate::ast::OnConflictAction::Update {
19004 assignments,
19005 where_,
19006 })
19007 }
19008
19009 fn parse_select_list(&mut self) -> Result<Vec<SelectItem>, ParseError> {
19010 let mut items = Vec::new();
19011 // v7.39 (round 341, V66) — PG's target list may be EMPTY
19012 // (`opt_target_list: target_list | /*EMPTY*/`): `SELECT FROM t`
19013 // answers one zero-column row per row of t, and a bare `SELECT`
19014 // answers a single zero-column row. SPG required at least one
19015 // item, so both were syntax errors. Recognised by the token that
19016 // follows — nothing that can start an expression appears here.
19017 if self.select_list_is_empty_here() {
19018 return Ok(items);
19019 }
19020 loop {
19021 items.push(self.parse_select_item()?);
19022 if matches!(self.peek(), Token::Comma) {
19023 self.advance();
19024 } else {
19025 break;
19026 }
19027 }
19028 Ok(items)
19029 }
19030
19031 /// Is the target list empty at this point — i.e. does the next token
19032 /// end the SELECT's item list rather than start an item?
19033 fn select_list_is_empty_here(&self) -> bool {
19034 match self.peek() {
19035 Token::From
19036 | Token::Where
19037 | Token::Group
19038 | Token::Having
19039 | Token::Order
19040 | Token::Limit
19041 | Token::Offset
19042 | Token::Semicolon
19043 | Token::RParen
19044 | Token::Union
19045 | Token::Except
19046 | Token::Eof => true,
19047 // `FETCH FIRST … ROWS ONLY` and `WINDOW w AS …` are spelled
19048 // with unreserved keywords, so they arrive as plain idents.
19049 Token::Ident(s) => {
19050 s.eq_ignore_ascii_case("fetch")
19051 || s.eq_ignore_ascii_case("window")
19052 || s.eq_ignore_ascii_case("intersect")
19053 }
19054 _ => false,
19055 }
19056 }
19057
19058 fn parse_select_item(&mut self) -> Result<SelectItem, ParseError> {
19059 if matches!(self.peek(), Token::Star) {
19060 self.advance();
19061 return Ok(SelectItem::Wildcard);
19062 }
19063 // v7.39 (read01 round 128) — qualified wildcard `qualifier.*`. Intercept
19064 // BEFORE `parse_expr`, which would treat `q.` as a qualified column and
19065 // choke on the `*` ("expected identifier, got Star"). The lookahead is
19066 // `<ident> . *` with nothing binding tighter.
19067 if let Token::Ident(q) | Token::QuotedIdent(q) = self.peek().clone() {
19068 if matches!(self.tokens.get(self.pos + 1), Some(Token::Dot))
19069 && matches!(self.tokens.get(self.pos + 2), Some(Token::Star))
19070 {
19071 self.advance(); // qualifier
19072 self.advance(); // .
19073 self.advance(); // *
19074 return Ok(SelectItem::QualifiedWildcard(q));
19075 }
19076 }
19077 let start_tok = self.pos;
19078 let expr = self.parse_expr(0)?;
19079 let end_tok = self.consumed_pos();
19080 // v7.39 (read01 round 69) — `(f(args)).*`: expand the RECORD a
19081 // multi-column function returns into columns. Marked here and lowered in
19082 // `parse_bare_select`, where the FROM clause is in hand.
19083 if matches!(self.peek(), Token::Dot)
19084 && matches!(self.tokens.get(self.pos + 1), Some(Token::Star))
19085 {
19086 self.advance(); // .
19087 self.advance(); // *
19088 return Ok(SelectItem::Expr {
19089 expr: Expr::FunctionCall {
19090 name: "__record_expand".to_string(),
19091 args: alloc::vec![expr],
19092 },
19093 alias: None,
19094 });
19095 }
19096 // v7.39.2 — MySQL lets a STRING name a projection item, with or
19097 // without `AS`: `SELECT 1 'x'`, `SELECT COUNT(*) 'total'`,
19098 // `SELECT 1 'a b'` (which is why one quotes it). SPG answered
19099 // `syntax error at or near "'x'"` to all of them.
19100 //
19101 // Only here, not in `parse_optional_alias`: that one also names
19102 // TABLES, and MySQL 9.7.2 refuses a string there — `FROM t 'ta'`
19103 // and `FROM t AS 'ta'` are both syntax errors, measured. And only
19104 // after the lexer's own rule has joined adjacent literals, or
19105 // `SELECT 'a' 'b'` would read as a literal aliased `b` where
19106 // MySQL answers the concatenation `ab`.
19107 if self.mysql_dialect {
19108 let at_as = matches!(self.peek(), Token::As)
19109 && matches!(self.tokens.get(self.pos + 1), Some(Token::String(_)));
19110 if at_as {
19111 self.advance();
19112 }
19113 if let Token::String(name) = self.peek().clone() {
19114 self.advance();
19115 return Ok(SelectItem::Expr {
19116 expr,
19117 alias: Some(name),
19118 });
19119 }
19120 }
19121 let alias = match self.parse_optional_alias()? {
19122 Some(a) => Some(a),
19123 None => self.mysql_item_label(&expr, start_tok, end_tok),
19124 };
19125 Ok(SelectItem::Expr { expr, alias })
19126 }
19127
19128 /// v7.39 (round 506) — the name MariaDB 11 gives a projection item that
19129 /// carries no `AS`, filled in here so every downstream path reports it
19130 /// without knowing the rule. `None` leaves the item un-aliased, which is
19131 /// what a PG session always gets.
19132 ///
19133 /// Measured against MariaDB 11, three rules and no more:
19134 ///
19135 /// | item | label | why |
19136 /// |------------------|------------|------------------------------|
19137 /// | `lbl.a` | `a` | a column reports its name |
19138 /// | `'it''s'` | `it's` | a string reports its VALUE |
19139 /// | `a + b` | `a + b` | anything else, source text |
19140 ///
19141 /// The third is why this lives in the parser at all: the label is the
19142 /// text the client WROTE, down to the spacing, so it cannot be printed
19143 /// back out of the parsed shape. `COUNT( * )` names itself `COUNT( * )`.
19144 ///
19145 /// Comments survive, and that is right: through a `mariadb` CLI both
19146 /// servers answer `a + b` for `SELECT a /* c */ + b`, but that is the
19147 /// CLIENT stripping the comment before it sends. Asked over the raw
19148 /// protocol, MariaDB answers `a /* c */ + b` — byte for byte what this
19149 /// produces.
19150 fn mysql_item_label(&self, expr: &Expr, start_tok: usize, end_tok: usize) -> Option<String> {
19151 if !self.mysql_dialect {
19152 return None;
19153 }
19154 match expr {
19155 // A column already reports its own name downstream; naming it
19156 // again here would only re-state the qualifier the label drops.
19157 Expr::Column(_) => None,
19158 // v7.39.3 — `SELECT 'a' 'b'` is ONE literal whose value is
19159 // `ab`, and MySQL 9.7.2 names the column `a`: the label is
19160 // the first segment as written, not the joined value
19161 // (measured). The lexer logs where it joined them.
19162 Expr::Literal(Literal::String(v)) => Some(
19163 self.merged_first_len(start_tok)
19164 .and_then(|n| v.get(..n))
19165 .map_or_else(|| v.clone(), String::from),
19166 ),
19167 _ => self.source_span(start_tok, end_tok).map(str::to_string),
19168 }
19169 }
19170
19171 /// v7.37.17 (17.6 siblings) — parse `(row), (row), …` after a
19172 /// consumed VALUES keyword. Each row lowers to a constant SELECT
19173 /// with PG's default column1..columnN names; subsequent rows
19174 /// chain as UNION ALL peers. Shared by the FROM-position
19175 /// `( VALUES … )` arm and the top-level bare VALUES statement.
19176 fn parse_values_rows_body(&mut self) -> Result<SelectStatement, ParseError> {
19177 let mut row_selects: Vec<SelectStatement> = Vec::new();
19178 loop {
19179 if !matches!(self.peek(), Token::LParen) {
19180 return Err(self.err(alloc::format!(
19181 "expected '(' to start a VALUES row, got {:?}",
19182 self.peek()
19183 )));
19184 }
19185 self.advance(); // (
19186 let mut items: Vec<SelectItem> = Vec::new();
19187 loop {
19188 let expr = self.parse_expr(0)?;
19189 items.push(SelectItem::Expr {
19190 expr,
19191 alias: Some(alloc::format!("column{}", items.len() + 1)),
19192 });
19193 match self.peek() {
19194 Token::Comma => {
19195 self.advance();
19196 }
19197 Token::RParen => break,
19198 other => {
19199 return Err(self.err(alloc::format!(
19200 "expected ',' or ')' in VALUES row, got {other:?}"
19201 )));
19202 }
19203 }
19204 }
19205 self.advance(); // )
19206 row_selects.push(SelectStatement {
19207 locking: None,
19208 ctes: Vec::new(),
19209 distinct: false,
19210 distinct_on: Vec::new(),
19211 items,
19212 from: None,
19213 where_: None,
19214 group_by: None,
19215 group_by_all: false,
19216 having: None,
19217 unions: Vec::new(),
19218 order_by: Vec::new(),
19219 limit: None,
19220 offset: None,
19221 limit_with_ties: false,
19222 window_check_exprs: Vec::new(),
19223 });
19224 if matches!(self.peek(), Token::Comma) {
19225 self.advance();
19226 continue;
19227 }
19228 break;
19229 }
19230 let mut head = row_selects.remove(0);
19231 head.unions = row_selects
19232 .into_iter()
19233 .map(|s| (UnionKind::All, s))
19234 .collect();
19235 Ok(head)
19236 }
19237
19238 fn parse_table_ref(&mut self) -> Result<TableRef, ParseError> {
19239 // v7.39 (round 621) — `FROM ONLY <table>` excludes a table's
19240 // children. It was read as a table NAMED `only`, so the query
19241 // failed on `relation "only" does not exist`.
19242 //
19243 // v7.39 (round 644) — and it is no longer a no-op. Round 621
19244 // absorbed the keyword, reasoning that SPG's children are
19245 // separate relations a plain scan does not descend into, so ONLY
19246 // already described the scan. That stopped being true when a
19247 // partition parent started unioning its children: measured,
19248 // `SELECT count(*) FROM ONLY <partitioned parent>` answered 2
19249 // where PG answers 0. The flag is carried now.
19250 let mut only = false;
19251 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("only"))
19252 && matches!(
19253 self.tokens.get(self.pos + 1),
19254 Some(Token::Ident(_) | Token::QuotedIdent(_))
19255 )
19256 {
19257 only = true;
19258 self.advance();
19259 }
19260 // `LATERAL generate_series(...)` / `LATERAL unnest(...)` —
19261 // for these SRFs the keyword is noise at parse time: the
19262 // join executor already substitutes outer-column references
19263 // into unnest_expr / generate_series_args per outer row
19264 // (v7.37.43-T4.5 substitute_outer_in_table_ref), and PG
19265 // licences the correlation even without the keyword. Absorb
19266 // it and fall through to the SRF arms below.
19267 // v7.39 (read01 round 69) — `LATERAL <fn>(args)` for ANY function, not
19268 // just the four builtin SRFs: a user set-returning function on a JOIN's
19269 // right side is the whole point of LATERAL. The keyword stays noise at
19270 // parse time — the join executor substitutes the outer row into the
19271 // call's arguments per outer row.
19272 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("lateral"))
19273 && matches!(
19274 self.tokens.get(self.pos + 1),
19275 // The json_each family has its OWN `LATERAL …` arm below, which
19276 // needs to see the keyword — absorbing it here would send those
19277 // calls down the generic table-function channel instead.
19278 Some(Token::Ident(s) | Token::QuotedIdent(s)) if !is_json_each_name(s)
19279 )
19280 && matches!(self.tokens.get(self.pos + 2), Some(Token::LParen))
19281 {
19282 self.advance(); // LATERAL
19283 }
19284 // v7.37.43-T4.5 — `LATERAL jsonb_each_text(<expr>)` —
19285 // set-returning function whose argument may reference a
19286 // preceding FROM item. We rewrite this to
19287 // `LATERAL (SELECT key, value FROM jsonb_each_text(<expr>)
19288 // AS __srf__) AS <alias>` so the existing LATERAL subquery
19289 // executor handles per-outer-row evaluation and the
19290 // SRF-primary jsonb_each_text path handles the inner
19291 // materialisation. Sentori 0067 backfill is the dogfood
19292 // shape.
19293 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("lateral"))
19294 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s) | Token::QuotedIdent(s)) if is_json_each_name(s))
19295 && matches!(self.tokens.get(self.pos + 2), Some(Token::LParen))
19296 {
19297 self.advance(); // LATERAL
19298 let each_fn = match self.peek() {
19299 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
19300 _ => unreachable!(),
19301 };
19302 self.advance(); // jsonb_each[_text] / json_each[_text]
19303 self.advance(); // (
19304 let arg = self.parse_expr(0)?;
19305 if !matches!(self.peek(), Token::RParen) {
19306 return Err(self.err(alloc::format!(
19307 "expected ')' after LATERAL {each_fn}() argument, got {:?}",
19308 self.peek()
19309 )));
19310 }
19311 self.advance();
19312 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
19313 let alias = alias_ident.clone().unwrap_or_else(|| each_fn.clone());
19314 // Synthesise: SELECT __srf__.key AS <key_alias>, __srf__.value AS <value_alias>
19315 // FROM jsonb_each_text(<arg>) AS __srf__
19316 // PG's `AS kv(key, value)` column-alias list maps
19317 // positions to names; default to (key, value) when
19318 // omitted (matching the SRF's natural column names).
19319 let srf_alias = "__srf__".to_string();
19320 let key_alias = column_aliases
19321 .first()
19322 .cloned()
19323 .unwrap_or_else(|| "key".to_string());
19324 let value_alias = column_aliases
19325 .get(1)
19326 .cloned()
19327 .unwrap_or_else(|| "value".to_string());
19328 let inner_select = crate::ast::SelectStatement {
19329 locking: None,
19330 ctes: Vec::new(),
19331 distinct: false,
19332 distinct_on: Vec::new(),
19333 items: alloc::vec![
19334 crate::ast::SelectItem::Expr {
19335 expr: crate::ast::Expr::Column(crate::ast::ColumnName {
19336 qualifier: Some(srf_alias.clone()),
19337 name: "key".to_string(),
19338 }),
19339 alias: Some(key_alias),
19340 },
19341 crate::ast::SelectItem::Expr {
19342 expr: crate::ast::Expr::Column(crate::ast::ColumnName {
19343 qualifier: Some(srf_alias.clone()),
19344 name: "value".to_string(),
19345 }),
19346 alias: Some(value_alias),
19347 },
19348 ],
19349 from: Some(crate::ast::FromClause {
19350 primary: TableRef {
19351 name: srf_alias.clone(),
19352 alias: Some(srf_alias.clone()),
19353 only: false,
19354 as_of_segment: None,
19355 unnest_expr: None,
19356 unnest_column_aliases: Vec::new(),
19357 with_ordinality: false,
19358 generate_series_args: None,
19359 lateral_subquery: None,
19360 jsonb_each_text_arg: Some((each_fn, Box::new(arg))),
19361 table_fn_call: None,
19362 rows_from: None,
19363 json_table: None,
19364 scalar_fn_item: false,
19365 },
19366 joins: Vec::new(),
19367 }),
19368 where_: None,
19369 group_by: None,
19370 group_by_all: false,
19371 having: None,
19372 unions: Vec::new(),
19373 order_by: Vec::new(),
19374 limit: None,
19375 offset: None,
19376 limit_with_ties: false,
19377 window_check_exprs: Vec::new(),
19378 };
19379 return Ok(TableRef {
19380 name: alias.clone(),
19381 alias: Some(alias),
19382 only: false,
19383 as_of_segment: None,
19384 unnest_expr: None,
19385 unnest_column_aliases: Vec::new(),
19386 with_ordinality: false,
19387 generate_series_args: None,
19388 lateral_subquery: Some(Box::new(inner_select)),
19389 jsonb_each_text_arg: None,
19390 table_fn_call: None,
19391 rows_from: None,
19392 json_table: None,
19393 scalar_fn_item: false,
19394 });
19395 }
19396 // v7.37.43-T4.5 — bare `CROSS JOIN jsonb_each_text(t.col)`
19397 // without an explicit `LATERAL` keyword is the same shape
19398 // PG accepts (SRF naturally licences lateral correlation).
19399 // We mirror the LATERAL rewrite when the argument syntactic-
19400 // ally references an outer column (Column { qualifier:
19401 // Some(_), … }). For simplicity we apply the rewrite
19402 // whenever the SRF directly follows JOIN/CROSS JOIN/comma
19403 // in the FROM-list — caller-side join parsing positions
19404 // this peek correctly.
19405 // (Implementation note: detection lives below; the LATERAL
19406 // branch above already covers the explicit form; the bare
19407 // form falls through to the plain SRF arm and the engine
19408 // treats it as a constant-arg SRF if no outer reference is
19409 // present.)
19410 // v7.17.0 Phase 3.P0-41 — `LATERAL ( SELECT … )` derived
19411 // table. Detect at the head so it claims precedence over
19412 // every other table-ref shape (unnest / generate_series /
19413 // bare ident); the lateral subquery itself follows the
19414 // regular SELECT grammar.
19415 // v7.37.17 (17.6 siblings) — `FROM ( VALUES (…), (…) ) [AS]
19416 // t(cols)`. Each row lowers to a constant SELECT with PG's
19417 // default column1..columnN names; subsequent rows chain as
19418 // UNION ALL peers. The result rides the derived-table
19419 // lateral_subquery channel — zero executor work.
19420 if matches!(self.peek(), Token::LParen)
19421 && matches!(self.tokens.get(self.pos + 1), Some(Token::Values))
19422 {
19423 self.advance(); // (
19424 self.advance(); // VALUES
19425 let head = self.parse_values_rows_body()?;
19426 if !matches!(self.peek(), Token::RParen) {
19427 return Err(self.err(alloc::format!(
19428 "expected ')' after VALUES list, got {:?}",
19429 self.peek()
19430 )));
19431 }
19432 self.advance();
19433 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
19434 let name = alias_ident.clone().unwrap_or_else(|| "values".to_string());
19435 return Ok(TableRef {
19436 name,
19437 alias: alias_ident,
19438 only: false,
19439 as_of_segment: None,
19440 unnest_expr: None,
19441 unnest_column_aliases: column_aliases,
19442 with_ordinality: false,
19443 generate_series_args: None,
19444 lateral_subquery: Some(Box::new(head)),
19445 jsonb_each_text_arg: None,
19446 table_fn_call: None,
19447 rows_from: None,
19448 json_table: None,
19449 scalar_fn_item: false,
19450 });
19451 }
19452 // v7.37.17 (17.6 siblings) — plain derived table:
19453 // `FROM ( SELECT … ) [AS] alias`. Rides the same
19454 // lateral_subquery channel the explicit LATERAL form uses —
19455 // an uncorrelated inner SELECT executes identically. The
19456 // inner parse carries UNION tails (they live on
19457 // SelectStatement.unions).
19458 // v7.37 D.20 — the derived-table inner may itself be a
19459 // parenthesized set-operation group (`FROM ((SELECT…) UNION
19460 // (SELECT…)) s`) or a CTE (`FROM (WITH … SELECT …) z`), not just a
19461 // bare `(SELECT …)`. parse_one_statement already routes a leading
19462 // `(` set-op group (its LParen arm) and a leading WITH
19463 // (parse_with_cte_then_select), so widen the second-token gate to
19464 // Select | LParen | WITH.
19465 // v7.39 (round 869) — `Table` joins that gate. `TABLE t` is
19466 // PG's spelling of `SELECT * FROM t` and is accepted wherever a
19467 // SELECT is, so `FROM (TABLE t) x` has to parse. The desugaring
19468 // has existed since the shorthand landed and `parse_bare_select`
19469 // already routes it ("valid anywhere a SELECT head is"); what was
19470 // missing is this second-token gate, and the CTE body's dispatch
19471 // below. Round 868 found both by putting the shorthand in a
19472 // subquery — the top-level forms had been the only ones tested.
19473 if matches!(self.peek(), Token::LParen)
19474 && (matches!(
19475 self.tokens.get(self.pos + 1),
19476 Some(Token::Select | Token::LParen | Token::Table)
19477 ) || matches!(self.tokens.get(self.pos + 1),
19478 Some(Token::Ident(s) | Token::QuotedIdent(s)) if s.eq_ignore_ascii_case("with")))
19479 {
19480 self.advance(); // (
19481 let inner = match self.parse_one_statement()? {
19482 Statement::Select(s) => s,
19483 other => {
19484 return Err(self.err(alloc::format!(
19485 "expected SELECT inside derived table ( … ), got {other:?}"
19486 )));
19487 }
19488 };
19489 if !matches!(self.peek(), Token::RParen) {
19490 return Err(self.err(alloc::format!(
19491 "expected ')' after derived-table subquery, got {:?}",
19492 self.peek()
19493 )));
19494 }
19495 self.advance();
19496 // `AS t(a, b)` column-alias list rides the
19497 // unnest_column_aliases field (same positional-rename
19498 // contract the unnest SRFs use).
19499 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
19500 let name = alias_ident
19501 .clone()
19502 .unwrap_or_else(|| "subquery".to_string());
19503 return Ok(TableRef {
19504 name,
19505 alias: alias_ident,
19506 only: false,
19507 as_of_segment: None,
19508 unnest_expr: None,
19509 unnest_column_aliases: column_aliases,
19510 with_ordinality: false,
19511 generate_series_args: None,
19512 lateral_subquery: Some(Box::new(inner)),
19513 jsonb_each_text_arg: None,
19514 table_fn_call: None,
19515 rows_from: None,
19516 json_table: None,
19517 scalar_fn_item: false,
19518 });
19519 }
19520 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("lateral"))
19521 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
19522 {
19523 self.advance(); // LATERAL
19524 self.advance(); // (
19525 // Parse the inner SELECT.
19526 let inner = match self.parse_one_statement()? {
19527 Statement::Select(s) => s,
19528 other => {
19529 return Err(self.err(alloc::format!(
19530 "expected SELECT inside LATERAL ( … ), got {other:?}"
19531 )));
19532 }
19533 };
19534 if !matches!(self.peek(), Token::RParen) {
19535 return Err(self.err(alloc::format!(
19536 "expected ')' after LATERAL subquery, got {:?}",
19537 self.peek()
19538 )));
19539 }
19540 self.advance();
19541 // v7.37 D.28 — `LATERAL (…) AS t(cols)` column-alias list (also how a
19542 // `(VALUES …) t(g)` derived table round-trips through view-body
19543 // Display, which renders on the lateral_subquery channel).
19544 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
19545 let name = alias_ident.clone().unwrap_or_else(|| "lateral".to_string());
19546 return Ok(TableRef {
19547 name,
19548 alias: alias_ident,
19549 only: false,
19550 as_of_segment: None,
19551 unnest_expr: None,
19552 unnest_column_aliases: column_aliases,
19553 with_ordinality: false,
19554 generate_series_args: None,
19555 lateral_subquery: Some(Box::new(inner)),
19556 jsonb_each_text_arg: None,
19557 table_fn_call: None,
19558 rows_from: None,
19559 json_table: None,
19560 scalar_fn_item: false,
19561 });
19562 }
19563 // v7.37.43-T4.5 — `jsonb_each_text(<expr>)` set-returning
19564 // function as a FROM item. Emits one row per (key, value)
19565 // pair in the JSONB object argument as TEXT columns. May
19566 // be wrapped in CROSS JOIN LATERAL when the argument
19567 // references a preceding FROM item (sentori migration
19568 // 0067 backfill shape: `CROSS JOIN LATERAL
19569 // jsonb_each_text(t.json_col) AS kv(key, value)`).
19570 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if is_json_each_name(s))
19571 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
19572 {
19573 let each_fn = match self.peek() {
19574 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
19575 _ => unreachable!(),
19576 };
19577 self.advance(); // jsonb_each[_text] / json_each[_text]
19578 self.advance(); // (
19579 let arg = self.parse_expr(0)?;
19580 if !matches!(self.peek(), Token::RParen) {
19581 return Err(self.err(alloc::format!(
19582 "expected ')' after {each_fn}() argument, got {:?}",
19583 self.peek()
19584 )));
19585 }
19586 self.advance();
19587 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
19588 let name = alias_ident.clone().unwrap_or_else(|| each_fn.clone());
19589 return Ok(TableRef {
19590 name,
19591 alias: alias_ident,
19592 only: false,
19593 as_of_segment: None,
19594 unnest_expr: None,
19595 // `AS t(k, v)` renames key/value positionally, same as the
19596 // LATERAL-position form already does.
19597 unnest_column_aliases: column_aliases,
19598 with_ordinality: false,
19599 generate_series_args: None,
19600 lateral_subquery: None,
19601 jsonb_each_text_arg: Some((each_fn, Box::new(arg))),
19602 table_fn_call: None,
19603 rows_from: None,
19604 json_table: None,
19605 scalar_fn_item: false,
19606 });
19607 }
19608 // `jsonb_to_recordset(J) AS t(a int, b text)` / `jsonb_to_record`
19609 // (+ json_ variants) — record-returning JSON functions with a
19610 // column-definition list. Desugar to a derived table that
19611 // projects each declared column from the JSON via `->>` + a cast,
19612 // over `jsonb_array_elements(J)` for the *set (per-element) form.
19613 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if is_json_to_record_name(s))
19614 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
19615 {
19616 return self.parse_json_to_record_from();
19617 }
19618 // v7.38 (T15) — `regexp_matches(s, pat[, flags])` as a FROM item. Each
19619 // row is a text[] of capture groups, so it cannot desugar to unnest
19620 // (that would flatten the array). Wrap it as a derived table
19621 // `(SELECT regexp_matches(args)) AS <alias>(<col>)` — the SELECT-list
19622 // SRF path already emits one text[] row per match. PG names the column
19623 // `regexp_matches`; an `AS a(col)` alias overrides it.
19624 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
19625 if s.eq_ignore_ascii_case("regexp_matches"))
19626 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
19627 {
19628 self.advance(); // fn name
19629 self.advance(); // (
19630 let mut fn_args: Vec<Expr> = Vec::new();
19631 loop {
19632 fn_args.push(self.parse_expr(0)?);
19633 if matches!(self.peek(), Token::Comma) {
19634 self.advance();
19635 continue;
19636 }
19637 break;
19638 }
19639 if !matches!(self.peek(), Token::RParen) {
19640 return Err(self.err(alloc::format!(
19641 "expected ')' after regexp_matches() arguments, got {:?}",
19642 self.peek()
19643 )));
19644 }
19645 self.advance();
19646 // v7.39 (read01 round 78) — WITH ORDINALITY sits BEFORE the alias
19647 // (`f(x) WITH ORDINALITY AS t(v, o)`), and this arm never looked for
19648 // it, so it died on the `with` token while every other table function
19649 // accepted it.
19650 let with_ordinality = self.absorb_with_ordinality();
19651 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
19652 let table_alias = alias_ident
19653 .clone()
19654 .unwrap_or_else(|| "regexp_matches".to_string());
19655 // PG names a single-column function's output column after the ALIAS
19656 // when one is given (`FROM regexp_matches(…) AS m` → column `m`), so
19657 // `m` reads as that column and not as a whole-row composite. Naming
19658 // it after the function regardless made `SELECT m[1] FROM … AS m`
19659 // subscript a record.
19660 let col_name = column_aliases
19661 .first()
19662 .cloned()
19663 .or_else(|| alias_ident.clone())
19664 .unwrap_or_else(|| "regexp_matches".to_string());
19665 let inner = crate::ast::SelectStatement {
19666 locking: None,
19667 ctes: Vec::new(),
19668 distinct: false,
19669 distinct_on: Vec::new(),
19670 items: alloc::vec![SelectItem::Expr {
19671 expr: Expr::FunctionCall {
19672 name: "regexp_matches".to_string(),
19673 args: fn_args,
19674 },
19675 alias: Some(col_name),
19676 }],
19677 from: None,
19678 where_: None,
19679 group_by: None,
19680 group_by_all: false,
19681 having: None,
19682 unions: Vec::new(),
19683 order_by: Vec::new(),
19684 limit: None,
19685 offset: None,
19686 limit_with_ties: false,
19687 window_check_exprs: Vec::new(),
19688 };
19689 return Ok(TableRef {
19690 name: table_alias.clone(),
19691 alias: Some(table_alias),
19692 only: false,
19693 as_of_segment: None,
19694 unnest_expr: None,
19695 unnest_column_aliases: column_aliases,
19696 with_ordinality,
19697 generate_series_args: None,
19698 lateral_subquery: Some(Box::new(inner)),
19699 jsonb_each_text_arg: None,
19700 table_fn_call: None,
19701 rows_from: None,
19702 json_table: None,
19703 // regexp_matches returns text[], a base type: `SELECT m FROM
19704 // regexp_matches(…) AS m` is the array, not a composite wrapping it.
19705 scalar_fn_item: true,
19706 });
19707 }
19708 // v7.37.17 (17.6 siblings) — `jsonb_array_elements[_text](<expr>)`
19709 // / json_ variants as a FROM item. Rewritten into
19710 // `unnest(<same fn>(<expr>))`: the scalar form returns the
19711 // elements as a TEXT array, and the existing unnest SRF path
19712 // materialises one row per element. PG's natural column name
19713 // is `value`; an `AS a(col)` column-alias list overrides it.
19714 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
19715 if s.eq_ignore_ascii_case("jsonb_array_elements")
19716 || s.eq_ignore_ascii_case("json_array_elements")
19717 || s.eq_ignore_ascii_case("jsonb_array_elements_text")
19718 || s.eq_ignore_ascii_case("json_array_elements_text")
19719 || s.eq_ignore_ascii_case("jsonb_object_keys")
19720 || s.eq_ignore_ascii_case("json_object_keys")
19721 || s.eq_ignore_ascii_case("jsonb_path_query")
19722 || s.eq_ignore_ascii_case("json_path_query")
19723 || s.eq_ignore_ascii_case("generate_subscripts")
19724 || s.eq_ignore_ascii_case("string_to_table")
19725 || s.eq_ignore_ascii_case("regexp_split_to_table"))
19726 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
19727 {
19728 let fn_name = match self.peek() {
19729 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
19730 _ => unreachable!(),
19731 };
19732 self.advance(); // fn name
19733 self.advance(); // (
19734 let mut fn_args: Vec<Expr> = Vec::new();
19735 loop {
19736 fn_args.push(self.parse_expr(0)?);
19737 if matches!(self.peek(), Token::Comma) {
19738 self.advance();
19739 continue;
19740 }
19741 break;
19742 }
19743 if !matches!(self.peek(), Token::RParen) {
19744 return Err(self.err(alloc::format!(
19745 "expected ')' after {fn_name}() arguments, got {:?}",
19746 self.peek()
19747 )));
19748 }
19749 self.advance();
19750 let with_ordinality = self.absorb_with_ordinality();
19751 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
19752 let name = alias_ident.clone().unwrap_or_else(|| fn_name.clone());
19753 // PG's natural column name: the array-elements SRFs
19754 // declare an OUT parameter `value`; jsonb_object_keys
19755 // and generate_subscripts have none, so the column is
19756 // named after the function. A bare table alias on a
19757 // single-column SRF renames the column too (PG: `FROM
19758 // generate_subscripts(a, 1) AS s` projects column s) —
19759 // except for the OUT-parameter SRFs, whose column stays
19760 // `value` under a bare alias.
19761 let natural_col = if fn_name.ends_with("_array_elements")
19762 || fn_name.ends_with("_array_elements_text")
19763 {
19764 "value".to_string()
19765 } else {
19766 alias_ident.clone().unwrap_or_else(|| fn_name.clone())
19767 };
19768 let mut srf_cols = alloc::vec![column_aliases.first().cloned().unwrap_or(natural_col)];
19769 // Keep any further entries — the second names the
19770 // ordinality column under WITH ORDINALITY.
19771 srf_cols.extend(column_aliases.into_iter().skip(1));
19772 // The *_to_table SRFs are row-streams over the existing
19773 // *_to_array scalars — map the call target; the display
19774 // name (alias / column defaults) keeps the SRF spelling.
19775 let call_name = match fn_name.as_str() {
19776 "string_to_table" => "string_to_array".to_string(),
19777 "regexp_split_to_table" => "regexp_split_to_array".to_string(),
19778 _ => fn_name,
19779 };
19780 // v7.38 (read01, T-srf/T-lateral) — an SRF argument that references a
19781 // preceding FROM item (bare or qualified column) is correlated;
19782 // route it through the per-outer-row lateral channel.
19783 let expr = crate::ast::Expr::FunctionCall {
19784 name: call_name,
19785 args: fn_args,
19786 };
19787 let correlated = Self::expr_has_any_column(&expr);
19788 let tref = TableRef {
19789 name,
19790 alias: alias_ident,
19791 only: false,
19792 as_of_segment: None,
19793 unnest_expr: Some(Box::new(expr)),
19794 unnest_column_aliases: srf_cols,
19795 with_ordinality,
19796 generate_series_args: None,
19797 lateral_subquery: None,
19798 jsonb_each_text_arg: None,
19799 table_fn_call: None,
19800 rows_from: None,
19801 json_table: None,
19802 // Each of these returns a BASE type (jsonb / text / int), so the item's
19803 // row type is that scalar: `SELECT j FROM jsonb_array_elements('[1]') j`
19804 // is `1`, not `(1)`. WITH ORDINALITY makes it a real two-column item.
19805 scalar_fn_item: !with_ordinality,
19806 };
19807 return Ok(if correlated {
19808 Self::wrap_correlated_srf(tref)
19809 } else {
19810 tref
19811 });
19812 }
19813 // `ROWS FROM ( srf(args) [, srf(args)]* )` — SQL-standard
19814 // explicit parallel-zip syntax. Each entry lowers to its
19815 // array-returning scalar form (unnest(x) → x itself; the
19816 // FROM-SRF rewrite family → their scalar array calls) and
19817 // the list rides the multi-arg unnest zip channel:
19818 // NULL-padded to the longest, WITH ORDINALITY appends the
19819 // counter. generate_series has no scalar array form and
19820 // errors honestly.
19821 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("rows"))
19822 && matches!(self.tokens.get(self.pos + 1), Some(Token::From))
19823 && matches!(self.tokens.get(self.pos + 2), Some(Token::LParen))
19824 {
19825 self.advance(); // ROWS
19826 self.advance(); // FROM
19827 self.advance(); // (
19828 let mut entries: Vec<Expr> = Vec::new();
19829 // v7.39 (read01 round 74) — the generic channel, filled in parallel.
19830 // Used only when some entry has no array form.
19831 let mut generic: Vec<(String, Vec<Expr>)> = Vec::new();
19832 loop {
19833 let fn_name = self.expect_ident_like()?.to_ascii_lowercase();
19834 if !matches!(self.peek(), Token::LParen) {
19835 return Err(self.err(alloc::format!(
19836 "expected '(' after {fn_name} in ROWS FROM, got {:?}",
19837 self.peek()
19838 )));
19839 }
19840 self.advance();
19841 let mut fn_args: Vec<Expr> = Vec::new();
19842 if !matches!(self.peek(), Token::RParen) {
19843 loop {
19844 fn_args.push(self.parse_expr(0)?);
19845 if matches!(self.peek(), Token::Comma) {
19846 self.advance();
19847 continue;
19848 }
19849 break;
19850 }
19851 }
19852 if !matches!(self.peek(), Token::RParen) {
19853 return Err(self.err(alloc::format!(
19854 "expected ')' after {fn_name}() arguments in ROWS FROM, got {:?}",
19855 self.peek()
19856 )));
19857 }
19858 self.advance();
19859 let entry = match fn_name.as_str() {
19860 "unnest" => {
19861 if fn_args.len() != 1 {
19862 return Err(
19863 self.err("unnest inside ROWS FROM takes exactly one array".into())
19864 );
19865 }
19866 fn_args.pop().expect("len checked")
19867 }
19868 "jsonb_array_elements"
19869 | "json_array_elements"
19870 | "jsonb_array_elements_text"
19871 | "json_array_elements_text"
19872 | "jsonb_object_keys"
19873 | "json_object_keys"
19874 | "generate_subscripts" => crate::ast::Expr::FunctionCall {
19875 name: fn_name,
19876 args: fn_args,
19877 },
19878 "string_to_table" => crate::ast::Expr::FunctionCall {
19879 name: "string_to_array".to_string(),
19880 args: fn_args,
19881 },
19882 "regexp_split_to_table" => crate::ast::Expr::FunctionCall {
19883 name: "regexp_split_to_array".to_string(),
19884 args: fn_args,
19885 },
19886 // v7.39 (read01 round 74) — an SRF with no array form
19887 // (`generate_series`, a user `RETURNS SETOF` function) has no
19888 // scalar expression to zip, so the WHOLE list switches to the
19889 // rows_from channel, which runs each function and zips the
19890 // rows themselves. The all-array case keeps the old lowering:
19891 // it is well-trodden and this must not disturb it.
19892 _ => {
19893 generic.push((fn_name, fn_args));
19894 if matches!(self.peek(), Token::Comma) {
19895 self.advance();
19896 continue;
19897 }
19898 break;
19899 }
19900 };
19901 generic.push((
19902 // The array-able entries carry their lowered expr along, so a
19903 // MIXED list still works: the engine sees the scalar array
19904 // form and unnests it.
19905 "__array".to_string(),
19906 alloc::vec![entry.clone()],
19907 ));
19908 entries.push(entry);
19909 if matches!(self.peek(), Token::Comma) {
19910 self.advance();
19911 continue;
19912 }
19913 break;
19914 }
19915 if !matches!(self.peek(), Token::RParen) {
19916 return Err(self.err(alloc::format!(
19917 "expected ')' to close ROWS FROM, got {:?}",
19918 self.peek()
19919 )));
19920 }
19921 self.advance();
19922 let with_ordinality = self.absorb_with_ordinality();
19923 let (alias_ident, unnest_column_aliases) = self.parse_optional_alias_with_columns()?;
19924 let name = alias_ident.clone().unwrap_or_else(|| "rows".to_string());
19925 // v7.39 (read01 round 74) — some entry had no array form, so the whole
19926 // list rides the generic channel.
19927 if generic.iter().any(|(n, _)| n != "__array") {
19928 let correlated = generic
19929 .iter()
19930 .any(|(_, a)| a.iter().any(Self::expr_has_any_column));
19931 let tref = TableRef {
19932 name,
19933 alias: alias_ident,
19934 only: false,
19935 as_of_segment: None,
19936 unnest_expr: None,
19937 unnest_column_aliases,
19938 with_ordinality,
19939 generate_series_args: None,
19940 lateral_subquery: None,
19941 jsonb_each_text_arg: None,
19942 table_fn_call: None,
19943 rows_from: Some(generic),
19944 json_table: None,
19945 scalar_fn_item: false,
19946 };
19947 return Ok(if correlated {
19948 Self::wrap_correlated_srf(tref)
19949 } else {
19950 tref
19951 });
19952 }
19953 let correlated = entries.iter().any(Self::expr_has_any_column);
19954 let expr = if entries.len() == 1 {
19955 entries.pop().expect("len checked")
19956 } else {
19957 crate::ast::Expr::FunctionCall {
19958 name: "__unnest_zip".to_string(),
19959 args: entries,
19960 }
19961 };
19962 let tref = TableRef {
19963 name,
19964 alias: alias_ident,
19965 only: false,
19966 as_of_segment: None,
19967 unnest_expr: Some(Box::new(expr)),
19968 unnest_column_aliases,
19969 with_ordinality,
19970 generate_series_args: None,
19971 lateral_subquery: None,
19972 jsonb_each_text_arg: None,
19973 table_fn_call: None,
19974 rows_from: None,
19975 json_table: None,
19976 scalar_fn_item: false,
19977 };
19978 return Ok(if correlated {
19979 Self::wrap_correlated_srf(tref)
19980 } else {
19981 tref
19982 });
19983 }
19984 // v7.11.7 — `FROM unnest(<expr>) [AS] <alias>` set-returning
19985 // source. Detect at the head before the bare-ident fallback;
19986 // unnest is not a reserved token.
19987 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("unnest"))
19988 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
19989 {
19990 self.advance(); // unnest
19991 self.advance(); // (
19992 let mut srf_args = alloc::vec![self.parse_expr(0)?];
19993 while matches!(self.peek(), Token::Comma) {
19994 self.advance();
19995 srf_args.push(self.parse_expr(0)?);
19996 }
19997 if !matches!(self.peek(), Token::RParen) {
19998 return Err(self.err(alloc::format!(
19999 "expected ')' after unnest() argument, got {:?}",
20000 self.peek()
20001 )));
20002 }
20003 self.advance();
20004 // Multi-arg unnest(a, b, …) zips the arrays in
20005 // parallel, NULL-padding to the longest (PG's ROWS
20006 // FROM shorthand). Lower onto the unnest channel as an
20007 // internal marker call the executors unpack.
20008 let expr = if srf_args.len() == 1 {
20009 srf_args.pop().expect("len checked")
20010 } else {
20011 crate::ast::Expr::FunctionCall {
20012 name: "__unnest_zip".to_string(),
20013 args: srf_args,
20014 }
20015 };
20016 let with_ordinality = self.absorb_with_ordinality();
20017 let (alias_ident, unnest_column_aliases) = self.parse_optional_alias_with_columns()?;
20018 let name = alias_ident.clone().unwrap_or_else(|| "unnest".to_string());
20019 let correlated = Self::expr_has_any_column(&expr);
20020 let tref = TableRef {
20021 name,
20022 alias: alias_ident,
20023 only: false,
20024 as_of_segment: None,
20025 unnest_expr: Some(Box::new(expr)),
20026 unnest_column_aliases,
20027 with_ordinality,
20028 generate_series_args: None,
20029 lateral_subquery: None,
20030 jsonb_each_text_arg: None,
20031 table_fn_call: None,
20032 rows_from: None,
20033 json_table: None,
20034 scalar_fn_item: false,
20035 };
20036 return Ok(if correlated {
20037 Self::wrap_correlated_srf(tref)
20038 } else {
20039 tref
20040 });
20041 }
20042 // v7.39 (round 205, JSON_TABLE epic) — `JSON_TABLE(doc, '$path'
20043 // COLUMNS (...))` has bespoke syntax (a COLUMNS clause the
20044 // generic table-fn arg parser can't read), so it is intercepted
20045 // here BEFORE the generic dispatch. The doc expr may reference
20046 // outer columns (implicit LATERAL) — same correlated-wrap rule.
20047 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
20048 if s.eq_ignore_ascii_case("json_table"))
20049 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
20050 {
20051 let tref = self.parse_json_table_ref()?;
20052 let correlated = tref
20053 .json_table
20054 .as_deref()
20055 .is_some_and(|jt| Self::expr_has_any_column(&jt.doc));
20056 return Ok(if correlated {
20057 Self::wrap_correlated_srf(tref)
20058 } else {
20059 tref
20060 });
20061 }
20062 // v7.39 (read01 partitionfuncs.c) — generic FROM-position table
20063 // functions dispatched by name (`pg_partition_tree('t')`,
20064 // `pg_partition_ancestors('t')`). Same head-detection shape as
20065 // unnest; the engine executor owns the row shape per function.
20066 // v7.39 (read01 round 65) — and a USER function in FROM position
20067 // (`FROM rows_of(2)`). The SRFs with their own FROM pipeline
20068 // (generate_series / unnest / the json_each family) keep it — their arms
20069 // sit further down, so they are excluded here by name rather than by
20070 // ordering. Anything else that is an ident followed by `(` is a table
20071 // function; the engine executor decides whether it is a builtin, a
20072 // set-returning user function, or an error.
20073 // 7.38.1 S5.1 — pg_dump spells its table functions
20074 // schema-qualified (`pg_catalog.pg_options_to_table(...)`);
20075 // strip the pg_catalog prefix here so the same head-detection
20076 // fires. Only pg_catalog: a user schema's `s.f(x)` keeps its
20077 // meaning.
20078 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("pg_catalog"))
20079 && matches!(self.tokens.get(self.pos + 1), Some(Token::Dot))
20080 && matches!(
20081 self.tokens.get(self.pos + 2),
20082 Some(Token::Ident(_) | Token::QuotedIdent(_))
20083 )
20084 && matches!(self.tokens.get(self.pos + 3), Some(Token::LParen))
20085 {
20086 self.advance(); // pg_catalog
20087 self.advance(); // .
20088 }
20089 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
20090 if !s.eq_ignore_ascii_case("generate_series")
20091 && !s.eq_ignore_ascii_case("unnest")
20092 && !is_json_each_name(s))
20093 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
20094 {
20095 // Body out-of-line — this parse sits on the FROM/subquery
20096 // recursion chain (debug frame-cliff discipline).
20097 // v7.39 (read01 round 69) — a call whose arguments reference an outer
20098 // column (`t JOIN LATERAL dbl(t.id)`) is CORRELATED: it runs once per
20099 // outer row, so it rides the lateral channel. Same rule the unnest
20100 // arm uses.
20101 let tref = self.parse_table_fn_ref()?;
20102 let correlated = tref
20103 .table_fn_call
20104 .as_deref()
20105 .is_some_and(|(_, args)| args.iter().any(Self::expr_has_any_column));
20106 return Ok(if correlated {
20107 Self::wrap_correlated_srf(tref)
20108 } else {
20109 tref
20110 });
20111 }
20112 // v7.17.0 Phase 3.10 — `FROM generate_series(start, stop
20113 // [, step])` set-returning source. Same shape as unnest:
20114 // detect at the head, parse the comma-separated arg list,
20115 // dispatch downstream through the engine's set-returning
20116 // path. Supports integer triplets (mailrs's `WITH row_no AS
20117 // (SELECT * FROM generate_series(1, N))` pattern) and
20118 // TIMESTAMP + INTERVAL triplets (the Tier-A audit's
20119 // date-range iteration pattern, which pre-3.10 had no
20120 // direct equivalent in SPG).
20121 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("generate_series"))
20122 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
20123 {
20124 self.advance(); // generate_series
20125 self.advance(); // (
20126 let mut args: Vec<Expr> = Vec::new();
20127 loop {
20128 args.push(self.parse_expr(0)?);
20129 if matches!(self.peek(), Token::Comma) {
20130 self.advance();
20131 continue;
20132 }
20133 break;
20134 }
20135 if !matches!(self.peek(), Token::RParen) {
20136 return Err(self.err(alloc::format!(
20137 "expected ')' after generate_series() arguments, got {:?}",
20138 self.peek()
20139 )));
20140 }
20141 self.advance();
20142 if args.len() < 2 || args.len() > 3 {
20143 return Err(self.err(alloc::format!(
20144 "generate_series() expects 2 or 3 arguments (start, stop [, step]); got {}",
20145 args.len()
20146 )));
20147 }
20148 let with_ordinality = self.absorb_with_ordinality();
20149 let (alias_ident, column_aliases) = self.parse_optional_alias_with_columns()?;
20150 let name = alias_ident
20151 .clone()
20152 .unwrap_or_else(|| "generate_series".to_string());
20153 let correlated = args.iter().any(Self::expr_has_any_column);
20154 let tref = TableRef {
20155 name,
20156 alias: alias_ident,
20157 only: false,
20158 as_of_segment: None,
20159 unnest_expr: None,
20160 unnest_column_aliases: column_aliases,
20161 with_ordinality,
20162 generate_series_args: Some(args),
20163 lateral_subquery: None,
20164 jsonb_each_text_arg: None,
20165 table_fn_call: None,
20166 rows_from: None,
20167 json_table: None,
20168 scalar_fn_item: false,
20169 };
20170 return Ok(if correlated {
20171 Self::wrap_correlated_srf(tref)
20172 } else {
20173 tref
20174 });
20175 }
20176 // v7.16.2 — preserve information_schema / pg_catalog
20177 // qualifiers (mailrs round-10 A.3). The generic
20178 // `expect_ident_like` strip silently drops the schema;
20179 // we want the engine to recognise these PG meta tables
20180 // and synthesise rows from the live catalog. Produce a
20181 // synthetic name (`__spg_info_columns` etc.) so the
20182 // engine's SELECT-side router can dispatch without
20183 // clashing with any user-defined `columns` table.
20184 let (name, meta_original) = if let Some((synth, orig)) = self.try_peek_meta_qualified() {
20185 (synth, Some(orig))
20186 } else if let Some((synth, orig)) = self.try_peek_meta_bare() {
20187 (synth, Some(orig))
20188 } else {
20189 (self.expect_ident_like()?, None)
20190 };
20191 // v6.10.2 — optional `AS OF SEGMENT '<id>'` cold-tier
20192 // time-travel clause. Parse BEFORE the alias so the
20193 // alias can still ride at the tail (`tbl AS OF SEGMENT
20194 // '5' alias`). `AS` is a reserved keyword token, while
20195 // `OF` and `SEGMENT` are bare idents.
20196 let as_of_segment = if matches!(self.peek(), Token::As)
20197 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(s) | Token::QuotedIdent(s)) if s.eq_ignore_ascii_case("of"))
20198 {
20199 self.advance(); // AS
20200 self.advance(); // OF
20201 let kw = match self.peek().clone() {
20202 Token::Ident(s) | Token::QuotedIdent(s) => s,
20203 other => {
20204 return Err(self.err(format!("expected SEGMENT after AS OF, got {other:?}")));
20205 }
20206 };
20207 if !kw.eq_ignore_ascii_case("segment") {
20208 return Err(self.err(format!(
20209 "expected SEGMENT after AS OF, got {kw:?}; v6.10.2 supports SEGMENT only"
20210 )));
20211 }
20212 self.advance();
20213 // Segment id literal — accept either a string or
20214 // integer for operator ergonomics.
20215 let id = match self.advance() {
20216 Token::String(s) => s
20217 .parse::<u32>()
20218 .map_err(|e| self.err(format!("AS OF SEGMENT id parse: {e}")))?,
20219 Token::Integer(n) => u32::try_from(n)
20220 .map_err(|e| self.err(format!("AS OF SEGMENT id parse: {e}")))?,
20221 other => {
20222 return Err(self.err(format!(
20223 "expected segment id literal after AS OF SEGMENT, got {other:?}"
20224 )));
20225 }
20226 };
20227 Some(id)
20228 } else {
20229 None
20230 };
20231 // TABLESAMPLE is not a reserved token — keep the bare-ident
20232 // alias rule from swallowing it (`FROM t TABLESAMPLE …`).
20233 let alias = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("tablesample"))
20234 {
20235 None
20236 } else {
20237 self.parse_optional_alias()?
20238 };
20239 // r1052 — a catalog name rewritten to its synthetic form keeps
20240 // the WRITTEN name as the relation's alias, so `pg_cast.oid`
20241 // still binds after `pg_cast` became `__spg_pg_cast`. PG
20242 // semantics: the visible name of `pg_catalog.pg_cast` IS
20243 // `pg_cast`. Without this, every table-name-qualified column
20244 // on a synthesised catalog answered "missing FROM-clause
20245 // entry" — which is the wall pg_dump hit on its first
20246 // pg_proc/pg_cast query.
20247 let alias = match (&alias, &meta_original) {
20248 (None, Some(orig)) if *orig != name => Some(orig.clone()),
20249 _ => alias,
20250 };
20251 // `TABLESAMPLE BERNOULLI(p) | SYSTEM(p)` follows the alias
20252 // (PG grammar). BERNOULLI lowers to a per-row
20253 // `random() < p/100` conjunct on the enclosing SELECT's
20254 // WHERE — exact row-level Bernoulli semantics. SYSTEM
20255 // shares the lowering: SPG has no page structure to
20256 // sample, and the row-level form returns the same expected
20257 // fraction. REPEATABLE(seed) promises a deterministic
20258 // sample SPG cannot honour yet — honest error rather than
20259 // a silently ignored seed.
20260 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("tablesample")) {
20261 self.advance();
20262 let method = self.expect_ident_like()?;
20263 if !method.eq_ignore_ascii_case("bernoulli") && !method.eq_ignore_ascii_case("system") {
20264 return Err(self.err(alloc::format!(
20265 "TABLESAMPLE method {method:?} not supported; use BERNOULLI or SYSTEM"
20266 )));
20267 }
20268 if !matches!(self.peek(), Token::LParen) {
20269 return Err(self.err(alloc::format!(
20270 "expected '(' after TABLESAMPLE {}, got {:?}",
20271 method.to_ascii_uppercase(),
20272 self.peek()
20273 )));
20274 }
20275 self.advance();
20276 let percent = self.parse_expr(0)?;
20277 if !matches!(self.peek(), Token::RParen) {
20278 return Err(self.err(alloc::format!(
20279 "expected ')' after TABLESAMPLE percentage, got {:?}",
20280 self.peek()
20281 )));
20282 }
20283 self.advance();
20284 // REPEATABLE(seed) → a deterministic per-row draw seeded by
20285 // `seed`, so the sample is stable across repeats and rescans.
20286 // Non-REPEATABLE keeps the non-deterministic `random()` draw.
20287 let mut sample_seed: Option<Expr> = None;
20288 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("repeatable")) {
20289 self.advance();
20290 if !matches!(self.peek(), Token::LParen) {
20291 return Err(self.err(alloc::format!(
20292 "expected '(' after REPEATABLE, got {:?}",
20293 self.peek()
20294 )));
20295 }
20296 self.advance();
20297 let seed = self.parse_expr(0)?;
20298 if !matches!(self.peek(), Token::RParen) {
20299 return Err(self.err(alloc::format!(
20300 "expected ')' after REPEATABLE seed, got {:?}",
20301 self.peek()
20302 )));
20303 }
20304 self.advance();
20305 sample_seed = Some(seed);
20306 }
20307 let draw = match sample_seed {
20308 Some(seed) => Expr::FunctionCall {
20309 name: "__tsm_fract".to_string(),
20310 args: alloc::vec![seed],
20311 },
20312 None => Expr::FunctionCall {
20313 name: "random".to_string(),
20314 args: Vec::new(),
20315 },
20316 };
20317 self.pending_sample_preds.push(Expr::Binary {
20318 lhs: Box::new(draw),
20319 op: crate::ast::BinOp::Lt,
20320 rhs: Box::new(Expr::Binary {
20321 lhs: Box::new(percent),
20322 op: crate::ast::BinOp::Div,
20323 rhs: Box::new(Expr::Literal(crate::ast::Literal::Float(100.0))),
20324 }),
20325 });
20326 }
20327 Ok(TableRef {
20328 name,
20329 alias,
20330 only,
20331 as_of_segment,
20332 unnest_expr: None,
20333 unnest_column_aliases: Vec::new(),
20334 with_ordinality: false,
20335 generate_series_args: None,
20336 lateral_subquery: None,
20337 jsonb_each_text_arg: None,
20338 table_fn_call: None,
20339 rows_from: None,
20340 json_table: None,
20341 scalar_fn_item: false,
20342 })
20343 }
20344
20345 /// v7.13.2 — mailrs round-6 S5. Like `parse_optional_alias`
20346 /// but also accepts `AS alias(col [, col, …])` — the
20347 /// PG-standard table-function column-list form. The column
20348 /// list is only honoured when paired with `UNNEST(...)` in
20349 /// the parent; other call sites currently discard it.
20350 /// True when the expression tree contains a qualified column
20351 /// reference (`t.col`) — the syntactic marker that an SRF
20352 /// argument correlates with a preceding FROM item.
20353 fn expr_has_qualified_column(e: &Expr) -> bool {
20354 match e {
20355 Expr::Column(c) => c.qualifier.is_some(),
20356 Expr::Binary { lhs, rhs, .. } => {
20357 Self::expr_has_qualified_column(lhs) || Self::expr_has_qualified_column(rhs)
20358 }
20359 Expr::Unary { expr, .. } => Self::expr_has_qualified_column(expr),
20360 Expr::Cast { expr, .. } => Self::expr_has_qualified_column(expr),
20361 Expr::FunctionCall { args, .. } => args.iter().any(Self::expr_has_qualified_column),
20362 Expr::Case {
20363 operand,
20364 branches,
20365 else_branch,
20366 } => {
20367 operand
20368 .as_deref()
20369 .is_some_and(Self::expr_has_qualified_column)
20370 || branches.iter().any(|(w, t)| {
20371 Self::expr_has_qualified_column(w) || Self::expr_has_qualified_column(t)
20372 })
20373 || else_branch
20374 .as_deref()
20375 .is_some_and(Self::expr_has_qualified_column)
20376 }
20377 _ => false,
20378 }
20379 }
20380
20381 /// v7.38 (read01, T-lateral) — like `expr_has_qualified_column` but also
20382 /// counts a bare (unqualified) column. A set-returning function has no
20383 /// input columns of its own, so ANY column in its arguments is an outer
20384 /// (correlated) reference — `generate_series(1, n)` correlates on `n`.
20385 fn expr_has_any_column(e: &Expr) -> bool {
20386 match e {
20387 Expr::Column(_) => true,
20388 Expr::Binary { lhs, rhs, .. } => {
20389 Self::expr_has_any_column(lhs) || Self::expr_has_any_column(rhs)
20390 }
20391 Expr::Unary { expr, .. } => Self::expr_has_any_column(expr),
20392 Expr::Cast { expr, .. } => Self::expr_has_any_column(expr),
20393 Expr::FunctionCall { args, .. } => args.iter().any(Self::expr_has_any_column),
20394 // v7.39 (round 759, F31-B8b) — a column INSIDE an array
20395 // constructor or subscript fell to the `_ => false` arm, so
20396 // `unnest(ARRAY[x, x + 1])` never wrapped into the lateral
20397 // channel and the eager peer eval answered `column "x" does
20398 // not exist` (the substitution walker already recurses both
20399 // shapes; only this detector was blind to them).
20400 Expr::Array(items) => items.iter().any(Self::expr_has_any_column),
20401 Expr::ArraySubscript { target, index } => {
20402 Self::expr_has_any_column(target) || Self::expr_has_any_column(index)
20403 }
20404 Expr::Case {
20405 operand,
20406 branches,
20407 else_branch,
20408 } => {
20409 operand.as_deref().is_some_and(Self::expr_has_any_column)
20410 || branches
20411 .iter()
20412 .any(|(w, t)| Self::expr_has_any_column(w) || Self::expr_has_any_column(t))
20413 || else_branch
20414 .as_deref()
20415 .is_some_and(Self::expr_has_any_column)
20416 }
20417 _ => false,
20418 }
20419 }
20420
20421 /// Wrap a correlated SRF table ref (`unnest(t.col)` /
20422 /// `generate_series(1, t.n)`) into the lateral_subquery
20423 /// channel: `SELECT * FROM <srf>` executes per outer row with
20424 /// outer references substituted (v7.37.43-T4.5 machinery).
20425 /// Uncorrelated SRFs stay on their plain channels.
20426 fn wrap_correlated_srf(srf: TableRef) -> TableRef {
20427 let name = srf.name.clone();
20428 let alias = srf.alias.clone();
20429 let inner = crate::ast::SelectStatement {
20430 locking: None,
20431 ctes: Vec::new(),
20432 distinct: false,
20433 distinct_on: Vec::new(),
20434 items: alloc::vec![crate::ast::SelectItem::Wildcard],
20435 from: Some(crate::ast::FromClause {
20436 primary: srf,
20437 joins: Vec::new(),
20438 }),
20439 where_: None,
20440 group_by: None,
20441 group_by_all: false,
20442 having: None,
20443 unions: Vec::new(),
20444 order_by: Vec::new(),
20445 limit: None,
20446 offset: None,
20447 limit_with_ties: false,
20448 window_check_exprs: Vec::new(),
20449 };
20450 TableRef {
20451 name,
20452 alias,
20453 only: false,
20454 as_of_segment: None,
20455 unnest_expr: None,
20456 unnest_column_aliases: Vec::new(),
20457 with_ordinality: false,
20458 generate_series_args: None,
20459 lateral_subquery: Some(Box::new(inner)),
20460 jsonb_each_text_arg: None,
20461 table_fn_call: None,
20462 rows_from: None,
20463 json_table: None,
20464 scalar_fn_item: false,
20465 }
20466 }
20467
20468 /// True when the expression tree contains an unresolved
20469 /// `OVER w` marker (see parse_over_clause).
20470 fn expr_has_named_window(e: &Expr) -> bool {
20471 match e {
20472 Expr::WindowFunction { partition_by, .. } => matches!(
20473 partition_by.as_slice(),
20474 [Expr::Column(c)] if matches!(
20475 c.qualifier.as_deref(),
20476 Some("__named_window__") | Some("__named_window_ref__")
20477 )
20478 ),
20479 Expr::Binary { lhs, rhs, .. } => {
20480 Self::expr_has_named_window(lhs) || Self::expr_has_named_window(rhs)
20481 }
20482 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } => Self::expr_has_named_window(expr),
20483 Expr::FunctionCall { args, .. } => args.iter().any(Self::expr_has_named_window),
20484 Expr::Case {
20485 operand,
20486 branches,
20487 else_branch,
20488 } => {
20489 operand.as_deref().is_some_and(Self::expr_has_named_window)
20490 || branches.iter().any(|(w, t)| {
20491 Self::expr_has_named_window(w) || Self::expr_has_named_window(t)
20492 })
20493 || else_branch
20494 .as_deref()
20495 .is_some_and(Self::expr_has_named_window)
20496 }
20497 _ => false,
20498 }
20499 }
20500
20501 /// v7.39 (round 705) — the NAMES the expression references through the
20502 /// `OVER w` markers, so `parse_bare_select` can tell which WINDOW
20503 /// definitions nothing referenced. Traversal mirrors
20504 /// `expr_has_named_window` above.
20505 fn collect_named_window_refs(e: &Expr, into: &mut Vec<String>) {
20506 match e {
20507 Expr::WindowFunction { partition_by, .. } => {
20508 if let [Expr::Column(c)] = partition_by.as_slice()
20509 && matches!(
20510 c.qualifier.as_deref(),
20511 Some("__named_window__") | Some("__named_window_ref__")
20512 )
20513 {
20514 into.push(c.name.clone());
20515 }
20516 }
20517 Expr::Binary { lhs, rhs, .. } => {
20518 Self::collect_named_window_refs(lhs, into);
20519 Self::collect_named_window_refs(rhs, into);
20520 }
20521 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } => {
20522 Self::collect_named_window_refs(expr, into);
20523 }
20524 Expr::FunctionCall { args, .. } => {
20525 for a in args {
20526 Self::collect_named_window_refs(a, into);
20527 }
20528 }
20529 Expr::Case {
20530 operand,
20531 branches,
20532 else_branch,
20533 } => {
20534 if let Some(o) = operand.as_deref() {
20535 Self::collect_named_window_refs(o, into);
20536 }
20537 for (w, t) in branches {
20538 Self::collect_named_window_refs(w, into);
20539 Self::collect_named_window_refs(t, into);
20540 }
20541 if let Some(eb) = else_branch.as_deref() {
20542 Self::collect_named_window_refs(eb, into);
20543 }
20544 }
20545 _ => {}
20546 }
20547 }
20548
20549 /// Inline named-window definitions into the `OVER w` markers.
20550 /// An unknown name errors (PG: window "w" does not exist).
20551 #[allow(clippy::type_complexity)]
20552 fn substitute_named_windows(
20553 e: &mut Expr,
20554 defs: &[(
20555 String,
20556 (
20557 Vec<Expr>,
20558 Vec<(Expr, bool, Option<bool>)>,
20559 Option<WindowFrame>,
20560 ),
20561 )],
20562 ) -> Result<(), String> {
20563 match e {
20564 Expr::WindowFunction {
20565 partition_by,
20566 order_by,
20567 frame,
20568 ..
20569 } => {
20570 // `is_copy` distinguishes `OVER (w1 …)` (a refinable copy)
20571 // from the bare `OVER w1` (a plain reference).
20572 let named = match partition_by.as_slice() {
20573 [Expr::Column(c)] => match c.qualifier.as_deref() {
20574 Some("__named_window__") => Some((c.name.clone(), false)),
20575 Some("__named_window_ref__") => Some((c.name.clone(), true)),
20576 _ => None,
20577 },
20578 _ => None,
20579 };
20580 if let Some((wname, is_copy)) = named {
20581 let Some((_, def)) = defs.iter().find(|(n, _)| n.eq_ignore_ascii_case(&wname))
20582 else {
20583 return Err(alloc::format!("window {wname:?} does not exist"));
20584 };
20585 if !is_copy {
20586 *partition_by = def.0.clone();
20587 *order_by = def.1.clone();
20588 *frame = def.2.clone();
20589 return Ok(());
20590 }
20591 // v7.39 (round 229) — PG's copy rules, probed against
20592 // 18.4: a copy inherits the partitioning, may supply an
20593 // ordering only when the base has none, and may not copy
20594 // a base that already carries a frame (its own frame
20595 // would be ambiguous with the inherited one).
20596 if !def.1.is_empty() && !order_by.is_empty() {
20597 return Err(alloc::format!(
20598 "cannot override ORDER BY clause of window \"{wname}\""
20599 ));
20600 }
20601 if def.2.is_some() {
20602 return Err(alloc::format!(
20603 "cannot copy window \"{wname}\" because it has a frame clause"
20604 ));
20605 }
20606 *partition_by = def.0.clone();
20607 if order_by.is_empty() {
20608 *order_by = def.1.clone();
20609 }
20610 }
20611 Ok(())
20612 }
20613 Expr::Binary { lhs, rhs, .. } => {
20614 Self::substitute_named_windows(lhs, defs)?;
20615 Self::substitute_named_windows(rhs, defs)
20616 }
20617 Expr::Unary { expr, .. } | Expr::Cast { expr, .. } => {
20618 Self::substitute_named_windows(expr, defs)
20619 }
20620 Expr::FunctionCall { args, .. } => {
20621 for a in args {
20622 Self::substitute_named_windows(a, defs)?;
20623 }
20624 Ok(())
20625 }
20626 Expr::Case {
20627 operand,
20628 branches,
20629 else_branch,
20630 } => {
20631 if let Some(op) = operand {
20632 Self::substitute_named_windows(op, defs)?;
20633 }
20634 for (w, t) in branches {
20635 Self::substitute_named_windows(w, defs)?;
20636 Self::substitute_named_windows(t, defs)?;
20637 }
20638 if let Some(el) = else_branch {
20639 Self::substitute_named_windows(el, defs)?;
20640 }
20641 Ok(())
20642 }
20643 _ => Ok(()),
20644 }
20645 }
20646
20647 /// SQL-standard `TABLE name` shorthand — builds the equivalent
20648 /// `SELECT * FROM name` head. Callers own set-op chain / tail
20649 /// composition.
20650 fn parse_table_shorthand(&mut self) -> Result<SelectStatement, ParseError> {
20651 debug_assert!(matches!(self.peek(), Token::Table));
20652 self.advance(); // TABLE
20653 let tname = self.expect_ident_like()?;
20654 Ok(SelectStatement {
20655 locking: None,
20656 ctes: Vec::new(),
20657 distinct: false,
20658 distinct_on: Vec::new(),
20659 items: alloc::vec![SelectItem::Wildcard],
20660 from: Some(FromClause {
20661 primary: TableRef {
20662 name: tname,
20663 alias: None,
20664 only: false,
20665 as_of_segment: None,
20666 unnest_expr: None,
20667 unnest_column_aliases: Vec::new(),
20668 with_ordinality: false,
20669 generate_series_args: None,
20670 lateral_subquery: None,
20671 jsonb_each_text_arg: None,
20672 table_fn_call: None,
20673 rows_from: None,
20674 json_table: None,
20675 scalar_fn_item: false,
20676 },
20677 joins: Vec::new(),
20678 }),
20679 where_: None,
20680 group_by: None,
20681 group_by_all: false,
20682 having: None,
20683 unions: Vec::new(),
20684 order_by: Vec::new(),
20685 limit: None,
20686 offset: None,
20687 limit_with_ties: false,
20688 window_check_exprs: Vec::new(),
20689 })
20690 }
20691
20692 /// `jsonb_to_recordset(J) AS t(c1 t1, c2 t2, …)` (and record / json_
20693 /// variants) → a derived table that reads each declared column out of
20694 /// the JSON with `(row ->> 'ci')::ti`. The *set form iterates
20695 /// `jsonb_array_elements(J)` (one row per element, column `value`);
20696 /// the scalar *record form projects a single row straight off `J`.
20697 /// Rides the existing lateral-subquery channel, so no new executor or
20698 /// AST is needed.
20699 fn parse_json_to_record_from(&mut self) -> Result<TableRef, ParseError> {
20700 use crate::ast::{
20701 BinOp, ColumnName, Expr, FromClause, Literal, SelectItem, SelectStatement,
20702 };
20703 let fn_name = match self.peek() {
20704 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
20705 _ => unreachable!("caller guarded is_json_to_record_name"),
20706 };
20707 self.advance(); // fn name
20708 self.advance(); // (
20709 let mut arg = self.parse_expr(0)?;
20710 // populate_record(base, json): the base only carries the record
20711 // type here — the JSON argument is the second expression.
20712 let mut base: Option<Expr> = None;
20713 if matches!(self.peek(), Token::Comma) {
20714 self.advance();
20715 base = Some(arg);
20716 arg = self.parse_expr(0)?;
20717 }
20718 if !matches!(self.peek(), Token::RParen) {
20719 return Err(self.err(alloc::format!(
20720 "expected ')' after {fn_name}() argument, got {:?}",
20721 self.peek()
20722 )));
20723 }
20724 self.advance(); // )
20725 let is_set = fn_name.ends_with("recordset");
20726 // `[AS] alias ( col type [, …] )` column-definition list.
20727 if matches!(self.peek(), Token::As) {
20728 self.advance();
20729 }
20730 let alias_opt = match self.peek() {
20731 Token::Ident(s) | Token::QuotedIdent(s) => {
20732 let a = s.clone();
20733 self.advance();
20734 Some(a)
20735 }
20736 _ => None,
20737 };
20738 // v7.39 (read01 round 76) — the populate family's canonical PG
20739 // spelling carries no column list at all: the row shape comes from
20740 // the BASE argument's declared type (`jsonb_populate_record(NULL::t,
20741 // j)`). The parser has no catalog, so hand the two arguments to the
20742 // engine's table-function channel, which does. Only `*_to_record*`
20743 // (whose base is bare `record`) genuinely requires the list.
20744 if !matches!(self.peek(), Token::LParen) {
20745 if let Some(base_expr) = base {
20746 let alias = alias_opt.unwrap_or_else(|| fn_name.clone());
20747 return Ok(TableRef {
20748 name: alias.clone(),
20749 alias: Some(alias),
20750 only: false,
20751 as_of_segment: None,
20752 unnest_expr: None,
20753 unnest_column_aliases: Vec::new(),
20754 with_ordinality: false,
20755 generate_series_args: None,
20756 lateral_subquery: None,
20757 jsonb_each_text_arg: None,
20758 table_fn_call: Some(Box::new((fn_name, alloc::vec![base_expr, arg]))),
20759 rows_from: None,
20760 json_table: None,
20761 scalar_fn_item: false,
20762 });
20763 }
20764 return Err(self.err(alloc::format!(
20765 "expected '(' to start the {fn_name} column-definition list, got {:?}",
20766 self.peek()
20767 )));
20768 }
20769 let Some(alias) = alias_opt else {
20770 return Err(self.err(alloc::format!(
20771 "{fn_name}(...) needs a column-definition list, e.g. AS t(a int, b text)"
20772 )));
20773 };
20774 self.advance(); // (
20775 let mut coldefs: Vec<(String, crate::ast::CastTarget)> = Vec::new();
20776 loop {
20777 let col = self.expect_ident_like()?;
20778 let ty = self.parse_cast_target()?;
20779 coldefs.push((col, ty));
20780 if matches!(self.peek(), Token::Comma) {
20781 self.advance();
20782 continue;
20783 }
20784 if matches!(self.peek(), Token::RParen) {
20785 self.advance();
20786 break;
20787 }
20788 return Err(self.err(alloc::format!(
20789 "expected ',' or ')' in {fn_name} column list, got {:?}",
20790 self.peek()
20791 )));
20792 }
20793 if coldefs.is_empty() {
20794 return Err(self.err(alloc::format!(
20795 "{fn_name} column-definition list must declare at least one column"
20796 )));
20797 }
20798 // Per column: (base ->> 'col')::type AS col. The base is the
20799 // per-element `value` column for the *set form, or the argument
20800 // itself for the scalar record form.
20801 let items: Vec<SelectItem> = coldefs
20802 .into_iter()
20803 .map(|(col, ty)| {
20804 let base = if is_set {
20805 Expr::Column(ColumnName {
20806 qualifier: None,
20807 name: "value".to_string(),
20808 })
20809 } else {
20810 arg.clone()
20811 };
20812 SelectItem::Expr {
20813 expr: Expr::Cast {
20814 expr: Box::new(Expr::Binary {
20815 lhs: Box::new(base),
20816 op: BinOp::JsonGetText,
20817 rhs: Box::new(Expr::Literal(Literal::String(col.clone()))),
20818 }),
20819 target: ty,
20820 },
20821 alias: Some(col),
20822 }
20823 })
20824 .collect();
20825 let from = if is_set {
20826 let elem_fn = if fn_name.starts_with("jsonb") {
20827 "jsonb_array_elements"
20828 } else {
20829 "json_array_elements"
20830 };
20831 Some(FromClause {
20832 primary: TableRef {
20833 name: "value".to_string(),
20834 alias: None,
20835 only: false,
20836 as_of_segment: None,
20837 unnest_expr: Some(Box::new(Expr::FunctionCall {
20838 name: elem_fn.to_string(),
20839 args: alloc::vec![arg],
20840 })),
20841 unnest_column_aliases: alloc::vec!["value".to_string()],
20842 with_ordinality: false,
20843 generate_series_args: None,
20844 lateral_subquery: None,
20845 jsonb_each_text_arg: None,
20846 table_fn_call: None,
20847 rows_from: None,
20848 json_table: None,
20849 scalar_fn_item: false,
20850 },
20851 joins: Vec::new(),
20852 })
20853 } else {
20854 None
20855 };
20856 let inner = SelectStatement {
20857 locking: None,
20858 ctes: Vec::new(),
20859 distinct: false,
20860 distinct_on: Vec::new(),
20861 items,
20862 from,
20863 where_: None,
20864 group_by: None,
20865 group_by_all: false,
20866 having: None,
20867 unions: Vec::new(),
20868 order_by: Vec::new(),
20869 limit: None,
20870 offset: None,
20871 limit_with_ties: false,
20872 window_check_exprs: Vec::new(),
20873 };
20874 Ok(TableRef {
20875 name: alias.clone(),
20876 alias: Some(alias),
20877 only: false,
20878 as_of_segment: None,
20879 unnest_expr: None,
20880 unnest_column_aliases: Vec::new(),
20881 with_ordinality: false,
20882 generate_series_args: None,
20883 lateral_subquery: Some(Box::new(inner)),
20884 jsonb_each_text_arg: None,
20885 table_fn_call: None,
20886 rows_from: None,
20887 json_table: None,
20888 scalar_fn_item: false,
20889 })
20890 }
20891
20892 /// Absorb `WITH ORDINALITY` after an SRF call in FROM position.
20893 /// Returns true when the clause was present. `WITH` alone (a
20894 /// CTE can never start here) is not enough — the ORDINALITY
20895 /// ident must follow, so a stray WITH still errors downstream.
20896 fn absorb_with_ordinality(&mut self) -> bool {
20897 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with"))
20898 && matches!(self.tokens.get(self.pos + 1),
20899 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("ordinality"))
20900 {
20901 self.advance();
20902 self.advance();
20903 true
20904 } else {
20905 false
20906 }
20907 }
20908
20909 /// v7.39 (read01 partitionfuncs.c) — parse a FROM-position table
20910 /// function reference (`pg_partition_tree('t') [AS a(c, …)]`).
20911 /// Out-of-line: the caller sits on the FROM recursion chain.
20912 #[inline(never)]
20913 fn parse_table_fn_ref(&mut self) -> Result<TableRef, ParseError> {
20914 let fn_name = match self.advance() {
20915 Token::Ident(s) | Token::QuotedIdent(s) => s.to_ascii_lowercase(),
20916 _ => unreachable!("caller peeked an ident"),
20917 };
20918 self.advance(); // (
20919 let mut args: Vec<Expr> = Vec::new();
20920 if !matches!(self.peek(), Token::RParen) {
20921 loop {
20922 args.push(self.parse_expr(0)?);
20923 if matches!(self.peek(), Token::Comma) {
20924 self.advance();
20925 continue;
20926 }
20927 break;
20928 }
20929 }
20930 if !matches!(self.peek(), Token::RParen) {
20931 return Err(self.err(alloc::format!(
20932 "expected ')' after {fn_name}() arguments, got {:?}",
20933 self.peek()
20934 )));
20935 }
20936 self.advance();
20937 // v7.39 (read01 round 68) — `f(args) WITH ORDINALITY AS a(x, n)`: the
20938 // counter column rides after the function's own, and the alias list
20939 // names it.
20940 let with_ordinality = self.absorb_with_ordinality();
20941 let (alias_ident, unnest_column_aliases) = self.parse_optional_alias_with_columns()?;
20942 let name = alias_ident.clone().unwrap_or_else(|| fn_name.clone());
20943 Ok(TableRef {
20944 name,
20945 alias: alias_ident,
20946 only: false,
20947 as_of_segment: None,
20948 unnest_expr: None,
20949 unnest_column_aliases,
20950 with_ordinality,
20951 generate_series_args: None,
20952 lateral_subquery: None,
20953 jsonb_each_text_arg: None,
20954 table_fn_call: Some(Box::new((fn_name, args))),
20955 rows_from: None,
20956 json_table: None,
20957 scalar_fn_item: false,
20958 })
20959 }
20960
20961 /// v7.39 (round 205, JSON_TABLE) — parse
20962 /// `JSON_TABLE(<doc>, '<row_path>' [PASSING …] COLUMNS (<coldefs>))
20963 /// [AS <alias>]`. The COLUMNS list is a recursive tree (NESTED
20964 /// PATH nests another COLUMNS). Out-of-line (FROM recursion chain).
20965 #[inline(never)]
20966 fn parse_json_table_ref(&mut self) -> Result<TableRef, ParseError> {
20967 self.advance(); // json_table
20968 self.advance(); // (
20969 let doc = Box::new(self.parse_expr(0)?);
20970 self.expect_comma_json_table()?;
20971 let row_path = self.parse_json_string_literal("JSON_TABLE row path")?;
20972 // Optional `PASSING <expr> AS <name> [, …]`.
20973 let mut passing: Vec<(String, Expr)> = Vec::new();
20974 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("passing")) {
20975 self.advance();
20976 loop {
20977 let e = self.parse_expr(0)?;
20978 if !matches!(self.peek(), Token::As) {
20979 return Err(self.err("expected AS after JSON_TABLE PASSING value".into()));
20980 }
20981 self.advance();
20982 let vname = match self.advance() {
20983 Token::Ident(s) | Token::QuotedIdent(s) => s,
20984 other => {
20985 return Err(self.err(alloc::format!(
20986 "expected PASSING variable name, got {other:?}"
20987 )));
20988 }
20989 };
20990 passing.push((vname, e));
20991 if matches!(self.peek(), Token::Comma) {
20992 self.advance();
20993 continue;
20994 }
20995 break;
20996 }
20997 }
20998 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("columns")) {
20999 return Err(self.err("expected COLUMNS in JSON_TABLE".into()));
21000 }
21001 self.advance();
21002 let columns = self.parse_json_table_columns()?;
21003 if !matches!(self.peek(), Token::RParen) {
21004 return Err(self.err(alloc::format!(
21005 "expected ')' to close JSON_TABLE, got {:?}",
21006 self.peek()
21007 )));
21008 }
21009 self.advance();
21010 let alias_ident = self.parse_optional_alias()?;
21011 let name = alias_ident
21012 .clone()
21013 .unwrap_or_else(|| String::from("json_table"));
21014 Ok(TableRef {
21015 name,
21016 alias: alias_ident,
21017 only: false,
21018 as_of_segment: None,
21019 unnest_expr: None,
21020 unnest_column_aliases: Vec::new(),
21021 with_ordinality: false,
21022 generate_series_args: None,
21023 lateral_subquery: None,
21024 jsonb_each_text_arg: None,
21025 table_fn_call: None,
21026 rows_from: None,
21027 json_table: Some(Box::new(crate::ast::JsonTable {
21028 doc,
21029 row_path,
21030 columns,
21031 passing,
21032 })),
21033 scalar_fn_item: false,
21034 })
21035 }
21036
21037 fn expect_comma_json_table(&mut self) -> Result<(), ParseError> {
21038 if !matches!(self.peek(), Token::Comma) {
21039 return Err(self.err(alloc::format!(
21040 "expected ',' after JSON_TABLE document, got {:?}",
21041 self.peek()
21042 )));
21043 }
21044 self.advance();
21045 Ok(())
21046 }
21047
21048 fn parse_json_string_literal(&mut self, what: &str) -> Result<String, ParseError> {
21049 match self.advance() {
21050 Token::String(s) => Ok(s),
21051 other => Err(self.err(alloc::format!(
21052 "expected {what} string literal, got {other:?}"
21053 ))),
21054 }
21055 }
21056
21057 /// v7.39 (round 205) — `( <coldef> [, <coldef>]* )`.
21058 #[inline(never)]
21059 fn parse_json_table_columns(
21060 &mut self,
21061 ) -> Result<alloc::vec::Vec<crate::ast::JsonTableColumn>, ParseError> {
21062 if !matches!(self.peek(), Token::LParen) {
21063 return Err(self.err("expected '(' after COLUMNS".into()));
21064 }
21065 self.advance();
21066 let mut cols = Vec::new();
21067 loop {
21068 cols.push(self.parse_json_table_one_column()?);
21069 if matches!(self.peek(), Token::Comma) {
21070 self.advance();
21071 continue;
21072 }
21073 break;
21074 }
21075 if !matches!(self.peek(), Token::RParen) {
21076 return Err(self.err(alloc::format!(
21077 "expected ')' after JSON_TABLE COLUMNS, got {:?}",
21078 self.peek()
21079 )));
21080 }
21081 self.advance();
21082 Ok(cols)
21083 }
21084
21085 fn parse_json_table_one_column(&mut self) -> Result<crate::ast::JsonTableColumn, ParseError> {
21086 use crate::ast::{JsonTableColumn, JsonTableOnBehavior};
21087 // NESTED [PATH] '<p>' COLUMNS (...)
21088 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("nested")) {
21089 self.advance();
21090 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("path")) {
21091 self.advance();
21092 }
21093 let path = self.parse_json_string_literal("NESTED PATH")?;
21094 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("columns")) {
21095 return Err(self.err("expected COLUMNS after NESTED PATH".into()));
21096 }
21097 self.advance();
21098 let columns = self.parse_json_table_columns()?;
21099 return Ok(JsonTableColumn::Nested { path, columns });
21100 }
21101 // <name> ...
21102 let name = match self.advance() {
21103 Token::Ident(s) | Token::QuotedIdent(s) => s,
21104 other => {
21105 return Err(self.err(alloc::format!("expected column name, got {other:?}")));
21106 }
21107 };
21108 // <name> FOR ORDINALITY
21109 if matches!(self.peek(), Token::For) {
21110 self.advance();
21111 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("ordinality")) {
21112 return Err(self.err("expected ORDINALITY after FOR".into()));
21113 }
21114 self.advance();
21115 return Ok(JsonTableColumn::Ordinality { name });
21116 }
21117 // <name> <type> [FORMAT JSON] {PATH '<p>' | EXISTS [PATH '<p>']} [WITH WRAPPER] [ON …]
21118 let ty = self.parse_column_type_name()?;
21119 let mut format_json = false;
21120 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("format")) {
21121 self.advance();
21122 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("json")) {
21123 return Err(self.err("expected JSON after FORMAT".into()));
21124 }
21125 self.advance();
21126 format_json = true;
21127 }
21128 let mut exists = false;
21129 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("exists")) {
21130 self.advance();
21131 exists = true;
21132 }
21133 let mut path = alloc::format!("$.{name}");
21134 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("path")) {
21135 self.advance();
21136 path = self.parse_json_string_literal("column PATH")?;
21137 }
21138 if !exists && matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("format")) {
21139 // `FORMAT JSON` after PATH (alternate placement).
21140 self.advance();
21141 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("json")) {
21142 self.advance();
21143 }
21144 format_json = true;
21145 }
21146 let mut wrapper = false;
21147 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with")) {
21148 self.advance();
21149 // optional CONDITIONAL/UNCONDITIONAL
21150 if matches!(self.peek(), Token::Ident(s)
21151 if s.eq_ignore_ascii_case("unconditional")
21152 || s.eq_ignore_ascii_case("conditional"))
21153 {
21154 self.advance();
21155 }
21156 if !matches!(self.peek(), Token::Ident(s)
21157 if s.eq_ignore_ascii_case("wrapper") || s.eq_ignore_ascii_case("array"))
21158 {
21159 return Err(self.err("expected WRAPPER after WITH".into()));
21160 }
21161 self.advance();
21162 // optional `ARRAY` after `WRAPPER`, or `WRAPPER` after `ARRAY`
21163 if matches!(self.peek(), Token::Ident(s)
21164 if s.eq_ignore_ascii_case("wrapper") || s.eq_ignore_ascii_case("array"))
21165 {
21166 self.advance();
21167 }
21168 wrapper = true;
21169 }
21170 // ON EMPTY / ON ERROR clauses (two, in any order).
21171 let mut on_empty = JsonTableOnBehavior::Null;
21172 let mut on_error = JsonTableOnBehavior::Null;
21173 for _ in 0..2 {
21174 let behavior = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("error"))
21175 {
21176 self.advance();
21177 Some(JsonTableOnBehavior::Error)
21178 } else if matches!(self.peek(), Token::Null) {
21179 self.advance();
21180 Some(JsonTableOnBehavior::Null)
21181 } else if matches!(self.peek(), Token::Default) {
21182 self.advance();
21183 Some(JsonTableOnBehavior::Default(Box::new(self.parse_expr(0)?)))
21184 } else {
21185 None
21186 };
21187 let Some(behavior) = behavior else { break };
21188 // `ON {EMPTY|ERROR}`
21189 if !matches!(self.peek(), Token::On) {
21190 return Err(self.err("expected ON after JSON_TABLE column behavior".into()));
21191 }
21192 self.advance();
21193 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("empty")) {
21194 self.advance();
21195 on_empty = behavior;
21196 } else if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("error")) {
21197 self.advance();
21198 on_error = behavior;
21199 } else {
21200 return Err(self.err("expected EMPTY or ERROR after ON".into()));
21201 }
21202 }
21203 Ok(JsonTableColumn::Regular {
21204 name,
21205 ty,
21206 path,
21207 exists,
21208 format_json,
21209 wrapper,
21210 on_empty,
21211 on_error,
21212 })
21213 }
21214
21215 fn parse_optional_alias_with_columns(
21216 &mut self,
21217 ) -> Result<(Option<String>, Vec<String>), ParseError> {
21218 let alias = self.parse_optional_alias()?;
21219 if alias.is_none() {
21220 return Ok((None, Vec::new()));
21221 }
21222 let mut cols: Vec<String> = Vec::new();
21223 if matches!(self.peek(), Token::LParen) {
21224 self.advance();
21225 while let Token::Ident(s) | Token::QuotedIdent(s) = self.peek().clone() {
21226 self.advance();
21227 cols.push(s);
21228 if matches!(self.peek(), Token::Comma) {
21229 self.advance();
21230 continue;
21231 }
21232 break;
21233 }
21234 if matches!(self.peek(), Token::RParen) {
21235 self.advance();
21236 }
21237 }
21238 Ok((alias, cols))
21239 }
21240
21241 /// v7.37.16 — parse a `left(str, n)` / `right(str, n)` function call
21242 /// whose keyword token was already consumed and whose `(` is the
21243 /// current token. Factored out of `parse_atom` (and marked
21244 /// `#[inline(never)]`) so its `Vec`/loop locals stay OFF the giant
21245 /// recursive `parse_atom` frame — inlining them there enlarges the
21246 /// per-nesting-level stack cost that `MAX_NEST_DEPTH` is tuned
21247 /// against, risking an overflow before the budget triggers.
21248 #[inline(never)]
21249 fn parse_lr_string_function_call(&mut self, name: &str) -> Result<Expr, ParseError> {
21250 self.advance(); // (
21251 let mut args = Vec::new();
21252 if !matches!(self.peek(), Token::RParen) {
21253 loop {
21254 args.push(self.parse_expr(0)?);
21255 match self.peek() {
21256 Token::Comma => {
21257 self.advance();
21258 }
21259 Token::RParen => break,
21260 other => {
21261 return Err(self.err(alloc::format!(
21262 "expected ',' or ')' in {name}() args, got {other:?}"
21263 )));
21264 }
21265 }
21266 }
21267 }
21268 self.advance(); // )
21269 Ok(Expr::FunctionCall {
21270 name: name.into(),
21271 args,
21272 })
21273 }
21274
21275 /// FROM-clause: a primary table reference plus zero-or-more joined
21276 /// peers expressed via either `, <table>` (cross-product, no ON) or
21277 /// `[INNER|LEFT|RIGHT [OUTER]|FULL [OUTER]|CROSS] JOIN <table> [ON expr]`.
21278 /// v1.10 keeps the join list flat (left-associative nested-loop
21279 /// semantics).
21280 fn parse_from_clause(&mut self) -> Result<FromClause, ParseError> {
21281 let primary = self.parse_table_ref()?;
21282 let primary_qual = primary
21283 .alias
21284 .clone()
21285 .unwrap_or_else(|| primary.name.clone());
21286 let joins = self.parse_from_joins(&primary_qual)?;
21287 Ok(FromClause { primary, joins })
21288 }
21289
21290 /// v7.39 (round 420) — the join tail of a FROM clause, factored out of
21291 /// [`Self::parse_from_clause`] so MySQL's multi-table UPDATE can read the
21292 /// SAME grammar after its target table has already been consumed.
21293 /// (`advance()` destroys the tokens it returns — `mem::replace(.., Eof)`
21294 /// — so re-parsing by rewinding `self.pos` is not possible; the tail must
21295 /// be parsed forward, once.)
21296 /// `left_primary_qual` is the qualifier (alias, else name) of whatever
21297 /// sits to the LEFT of the first join — the FROM primary, or the UPDATE
21298 /// target in the MySQL multi-table form. It only feeds the `USING (…)`
21299 /// desugaring, which needs a name for the left side of each equality.
21300 fn parse_from_joins(&mut self, left_primary_qual: &str) -> Result<Vec<FromJoin>, ParseError> {
21301 let mut joins = Vec::new();
21302 loop {
21303 // `, <table>` — cross-product with no ON.
21304 if matches!(self.peek(), Token::Comma) {
21305 self.advance();
21306 let table = self.parse_table_ref()?;
21307 joins.push(FromJoin {
21308 kind: JoinKind::Cross,
21309 table,
21310 on: None,
21311 using_cols: None,
21312 natural: false,
21313 });
21314 continue;
21315 }
21316 // v7.37.16 — optional leading `NATURAL` before the join
21317 // kind: `NATURAL JOIN`, `NATURAL LEFT JOIN`, etc. NATURAL is
21318 // not a lexer keyword (it arrives as a bare Ident), so match
21319 // it case-insensitively here. When present, no ON/USING
21320 // clause is allowed — the common columns are resolved at
21321 // execution time.
21322 let natural = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("natural"));
21323 if natural {
21324 self.advance();
21325 }
21326 // Explicit JOIN syntax. Accept INNER JOIN, LEFT [OUTER] JOIN,
21327 // CROSS JOIN, and bare JOIN (defaults to INNER).
21328 let kind =
21329 match self.peek() {
21330 Token::Inner => {
21331 self.advance();
21332 if !matches!(self.peek(), Token::Join) {
21333 return Err(self
21334 .err(format!("expected JOIN after INNER, got {:?}", self.peek())));
21335 }
21336 self.advance();
21337 JoinKind::Inner
21338 }
21339 Token::Left => {
21340 self.advance();
21341 if matches!(self.peek(), Token::Outer) {
21342 self.advance();
21343 }
21344 if !matches!(self.peek(), Token::Join) {
21345 return Err(self.err(format!(
21346 "expected JOIN after LEFT [OUTER], got {:?}",
21347 self.peek()
21348 )));
21349 }
21350 self.advance();
21351 JoinKind::Left
21352 }
21353 Token::Cross => {
21354 self.advance();
21355 if !matches!(self.peek(), Token::Join) {
21356 return Err(self
21357 .err(format!("expected JOIN after CROSS, got {:?}", self.peek())));
21358 }
21359 self.advance();
21360 JoinKind::Cross
21361 }
21362 // v7.37.16 — RIGHT [OUTER] JOIN. OUTER is optional noise.
21363 Token::Right => {
21364 self.advance();
21365 if matches!(self.peek(), Token::Outer) {
21366 self.advance();
21367 }
21368 if !matches!(self.peek(), Token::Join) {
21369 return Err(self.err(format!(
21370 "expected JOIN after RIGHT [OUTER], got {:?}",
21371 self.peek()
21372 )));
21373 }
21374 self.advance();
21375 JoinKind::Right
21376 }
21377 // v7.37.16 — FULL [OUTER] JOIN. OUTER is optional noise.
21378 Token::Full => {
21379 self.advance();
21380 if matches!(self.peek(), Token::Outer) {
21381 self.advance();
21382 }
21383 if !matches!(self.peek(), Token::Join) {
21384 return Err(self.err(format!(
21385 "expected JOIN after FULL [OUTER], got {:?}",
21386 self.peek()
21387 )));
21388 }
21389 self.advance();
21390 JoinKind::FullOuter
21391 }
21392 Token::Join => {
21393 self.advance();
21394 JoinKind::Inner
21395 }
21396 _ => break,
21397 };
21398 let table = self.parse_table_ref()?;
21399 // v7.37.7 C.1 — USING (col_list) sugar. Desugars to
21400 // `prev_table.col1 = table.col1 AND prev_table.col2 = table.col2 …`
21401 // where prev_table is the most-recent left-side table
21402 // (the previous join's table if any, else the FROM primary).
21403 // PG semantics around column merging are richer (USING'd
21404 // cols become deduplicated single output columns); for
21405 // sugar purposes the predicate-only form covers the
21406 // baseline corpus shape and chained `… JOIN x USING (k)
21407 // JOIN y USING (k)` calls.
21408 // v7.37.16 — NATURAL joins carry no ON/USING clause; the
21409 // common columns resolve at execution time.
21410 if natural {
21411 joins.push(FromJoin {
21412 kind,
21413 table,
21414 on: None,
21415 using_cols: None,
21416 natural: true,
21417 });
21418 continue;
21419 }
21420 let using_match = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("using"));
21421 // v7.37.16 — capture the USING column list (in addition to
21422 // the ON desugar below) so the executor can perform PG's
21423 // column-merge on the output side.
21424 let mut using_cols: Option<Vec<String>> = None;
21425 let on = if matches!(self.peek(), Token::On) {
21426 self.advance();
21427 Some(self.parse_expr(0)?)
21428 } else if using_match {
21429 self.advance();
21430 if !matches!(self.peek(), Token::LParen) {
21431 return Err(
21432 self.err(format!("expected '(' after USING, got {:?}", self.peek()))
21433 );
21434 }
21435 self.advance();
21436 let mut cols: Vec<String> = Vec::new();
21437 loop {
21438 match self.peek().clone() {
21439 Token::Ident(s) | Token::QuotedIdent(s) => {
21440 self.advance();
21441 cols.push(s);
21442 }
21443 other => {
21444 return Err(self.err(format!(
21445 "expected column name inside USING (…), got {other:?}"
21446 )));
21447 }
21448 }
21449 match self.peek() {
21450 Token::Comma => {
21451 self.advance();
21452 continue;
21453 }
21454 Token::RParen => {
21455 self.advance();
21456 break;
21457 }
21458 other => {
21459 return Err(self.err(format!(
21460 "expected ',' or ')' inside USING (…), got {other:?}"
21461 )));
21462 }
21463 }
21464 }
21465 if cols.is_empty() {
21466 return Err(self.err("USING (…) requires at least one column".to_string()));
21467 }
21468 using_cols = Some(cols.clone());
21469 // Pick the left-side alias: prev join's table if any,
21470 // else FROM primary. Use alias when present, else
21471 // table name (PG-equivalent qualifier).
21472 let left_qual: String = joins
21473 .last()
21474 .map(|j| {
21475 j.table
21476 .alias
21477 .clone()
21478 .unwrap_or_else(|| j.table.name.clone())
21479 })
21480 .unwrap_or_else(|| alloc::string::String::from(left_primary_qual));
21481 let right_qual = table.alias.clone().unwrap_or_else(|| table.name.clone());
21482 let mut iter = cols.into_iter().map(|c| Expr::Binary {
21483 lhs: alloc::boxed::Box::new(Expr::Column(crate::ast::ColumnName {
21484 qualifier: Some(left_qual.clone()),
21485 name: c.clone(),
21486 })),
21487 op: crate::ast::BinOp::Eq,
21488 rhs: alloc::boxed::Box::new(Expr::Column(crate::ast::ColumnName {
21489 qualifier: Some(right_qual.clone()),
21490 name: c,
21491 })),
21492 });
21493 let first = iter.next().expect("at least one col");
21494 Some(iter.fold(first, |acc, pred| Expr::Binary {
21495 lhs: alloc::boxed::Box::new(acc),
21496 op: crate::ast::BinOp::And,
21497 rhs: alloc::boxed::Box::new(pred),
21498 }))
21499 } else if kind == JoinKind::Cross {
21500 None
21501 } else {
21502 return Err(self.err(format!(
21503 "expected ON or USING after {:?} JOIN, got {:?}",
21504 kind,
21505 self.peek()
21506 )));
21507 };
21508 joins.push(FromJoin {
21509 kind,
21510 table,
21511 on,
21512 using_cols,
21513 natural: false,
21514 });
21515 }
21516 Ok(joins)
21517 }
21518
21519 /// Optional alias after an expression or table:
21520 /// `AS <ident>` is unambiguous; a bare `<ident>` directly after is also
21521 /// accepted (PG-style implicit alias). Returns `None` if the next token
21522 /// is not alias-shaped (e.g. comma, FROM, WHERE, semicolon, EOF, operator).
21523 fn parse_optional_alias(&mut self) -> Result<Option<String>, ParseError> {
21524 if matches!(self.peek(), Token::As) {
21525 self.advance();
21526 // v7.39 (round 340, V56) — after AS the next token MUST be an
21527 // identifier. This used to return None and "let the caller
21528 // surface the error on the next expectation", but when AS is
21529 // the LAST token there is no next expectation: `SELECT 1 AS`
21530 // parsed clean and silently dropped the alias. PG rejects it.
21531 if let Token::Ident(_) | Token::QuotedIdent(_) = self.peek() {
21532 return self.expect_ident_like().map(Some);
21533 }
21534 return Err(self.err(alloc::format!(
21535 "expected an alias after AS, got {:?}",
21536 self.peek()
21537 )));
21538 }
21539 // v7.17.0 Phase 1.3 — implicit alias (no `AS`). PG's
21540 // grammar reserves a long list of follow-keywords from the
21541 // alias slot. SPG's bareword approximation: skip a small
21542 // set of idents that would otherwise be swallowed as the
21543 // table alias and break trailing clauses like CREATE
21544 // MATERIALIZED VIEW … WITH [NO] DATA or future ON
21545 // CONFLICT WHERE shapes.
21546 if let Token::Ident(s) | Token::QuotedIdent(s) = self.peek() {
21547 if is_alias_stopword(s) {
21548 return Ok(None);
21549 }
21550 return Ok(self.expect_ident_like().ok());
21551 }
21552 Ok(None)
21553 }
21554
21555 /// Pratt loop. `min_prec` is the minimum binary-op precedence we'll accept.
21556 fn parse_expr(&mut self, min_prec: u8) -> Result<Expr, ParseError> {
21557 // v7.30.2 (mailrs round-25 ask 2) — nesting budget: a parse
21558 // error beats a stack overflow (an overflow aborts the
21559 // embedding host process).
21560 self.enter_nested()?;
21561 let r = self.parse_expr_inner(min_prec);
21562 self.nest_depth -= 1;
21563 r
21564 }
21565
21566 /// `OPERATOR([schema.]<op>)` — PG's explicit-operator spelling.
21567 /// When the upcoming tokens form one, return the underlying
21568 /// operator token and the position just past the closing paren
21569 /// so the binary loop can dispatch on the plain operator.
21570 fn peek_explicit_operator(&self) -> Option<(usize, Token)> {
21571 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("operator")) {
21572 return None;
21573 }
21574 if !matches!(self.tokens.get(self.pos + 1), Some(Token::LParen)) {
21575 return None;
21576 }
21577 let mut i = self.pos + 2;
21578 // Optional schema qualifier (pg_catalog.<op> etc.).
21579 if matches!(self.tokens.get(i), Some(Token::Ident(_)))
21580 && matches!(self.tokens.get(i + 1), Some(Token::Dot))
21581 {
21582 i += 2;
21583 }
21584 let op_tok = self.tokens.get(i)?.clone();
21585 if !matches!(self.tokens.get(i + 1), Some(Token::RParen)) {
21586 return None;
21587 }
21588 Some((i + 2, op_tok))
21589 }
21590
21591 /// PG operator symbols that lower onto function calls in
21592 /// binary position: `~` / `~*` / `!~` / `!~*` (regex match
21593 /// family → regexp_like, comparison rung), `^@` (starts_with,
21594 /// comparison rung), `^` (power, tighter than `*`), `#`
21595 /// (integer XOR via `(a|b) - (a&b)` — the AND bits are a
21596 /// subset of the OR bits so the subtraction never borrows).
21597 fn try_symbol_operator(
21598 &mut self,
21599 lhs: &Expr,
21600 min_prec: u8,
21601 ) -> Result<Option<Expr>, ParseError> {
21602 enum Sym {
21603 Regex { ci: bool, negated: bool },
21604 Like { ci: bool, negated: bool },
21605 StartsWith,
21606 Power,
21607 Xor,
21608 RangeAdjacent,
21609 }
21610 // v7.39 (IS-precedence knife) — the low-precedence postfix
21611 // predicates ride this existing leaf call (zero new frame slots
21612 // on the nesting chain).
21613 if let Some(e) = self.parse_postfix_predicate(lhs, min_prec)? {
21614 return Ok(Some(e));
21615 }
21616 let (sym, prec): (Sym, u8) = match self.peek() {
21617 Token::Tilde => (
21618 Sym::Regex {
21619 ci: false,
21620 negated: false,
21621 },
21622 5,
21623 ),
21624 Token::TildeStar => (
21625 Sym::Regex {
21626 ci: true,
21627 negated: false,
21628 },
21629 5,
21630 ),
21631 Token::NotTilde => (
21632 Sym::Regex {
21633 ci: false,
21634 negated: true,
21635 },
21636 5,
21637 ),
21638 Token::NotTildeStar => (
21639 Sym::Regex {
21640 ci: true,
21641 negated: true,
21642 },
21643 5,
21644 ),
21645 // v7.37 D.25 — PG operator spellings of LIKE/ILIKE.
21646 Token::DoubleTilde => (
21647 Sym::Like {
21648 ci: false,
21649 negated: false,
21650 },
21651 5,
21652 ),
21653 Token::DoubleTildeStar => (
21654 Sym::Like {
21655 ci: true,
21656 negated: false,
21657 },
21658 5,
21659 ),
21660 Token::NotDoubleTilde => (
21661 Sym::Like {
21662 ci: false,
21663 negated: true,
21664 },
21665 5,
21666 ),
21667 Token::NotDoubleTildeStar => (
21668 Sym::Like {
21669 ci: true,
21670 negated: true,
21671 },
21672 5,
21673 ),
21674 Token::CaretAt => (Sym::StartsWith, 5),
21675 // PG `^` is exponentiation; MySQL `^` is bitwise XOR (and binds
21676 // tighter than `* / & |`, which the prec-9 rung preserves —
21677 // v7.39 round 407: +1 from the pre-XOR ladder's rung 8).
21678 Token::Caret if self.mysql_dialect => (Sym::Xor, 9),
21679 Token::Caret => (Sym::Power, 9),
21680 // v7.39 (round 760, F31-B1) — `#` is a generic operator too:
21681 // PG answers `5 # 3 + 1` as `5 # 4` = 1 (additive first).
21682 Token::Hash => (Sym::Xor, 6),
21683 Token::Adjacent => (Sym::RangeAdjacent, 5),
21684 _ => return Ok(None),
21685 };
21686 if prec < min_prec {
21687 return Ok(None);
21688 }
21689 self.advance();
21690 let rhs = self.parse_expr(prec + 1)?;
21691 let out = match sym {
21692 Sym::Regex { ci, negated } => {
21693 let mut args = alloc::vec![lhs.clone(), rhs];
21694 if ci {
21695 args.push(Expr::Literal(Literal::String(String::from("i"))));
21696 }
21697 maybe_not(
21698 Expr::FunctionCall {
21699 name: String::from("regexp_like"),
21700 args,
21701 },
21702 negated,
21703 )
21704 }
21705 Sym::Like { ci, negated } => Expr::Like {
21706 expr: alloc::boxed::Box::new(lhs.clone()),
21707 pattern: alloc::boxed::Box::new(rhs),
21708 negated,
21709 case_insensitive: ci,
21710 },
21711 Sym::StartsWith => Expr::FunctionCall {
21712 name: String::from("starts_with"),
21713 args: alloc::vec![lhs.clone(), rhs],
21714 },
21715 Sym::Power => Expr::FunctionCall {
21716 name: String::from("power"),
21717 args: alloc::vec![lhs.clone(), rhs],
21718 },
21719 // `#` bitwise XOR — a real operator now (was desugared to
21720 // `(a|b)-(a&b)`, algebraically identical for integers but
21721 // undefined for bit strings; the direct op handles both).
21722 Sym::Xor => Expr::Binary {
21723 lhs: Box::new(lhs.clone()),
21724 op: BinOp::BitXor,
21725 rhs: Box::new(rhs),
21726 },
21727 // range `-|-` "is adjacent to" — lowered to a catalog function.
21728 Sym::RangeAdjacent => Expr::FunctionCall {
21729 name: String::from("range_adjacent"),
21730 args: alloc::vec![lhs.clone(), rhs],
21731 },
21732 };
21733 Ok(Some(out))
21734 }
21735
21736 /// v7.39 (IS-precedence knife) — the LOW-precedence postfix
21737 /// predicates, moved out of the tight postfix-cast loop: PG binds
21738 /// `IS [NOT] NULL/TRUE/FALSE/UNKNOWN/DISTINCT FROM/JSON/NORMALIZED`
21739 /// looser than EVERY binary operator (only NOT/AND/OR are looser),
21740 /// and BETWEEN/IN/LIKE/ILIKE/SIMILAR at the comparison rung — so
21741 /// `1 + 1 IS NULL` is `(1+1) IS NULL`, not `1 + (1 IS NULL)`.
21742 /// Returns Ok(consumed expr) when a predicate fired, Err(expr back)
21743 /// when nothing at this position belongs to the family. Out-of-line
21744 /// (`inline(never)`): the caller sits on the per-nesting-level frame
21745 /// chain that MAX_NEST_DEPTH is tuned against.
21746 #[inline(never)]
21747 fn parse_postfix_predicate(
21748 &mut self,
21749 lhs: &Expr,
21750 min_prec: u8,
21751 ) -> Result<Option<Expr>, ParseError> {
21752 // Reached through try_symbol_operator (an existing leaf call of
21753 // the binary loop) so NO new stack slots land on the per-nesting
21754 // frame chain; the lhs clones only when a predicate actually
21755 // consumes it.
21756 match self.peek() {
21757 // v7.39 (round 407) — IS is rung 4, the BETWEEN/IN/LIKE
21758 // comparison family rung 5 (each +1 from the pre-XOR ladder).
21759 Token::Is if min_prec <= 4 => {}
21760 Token::Between | Token::In | Token::Like if min_prec <= 5 => {}
21761 Token::Not
21762 if min_prec <= 5
21763 && matches!(
21764 self.tokens.get(self.pos + 1),
21765 Some(Token::Between | Token::In | Token::Like)
21766 ) => {}
21767 Token::Not | Token::Ident(_)
21768 if min_prec <= 5
21769 && (matches!(self.peek(), Token::Ident(s)
21770 if s.eq_ignore_ascii_case("ilike")
21771 || (self.mysql_dialect
21772 && (s.eq_ignore_ascii_case("regexp")
21773 || s.eq_ignore_ascii_case("rlike")))
21774 || (s.eq_ignore_ascii_case("similar")
21775 && matches!(self.tokens.get(self.pos + 1), Some(Token::To))))
21776 || (matches!(self.peek(), Token::Not)
21777 && matches!(self.tokens.get(self.pos + 1),
21778 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("ilike")
21779 || (self.mysql_dialect
21780 && (s.eq_ignore_ascii_case("regexp")
21781 || s.eq_ignore_ascii_case("rlike")))
21782 || s.eq_ignore_ascii_case("similar")))) => {}
21783 _ => return Ok(None),
21784 }
21785 let mut expr = lhs.clone();
21786 // IS family: rung 4 (NOT's operand parses at 4, so `NOT x IS NULL`
21787 // still groups as NOT (x IS NULL); OR/XOR/AND at 1-3 stay outside).
21788 if min_prec <= 4 {
21789 if matches!(self.peek(), Token::Is) {
21790 self.advance();
21791 let negated = if matches!(self.peek(), Token::Not) {
21792 self.advance();
21793 true
21794 } else {
21795 false
21796 };
21797 // v7.9.27b — `IS [NOT] DISTINCT FROM <rhs>`.
21798 // mailrs pg_dump.
21799 if matches!(self.peek(), Token::Distinct) {
21800 self.advance();
21801 if !matches!(self.peek(), Token::From) {
21802 return Err(self.err(format!(
21803 "expected FROM after IS{} DISTINCT, got {:?}",
21804 if negated { " NOT" } else { "" },
21805 self.peek()
21806 )));
21807 }
21808 self.advance();
21809 // Right-hand side: parse at the same precedence
21810 // tier as comparison (rung 5) so `x IS DISTINCT FROM a + b`
21811 // groups as `x IS DISTINCT FROM (a + b)`.
21812 let rhs = self.parse_expr(5)?;
21813 let op = if negated {
21814 BinOp::IsNotDistinctFrom
21815 } else {
21816 BinOp::IsDistinctFrom
21817 };
21818 expr = Expr::Binary {
21819 op,
21820 lhs: Box::new(expr),
21821 rhs: Box::new(rhs),
21822 };
21823 {
21824 return Ok(Some(expr));
21825 }
21826 }
21827 // v7.37.17 (17.6 siblings) — SQL:2016 / PG 16
21828 // `IS [NOT] JSON [VALUE|OBJECT|ARRAY|SCALAR]`.
21829 // Lowers onto pg_is_json(x, kind); NOT wraps the
21830 // call in a logical negation.
21831 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
21832 if s.eq_ignore_ascii_case("json"))
21833 {
21834 self.advance(); // JSON
21835 let kind = match self.peek() {
21836 Token::Ident(s) | Token::QuotedIdent(s)
21837 if matches!(
21838 s.to_ascii_lowercase().as_str(),
21839 "value" | "object" | "array" | "scalar"
21840 ) =>
21841 {
21842 let k = s.to_ascii_lowercase();
21843 self.advance();
21844 k
21845 }
21846 _ => "value".to_string(),
21847 };
21848 let call = Expr::FunctionCall {
21849 name: "pg_is_json".to_string(),
21850 args: alloc::vec![expr, Expr::Literal(Literal::String(kind)),],
21851 };
21852 expr = if negated {
21853 Expr::Unary {
21854 op: UnOp::Not,
21855 expr: Box::new(call),
21856 }
21857 } else {
21858 call
21859 };
21860 {
21861 return Ok(Some(expr));
21862 }
21863 }
21864 // v7.38 (read01 sweep) — SQL:2016 `x IS [NOT] [form]
21865 // NORMALIZED` (form ∈ NFC/NFD/NFKC/NFKD, default NFC).
21866 // Lowers onto is_normalized(x [, 'FORM']); NOT negates.
21867 {
21868 let form_kw = match self.peek() {
21869 Token::Ident(s) | Token::QuotedIdent(s)
21870 if matches!(
21871 s.to_ascii_uppercase().as_str(),
21872 "NFC" | "NFD" | "NFKC" | "NFKD"
21873 ) && matches!(
21874 self.tokens.get(self.pos + 1),
21875 Some(Token::Ident(n) | Token::QuotedIdent(n))
21876 if n.eq_ignore_ascii_case("normalized")
21877 ) =>
21878 {
21879 Some(s.to_ascii_uppercase())
21880 }
21881 _ => None,
21882 };
21883 let bare_normalized = form_kw.is_none()
21884 && matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
21885 if s.eq_ignore_ascii_case("normalized"));
21886 if form_kw.is_some() || bare_normalized {
21887 if form_kw.is_some() {
21888 self.advance(); // form keyword
21889 }
21890 self.advance(); // NORMALIZED
21891 let mut args = alloc::vec![expr];
21892 if let Some(f) = form_kw {
21893 args.push(Expr::Literal(Literal::String(f)));
21894 }
21895 let call = Expr::FunctionCall {
21896 name: "is_normalized".to_string(),
21897 args,
21898 };
21899 expr = if negated {
21900 Expr::Unary {
21901 op: UnOp::Not,
21902 expr: Box::new(call),
21903 }
21904 } else {
21905 call
21906 };
21907 {
21908 return Ok(Some(expr));
21909 }
21910 }
21911 }
21912 // `x IS [NOT] TRUE | FALSE | UNKNOWN` — the
21913 // three-valued boolean tests. IS TRUE/FALSE never
21914 // return NULL, so they lower to CASE forms whose
21915 // ELSE catches the NULL branch; IS UNKNOWN on a
21916 // boolean is exactly IS NULL.
21917 if matches!(self.peek(), Token::True | Token::False)
21918 || matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("unknown"))
21919 {
21920 let tok = self.advance();
21921 let test = match tok {
21922 Token::True => Some(true),
21923 Token::False => Some(false),
21924 _ => None, // UNKNOWN
21925 };
21926 // v7.39 (round 328, V45) — kept as what the user
21927 // wrote. These used to be lowered here into `CASE` /
21928 // `IS NULL`; the semantics were right but the AST no
21929 // longer knew the form, so `CHECK ((a > 1) IS TRUE)`
21930 // was echoed back as
21931 // `CHECK ((CASE WHEN (a > 1) THEN TRUE ELSE FALSE END))`.
21932 expr = Expr::BoolTest {
21933 expr: Box::new(expr),
21934 value: test,
21935 negated,
21936 };
21937 {
21938 return Ok(Some(expr));
21939 }
21940 }
21941 if !matches!(self.peek(), Token::Null) {
21942 return Err(self.err(format!(
21943 "expected NULL, DISTINCT, JSON, TRUE, FALSE or UNKNOWN after IS{}, got {:?}",
21944 if negated { " NOT" } else { "" },
21945 self.peek()
21946 )));
21947 }
21948 self.advance();
21949 expr = Expr::IsNull {
21950 expr: Box::new(expr),
21951 negated,
21952 };
21953 {
21954 return Ok(Some(expr));
21955 }
21956 }
21957 }
21958 // BETWEEN / IN / LIKE / ILIKE / SIMILAR: comparison rung (5).
21959 if min_prec <= 5 {
21960 // `x [NOT] BETWEEN a AND b`, `x [NOT] IN (...)`, `x [NOT] LIKE p`.
21961 // Look one token ahead so a stray `NOT` not followed by any of
21962 // these flows through to the early return below untouched.
21963 let negated = if matches!(self.peek(), Token::Not) {
21964 let next = self.tokens.get(self.pos + 1);
21965 matches!(next, Some(Token::Between | Token::In | Token::Like))
21966 || matches!(next, Some(Token::Ident(s)) if s.eq_ignore_ascii_case("ilike")
21967 || (self.mysql_dialect
21968 && (s.eq_ignore_ascii_case("regexp") || s.eq_ignore_ascii_case("rlike")))
21969 || s.eq_ignore_ascii_case("similar"))
21970 } else {
21971 false
21972 };
21973 if negated {
21974 self.advance();
21975 }
21976 if matches!(self.peek(), Token::Between) {
21977 expr = self.parse_between_tail(expr, negated)?;
21978 {
21979 return Ok(Some(expr));
21980 }
21981 }
21982 if matches!(self.peek(), Token::In) {
21983 if self.suppress_in_tail && !negated {
21984 // POSITION(sub IN str) — IN belongs to the
21985 // enclosing function syntax; stop here.
21986 {
21987 return Ok(None);
21988 }
21989 }
21990 expr = self.parse_in_tail(expr, negated)?;
21991 {
21992 return Ok(Some(expr));
21993 }
21994 }
21995 // v7.39 (read01 regexp.c) — `x [NOT] SIMILAR TO p [ESCAPE e]`
21996 // lowers onto the internal __similar_to(expr, pat[, esc]) call
21997 // (the SQL→regex transform runs inside, in the backtracking-
21998 // friendly shape SPG's matcher needs).
21999 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("similar"))
22000 && matches!(self.tokens.get(self.pos + 1), Some(Token::To))
22001 {
22002 self.advance(); // SIMILAR
22003 self.advance(); // TO
22004 let pattern = self.parse_expr(6)?;
22005 let mut args = alloc::vec![expr, pattern];
22006 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("escape")) {
22007 self.advance();
22008 args.push(self.parse_expr(6)?);
22009 }
22010 let call = Expr::FunctionCall {
22011 name: "__similar_to".to_string(),
22012 args,
22013 };
22014 expr = maybe_not(call, negated);
22015 {
22016 return Ok(Some(expr));
22017 }
22018 }
22019 if matches!(self.peek(), Token::Like) {
22020 self.advance();
22021 // `x [NOT] LIKE ANY/ALL (ARRAY[...])` — quantified LIKE.
22022 if let Some(q) = self.try_like_any_all(&expr, negated, false)? {
22023 expr = q;
22024 {
22025 return Ok(Some(expr));
22026 }
22027 }
22028 // Pattern at the same precedence as other comparison RHSes —
22029 // 5 leaves AND/OR alone so `a LIKE 'x%' AND b` parses right.
22030 let mut pattern = self.parse_expr(6)?;
22031 // `ESCAPE 'c'` — rewrite a literal pattern to the
22032 // default backslash escape at parse time. Custom
22033 // escapes on non-literal patterns would need
22034 // matcher support; error honestly.
22035 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("escape")) {
22036 self.advance();
22037 let esc = self.parse_expr(6)?;
22038 pattern = Self::rewrite_like_escape(pattern, esc).map_err(|m| self.err(m))?;
22039 }
22040 expr = Expr::Like {
22041 expr: Box::new(expr),
22042 pattern: Box::new(pattern),
22043 negated,
22044 case_insensitive: false,
22045 };
22046 {
22047 return Ok(Some(expr));
22048 }
22049 }
22050 // v7.25 (round-17) — ILIKE: case-insensitive LIKE. The
22051 // keyword reaches us as a plain identifier.
22052 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("ilike")) {
22053 self.advance();
22054 if let Some(q) = self.try_like_any_all(&expr, negated, true)? {
22055 expr = q;
22056 {
22057 return Ok(Some(expr));
22058 }
22059 }
22060 let pattern = self.parse_expr(6)?;
22061 expr = Expr::Like {
22062 expr: Box::new(expr),
22063 pattern: Box::new(pattern),
22064 negated,
22065 case_insensitive: true,
22066 };
22067 {
22068 return Ok(Some(expr));
22069 }
22070 }
22071 // v7.39 (round 380) — MySQL's REGEXP / RLIKE regex-match
22072 // operator (RLIKE is the alias). It is a keyword, not `~`, and
22073 // matches case-insensitively under the default collation, so it
22074 // lowers onto the same `regexp_like(expr, pattern, 'i')` the
22075 // `~*` operator uses, wrapped in NOT when negated.
22076 if self.mysql_dialect
22077 && matches!(self.peek(), Token::Ident(s)
22078 if s.eq_ignore_ascii_case("regexp") || s.eq_ignore_ascii_case("rlike"))
22079 {
22080 self.advance();
22081 let pattern = self.parse_expr(6)?;
22082 let call = Expr::FunctionCall {
22083 name: String::from("regexp_like"),
22084 args: alloc::vec![
22085 expr,
22086 pattern,
22087 Expr::Literal(Literal::String(String::from("i"))),
22088 ],
22089 };
22090 return Ok(Some(maybe_not(call, negated)));
22091 }
22092 }
22093 let _ = expr;
22094 Ok(None)
22095 }
22096
22097 fn parse_expr_inner(&mut self, min_prec: u8) -> Result<Expr, ParseError> {
22098 let mut lhs = self.parse_unary()?;
22099 let mut chain_len = 0usize;
22100 loop {
22101 // OPERATOR([schema.]op) reduces to its underlying
22102 // operator token before the normal dispatch.
22103 let explicit = self.peek_explicit_operator();
22104 let dispatch = match &explicit {
22105 Some((_, tok)) => self.binop_here(tok),
22106 None => self.binop_here(self.peek()),
22107 };
22108 let Some((op, prec)) = dispatch else {
22109 // v7.39 (round 539) — `OPERATOR(pg_catalog.~)` and the rest
22110 // of the symbol family. `binop_here` answers None for them
22111 // because they lower onto function calls rather than a
22112 // BinOp, and the fallback below reads `self.peek()` — the
22113 // word OPERATOR, not the operator. `pg_dump` writes every
22114 // catalog predicate this way, so its first query failed
22115 // and no dump ran:
22116 //
22117 // AND c.relname OPERATOR(pg_catalog.~) '^(t)$'
22118 //
22119 // Collapsing the wrapper to the operator it names puts the
22120 // token where the fallback already looks.
22121 if let Some((next, op_tok)) = explicit {
22122 self.tokens.splice(self.pos..next, [op_tok]);
22123 }
22124 if let Some(e) = self.try_symbol_operator(&lhs, min_prec)? {
22125 lhs = e;
22126 chain_len += 1;
22127 if chain_len > MAX_BINARY_CHAIN {
22128 return Err(self.err(alloc::format!(
22129 "more than {MAX_BINARY_CHAIN} chained binary operators"
22130 )));
22131 }
22132 continue;
22133 }
22134 break;
22135 };
22136 if prec < min_prec {
22137 break;
22138 }
22139 // v7.30.2 (mailrs round-25 ask 2) — the chain builds
22140 // iteratively but evaluates and drops recursively;
22141 // depth beyond the budget overflows worker stacks.
22142 chain_len += 1;
22143 if chain_len > MAX_BINARY_CHAIN {
22144 return Err(self.err(alloc::format!(
22145 "more than {MAX_BINARY_CHAIN} chained binary operators; rewrite long OR-equality chains as IN (…)"
22146 )));
22147 }
22148 match explicit {
22149 Some((end_pos, _)) => self.pos = end_pos,
22150 None => {
22151 self.advance();
22152 }
22153 }
22154 // v7.10.12 — `x <op> ANY(arr)` / `x <op> ALL(arr)`.
22155 // ANY is a bare ident; ALL is a reserved Token. Both
22156 // require an immediate `(` to disambiguate from
22157 // identifier columns named `any` / `all`.
22158 let any_kind = match self.peek() {
22159 Token::All if matches!(self.tokens.get(self.pos + 1), Some(Token::LParen)) => {
22160 Some(false)
22161 }
22162 Token::Ident(s) | Token::QuotedIdent(s)
22163 if (s.eq_ignore_ascii_case("any")
22164 || s.eq_ignore_ascii_case("some")
22165 || s.eq_ignore_ascii_case("all"))
22166 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen)) =>
22167 {
22168 Some(!s.eq_ignore_ascii_case("all"))
22169 }
22170 _ => None,
22171 };
22172 if let Some(is_any) = any_kind {
22173 lhs = self.parse_any_all_rhs(lhs, op, is_any)?;
22174 continue;
22175 }
22176 let rhs = self.parse_expr(prec + 1)?;
22177 lhs = Expr::Binary {
22178 lhs: Box::new(lhs),
22179 op,
22180 rhs: Box::new(rhs),
22181 };
22182 }
22183 Ok(lhs)
22184 }
22185
22186 /// `x <op> ANY (…)` / `ALL (…)`, both the quantified-subquery form
22187 /// and the array form.
22188 ///
22189 /// `#[inline(never)]` and out of `parse_expr_inner`, which sits on the
22190 /// frame chain `MAX_NEST_DEPTH` is tuned against: a debug build gives
22191 /// this block's `Expr` temporaries and four `format!` sites slots in
22192 /// that frame on every level of `((((1))))`, which never reaches it.
22193 #[inline(never)]
22194 fn parse_any_all_rhs(
22195 &mut self,
22196 lhs: Expr,
22197 op: BinOp,
22198 is_any: bool,
22199 ) -> Result<Expr, ParseError> {
22200 self.advance(); // ident
22201 self.advance(); // (
22202 // `x op ANY (SELECT …)` — the quantified-subquery
22203 // form. `= ANY` is exactly IN; the other operators
22204 // lower onto EXISTS over the subquery as a derived
22205 // table, comparing against its single projection
22206 // aliased __v (x's columns resolve correlated).
22207 // ALL is the negated-EXISTS complement; a NULL
22208 // element makes PG return NULL where this lowering
22209 // returns true — the NOT NULL column case (the
22210 // practical one) is exact.
22211 if matches!(self.peek(), Token::Select) || self.peek_is_with_kw() {
22212 // v7.39 (round 153) — `ANY (WITH … SELECT …)` is
22213 // legal PG too (round-151 sibling). Out-of-line
22214 // (#[inline(never)] helper) — this sits on
22215 // parse_expr's recursive frame and the two-armed
22216 // SELECT temporary blew the nesting-budget stack.
22217 let mut sub = self.parse_any_all_select_body()?;
22218 if !matches!(self.peek(), Token::RParen) {
22219 return Err(self.err(alloc::format!(
22220 "expected ')' after ANY/ALL subquery, got {:?}",
22221 self.peek()
22222 )));
22223 }
22224 self.advance();
22225 if sub.items.len() != 1 {
22226 return Err(self.err(alloc::format!(
22227 "ANY/ALL subquery must return one column, got {}",
22228 sub.items.len()
22229 )));
22230 }
22231 if is_any && matches!(op, BinOp::Eq) {
22232 return Ok(Expr::InSubquery {
22233 expr: Box::new(lhs),
22234 subquery: Box::new(sub),
22235 negated: false,
22236 });
22237 }
22238 // The engine's subquery resolvers materialise
22239 // the single-column result into an ARRAY the
22240 // existing AnyAll three-valued eval consumes.
22241 return Ok(Expr::AnyAll {
22242 expr: Box::new(lhs),
22243 op,
22244 array: Box::new(Expr::ScalarSubquery(Box::new(sub))),
22245 is_any,
22246 });
22247 }
22248 let arr = self.parse_expr(0)?;
22249 if !matches!(self.peek(), Token::RParen) {
22250 return Err(self.err(alloc::format!(
22251 "expected ')' after ANY/ALL argument, got {:?}",
22252 self.peek()
22253 )));
22254 }
22255 self.advance();
22256 Ok(Expr::AnyAll {
22257 expr: Box::new(lhs),
22258 op,
22259 array: Box::new(arr),
22260 is_any,
22261 })
22262 }
22263
22264 /// v7.39 (read01 geo_ops.c) — prefix `@@` (center-of). Out-of-line
22265 /// from `parse_unary` (see the frame-budget note at MAX_NEST_DEPTH).
22266 #[inline(never)]
22267 fn parse_prefix_center(&mut self) -> Result<Expr, ParseError> {
22268 self.advance();
22269 let e = self.parse_expr(9)?;
22270 Ok(build_center_call(e))
22271 }
22272
22273 /// v7.39 (round 508) — a prefix operator that IS a function: `@ x` is
22274 /// `abs(x)`, `# p` is `npoints(p)`, `@-@ p` is `length(p)`. Binds like
22275 /// unary minus.
22276 ///
22277 /// `#[inline(never)]` for the same reason as its neighbours: parse_unary
22278 /// sits on the recursive frame chain MAX_NEST_DEPTH is tuned against, so
22279 /// the Expr-sized local stays out of that frame.
22280 #[inline(never)]
22281 fn parse_prefix_call(&mut self, name: &str) -> Result<Expr, ParseError> {
22282 self.advance();
22283 let e = self.parse_expr(9)?;
22284 Ok(Expr::FunctionCall {
22285 name: alloc::string::String::from(name),
22286 args: alloc::vec![e],
22287 })
22288 }
22289
22290 /// v7.39 (read01 geo_ops.c) — prefix `?|` (vertical) / `?-`
22291 /// (horizontal). Out-of-line from `parse_unary` (frame budget).
22292 #[inline(never)]
22293 fn parse_prefix_geom_axis(&mut self, vertical: bool) -> Result<Expr, ParseError> {
22294 self.advance();
22295 let e = self.parse_expr(9)?;
22296 Ok(Expr::FunctionCall {
22297 name: alloc::string::String::from(if vertical {
22298 "isvertical"
22299 } else {
22300 "ishorizontal"
22301 }),
22302 args: alloc::vec![e],
22303 })
22304 }
22305
22306 /// v7.39 (round 355, M13) — `BINARY <expr>`, lowered onto the same
22307 /// cast the `CAST(x AS BINARY)` spelling produces. It binds tightly:
22308 /// MariaDB reads `BINARY 1 + 1` as `(BINARY 1) + 1` = 2.
22309 #[inline(never)]
22310 fn parse_binary_prefix(&mut self) -> Result<Expr, ParseError> {
22311 self.advance();
22312 let e = self.parse_expr(9)?;
22313 Ok(Expr::Cast {
22314 expr: Box::new(e),
22315 target: CastTarget::Named("binary".to_string()),
22316 })
22317 }
22318
22319 /// The prefix operators that share one shape: take the token, parse
22320 /// an operand at `prec`, wrap it.
22321 ///
22322 /// `#[inline(never)]`, and one function instead of five arms, for the
22323 /// reason the neighbouring `parse_prefix_*` helpers give: `parse_unary`
22324 /// sits on the frame chain `MAX_NEST_DEPTH` is tuned against, and a
22325 /// debug build gives EVERY arm's locals a slot in the frame, whichever
22326 /// arm runs. `((((1))))` reaches none of these arms and was carrying
22327 /// five `Expr`-sized locals per level for them anyway.
22328 #[inline(never)]
22329 fn parse_unary_op(&mut self, op: UnOp, prec: u8) -> Result<Expr, ParseError> {
22330 self.advance();
22331 let e = self.parse_expr(prec)?;
22332 Ok(Expr::Unary {
22333 op,
22334 expr: Box::new(e),
22335 })
22336 }
22337
22338 /// Unary minus. Out-of-line for the frame reason on `parse_unary_op`,
22339 /// and separate from it because of the literal folding below and the
22340 /// `format!` temporaries that folding needs.
22341 #[inline(never)]
22342 fn parse_prefix_minus(&mut self) -> Result<Expr, ParseError> {
22343 self.advance();
22344 // v7.39 (round 549) — fold the sign into an integer literal that
22345 // only fits once it is negative.
22346 //
22347 // `9223372036854775808` is one past i64::MAX, so the lexer hands
22348 // it over as a NUMERIC and `-` on a numeric stays numeric. PG
22349 // folds the sign first, so `-9223372036854775808` is a bigint
22350 // there — and `-9223372036854775808 - 1` raises "bigint out of
22351 // range" where SPG quietly answered -9223372036854775809, a value
22352 // no bigint can hold. The arithmetic itself was already checked;
22353 // only the literal's type was wrong.
22354 if let Token::Numeric(lit) = self.peek()
22355 && let Ok(folded) = alloc::format!("-{lit}").parse::<i64>()
22356 {
22357 self.advance();
22358 return Ok(Expr::Literal(Literal::Integer(folded)));
22359 }
22360 // Unary minus binds tighter than `*`/`/` (now at prec 7 after
22361 // `<->` slotted into 5 and arithmetic shifted up).
22362 let e = self.parse_expr(9)?;
22363 Ok(Expr::Unary {
22364 op: UnOp::Neg,
22365 expr: Box::new(e),
22366 })
22367 }
22368
22369 /// tsquery `!!` prefix negation, lowered to the catalog function.
22370 /// Binds like unary minus. Out-of-line for the frame reason on
22371 /// `parse_unary_op`.
22372 #[inline(never)]
22373 fn parse_prefix_tsquery_not(&mut self) -> Result<Expr, ParseError> {
22374 self.advance();
22375 let e = self.parse_expr(9)?;
22376 Ok(Expr::FunctionCall {
22377 name: String::from("tsquery_not"),
22378 args: alloc::vec![e],
22379 })
22380 }
22381
22382 fn parse_unary(&mut self) -> Result<Expr, ParseError> {
22383 match self.peek() {
22384 // NOT binds tighter than AND / XOR / OR but looser than
22385 // comparisons — its operand takes everything ≥ the comparison
22386 // rung (4), leaving AND (3) / XOR (2) / OR (1) outside so
22387 // `NOT a AND b` groups as `(NOT a) AND b`. (v7.39 round 407:
22388 // was rung 3, behaviour-identical when 3 was unused; AND now
22389 // occupies 3, so this must be 4 to keep NOT tighter than AND.)
22390 Token::Not => self.parse_unary_op(UnOp::Not, 4),
22391 // v7.39 (round 355, M13) — MySQL's `BINARY <expr>` prefix.
22392 // The body is out-of-line: `parse_unary` is one of the three
22393 // frames the parser's MAX_NEST_DEPTH is tuned against, and an
22394 // inline arm here overflowed the native stack in
22395 // `nesting_budget_errors_cleanly` — the guard test caught it,
22396 // exactly as the eval-side cliff did in rounds 346 and 351.
22397 Token::Ident(w) if self.mysql_dialect && w.eq_ignore_ascii_case("binary") => {
22398 self.parse_binary_prefix()
22399 }
22400 // v7.39 (round 353, M10) — MySQL's `!`. It binds TIGHTER than
22401 // arithmetic, unlike NOT: MariaDB answers 1 for `!1 + 1`
22402 // (`(!1)+1`) and 0 for `NOT 1 + 1` (`NOT (1+1)`), measured.
22403 Token::Bang => self.parse_unary_op(UnOp::Not, 9),
22404 Token::Minus => self.parse_prefix_minus(),
22405 // v7.39 (round 507) — unary `+`, which SPG did not have. `+1`
22406 // worked only because the lexer reads it as one signed literal;
22407 // `+ 1`, `+a`, `+(1)` and `1 + +1` were syntax errors, and both
22408 // PG18 and MariaDB take all of them. Binds like unary minus.
22409 Token::Plus => self.parse_unary_op(UnOp::Plus, 9),
22410 // Bitwise NOT binds like unary minus.
22411 Token::Tilde => self.parse_unary_op(UnOp::BitNot, 9),
22412 // v7.39 (read01 geo_ops.c) — prefix `@@` is PG's geometric
22413 // "center of" operator; desugars to center(x). The whole arm
22414 // is out-of-line: parse_unary sits on the per-nesting-level
22415 // frame chain that MAX_NEST_DEPTH is tuned against, so no
22416 // Expr-sized local may live in this frame.
22417 Token::TsMatch => self.parse_prefix_center(),
22418 // v7.39 (round 508) — the prefix operators that are named
22419 // functions in disguise: `@ x` is abs, `# p` is npoints, `@-@ p`
22420 // is length. Out-of-line for the same nesting-frame reason as
22421 // parse_prefix_center — parse_unary sits on the recursive cycle
22422 // MAX_NEST_DEPTH is tuned against, so no Expr-sized local may
22423 // live in this frame.
22424 Token::At => self.parse_prefix_call("abs"),
22425 Token::Hash => self.parse_prefix_call("npoints"),
22426 Token::AtMinusAt => self.parse_prefix_call("length"),
22427 // v7.39 (read01 geo_ops.c) — prefix `?|` / `?-`: "is vertical" /
22428 // "is horizontal" (lseg / line); desugars to the existing
22429 // isvertical()/ishorizontal() functions. Out-of-line for the
22430 // same nesting-frame reason as parse_prefix_center.
22431 Token::JsonKeysAny => self.parse_prefix_geom_axis(true),
22432 Token::GeomHoriz => self.parse_prefix_geom_axis(false),
22433 Token::DoubleBang => self.parse_prefix_tsquery_not(),
22434 _ => self.parse_atom(),
22435 }
22436 }
22437
22438 /// Parse a parenthesised scalar subquery body after the caller has consumed
22439 /// `(` and confirmed the next token is SELECT (or WITH, when `is_with`).
22440 /// v7.37 D.43 — `#[inline(never)]` keeps the large `Statement` local and the
22441 /// SELECT/WITH parse machinery off `parse_atom`'s stack frame; parse_atom sits
22442 /// on the recursive `((…))` cycle whose depth budget is tuned to that frame.
22443 /// v7.39 (read01 round 105) — is the current position `( <subquery-start>`,
22444 /// i.e. an `ARRAY(<subquery>)` and not `ARRAY[...]`? A subquery starts with
22445 /// SELECT, VALUES, or WITH (WITH lexes as a bare ident).
22446 /// `#[inline(never)]`: keeps this guard's locals off parse_atom's frame,
22447 /// which sits on the recursive nesting-budget cycle (a few extra bytes there
22448 /// tips the deep-nesting test into a stack overflow).
22449 #[inline(never)]
22450 fn array_subquery_ahead(&self) -> bool {
22451 if !matches!(self.peek(), Token::LParen) {
22452 return false;
22453 }
22454 matches!(
22455 self.tokens.get(self.pos + 1),
22456 Some(Token::Select | Token::Values)
22457 ) || matches!(
22458 self.tokens.get(self.pos + 1),
22459 Some(Token::Ident(w) | Token::QuotedIdent(w)) if w.eq_ignore_ascii_case("with")
22460 )
22461 }
22462
22463 /// v7.10.10 — `ARRAY[expr, …]` literal body. The `array` ident is consumed
22464 /// and the current token is `[`. `#[inline(never)]` so its `Vec`/loop
22465 /// locals stay off parse_atom's recursive frame (round 105).
22466 #[inline(never)]
22467 fn parse_array_literal_body(&mut self) -> Result<Expr, ParseError> {
22468 self.advance(); // consume `[`
22469 let mut items: Vec<Expr> = Vec::new();
22470 if !matches!(self.peek(), Token::RBracket) {
22471 loop {
22472 // Inside `ARRAY[...]`, a nested `[...]` is a sub-array
22473 // (`ARRAY[[1,2],[3,4]]`), not a pgvector literal.
22474 if matches!(self.peek(), Token::LBracket) {
22475 items.push(self.parse_array_bracket_body()?);
22476 } else {
22477 items.push(self.parse_expr(0)?);
22478 }
22479 match self.peek() {
22480 Token::Comma => {
22481 self.advance();
22482 }
22483 Token::RBracket => break,
22484 other => {
22485 return Err(self.err(alloc::format!(
22486 "expected ',' or ']' in ARRAY literal, got {other:?}"
22487 )));
22488 }
22489 }
22490 }
22491 }
22492 self.advance(); // consume `]`
22493 Ok(Expr::Array(items))
22494 }
22495
22496 /// v7.39 (read01 round 105) — parse `ARRAY(<subquery>)`. The `array` ident
22497 /// is already consumed; the current token is `(`. Desugars to a scalar
22498 /// subquery `SELECT array_agg(c) FROM (<subquery>) AS t(c)`, which collects
22499 /// the subquery's single-column rows in order — reusing the existing
22500 /// ScalarSubquery machinery rather than adding an AST node. `#[inline(never)]`
22501 /// keeps the large `Statement` local off parse_atom's recursive frame.
22502 #[inline(never)]
22503 fn parse_array_subquery(&mut self) -> Result<Expr, ParseError> {
22504 self.advance(); // consume `(`
22505 let is_with = matches!(self.peek(), Token::Ident(w) | Token::QuotedIdent(w)
22506 if w.eq_ignore_ascii_case("with"));
22507 let sub = if is_with {
22508 self.advance(); // WITH
22509 self.parse_with_cte_then_select()?
22510 } else {
22511 self.parse_select_stmt()?
22512 };
22513 if !matches!(self.peek(), Token::RParen) {
22514 return Err(self.err(alloc::format!(
22515 "expected ')' to close ARRAY(subquery), got {:?}",
22516 self.peek()
22517 )));
22518 }
22519 self.advance(); // consume `)`
22520 // Reuse the parser to build the array_agg wrapper from the subquery's
22521 // canonical text — avoids hand-constructing the derived-table AST.
22522 let wrapper = alloc::format!(
22523 "SELECT array_agg(\"__spg_arr_c\") FROM ({sub}) AS \"__spg_arr_t\"(\"__spg_arr_c\")"
22524 );
22525 let stmt = parse_statement(&wrapper)
22526 .map_err(|e| self.err(alloc::format!("ARRAY(subquery): {}", e.message)))?;
22527 let Statement::Select(sel) = stmt else {
22528 return Err(self.err("ARRAY(subquery) did not desugar to a SELECT".into()));
22529 };
22530 Ok(Expr::ScalarSubquery(alloc::boxed::Box::new(sel)))
22531 }
22532
22533 #[inline(never)]
22534 fn parse_paren_scalar_subquery(&mut self, is_with: bool) -> Result<Expr, ParseError> {
22535 let inner = if is_with {
22536 self.advance(); // WITH
22537 self.parse_with_cte_then_select()?
22538 } else {
22539 self.parse_select_stmt()?
22540 };
22541 match self.advance() {
22542 Token::RParen => {
22543 let Statement::Select(s) = inner else {
22544 return Err(ParseError {
22545 message: "scalar subquery body must be a SELECT".into(),
22546 token_pos: self.consumed_pos(),
22547 });
22548 };
22549 Ok(Expr::ScalarSubquery(Box::new(s)))
22550 }
22551 other => Err(ParseError {
22552 message: format!("expected ')' after scalar subquery, got {other:?}"),
22553 token_pos: self.consumed_pos(),
22554 }),
22555 }
22556 }
22557
22558 /// `B'1010'` / `X'1F'` bit-string (PG) or binary-string (MySQL)
22559 /// literals. The lexer splits them into an ident + string; recombine
22560 /// here. Out-of-line and returning `Option` so `parse_atom` — the
22561 /// recursive frame the 768 KiB stack budget is tuned against — pays no
22562 /// frame for the `body` / `bits` strings and their char loops (the
22563 /// round-367 frame cliff, M20).
22564 #[inline(never)]
22565 fn try_parse_bit_string_literal(&mut self) -> Option<Result<Expr, ParseError>> {
22566 let is_hex = match self.peek() {
22567 Token::Ident(p) if p.eq_ignore_ascii_case("x") => true,
22568 Token::Ident(p) if p.eq_ignore_ascii_case("b") => false,
22569 _ => return None,
22570 };
22571 if !matches!(self.tokens.get(self.pos + 1), Some(Token::String(_))) {
22572 return None;
22573 }
22574 // v7.39.3 — where the LITERAL starts, because the errors below
22575 // are about the literal and both engines point at it. `err`
22576 // reports the CURRENT token, which by then is the one after the
22577 // string: `SELECT x'123'` pointed at Eof, so the MySQL wire's
22578 // `near '…'` snippet — which runs from the reported position to
22579 // the end — came out empty where MySQL 9.7.2 says `near
22580 // 'x'123''`.
22581 let lit_pos = self.pos;
22582 self.advance();
22583 let Token::String(body) = self.advance() else {
22584 unreachable!("guarded above");
22585 };
22586 // v7.39 (round 367, M20) — in the MySQL dialect `X'…'` and `b'…'`
22587 // are BINARY STRINGS, not PG bit strings. `X'41'` is the byte 0x41
22588 // (hex pairs, even count required — MariaDB errors on an odd
22589 // count); `b'1010'` packs its bits big-endian, left-padded to a
22590 // byte. Lower both onto the bytea cast.
22591 if self.mysql_dialect {
22592 if is_hex {
22593 if body.len() % 2 == 1 {
22594 return Some(Err(self.err_at(
22595 lit_pos,
22596 alloc::format!("invalid hex string literal X'{body}': odd digit count"),
22597 )));
22598 }
22599 for c in body.chars() {
22600 if !c.is_ascii_hexdigit() {
22601 return Some(Err(self.err_at(
22602 lit_pos,
22603 alloc::format!("invalid hexadecimal digit {c:?} in X'…'"),
22604 )));
22605 }
22606 }
22607 return Some(self.finish_postfix_casts(hex_literal_to_bytea_expr(&body)));
22608 }
22609 if let Some(bad) = body.chars().find(|c| *c != '0' && *c != '1') {
22610 return Some(Err(self.err_at(
22611 lit_pos,
22612 alloc::format!("invalid binary digit {bad:?} in b'…'"),
22613 )));
22614 }
22615 return Some(self.finish_postfix_casts(bits_literal_to_bytea_expr(&body)));
22616 }
22617 let bits = if is_hex {
22618 let mut out = String::with_capacity(body.len() * 4);
22619 for c in body.chars() {
22620 let Some(d) = c.to_digit(16) else {
22621 // v7.39.3 — PostgreSQL 18.6's own sentence, and its
22622 // own quoting: `"g" is not a valid hexadecimal
22623 // digit` (measured, with the caret on the literal).
22624 return Some(Err(self.err_at(
22625 lit_pos,
22626 alloc::format!("\"{c}\" is not a valid hexadecimal digit"),
22627 )));
22628 };
22629 out.push_str(&alloc::format!("{d:04b}"));
22630 }
22631 out
22632 } else {
22633 if let Some(bad) = body.chars().find(|c| *c != '0' && *c != '1') {
22634 return Some(Err(self.err_at(
22635 lit_pos,
22636 alloc::format!("\"{bad}\" is not a valid binary digit"),
22637 )));
22638 }
22639 body
22640 };
22641 // Route through the postfix-cast loop so a chained cast like
22642 // `B'1010'::int` attaches onto the implicit `::bit` cast instead
22643 // of erroring at the `::`.
22644 // v7.39 (read01 varbit.c) — a distinct internal target: a B'...'
22645 // literal keeps its exact length, while an explicit `::bit` cast is
22646 // bit(1) with pad/truncate semantics (PG).
22647 Some(self.finish_postfix_casts(Expr::Cast {
22648 expr: Box::new(Expr::Literal(Literal::String(bits))),
22649 target: CastTarget::Named("__bit_literal".to_string()),
22650 }))
22651 }
22652
22653 fn parse_atom(&mut self) -> Result<Expr, ParseError> {
22654 if let Some(res) = self.try_parse_bit_string_literal() {
22655 return res;
22656 }
22657 let tok_pos = self.pos;
22658 match self.advance() {
22659 Token::Integer(n) => Ok(Expr::Literal(Literal::Integer(n))),
22660 Token::Float(x) => Ok(Expr::Literal(Literal::Float(x))),
22661 // v7.38 (read01) — dotted / over-i64 literal → exact NUMERIC (PG),
22662 // carrying the source mantissa + scale so no precision is lost. A
22663 // literal too wide for i128 falls back to double precision.
22664 // Out-of-line (#[inline(never)]) — this arm sits on the
22665 // parse_expr recursion chain; its expansion locals must not
22666 // widen the recursive frame (debug frame-cliff discipline).
22667 Token::Numeric(s) => match numeric_token_to_literal(s) {
22668 Ok(lit) => Ok(Expr::Literal(lit)),
22669 Err(msg) => Err(self.err(msg)),
22670 },
22671 Token::String(s) => Ok(Expr::Literal(Literal::String(s))),
22672 // v7.39 (round 367, M20) — a MySQL `0x…` binary-string literal
22673 // (the lexer only emits this token in the MySQL dialect). Lower
22674 // onto the existing bytea cast; out-of-line to keep this arm off
22675 // the parse recursion frame.
22676 Token::HexBytes(s) => Ok(hex_literal_to_bytea_expr(&s)),
22677 Token::True => Ok(Expr::Literal(Literal::Bool(true))),
22678 Token::False => Ok(Expr::Literal(Literal::Bool(false))),
22679 Token::Null => Ok(Expr::Literal(Literal::Null)),
22680 // v6.1.1 — `$N` placeholder. The actual Value lookup
22681 // happens in the engine eval path against the prepared-
22682 // statement bind buffer.
22683 Token::Placeholder(n) => Ok(Expr::Placeholder(n)),
22684 Token::LParen => {
22685 // v4.10: `(SELECT ...)` in expression position is a
22686 // scalar subquery; otherwise it's a parenthesised
22687 // expression. Peek for SELECT keyword to dispatch.
22688 // v7.37 D.43 — also accept `(WITH [RECURSIVE] … SELECT …)`; WITH
22689 // lexes as Ident("with") (not a reserved token). The subquery body
22690 // is parsed in `parse_paren_scalar_subquery` (marked #[inline(never)]
22691 // so its large `Statement` local stays out of parse_atom's stack
22692 // frame — parse_atom is on the recursive `((…))` cycle and the
22693 // nesting budget is tuned to its frame size).
22694 let is_with = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
22695 if s.eq_ignore_ascii_case("with"));
22696 if matches!(self.peek(), Token::Select) || is_with {
22697 self.parse_paren_scalar_subquery(is_with)
22698 } else {
22699 let e = self.parse_expr(0)?;
22700 // `(a, b, …)` — a row constructor. Valid only
22701 // in front of a comparison operator or [NOT]
22702 // IN; both expand at parse time (lexicographic
22703 // comparison / OR'd row equalities).
22704 if matches!(self.peek(), Token::Comma) {
22705 let mut row = alloc::vec![e];
22706 while matches!(self.peek(), Token::Comma) {
22707 self.advance();
22708 row.push(self.parse_expr(0)?);
22709 }
22710 if !matches!(self.peek(), Token::RParen) {
22711 return Err(self.err(alloc::format!(
22712 "expected ')' after row constructor, got {:?}",
22713 self.peek()
22714 )));
22715 }
22716 self.advance();
22717 // A bare `(a, b, …)` row constructor can carry postfix
22718 // (`::text`, `.field`) just like `ROW(a, b, …)`; the
22719 // early return here skips parse_atom's tail postfix
22720 // pass, so fold casts in explicitly. For the
22721 // comparison / predicate forms nothing postfix follows,
22722 // so this is a no-op.
22723 return self
22724 .parse_row_comparison_tail(row)
22725 .and_then(|e| self.finish_postfix_casts(e));
22726 }
22727 match self.advance() {
22728 Token::RParen => Ok(e),
22729 other => Err(ParseError {
22730 message: format!("expected ')', got {other:?}"),
22731 token_pos: self.consumed_pos(),
22732 }),
22733 }
22734 }
22735 }
22736 Token::LBracket => self.parse_vector_literal_body(),
22737 Token::Extract => self.parse_extract_atom(),
22738 Token::Interval => self.parse_interval_atom(),
22739 // `LEFT` / `RIGHT` are reserved-keyword tokens because the
22740 // grammar dedicates arms for `LEFT [OUTER] JOIN` /
22741 // `RIGHT [OUTER] JOIN`. When followed by `(` we're in
22742 // expression position calling the PG `left(string, n)` /
22743 // `right(string, n)` function; rebuild the AST as a regular
22744 // function call so the engine's apply_function dispatch picks
22745 // it up. Delegated to a #[inline(never)] helper so its locals
22746 // don't bloat this recursive `parse_atom` frame (the nesting
22747 // budget in `enter_nested` is tuned to parse_atom's size).
22748 Token::Left if matches!(self.peek(), Token::LParen) => {
22749 self.parse_lr_string_function_call("left")
22750 }
22751 Token::Right if matches!(self.peek(), Token::LParen) => {
22752 self.parse_lr_string_function_call("right")
22753 }
22754 // v4.10: EXISTS / NOT EXISTS. EXISTS isn't a reserved
22755 // token; we match on the bare ident. NOT is a token
22756 // (consumed in the comparison rung), but `EXISTS (...)`
22757 // at the top of an expression starts here.
22758 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("exists") => {
22759 self.parse_exists_atom(false)
22760 }
22761 // v7.13.0 — `CASE [<operand>] WHEN <cond> THEN <val>
22762 // [WHEN ...] [ELSE <val>] END` (mailrs round-5 G9).
22763 // CASE is a bare ident; we dispatch on lowercase match.
22764 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("case") => {
22765 self.parse_case_atom()
22766 }
22767 // v7.37.17 (17.6 siblings) — PG typed datetime literals:
22768 // `DATE '2003-01-02'` / `TIMESTAMP '…'` / `TIMESTAMPTZ
22769 // '…'`. Lower onto the ::cast node so the existing
22770 // runtime text→date/timestamp paths do the parsing. The
22771 // string must follow immediately, else the ident stays a
22772 // plain column reference.
22773 Token::Ident(s)
22774 if typed_literal_cast_target(&s.to_ascii_lowercase()).is_some()
22775 && matches!(self.peek(), Token::String(_)) =>
22776 {
22777 let target =
22778 typed_literal_cast_target(&s.to_ascii_lowercase()).expect("guard checked");
22779 let Token::String(lit) = self.advance() else {
22780 unreachable!("peek guaranteed a string token");
22781 };
22782 Ok(Expr::Cast {
22783 expr: Box::new(Expr::Literal(Literal::String(lit))),
22784 target,
22785 })
22786 }
22787 // v7.39 (round 221) — the SQL-standard long spellings:
22788 // `TIME [WITHOUT|WITH] TIME ZONE '…'` / `TIMESTAMP [WITHOUT|WITH]
22789 // TIME ZONE '…'`. Consume the modifier and lower to the same
22790 // typed-literal cast (`timetz` / `timestamptz` for WITH).
22791 Token::Ident(s)
22792 if (s.eq_ignore_ascii_case("time") || s.eq_ignore_ascii_case("timestamp"))
22793 && matches!(self.peek(), Token::Ident(w) if w.eq_ignore_ascii_case("with")
22794 || w.eq_ignore_ascii_case("without"))
22795 && matches!(self.tokens.get(self.pos + 1), Some(Token::Ident(t)) if t.eq_ignore_ascii_case("time"))
22796 && matches!(self.tokens.get(self.pos + 2), Some(Token::Ident(z)) if z.eq_ignore_ascii_case("zone"))
22797 && matches!(self.tokens.get(self.pos + 3), Some(Token::String(_))) =>
22798 {
22799 let with_tz = matches!(self.peek(), Token::Ident(w) if w.eq_ignore_ascii_case("with"));
22800 self.advance(); // WITH / WITHOUT
22801 self.advance(); // TIME
22802 self.advance(); // ZONE
22803 let Token::String(lit) = self.advance() else {
22804 unreachable!("guard checked a string token");
22805 };
22806 let base = s.to_ascii_lowercase();
22807 let target = match (base.as_str(), with_tz) {
22808 ("time", true) => CastTarget::Named(alloc::string::String::from("timetz")),
22809 ("time", false) => CastTarget::Named(alloc::string::String::from("time")),
22810 (_, true) => CastTarget::Timestamptz,
22811 (_, false) => CastTarget::Timestamp,
22812 };
22813 Ok(Expr::Cast {
22814 expr: Box::new(Expr::Literal(Literal::String(lit))),
22815 target,
22816 })
22817 }
22818 // v7.39 (read01 round 105) — `ARRAY(<subquery>)` constructor:
22819 // gathers the subquery's single-column rows (in its row order)
22820 // into an array. Desugared to `array_agg` over the subquery as a
22821 // derived table; out-of-line to keep parse_atom's frame small (it
22822 // sits on the recursive nesting-budget cycle).
22823 Token::Ident(s) | Token::QuotedIdent(s)
22824 if s.eq_ignore_ascii_case("array") && self.array_subquery_ahead() =>
22825 {
22826 self.parse_array_subquery()
22827 }
22828 // v7.10.10 — `ARRAY[expr, expr, …]` constructor. ARRAY
22829 // is not a reserved token; we match by case-insensitive
22830 // ident. The opening `[` must follow immediately. v7.39 (read01
22831 // round 105) — the body moved out-of-line so its `Vec`/loop locals
22832 // leave parse_atom's frame (which sits on the nesting-budget cycle).
22833 Token::Ident(s) | Token::QuotedIdent(s)
22834 if s.eq_ignore_ascii_case("array") && matches!(self.peek(), Token::LBracket) =>
22835 {
22836 self.parse_array_literal_body()
22837 }
22838 // v7.17.0 Phase 2.2 — MySQL `MATCH(col, ...) AGAINST
22839 // ('term' [IN BOOLEAN MODE | IN NATURAL LANGUAGE MODE])`.
22840 // We special-case before the generic ident dispatch so
22841 // the AGAINST clause never reaches the function-call
22842 // loop (which would mis-read `(cols) AGAINST` as a
22843 // call with no trailing modifier). The shape is
22844 // rewritten to a Boolean OR over per-column
22845 // `to_tsvector('simple', col) @@ plainto_tsquery('simple',
22846 // term)` so the existing FTS evaluator handles
22847 // semantics — the fulltext-GIN built at CREATE TABLE
22848 // time is currently a "real index that survives dump
22849 // round-trip"; the planner hook that actually uses
22850 // it for posting-list intersection lands in a later
22851 // sub-phase (Phase 2.2b) without touching this surface.
22852 Token::Ident(s) | Token::QuotedIdent(s)
22853 if s.eq_ignore_ascii_case("match") && matches!(self.peek(), Token::LParen) =>
22854 {
22855 self.parse_match_against_atom()
22856 }
22857 Token::Ident(s) | Token::QuotedIdent(s) => self.finish_ident_atom(s),
22858 // v7.37.43-T4 — PG-unreserved keywords are legal column /
22859 // alias names in expression context too. `release` appears
22860 // in sentori `0003_partition_events.sql` as both a column
22861 // reference (SELECT … release …) and an INSERT column list
22862 // entry. Mirrors `expect_ident_like`'s expansion of the
22863 // identifier set.
22864 other if unreserved_keyword_text(&other).is_some() => {
22865 let s = unreserved_keyword_text(&other).unwrap();
22866 self.finish_ident_atom(s)
22867 }
22868 // v7.39 (round 331, V50) — `@@var` in an EXPRESSION. It parsed
22869 // only inside `SET` before, so `SELECT @@autocommit` — which
22870 // every MySQL connector asks at handshake — was a parse error.
22871 // MariaDB accepts the bare, `@@session.` and `@@global.`
22872 // spellings alike and answers from the session's own value.
22873 Token::SessionVar(v) => {
22874 // v7.39 (round 430) — ONE `@` is a MySQL USER variable, which
22875 // has nothing to do with a `@@` engine setting: its own
22876 // per-session namespace, and an unset one reads NULL instead
22877 // of raising. Stripping every `@` (as this did) made `@x` and
22878 // `@@x` the same node, so `SELECT @x` answered "Unknown
22879 // system variable".
22880 Ok(variable_ref_atom(&v))
22881 }
22882 other => Err(ParseError {
22883 message: format!("unexpected token {other:?} in expression"),
22884 token_pos: tok_pos,
22885 }),
22886 }
22887 // After parsing the atom, fold any postfix `::vector` casts.
22888 .and_then(|atom| self.finish_postfix_casts(atom))
22889 }
22890
22891 /// Postfix operators on an atom: `::TYPE` cast and `IS [NOT] NULL`.
22892 /// Both bind tighter than any binary op.
22893 /// Shared cast-target parser for postfix `::TYPE` and the
22894 /// standard `CAST(expr AS TYPE)` form (v7.25, round-17).
22895 /// If the next tokens are `( N )`, consume them and return the canonical
22896 /// `base(N)` name so a temporal cast (`::timestamp(2)`) carries its
22897 /// fractional-seconds precision into `CastTarget::Named`; otherwise `None`.
22898 fn consume_temporal_typmod(&mut self, base: &str) -> Option<alloc::string::String> {
22899 if !matches!(self.peek(), Token::LParen) {
22900 return None;
22901 }
22902 self.advance(); // (
22903 let n = match self.advance() {
22904 Token::Integer(n) => n,
22905 _ => return Some(base.to_string()), // malformed → drop precision
22906 };
22907 if matches!(self.peek(), Token::RParen) {
22908 self.advance();
22909 }
22910 Some(alloc::format!("{base}({n})"))
22911 }
22912
22913 fn parse_cast_target(&mut self) -> Result<CastTarget, ParseError> {
22914 // r1052 — `::pg_catalog.regproc` and friends: pg_dump
22915 // schema-qualifies every cast target, and `pg_catalog.X` names
22916 // exactly the builtin type X. Consume the qualifier and let
22917 // the ordinary target parse decide.
22918 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("pg_catalog"))
22919 && matches!(self.tokens.get(self.pos + 1), Some(Token::Dot))
22920 {
22921 self.advance();
22922 self.advance();
22923 }
22924 let target = match self.advance() {
22925 Token::Ident(s) => match s.to_ascii_lowercase().as_str() {
22926 "int" | "integer" | "int4" => {
22927 if matches!(self.peek(), Token::LBracket)
22928 && matches!(self.tokens.get(self.pos + 1), Some(Token::RBracket))
22929 {
22930 self.advance();
22931 self.advance();
22932 CastTarget::IntArray
22933 } else {
22934 CastTarget::Int
22935 }
22936 }
22937 "bigint" | "int8" => {
22938 if matches!(self.peek(), Token::LBracket)
22939 && matches!(self.tokens.get(self.pos + 1), Some(Token::RBracket))
22940 {
22941 self.advance();
22942 self.advance();
22943 CastTarget::BigIntArray
22944 } else {
22945 CastTarget::BigInt
22946 }
22947 }
22948 "float" | "double" => CastTarget::Float,
22949 "text" => {
22950 // v7.10.11 — `::TEXT[]` widens to TextArray.
22951 if matches!(self.peek(), Token::LBracket)
22952 && matches!(self.tokens.get(self.pos + 1), Some(Token::RBracket))
22953 {
22954 self.advance();
22955 self.advance();
22956 CastTarget::TextArray
22957 } else {
22958 CastTarget::Text
22959 }
22960 }
22961 "bool" | "boolean" => CastTarget::Bool,
22962 "vector" => CastTarget::Vector,
22963 "date" => CastTarget::Date,
22964 // v7.38 (read01) — `::timestamp(N)` carries its fractional-
22965 // seconds precision through the Named path (the engine rounds
22966 // the sub-second field); bare `::timestamp` keeps the fast arm.
22967 "timestamp" | "datetime" => match self.consume_temporal_typmod("timestamp") {
22968 Some(named) => CastTarget::Named(named),
22969 None => CastTarget::Timestamp,
22970 },
22971 "timestamptz" => match self.consume_temporal_typmod("timestamptz") {
22972 Some(named) => CastTarget::Named(named),
22973 None => CastTarget::Timestamptz,
22974 },
22975 "interval" => CastTarget::Interval,
22976 "json" => CastTarget::Json,
22977 "jsonb" => CastTarget::Jsonb,
22978 // v7.39 (round 694) — these have dedicated CastTarget
22979 // variants, so they never reached the postfix `[]` handling
22980 // further down and `::regtype[]` was a SYNTAX error at the
22981 // `]`. PG has an array type for every scalar; take the
22982 // suffix here and hand the canonical `<ty>_array` name to
22983 // the engine, the same shape every other array cast uses.
22984 "regtype" if self.peek_postfix_array_brackets() => {
22985 self.advance();
22986 self.advance();
22987 CastTarget::Named(alloc::string::String::from("regtype_array"))
22988 }
22989 "regclass" if self.peek_postfix_array_brackets() => {
22990 self.advance();
22991 self.advance();
22992 CastTarget::Named(alloc::string::String::from("regclass_array"))
22993 }
22994 "regtype" => CastTarget::RegType,
22995 "regclass" => CastTarget::RegClass,
22996 // v7.12.0 — `::tsvector` / `::tsquery`.
22997 // Engine decodes the LHS text via the PG
22998 // external form parser.
22999 // v7.39 (round 352, M8) — MySQL's own cast targets.
23000 // `CAST(x AS SIGNED)` / `UNSIGNED`, with the optional
23001 // `INTEGER` / `INT` tail MariaDB also accepts. PG has no
23002 // such type, so they are taken only in that dialect and
23003 // fall through to the "type does not exist" arm otherwise.
23004 "signed" | "unsigned" if self.mysql_dialect => {
23005 if matches!(self.peek(), Token::Ident(k)
23006 if k.eq_ignore_ascii_case("integer") || k.eq_ignore_ascii_case("int"))
23007 {
23008 self.advance();
23009 }
23010 CastTarget::Named(s.to_ascii_lowercase())
23011 }
23012 // v7.39 (round 352, M8) — `CAST(x AS CHAR)` is UNBOUNDED
23013 // in MySQL: MariaDB answers '123' where the SQL-standard
23014 // reading (PG's, and SPG's) is `char(1)` and answers '1'.
23015 // Truncating a number to its first digit is a wrong answer
23016 // with no error, so the MySQL session gets MySQL's reading.
23017 "char" if self.mysql_dialect && !matches!(self.peek(), Token::LParen) => {
23018 CastTarget::Text
23019 }
23020 "tsvector" => CastTarget::TsVector,
23021 "tsquery" => CastTarget::TsQuery,
23022 // v7.17.0 — `::uuid`. Engine decodes the LHS
23023 // text via `spg_storage::parse_uuid_str`.
23024 "uuid" => CastTarget::Uuid,
23025 // v7.18 — `::bytea`. Engine decodes the LHS
23026 // text via the PG hex form (`'\xdeadbeef'`)
23027 // or escape form (`'\\x05\\x00'`). Closes
23028 // mailrs D-pre #3 reverse-acceptance gap.
23029 "bytea" => CastTarget::Bytea,
23030 // v7.37.5 ship triage — generic typed-cast escape.
23031 // Anything the long-tail PG type ident table knows
23032 // about(network/bit/geometry/multirange/etc.)flows
23033 // through `CastTarget::Named(canonical)`; the engine
23034 // resolves via `column_type_to_data_type` and dispatches
23035 // through the typed `coerce_value` path. Truly
23036 // unrecognised idents still hit the error arm below
23037 // because the engine rejects them.
23038 other => {
23039 // Optional `(N[, M])` precision args — `::numeric(10,2)`,
23040 // `::varchar(255)`, etc. Capture into the canonical
23041 // `name(p,s)` form so `type_name_to_data_type` can
23042 // reconstruct the `DataType::Numeric { precision,
23043 // scale }` (and similar param-carrying types).
23044 let mut name = other.to_string();
23045 // v7.39 (round 281) — `::bit varying(3)` is two
23046 // words; fold the tail in so the typmod reaches the
23047 // type resolver instead of tripping the parser.
23048 if name.eq_ignore_ascii_case("bit")
23049 && matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("varying"))
23050 {
23051 self.advance();
23052 name = alloc::string::String::from("varbit");
23053 }
23054 // v7.39 (round 613) — `::character varying` is the same
23055 // two-word shape and had no fold, so the `varying` was
23056 // left behind and the cast became a bare `character`,
23057 // which is `char(1)`: `'ab'::CHARACTER VARYING` answered
23058 // `a` where PG answers `ab`. Silently, and for a spelling
23059 // pg_dump writes.
23060 if name.eq_ignore_ascii_case("character")
23061 && matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("varying"))
23062 {
23063 self.advance();
23064 name = alloc::string::String::from("varchar");
23065 }
23066 if matches!(self.peek(), Token::LParen) {
23067 let mut buf = alloc::string::String::from("(");
23068 let mut depth = 0usize;
23069 loop {
23070 match self.advance() {
23071 Token::LParen => {
23072 depth += 1;
23073 if depth > 1 {
23074 buf.push('(');
23075 }
23076 }
23077 Token::RParen => {
23078 depth -= 1;
23079 if depth == 0 {
23080 buf.push(')');
23081 break;
23082 }
23083 buf.push(')');
23084 }
23085 Token::Comma => buf.push(','),
23086 Token::Integer(n) => buf.push_str(&alloc::format!("{n}")),
23087 // v7.39 (round 273) — a minus used to fall
23088 // into the catch-all below and vanish, so
23089 // `::numeric(10,-2)` reached the engine as
23090 // the text `numeric(10,2)` and silently
23091 // rounded to two DECIMALS instead of to
23092 // hundreds. A dropped token is not a
23093 // no-op when it carries a sign.
23094 Token::Minus => buf.push('-'),
23095 Token::Eof => break,
23096 _ => {}
23097 }
23098 }
23099 name.push_str(&buf);
23100 }
23101 // Optional postfix `[]` widens to the array form —
23102 // `::BOOL[]`, `::NUMERIC[]`, `::SMALLINT[]`, etc.
23103 // The engine's `type_name_to_data_type` recognises
23104 // the canonical `<ty>_array` form.
23105 if matches!(self.peek(), Token::LBracket)
23106 && matches!(self.tokens.get(self.pos + 1), Some(Token::RBracket))
23107 {
23108 self.advance();
23109 self.advance();
23110 name.push_str("_array");
23111 }
23112 CastTarget::Named(name)
23113 }
23114 },
23115 Token::Interval => CastTarget::Interval,
23116 // v7.39 — a quoted type name: `::"char"` is PG's 1-byte
23117 // "char" (oid 18, SPG Char1 — distinct from bare `char`
23118 // = char(1)); other quoted names resolve like idents.
23119 Token::QuotedIdent(q) => {
23120 if q.eq_ignore_ascii_case("char") {
23121 CastTarget::Named("char1".into())
23122 } else {
23123 CastTarget::Named(q.to_ascii_lowercase())
23124 }
23125 }
23126 other => {
23127 return Err(ParseError {
23128 message: format!("expected type ident after `::`, got {other:?}"),
23129 token_pos: self.consumed_pos(),
23130 });
23131 }
23132 };
23133 // v7.37.5 ship triage — postfix `[]` widens a scalar cast
23134 // target to its array sibling. Closed-enum arms (Bool /
23135 // SmallInt / Numeric / Float / Date / …) didn't carry the
23136 // explicit widening that Text / Int / BigInt did, so
23137 // `::BOOL[]` / `::NUMERIC[]` etc. surfaced as a parse
23138 // error. The widening here mirrors the per-arm Text /
23139 // Int / BigInt logic above + folds the new ζ-A first-class
23140 // types through `CastTarget::Named("<ty>_array")`.
23141 if matches!(self.peek(), Token::LBracket)
23142 && matches!(self.tokens.get(self.pos + 1), Some(Token::RBracket))
23143 {
23144 let widened = match &target {
23145 CastTarget::Bool => Some(CastTarget::Named("bool_array".to_string())),
23146 CastTarget::Date => Some(CastTarget::Named("date_array".to_string())),
23147 // v7.39 (round 326, V43) — the two temporal types stay
23148 // distinct. Both used to widen to `timestamptz_array`, so
23149 // `::timestamp[]` named the wrong target in its own error
23150 // message and lost the zone-less identity on the way.
23151 CastTarget::Timestamp => Some(CastTarget::Named("timestamp_array".to_string())),
23152 CastTarget::Timestamptz => Some(CastTarget::Named("timestamptz_array".to_string())),
23153 CastTarget::Uuid => Some(CastTarget::Named("uuid_array".to_string())),
23154 CastTarget::Json | CastTarget::Jsonb => {
23155 Some(CastTarget::Named("jsonb_array".to_string()))
23156 }
23157 CastTarget::Bytea => Some(CastTarget::Named("bytea_array".to_string())),
23158 CastTarget::Interval => Some(CastTarget::Named("interval_array".to_string())),
23159 CastTarget::Float => Some(CastTarget::Named("float_array".to_string())),
23160 CastTarget::Named(name) => {
23161 let mut a = name.clone();
23162 a.push_str("_array");
23163 Some(CastTarget::Named(a))
23164 }
23165 // Int / BigInt / Text / Vector / TsVector / TsQuery /
23166 // RegType / RegClass / TextArray / IntArray /
23167 // BigIntArray already finalised — leave as is.
23168 _ => None,
23169 };
23170 if let Some(w) = widened {
23171 self.advance();
23172 self.advance();
23173 return Ok(w);
23174 }
23175 }
23176 Ok(target)
23177 }
23178
23179 fn finish_postfix_casts(&mut self, mut expr: Expr) -> Result<Expr, ParseError> {
23180 loop {
23181 // v7.38 (read01, T9) — composite field access `(expr).field`.
23182 // A bare `a.b` is consumed as a qualified column inside the ident
23183 // atom, so a Dot only survives to this postfix position when the
23184 // base was a parenthesised expression (`(e).id`, `(row(1,2)).f1`).
23185 // `.*` whole-row expansion is not handled here (projection-level).
23186 if matches!(self.peek(), Token::Dot)
23187 && matches!(
23188 self.tokens.get(self.pos + 1),
23189 Some(Token::Ident(_) | Token::QuotedIdent(_))
23190 )
23191 {
23192 self.advance(); // .
23193 let field = match self.advance() {
23194 Token::Ident(s) | Token::QuotedIdent(s) => s,
23195 other => {
23196 return Err(
23197 self.err(format!("expected a field name after '.', got {other:?}"))
23198 );
23199 }
23200 };
23201 expr = Expr::FieldAccess {
23202 base: Box::new(expr),
23203 field,
23204 };
23205 continue;
23206 }
23207 if matches!(self.peek(), Token::DoubleColon) {
23208 self.advance();
23209 // v7.9.25 / v7.9.26 — broaden the postfix `::` cast
23210 // target set to include INTERVAL (reserved Token),
23211 // TIMESTAMPTZ, and PG catalog regtype / regclass.
23212 // mailrs follow-up H3a + H3b.
23213 let target = self.parse_cast_target()?;
23214 expr = Expr::Cast {
23215 expr: Box::new(expr),
23216 target,
23217 };
23218 continue;
23219 }
23220 // v7.10.12 — `arr[i]` subscript. PG 1-based; engine
23221 // returns NULL for out-of-range. Multiple subscripts
23222 // chain: `a[i][j]` parses left-to-right.
23223 if matches!(self.peek(), Token::LBracket) {
23224 self.advance();
23225 // `[lo:hi]` / `[:hi]` / `[lo:]` — array slice. A
23226 // bare index stays a subscript.
23227 let lo = if matches!(self.peek(), Token::Colon) {
23228 None
23229 } else {
23230 Some(self.parse_expr(0)?)
23231 };
23232 if matches!(self.peek(), Token::Colon) {
23233 self.advance();
23234 let hi = if matches!(self.peek(), Token::RBracket) {
23235 None
23236 } else {
23237 Some(Box::new(self.parse_expr(0)?))
23238 };
23239 if !matches!(self.peek(), Token::RBracket) {
23240 return Err(self.err(alloc::format!(
23241 "expected ']' after array slice, got {:?}",
23242 self.peek()
23243 )));
23244 }
23245 self.advance();
23246 expr = Expr::ArraySlice {
23247 target: Box::new(expr),
23248 lo: lo.map(Box::new),
23249 hi,
23250 };
23251 continue;
23252 }
23253 let index = lo.expect("non-colon branch parsed an index");
23254 if !matches!(self.peek(), Token::RBracket) {
23255 return Err(self.err(alloc::format!(
23256 "expected ']' after array index, got {:?}",
23257 self.peek()
23258 )));
23259 }
23260 self.advance();
23261 expr = Expr::ArraySubscript {
23262 target: Box::new(expr),
23263 index: Box::new(index),
23264 };
23265 continue;
23266 }
23267 // `expr AT TIME ZONE zone` — lowers to PG's own function
23268 // form timezone(zone, expr); the scalar implements the
23269 // offset shift (named zones error there — no tzdata).
23270 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("at"))
23271 && matches!(self.tokens.get(self.pos + 1),
23272 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("time"))
23273 && matches!(self.tokens.get(self.pos + 2),
23274 Some(Token::Ident(s)) if s.eq_ignore_ascii_case("zone"))
23275 {
23276 self.advance(); // AT
23277 self.advance(); // TIME
23278 self.advance(); // ZONE
23279 // Zone at comparison precedence so AND/OR stay out.
23280 let zone = self.parse_expr(6)?;
23281 expr = Expr::FunctionCall {
23282 name: "timezone".to_string(),
23283 args: alloc::vec![zone, expr],
23284 };
23285 continue;
23286 }
23287 // `expr COLLATE "name"` — SPG's single text ordering IS
23288 // byte order, i.e. the C collation. The byte-order
23289 // spellings absorb as no-ops; a locale collation would
23290 // silently sort differently from PG, so it errors
23291 // honestly instead.
23292 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("collate")) {
23293 self.advance();
23294 let mut cname = match self.advance() {
23295 Token::Ident(s) | Token::QuotedIdent(s) | Token::String(s) => s,
23296 other => {
23297 return Err(self.err(alloc::format!(
23298 "expected collation name after COLLATE, got {other:?}"
23299 )));
23300 }
23301 };
23302 // v7.39 (round 539) — a SCHEMA-QUALIFIED collation, which
23303 // is how `pg_dump` writes the default one:
23304 // `… COLLATE pg_catalog.default`. Reading a single token
23305 // left the SCHEMA as the name, so the clause was refused
23306 // as an unsupported locale collation and no dump ran.
23307 if matches!(self.peek(), Token::Dot) {
23308 // v7.39.2 — the qualifier is DROPPED (SPG is single
23309 // schema) but it is checked first. PostgreSQL 18.6
23310 // answers `schema "nosuch_schema" does not exist` for
23311 // one it has never heard of, and dropping it unread
23312 // meant `COLLATE nosuch_schema."C"` succeeded here —
23313 // a name that names nothing, accepted.
23314 let schema = cname.to_ascii_lowercase();
23315 if !matches!(
23316 schema.as_str(),
23317 "pg_catalog" | "public" | "information_schema"
23318 ) {
23319 return Err(self.err(alloc::format!("schema \"{cname}\" does not exist")));
23320 }
23321 self.advance();
23322 cname = match self.advance() {
23323 Token::Ident(s) | Token::QuotedIdent(s) | Token::String(s) => s,
23324 // `default` lexes as a KEYWORD, and it is the name
23325 // pg_dump writes — the same trap round 535 hit with
23326 // TABLE / INDEX / FULL.
23327 Token::Default => alloc::string::String::from("default"),
23328 other => {
23329 return Err(self.err(alloc::format!(
23330 "expected collation name after COLLATE, got {other:?}"
23331 )));
23332 }
23333 };
23334 }
23335 let lc = cname.to_ascii_lowercase();
23336 // v7.39 (round 371, M4 P4b) — a per-expression MySQL
23337 // collation override. `… COLLATE utf8mb4_bin` (any `_bin`
23338 // family / `binary`) forces byte-wise, which is exactly
23339 // what `BINARY expr` does — lower onto that so every fold
23340 // site (comparison, LIKE, ORDER BY) suppresses via
23341 // `is_binary_coerced`. A `_ci` family override folds, and
23342 // under the MySQL dialect the default already folds, so it
23343 // absorbs as a no-op; likewise the C / byte-order spellings.
23344 // v7.39.2 — against MySQL's own list, not against the
23345 // shape of the name. `nosuch_bin` took this shortcut and
23346 // became a BINARY cast; `nosuch_ci` took the one below
23347 // and was absorbed as a no-op. Either way the client
23348 // named a collation that does not exist and was told
23349 // nothing. An unknown name now falls through to the
23350 // node, and the engine refuses it.
23351 let real = crate::charset::is_mysql_collation(&lc);
23352 if self.mysql_dialect && real && (lc.ends_with("_bin") || lc == "binary") {
23353 expr = Expr::Cast {
23354 expr: alloc::boxed::Box::new(expr),
23355 target: CastTarget::Named("binary".to_string()),
23356 };
23357 continue;
23358 }
23359 let mysql_ci = self.mysql_dialect
23360 && ((real && lc.ends_with("_ci"))
23361 || matches!(lc.as_str(), "case_insensitive" | "nocase"));
23362 // v7.39 (round 691/692) — inside an ORDER BY key EVERY name
23363 // goes to the lowering channel, the byte-order spellings
23364 // included. Round 691 recorded only the names the old
23365 // allow-list rejected, which left `ORDER BY a COLLATE "C"`
23366 // absorbed as a no-op — and once a column could declare a
23367 // collation, absorbing the clause meant the COLUMN's
23368 // collation won where the query had asked for bytes.
23369 if self.in_order_by_key && !mysql_ci {
23370 self.order_key_collation = Some(cname);
23371 continue;
23372 }
23373 // v7.39.2 — the clause becomes a NODE rather than being
23374 // refused or absorbed.
23375 //
23376 // What stood here refused the locale names and SILENTLY
23377 // DROPPED the byte-order ones, so `'a' COLLATE "C" < 'B'`
23378 // answered `t` where PostgreSQL 18.6 answers `f`: the one
23379 // family it let through is the one where dropping it
23380 // changes the answer. Absorbing is only correct when the
23381 // collation asked for is the one the comparison would use
23382 // anyway, and that depends on the DATABASE — which the
23383 // parser cannot see. So it rides along and the engine,
23384 // which can, decides.
23385 //
23386 // `collate_derive` already modelled `Explicit(name)` and
23387 // had no way to be handed one.
23388 // v7.39.2 — a MySQL spelling does not exist on the
23389 // PostgreSQL wire, and THIS is where the wire is known.
23390 //
23391 // The check lived in the evaluator first and asked
23392 // `ctx.mysql_dialect`, which the INSERT path builds as a
23393 // hard-coded `false` — so `INSERT … VALUES (_utf8mb4'x')`
23394 // in a MySQL session was refused for a collation that
23395 // does not exist on a wire it was not on. Making that
23396 // context truthful would change INSERT-time evaluation
23397 // in other ways as a side effect; the parser already
23398 // gates the introducer on the same flag and is the
23399 // honest place to ask.
23400 if !self.mysql_dialect
23401 && (lc.ends_with("_ci")
23402 || lc.ends_with("_cs")
23403 || lc.ends_with("_bin")
23404 || lc == "binary"
23405 || matches!(lc.as_str(), "case_insensitive" | "nocase"))
23406 {
23407 return Err(self.err(alloc::format!(
23408 "collation \"{cname}\" for encoding \"UTF8\" does not exist"
23409 )));
23410 }
23411 // v7.39.3 — the node is built for EVERY name, `_ci`
23412 // included.
23413 //
23414 // A MySQL `_ci` spelling used to be absorbed here on the
23415 // reasoning that a MySQL session folds anyway, so the
23416 // clause asked for what it would have got. That stopped
23417 // being true when the fold learned to read the session's
23418 // collation NAME: under `SET NAMES utf8mb4 COLLATE
23419 // utf8mb4_bin`, `'AB' COLLATE utf8mb4_general_ci = 'ab'`
23420 // is 1 on MySQL 9.7.2 and was 0 here, because the clause
23421 // that would have made it 1 had been dropped in the
23422 // parser. Absorbing is only ever correct when the
23423 // collation asked for is the one the comparison would use
23424 // anyway, and the parser cannot know that — the same
23425 // reasoning already written above for the byte-order
23426 // spellings, applied to the family it had exempted.
23427 expr = Expr::Collate {
23428 expr: alloc::boxed::Box::new(expr),
23429 collation: cname,
23430 };
23431 continue;
23432 }
23433 return Ok(expr);
23434 }
23435 }
23436
23437 /// v7.39 (round 696) — a comma-separated list of bare names, stopping at
23438 /// the first token that is not one. Schema qualifiers collapse to the
23439 /// last part, which is what every other name path here does (SPG is
23440 /// single-schema).
23441 fn take_comma_separated_names(&mut self) -> Vec<String> {
23442 let mut out = Vec::new();
23443 while let Token::Ident(n) | Token::QuotedIdent(n) = self.peek().clone() {
23444 self.advance();
23445 let mut last = n;
23446 while matches!(self.peek(), Token::Dot) {
23447 self.advance();
23448 if let Token::Ident(t) | Token::QuotedIdent(t) = self.advance() {
23449 last = t;
23450 }
23451 }
23452 out.push(last);
23453 if matches!(self.peek(), Token::Comma) {
23454 self.advance();
23455 } else {
23456 break;
23457 }
23458 }
23459 out
23460 }
23461
23462 /// v7.39 (round 694) — is the next token pair a postfix `[]`?
23463 ///
23464 /// The general cast-target path tests this inline; the types with their
23465 /// own `CastTarget` variant need it as a guard on their match arm,
23466 /// which is what this exists for.
23467 fn peek_postfix_array_brackets(&self) -> bool {
23468 matches!(self.peek(), Token::LBracket)
23469 && matches!(self.tokens.get(self.pos + 1), Some(Token::RBracket))
23470 }
23471
23472 /// Parse the operator tail after a `(a, b, …)` row constructor
23473 /// and expand at parse time. `=` is the conjunction of element
23474 /// equalities; `<>` its negation; the order operators expand
23475 /// lexicographically; `[NOT] IN ( (row), … )` ORs the row
23476 /// equalities. Anything else (a bare row value, a subquery
23477 /// RHS) errors honestly — SPG has no composite runtime value.
23478 fn parse_row_comparison_tail(&mut self, row: Vec<Expr>) -> Result<Expr, ParseError> {
23479 fn row_eq(lhs: &[Expr], rhs: &[Expr]) -> Expr {
23480 let mut it = lhs.iter().zip(rhs.iter()).map(|(l, r)| Expr::Binary {
23481 lhs: Box::new(l.clone()),
23482 op: BinOp::Eq,
23483 rhs: Box::new(r.clone()),
23484 });
23485 let first = it.next().expect("row has at least two elements");
23486 it.fold(first, |acc, e| Expr::Binary {
23487 lhs: Box::new(acc),
23488 op: BinOp::And,
23489 rhs: Box::new(e),
23490 })
23491 }
23492 // Lexicographic (a,b) OP (c,d):
23493 // a STRICT c OR (a = c AND (b OP d)) — recursing right.
23494 fn row_lex(lhs: &[Expr], rhs: &[Expr], strict: BinOp, last: BinOp) -> Expr {
23495 if lhs.len() == 1 {
23496 return Expr::Binary {
23497 lhs: Box::new(lhs[0].clone()),
23498 op: last,
23499 rhs: Box::new(rhs[0].clone()),
23500 };
23501 }
23502 let head_strict = Expr::Binary {
23503 lhs: Box::new(lhs[0].clone()),
23504 op: strict,
23505 rhs: Box::new(rhs[0].clone()),
23506 };
23507 let head_eq = Expr::Binary {
23508 lhs: Box::new(lhs[0].clone()),
23509 op: BinOp::Eq,
23510 rhs: Box::new(rhs[0].clone()),
23511 };
23512 Expr::Binary {
23513 lhs: Box::new(head_strict),
23514 op: BinOp::Or,
23515 rhs: Box::new(Expr::Binary {
23516 lhs: Box::new(head_eq),
23517 op: BinOp::And,
23518 rhs: Box::new(row_lex(&lhs[1..], &rhs[1..], strict, last)),
23519 }),
23520 }
23521 }
23522 let negated_in = if matches!(self.peek(), Token::Not)
23523 && matches!(self.tokens.get(self.pos + 1), Some(Token::In))
23524 {
23525 self.advance();
23526 true
23527 } else {
23528 false
23529 };
23530 if matches!(self.peek(), Token::In) {
23531 self.advance();
23532 if !matches!(self.peek(), Token::LParen) {
23533 return Err(self.err(alloc::format!(
23534 "expected '(' after row IN, got {:?}",
23535 self.peek()
23536 )));
23537 }
23538 self.advance();
23539 // `(a, b) [NOT] IN (SELECT x, y)` — a multi-column subquery,
23540 // not a list of literal rows. Row-vs-list decomposes to
23541 // OR-of-AND above, but the subquery's rows are only known at
23542 // runtime, so keep it as a RowInSubquery node.
23543 if matches!(self.peek(), Token::Select) {
23544 let inner = self.parse_select_stmt()?;
23545 if !matches!(self.peek(), Token::RParen) {
23546 return Err(self.err(alloc::format!(
23547 "expected ')' after row IN-subquery, got {:?}",
23548 self.peek()
23549 )));
23550 }
23551 self.advance();
23552 let Statement::Select(s) = inner else {
23553 unreachable!("parse_select_stmt always returns Statement::Select")
23554 };
23555 return Ok(Expr::RowInSubquery {
23556 row,
23557 subquery: Box::new(s),
23558 negated: negated_in,
23559 });
23560 }
23561 let mut alternatives: Vec<Expr> = Vec::new();
23562 loop {
23563 // Optional ROW keyword before the paren row.
23564 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("row"))
23565 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
23566 {
23567 self.advance();
23568 }
23569 if !matches!(self.peek(), Token::LParen) {
23570 return Err(self.err(alloc::format!(
23571 "expected '(' to open a row inside IN, got {:?}",
23572 self.peek()
23573 )));
23574 }
23575 self.advance();
23576 let mut rhs = alloc::vec![self.parse_expr(0)?];
23577 while matches!(self.peek(), Token::Comma) {
23578 self.advance();
23579 rhs.push(self.parse_expr(0)?);
23580 }
23581 if !matches!(self.peek(), Token::RParen) {
23582 return Err(self.err(alloc::format!(
23583 "expected ')' after row inside IN, got {:?}",
23584 self.peek()
23585 )));
23586 }
23587 self.advance();
23588 if rhs.len() != row.len() {
23589 return Err(self.err(alloc::format!(
23590 "row IN arity mismatch: left has {}, right has {}",
23591 row.len(),
23592 rhs.len()
23593 )));
23594 }
23595 alternatives.push(row_eq(&row, &rhs));
23596 if matches!(self.peek(), Token::Comma) {
23597 self.advance();
23598 continue;
23599 }
23600 break;
23601 }
23602 if !matches!(self.peek(), Token::RParen) {
23603 return Err(self.err(alloc::format!(
23604 "expected ')' to close row IN list, got {:?}",
23605 self.peek()
23606 )));
23607 }
23608 self.advance();
23609 let mut it = alternatives.into_iter();
23610 let first = it.next().expect("IN list has at least one row");
23611 let combined = it.fold(first, |acc, e| Expr::Binary {
23612 lhs: Box::new(acc),
23613 op: BinOp::Or,
23614 rhs: Box::new(e),
23615 });
23616 return Ok(maybe_not(combined, negated_in));
23617 }
23618 // SQL-standard `(S1, E1) OVERLAPS (S2, E2)` — true when the
23619 // two periods share at least one time point. Each pair is
23620 // normalised with least/greatest (PG accepts the endpoints
23621 // in either order), then lowered to the standard
23622 // `start1 < end2 AND start2 < end1` form.
23623 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("overlaps")) {
23624 if row.len() != 2 {
23625 return Err(self.err(alloc::format!(
23626 "OVERLAPS needs (start, end) pairs; left side has {} elements",
23627 row.len()
23628 )));
23629 }
23630 self.advance();
23631 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("row"))
23632 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
23633 {
23634 self.advance();
23635 }
23636 if !matches!(self.peek(), Token::LParen) {
23637 return Err(self.err(alloc::format!(
23638 "expected '(' after OVERLAPS, got {:?}",
23639 self.peek()
23640 )));
23641 }
23642 self.advance();
23643 let r0 = self.parse_expr(0)?;
23644 if !matches!(self.peek(), Token::Comma) {
23645 return Err(self.err(alloc::format!(
23646 "OVERLAPS needs (start, end) on the right, got {:?}",
23647 self.peek()
23648 )));
23649 }
23650 self.advance();
23651 let r1 = self.parse_expr(0)?;
23652 if !matches!(self.peek(), Token::RParen) {
23653 return Err(self.err(alloc::format!(
23654 "expected ')' after OVERLAPS pair, got {:?}",
23655 self.peek()
23656 )));
23657 }
23658 self.advance();
23659 let pair_fn = |name: &str, a: &Expr, b: &Expr| Expr::FunctionCall {
23660 name: String::from(name),
23661 args: alloc::vec![a.clone(), b.clone()],
23662 };
23663 let lt = |lhs: Expr, rhs: Expr| Expr::Binary {
23664 lhs: Box::new(lhs),
23665 op: BinOp::Lt,
23666 rhs: Box::new(rhs),
23667 };
23668 return Ok(Expr::Binary {
23669 lhs: Box::new(lt(
23670 pair_fn("least", &row[0], &row[1]),
23671 pair_fn("greatest", &r0, &r1),
23672 )),
23673 op: BinOp::And,
23674 rhs: Box::new(lt(
23675 pair_fn("least", &r0, &r1),
23676 pair_fn("greatest", &row[0], &row[1]),
23677 )),
23678 });
23679 }
23680 // `(a, b, …) IS [NOT] NULL` — the SQL row null predicate. Per
23681 // PG, `IS NULL` is true only when EVERY field is NULL, and
23682 // `IS NOT NULL` is true only when every field is non-NULL — the
23683 // latter is NOT the negation of the former (a mixed row is
23684 // neither). Desugar to an AND chain of per-field `IS [NOT] NULL`,
23685 // which reproduces exactly that all-fields semantics.
23686 if matches!(self.peek(), Token::Is) {
23687 self.advance();
23688 let negated = if matches!(self.peek(), Token::Not) {
23689 self.advance();
23690 true
23691 } else {
23692 false
23693 };
23694 if !matches!(self.peek(), Token::Null) {
23695 return Err(self.err(alloc::format!(
23696 "expected NULL after row IS [NOT], got {:?}",
23697 self.peek()
23698 )));
23699 }
23700 self.advance();
23701 let mut it = row.iter().map(|e| Expr::IsNull {
23702 expr: Box::new(e.clone()),
23703 negated,
23704 });
23705 let first = it.next().expect("row has at least two elements");
23706 return Ok(it.fold(first, |acc, e| Expr::Binary {
23707 lhs: Box::new(acc),
23708 op: BinOp::And,
23709 rhs: Box::new(e),
23710 }));
23711 }
23712 let op = match self.peek() {
23713 Token::Eq => BinOp::Eq,
23714 Token::NotEq => BinOp::NotEq,
23715 Token::Lt => BinOp::Lt,
23716 Token::LtEq => BinOp::LtEq,
23717 Token::Gt => BinOp::Gt,
23718 Token::GtEq => BinOp::GtEq,
23719 // v7.38 (read01, composite) — a bare `(a, b, …)` not followed by a
23720 // comparison / [NOT] IN / IS [NOT] NULL / OVERLAPS is a row (record)
23721 // constructor value, identical to the `ROW(a, b, …)` keyword form:
23722 // `(1,'a')::text` → `(1,a)`, `SELECT (1,2,3)` → `(1,2,3)`. Postfix
23723 // (`::text`, `.field`) applies at the caller just as it does for the
23724 // ROW(...) node. All the comparison / predicate forms returned above.
23725 _ => {
23726 return Ok(Expr::FunctionCall {
23727 name: String::from("row"),
23728 args: row,
23729 });
23730 }
23731 };
23732 self.advance();
23733 // Optional ROW keyword before the paren row.
23734 if matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("row"))
23735 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen))
23736 {
23737 self.advance();
23738 }
23739 if !matches!(self.peek(), Token::LParen) {
23740 return Err(self.err(alloc::format!(
23741 "expected '(' to open the right-hand row, got {:?}",
23742 self.peek()
23743 )));
23744 }
23745 self.advance();
23746 // `(a, b) <op> (SELECT x, y)` — compare against a single-row
23747 // subquery. Kept as a node (the subquery's row is a runtime value);
23748 // the literal-RHS form below still decomposes at parse time.
23749 if matches!(self.peek(), Token::Select) {
23750 let inner = self.parse_select_stmt()?;
23751 if !matches!(self.peek(), Token::RParen) {
23752 return Err(self.err(alloc::format!(
23753 "expected ')' after row comparison subquery, got {:?}",
23754 self.peek()
23755 )));
23756 }
23757 self.advance();
23758 let Statement::Select(s) = inner else {
23759 unreachable!("parse_select_stmt always returns Statement::Select")
23760 };
23761 return Ok(Expr::RowCmpSubquery {
23762 row,
23763 op,
23764 subquery: Box::new(s),
23765 });
23766 }
23767 let mut rhs = alloc::vec![self.parse_expr(0)?];
23768 while matches!(self.peek(), Token::Comma) {
23769 self.advance();
23770 rhs.push(self.parse_expr(0)?);
23771 }
23772 if !matches!(self.peek(), Token::RParen) {
23773 return Err(self.err(alloc::format!(
23774 "expected ')' after right-hand row, got {:?}",
23775 self.peek()
23776 )));
23777 }
23778 self.advance();
23779 if rhs.len() != row.len() {
23780 // v7.39 (round 239) — PG's wording (42601).
23781 return Err(self.err("unequal number of entries in row expressions".to_string()));
23782 }
23783 Ok(match op {
23784 BinOp::Eq => row_eq(&row, &rhs),
23785 BinOp::NotEq => maybe_not(row_eq(&row, &rhs), true),
23786 BinOp::Lt => row_lex(&row, &rhs, BinOp::Lt, BinOp::Lt),
23787 BinOp::LtEq => row_lex(&row, &rhs, BinOp::Lt, BinOp::LtEq),
23788 BinOp::Gt => row_lex(&row, &rhs, BinOp::Gt, BinOp::Gt),
23789 BinOp::GtEq => row_lex(&row, &rhs, BinOp::Gt, BinOp::GtEq),
23790 _ => unreachable!("op restricted above"),
23791 })
23792 }
23793
23794 /// `LIKE p ESCAPE 'c'` — rewrite the pattern so the custom
23795 /// escape character becomes the matcher's default backslash:
23796 /// `c%` (escaped wildcard) → `\%`, `cc` (literal escape char)
23797 /// → the char itself, and any pre-existing backslash escapes
23798 /// itself so it stays literal. Both operands must be string
23799 /// literals — a runtime pattern would need matcher support.
23800 fn rewrite_like_escape(pattern: Expr, esc: Expr) -> Result<Expr, String> {
23801 let (Expr::Literal(Literal::String(p)), Expr::Literal(Literal::String(e))) =
23802 (&pattern, &esc)
23803 else {
23804 return Err(
23805 "LIKE ... ESCAPE requires string-literal pattern and escape \
23806 (runtime escape characters are not supported yet)"
23807 .into(),
23808 );
23809 };
23810 // v7.38 (read01 P6.18) — PG accepts `ESCAPE ''` to mean "no escape
23811 // character" (every `%`/`_` is a wildcard, nothing is escaped). Only a
23812 // multi-character escape is an error.
23813 let esc_ch: Option<char> = {
23814 let mut ch_iter = e.chars();
23815 match (ch_iter.next(), ch_iter.next()) {
23816 (Some(c), None) => Some(c),
23817 (None, _) => None,
23818 (Some(_), Some(_)) => {
23819 return Err(alloc::format!(
23820 "ESCAPE must be a single character, got {e:?}"
23821 ));
23822 }
23823 }
23824 };
23825 let mut out = String::with_capacity(p.len() + 4);
23826 let mut chars = p.chars();
23827 while let Some(c) = chars.next() {
23828 if Some(c) == esc_ch {
23829 match chars.next() {
23830 // Escaped wildcard / escaped escape → keep the
23831 // next char literal via backslash.
23832 Some(next) => {
23833 out.push('\\');
23834 out.push(next);
23835 }
23836 None => {
23837 return Err("LIKE pattern ends with the escape character".into());
23838 }
23839 }
23840 } else if c == '\\' && esc_ch != Some('\\') {
23841 // A raw backslash is literal under a custom (or absent) escape
23842 // — escape it for the backslash-based matcher.
23843 out.push('\\');
23844 out.push('\\');
23845 } else {
23846 out.push(c);
23847 }
23848 }
23849 Ok(Expr::Literal(Literal::String(out)))
23850 }
23851
23852 /// `x [NOT] LIKE ANY/ALL (ARRAY[p1, p2, …])` — quantified pattern
23853 /// match. Desugars to an OR (ANY) / AND (ALL) chain of per-element
23854 /// `x [NOT] LIKE pi`, which reproduces PG's three-valued semantics
23855 /// exactly (a NULL pattern makes an element NULL; `false OR NULL` =
23856 /// NULL, `true AND NULL` = NULL, …). ANY over an empty array is
23857 /// FALSE, ALL over empty is TRUE. Returns `None` when the token after
23858 /// LIKE is not `ANY(`/`ALL(`, so the caller falls back to a plain
23859 /// pattern. Only a literal `ARRAY[...]` is accepted today — a runtime
23860 /// array expression errors honestly rather than silently mismatching.
23861 fn try_like_any_all(
23862 &mut self,
23863 base: &Expr,
23864 negated: bool,
23865 case_insensitive: bool,
23866 ) -> Result<Option<Expr>, ParseError> {
23867 let is_any = match self.peek() {
23868 Token::All if matches!(self.tokens.get(self.pos + 1), Some(Token::LParen)) => false,
23869 Token::Ident(s)
23870 if s.eq_ignore_ascii_case("any")
23871 && matches!(self.tokens.get(self.pos + 1), Some(Token::LParen)) =>
23872 {
23873 true
23874 }
23875 _ => return Ok(None),
23876 };
23877 self.advance(); // ANY / ALL
23878 self.advance(); // '('
23879 let arr = self.parse_expr(0)?;
23880 if !matches!(self.peek(), Token::RParen) {
23881 return Err(self.err(format!(
23882 "expected ')' after LIKE {} argument, got {:?}",
23883 if is_any { "ANY" } else { "ALL" },
23884 self.peek()
23885 )));
23886 }
23887 self.advance(); // ')'
23888 let Expr::Array(items) = arr else {
23889 return Err(self.err(
23890 "LIKE ANY/ALL currently requires a literal ARRAY[...] of patterns".to_string(),
23891 ));
23892 };
23893 let mut clauses = items.into_iter().map(|p| Expr::Like {
23894 expr: Box::new(base.clone()),
23895 pattern: Box::new(p),
23896 negated,
23897 case_insensitive,
23898 });
23899 let Some(first) = clauses.next() else {
23900 // ANY(empty) = FALSE, ALL(empty) = TRUE.
23901 return Ok(Some(Expr::Literal(Literal::Bool(!is_any))));
23902 };
23903 let op = if is_any { BinOp::Or } else { BinOp::And };
23904 let combined = clauses.fold(first, |acc, c| Expr::Binary {
23905 lhs: Box::new(acc),
23906 op,
23907 rhs: Box::new(c),
23908 });
23909 Ok(Some(combined))
23910 }
23911
23912 /// `x BETWEEN low AND high` → `(x >= low) AND (x <= high)`, wrapped in
23913 /// `NOT` when `negated`. Bounds parse at precedence 5 so the trailing
23914 /// `AND` is not swallowed.
23915 fn parse_between_tail(&mut self, expr: Expr, negated: bool) -> Result<Expr, ParseError> {
23916 self.advance(); // BETWEEN
23917 // SYMMETRIC — the bounds may arrive in either order; both
23918 // orientations OR together. ASYMMETRIC is the default and
23919 // absorbs as noise.
23920 let symmetric = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("symmetric"))
23921 {
23922 self.advance();
23923 true
23924 } else {
23925 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("asymmetric")) {
23926 self.advance();
23927 }
23928 false
23929 };
23930 let low = self.parse_expr(6)?;
23931 if !matches!(self.peek(), Token::And) {
23932 return Err(self.err(format!(
23933 "expected AND after BETWEEN low bound, got {:?}",
23934 self.peek()
23935 )));
23936 }
23937 self.advance();
23938 let high = self.parse_expr(6)?;
23939 let target = Box::new(expr);
23940 let range = |lo: Expr, hi: Expr| Expr::Binary {
23941 lhs: Box::new(Expr::Binary {
23942 lhs: target.clone(),
23943 op: BinOp::GtEq,
23944 rhs: Box::new(lo),
23945 }),
23946 op: BinOp::And,
23947 rhs: Box::new(Expr::Binary {
23948 lhs: target.clone(),
23949 op: BinOp::LtEq,
23950 rhs: Box::new(hi),
23951 }),
23952 };
23953 let combined = if symmetric {
23954 Expr::Binary {
23955 lhs: Box::new(range(low.clone(), high.clone())),
23956 op: BinOp::Or,
23957 rhs: Box::new(range(high, low)),
23958 }
23959 } else {
23960 range(low, high)
23961 };
23962 Ok(maybe_not(combined, negated))
23963 }
23964
23965 /// `x IN (a, b, c)` → chained OR of equalities. Empty list collapses
23966 /// to FALSE (TRUE under NOT IN), matching standard SQL semantics.
23967 /// v4.11: parse `WITH name AS (SELECT ...) [, ...] SELECT ...`.
23968 /// Caller already consumed the leading `WITH` ident.
23969 /// v7.38 (read01) — recursive-CTE well-formedness. PG rejects ORDER BY
23970 /// / LIMIT / OFFSET anywhere in a recursive query, and a recursive
23971 /// self-reference that appears more than once in a single term.
23972 fn validate_recursive_cte(&self, cte: &crate::ast::Cte) -> Result<(), ParseError> {
23973 use crate::ast::{CteBody, SelectStatement};
23974 if !cte.recursive {
23975 return Ok(());
23976 }
23977 let CteBody::Select(body) = &cte.body else {
23978 return Ok(());
23979 };
23980 // A recursive CTE body is `base UNION [ALL] recursive [UNION …]`;
23981 // check the anchor and every peer term.
23982 let has_order = |s: &SelectStatement| !s.order_by.is_empty();
23983 let has_limit = |s: &SelectStatement| s.limit.is_some() || s.offset.is_some();
23984 if has_order(body) || body.unions.iter().any(|(_, u)| has_order(u)) {
23985 return Err(self.err(String::from(
23986 "ORDER BY in a recursive query is not implemented",
23987 )));
23988 }
23989 if has_limit(body) || body.unions.iter().any(|(_, u)| has_limit(u)) {
23990 return Err(self.err(String::from(
23991 "LIMIT in a recursive query is not implemented",
23992 )));
23993 }
23994 let self_refs = |s: &SelectStatement| -> usize {
23995 let Some(from) = &s.from else {
23996 return 0;
23997 };
23998 let mut n = usize::from(from.primary.name.eq_ignore_ascii_case(&cte.name));
23999 for j in &from.joins {
24000 if j.table.name.eq_ignore_ascii_case(&cte.name) {
24001 n += 1;
24002 }
24003 }
24004 n
24005 };
24006 if body.unions.iter().any(|(_, u)| self_refs(u) > 1) {
24007 return Err(self.err(alloc::format!(
24008 "recursive reference to query \"{}\" must not appear more than once",
24009 cte.name
24010 )));
24011 }
24012 // v7.39 (round 145, parse_cte.c) — the remaining well-formedness rules
24013 // apply only when the body actually references itself (a non-self-
24014 // referencing CTE under WITH RECURSIVE may use any set-op shape).
24015 let anchor_refs = self_refs(body);
24016 let union_refs = body.unions.iter().any(|(_, u)| self_refs(u) > 0);
24017 if anchor_refs > 0 || union_refs {
24018 // Shape: the top level must be UNION [ALL] arms only. A self-ref
24019 // under INTERSECT / EXCEPT (or with no set-op at all) is PG's
24020 // "does not have the form" error — SPG used to compute a value.
24021 if body.unions.is_empty()
24022 || body.unions.iter().any(|(k, _)| {
24023 !matches!(
24024 k,
24025 crate::ast::UnionKind::Distinct | crate::ast::UnionKind::All
24026 )
24027 })
24028 {
24029 return Err(self.err(alloc::format!(
24030 "recursive query \"{}\" does not have the form non-recursive-term \
24031 UNION [ALL] recursive-term",
24032 cte.name
24033 )));
24034 }
24035 if anchor_refs > 0 {
24036 return Err(self.err(alloc::format!(
24037 "recursive reference to query \"{}\" must not appear within its non-recursive term",
24038 cte.name
24039 )));
24040 }
24041 }
24042 let is_self = |t: &crate::ast::TableRef| t.name.eq_ignore_ascii_case(&cte.name);
24043 for (_, u) in &body.unions {
24044 if self_refs(u) == 0 {
24045 continue;
24046 }
24047 // The self-reference must not sit on the nullable side of an outer
24048 // join (LEFT: right side; RIGHT: everything before it; FULL: both).
24049 if let Some(from) = &u.from {
24050 for (i, j) in from.joins.iter().enumerate() {
24051 let left_has_self = is_self(&from.primary)
24052 || from.joins[..i].iter().any(|pj| is_self(&pj.table));
24053 let violated = match j.kind {
24054 crate::ast::JoinKind::Left => is_self(&j.table),
24055 crate::ast::JoinKind::Right => left_has_self,
24056 crate::ast::JoinKind::FullOuter => is_self(&j.table) || left_has_self,
24057 _ => false,
24058 };
24059 if violated {
24060 return Err(self.err(alloc::format!(
24061 "recursive reference to query \"{}\" must not appear within an outer join",
24062 cte.name
24063 )));
24064 }
24065 }
24066 }
24067 // No aggregates at the top level of the recursive term (SPG used
24068 // to run them and surface a misleading downstream error).
24069 let mut items_and_having: Vec<&Expr> = Vec::new();
24070 for it in &u.items {
24071 if let crate::ast::SelectItem::Expr { expr, .. } = it {
24072 items_and_having.push(expr);
24073 }
24074 }
24075 if let Some(h) = &u.having {
24076 items_and_having.push(h);
24077 }
24078 for e in items_and_having {
24079 if expr_has_toplevel_aggregate(e) {
24080 return Err(self.err(String::from(
24081 "aggregate functions are not allowed in a recursive query's recursive term",
24082 )));
24083 }
24084 }
24085 }
24086 // A self-reference inside a sublink expression (EXISTS / IN / scalar
24087 // subquery) anywhere in the body is rejected; a plain FROM derived
24088 // table is legal in PG and untouched here.
24089 let mut all_terms: Vec<&SelectStatement> = alloc::vec![body];
24090 all_terms.extend(body.unions.iter().map(|(_, u)| u));
24091 for term in all_terms {
24092 if select_has_self_ref_in_sublink(term, &cte.name) {
24093 return Err(self.err(alloc::format!(
24094 "recursive reference to query \"{}\" must not appear within a subquery",
24095 cte.name
24096 )));
24097 }
24098 }
24099 Ok(())
24100 }
24101
24102 /// v7.38 (read01 U16) — desugar a CTE's SEARCH / CYCLE clause into
24103 /// extra body columns, mirroring PG's `rewriteSearchAndCycle`. Runs
24104 /// right after parse so the engine sees a plain recursive CTE with the
24105 /// tracking columns already projected. DEPTH FIRST and CYCLE are
24106 /// supported; BREADTH FIRST needs numeric-composite ordering SPG's
24107 /// text-rendered rows can't provide, and errors honestly.
24108 fn desugar_cte_search_cycle(&self, cte: &mut crate::ast::Cte) -> Result<(), ParseError> {
24109 use crate::ast::{BinOp, ColumnName, CteBody, Expr, Literal, SelectItem, UnOp};
24110 if cte.search.is_none() && cte.cycle.is_none() {
24111 return Ok(());
24112 }
24113 let cte_name = cte.name.clone();
24114 let col_names = cte.column_overrides.clone();
24115 if col_names.is_empty() {
24116 return Err(
24117 self.err("SEARCH / CYCLE requires an explicit WITH name(cols) column list".into())
24118 );
24119 }
24120 let search = cte.search.take();
24121 let cycle = cte.cycle.take();
24122 let mut extra_cols: Vec<String> = Vec::new();
24123 let col_ref = |name: &str| {
24124 Expr::Column(ColumnName {
24125 qualifier: Some(cte_name.clone()),
24126 name: name.to_string(),
24127 })
24128 };
24129 // Position of a SEARCH/CYCLE column within the CTE's column list.
24130 let pos_of = |name: &str| -> Result<usize, ParseError> {
24131 col_names
24132 .iter()
24133 .position(|c| c.eq_ignore_ascii_case(name))
24134 .ok_or_else(|| {
24135 self.err(format!("SEARCH/CYCLE column {name:?} is not a CTE column"))
24136 })
24137 };
24138 let row_of = |items: &[SelectItem], positions: &[usize]| -> Result<Expr, ParseError> {
24139 let mut args = Vec::with_capacity(positions.len());
24140 for &p in positions {
24141 match items.get(p) {
24142 Some(SelectItem::Expr { expr, .. }) => args.push(expr.clone()),
24143 _ => {
24144 return Err(self.err(
24145 "SEARCH/CYCLE column maps to a non-expression select item".into(),
24146 ));
24147 }
24148 }
24149 }
24150 Ok(Expr::FunctionCall {
24151 name: "row".into(),
24152 args,
24153 })
24154 };
24155 let CteBody::Select(body) = &mut cte.body else {
24156 return Err(self.err("SEARCH / CYCLE requires a SELECT CTE body".into()));
24157 };
24158 if body.unions.is_empty() {
24159 return Err(self.err("SEARCH / CYCLE requires a recursive (UNION) CTE".into()));
24160 }
24161 let rec = body.unions.len() - 1; // recursive term = last UNION peer
24162
24163 if let Some(srch) = search {
24164 // v7.38 (T31) — SEARCH's SET column is ORDER BY'd, and PG's key is a
24165 // `record[]` (DEPTH) or `(depth, keys…)` record (BREADTH). SPG has
24166 // no typed `record[]`, but element-wise array ORDER BY is correct
24167 // (`[1,2] < [1,10] < [2]`), so a SINGLE scalar BY column maps
24168 // exactly onto a typed array: DEPTH is the root→node path
24169 // `array_append(parent, key)`, BREADTH is `[depth, key]`. This
24170 // orders numerically (multi-digit keys included), matching PG.
24171 //
24172 // A multi-column BY would need a record[] to keep the per-node key
24173 // tuple orderable, which SPG can't express — error honestly there
24174 // rather than mis-order.
24175 if srch.by_columns.len() != 1 {
24176 return Err(self.err(
24177 "SEARCH … BY with multiple columns needs typed record[] ordering \
24178 SPG doesn't have yet; a single BY column is supported"
24179 .into(),
24180 ));
24181 }
24182 let key_pos = pos_of(&srch.by_columns[0])?;
24183 let base_key = match body.items.get(key_pos) {
24184 Some(SelectItem::Expr { expr, .. }) => expr.clone(),
24185 _ => {
24186 return Err(
24187 self.err("SEARCH BY column maps to a non-expression select item".into())
24188 );
24189 }
24190 };
24191 let rec_key = match body.unions[rec].1.items.get(key_pos) {
24192 Some(SelectItem::Expr { expr, .. }) => expr.clone(),
24193 _ => {
24194 return Err(
24195 self.err("SEARCH BY column maps to a non-expression select item".into())
24196 );
24197 }
24198 };
24199 if srch.depth_first {
24200 // base: ARRAY[key]; rec: array_append(cte.set, key).
24201 body.items.push(SelectItem::Expr {
24202 expr: Expr::Array(alloc::vec![base_key]),
24203 alias: Some(srch.set_column.clone()),
24204 });
24205 body.unions[rec].1.items.push(SelectItem::Expr {
24206 expr: Expr::FunctionCall {
24207 name: "array_append".into(),
24208 args: alloc::vec![col_ref(&srch.set_column), rec_key],
24209 },
24210 alias: Some(srch.set_column.clone()),
24211 });
24212 } else {
24213 // BREADTH: [depth, key]; depth starts at 0 and increments. The
24214 // leading depth element dominates the element-wise comparison,
24215 // so shallower rows sort first, then by key — PG's (depth, key).
24216 body.items.push(SelectItem::Expr {
24217 expr: Expr::Array(alloc::vec![Expr::Literal(Literal::Integer(0)), base_key,]),
24218 alias: Some(srch.set_column.clone()),
24219 });
24220 // rec depth = cte.set[1] + 1.
24221 let parent_depth = Expr::ArraySubscript {
24222 target: Box::new(col_ref(&srch.set_column)),
24223 index: Box::new(Expr::Literal(Literal::Integer(1))),
24224 };
24225 body.unions[rec].1.items.push(SelectItem::Expr {
24226 expr: Expr::Array(alloc::vec![
24227 Expr::Binary {
24228 lhs: Box::new(parent_depth),
24229 op: BinOp::Add,
24230 rhs: Box::new(Expr::Literal(Literal::Integer(1))),
24231 },
24232 rec_key,
24233 ]),
24234 alias: Some(srch.set_column.clone()),
24235 });
24236 }
24237 extra_cols.push(srch.set_column);
24238 }
24239
24240 if let Some(cyc) = cycle {
24241 let positions: Vec<usize> = cyc
24242 .columns
24243 .iter()
24244 .map(|c| pos_of(c))
24245 .collect::<Result<_, _>>()?;
24246 // v7.38 (read01, T9) — ROW(cols) is now a first-class composite, so
24247 // cast it to text for the cycle path: membership only needs equality,
24248 // and the record text form gives SPG a TextArray path (SPG has no
24249 // typed record[] array). Cycle detection is unaffected.
24250 let base_row = Expr::Cast {
24251 expr: Box::new(row_of(&body.items, &positions)?),
24252 target: CastTarget::Text,
24253 };
24254 let rec_row = Expr::Cast {
24255 expr: Box::new(row_of(&body.unions[rec].1.items, &positions)?),
24256 target: CastTarget::Text,
24257 };
24258 let mark = cyc.mark_value.clone().unwrap_or(Literal::Bool(true));
24259 let dflt = cyc.default_value.clone().unwrap_or(Literal::Bool(false));
24260 // base: <default> AS mark, ARRAY[ROW(cols)] AS path.
24261 body.items.push(SelectItem::Expr {
24262 expr: Expr::Literal(dflt.clone()),
24263 alias: Some(cyc.mark_column.clone()),
24264 });
24265 body.items.push(SelectItem::Expr {
24266 expr: Expr::Array(alloc::vec![base_row]),
24267 alias: Some(cyc.path_column.clone()),
24268 });
24269 // rec mark: ROW(cols) already in the path → cycle.
24270 let hit = Expr::AnyAll {
24271 expr: Box::new(rec_row.clone()),
24272 op: BinOp::Eq,
24273 array: Box::new(col_ref(&cyc.path_column)),
24274 is_any: true,
24275 };
24276 let mark_expr = if cyc.mark_value.is_some() || cyc.default_value.is_some() {
24277 Expr::Case {
24278 operand: None,
24279 branches: alloc::vec![(hit, Expr::Literal(mark))],
24280 else_branch: Some(Box::new(Expr::Literal(dflt))),
24281 }
24282 } else {
24283 hit
24284 };
24285 body.unions[rec].1.items.push(SelectItem::Expr {
24286 expr: mark_expr,
24287 alias: Some(cyc.mark_column.clone()),
24288 });
24289 // rec path: array_append(cte.path, ROW(cols)).
24290 body.unions[rec].1.items.push(SelectItem::Expr {
24291 expr: Expr::FunctionCall {
24292 name: "array_append".into(),
24293 args: alloc::vec![col_ref(&cyc.path_column), rec_row],
24294 },
24295 alias: Some(cyc.path_column.clone()),
24296 });
24297 // rec WHERE: AND NOT cte.mark — stop expanding a cycled row.
24298 let stop = Expr::Unary {
24299 op: UnOp::Not,
24300 expr: Box::new(col_ref(&cyc.mark_column)),
24301 };
24302 let w = &mut body.unions[rec].1.where_;
24303 *w = Some(match w.take() {
24304 Some(prev) => Expr::Binary {
24305 lhs: Box::new(prev),
24306 op: BinOp::And,
24307 rhs: Box::new(stop),
24308 },
24309 None => stop,
24310 });
24311 extra_cols.push(cyc.mark_column);
24312 extra_cols.push(cyc.path_column);
24313 }
24314 cte.column_overrides.extend(extra_cols);
24315 Ok(())
24316 }
24317
24318 /// v7.38 (read01 U16) — `SEARCH { DEPTH | BREADTH } FIRST BY col [,
24319 /// col…] SET seqcol`. Returns None when the next token isn't SEARCH.
24320 fn parse_cte_search_clause(&mut self) -> Result<Option<crate::ast::SearchClause>, ParseError> {
24321 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("search")) {
24322 return Ok(None);
24323 }
24324 self.advance(); // SEARCH
24325 let depth_first = match self.peek() {
24326 Token::Ident(s) if s.eq_ignore_ascii_case("depth") => true,
24327 Token::Ident(s) if s.eq_ignore_ascii_case("breadth") => false,
24328 other => {
24329 return Err(self.err(format!(
24330 "expected DEPTH or BREADTH after SEARCH, got {other:?}"
24331 )));
24332 }
24333 };
24334 self.advance();
24335 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("first")) {
24336 return Err(self.err(format!(
24337 "expected FIRST after SEARCH mode, got {:?}",
24338 self.peek()
24339 )));
24340 }
24341 self.advance();
24342 if !self.peek_is_by() {
24343 return Err(self.err(format!(
24344 "expected BY after SEARCH … FIRST, got {:?}",
24345 self.peek()
24346 )));
24347 }
24348 self.advance();
24349 let mut by_columns = alloc::vec![self.expect_ident_like()?];
24350 while matches!(self.peek(), Token::Comma) {
24351 self.advance();
24352 by_columns.push(self.expect_ident_like()?);
24353 }
24354 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("set")) {
24355 return Err(self.err(format!(
24356 "expected SET in SEARCH clause, got {:?}",
24357 self.peek()
24358 )));
24359 }
24360 self.advance();
24361 let set_column = self.expect_ident_like()?;
24362 Ok(Some(crate::ast::SearchClause {
24363 depth_first,
24364 by_columns,
24365 set_column,
24366 }))
24367 }
24368
24369 /// v7.38 (read01 U16) — `CYCLE col [, col…] SET markcol [TO v DEFAULT w]
24370 /// USING pathcol`. Returns None when the next token isn't CYCLE.
24371 fn parse_cte_cycle_clause(&mut self) -> Result<Option<crate::ast::CycleClause>, ParseError> {
24372 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("cycle")) {
24373 return Ok(None);
24374 }
24375 self.advance(); // CYCLE
24376 let mut columns = alloc::vec![self.expect_ident_like()?];
24377 while matches!(self.peek(), Token::Comma) {
24378 self.advance();
24379 columns.push(self.expect_ident_like()?);
24380 }
24381 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("set")) {
24382 return Err(self.err(format!(
24383 "expected SET in CYCLE clause, got {:?}",
24384 self.peek()
24385 )));
24386 }
24387 self.advance();
24388 let mark_column = self.expect_ident_like()?;
24389 let mut mark_value = None;
24390 let mut default_value = None;
24391 if matches!(self.peek(), Token::To) {
24392 self.advance();
24393 mark_value = Some(self.parse_cycle_literal()?);
24394 if !matches!(self.peek(), Token::Default) {
24395 return Err(self.err(format!(
24396 "expected DEFAULT after CYCLE … TO, got {:?}",
24397 self.peek()
24398 )));
24399 }
24400 self.advance();
24401 default_value = Some(self.parse_cycle_literal()?);
24402 }
24403 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
24404 return Err(self.err(format!(
24405 "expected USING in CYCLE clause, got {:?}",
24406 self.peek()
24407 )));
24408 }
24409 self.advance();
24410 let path_column = self.expect_ident_like()?;
24411 Ok(Some(crate::ast::CycleClause {
24412 columns,
24413 mark_column,
24414 mark_value,
24415 default_value,
24416 path_column,
24417 }))
24418 }
24419
24420 /// The mark / default value in a CYCLE `TO v DEFAULT w` — a bare
24421 /// literal (string / bool / number) in PG.
24422 fn parse_cycle_literal(&mut self) -> Result<crate::ast::Literal, ParseError> {
24423 match self.parse_expr(0)? {
24424 Expr::Literal(l) => Ok(l),
24425 other => Err(self.err(format!(
24426 "CYCLE mark/default value must be a literal, got {other:?}"
24427 ))),
24428 }
24429 }
24430
24431 fn parse_with_cte_then_select(&mut self) -> Result<Statement, ParseError> {
24432 // v4.22: WITH RECURSIVE — optional keyword right after WITH.
24433 // Comes through as an identifier; consume it if present and
24434 // mark every CTE in the clause as recursive (PG semantics —
24435 // the flag is per-WITH, not per-CTE).
24436 let mut recursive = false;
24437 if let Token::Ident(s) | Token::QuotedIdent(s) = self.peek()
24438 && s.eq_ignore_ascii_case("recursive")
24439 {
24440 self.advance();
24441 recursive = true;
24442 }
24443 let mut ctes = Vec::new();
24444 loop {
24445 let name = self.expect_ident_like()?;
24446 // v4.22: optional column-name list — `WITH t(a,b,c) AS ...`.
24447 // PG uses these to rename the body's output columns; we
24448 // do the same below by overriding `columns[i].name`.
24449 let column_overrides: Vec<String> = if matches!(self.peek(), Token::LParen) {
24450 self.advance();
24451 let mut names = Vec::new();
24452 loop {
24453 names.push(self.expect_ident_like()?);
24454 if matches!(self.peek(), Token::Comma) {
24455 self.advance();
24456 continue;
24457 }
24458 break;
24459 }
24460 if !matches!(self.peek(), Token::RParen) {
24461 return Err(self.err(format!(
24462 "expected ')' to close CTE column list, got {:?}",
24463 self.peek()
24464 )));
24465 }
24466 self.advance();
24467 names
24468 } else {
24469 Vec::new()
24470 };
24471 // AS is a reserved Token::As (used by SELECT-item / FROM
24472 // aliasing) — handle it specially rather than as a bare
24473 // ident.
24474 if !matches!(self.peek(), Token::As) {
24475 return Err(self.err(format!(
24476 "expected AS after CTE name {name:?}, got {:?}",
24477 self.peek()
24478 )));
24479 }
24480 self.advance();
24481 // v7.37.17 (17.6 siblings) — PG 12+ `AS [NOT]
24482 // MATERIALIZED` optimizer hints. SPG materialises every
24483 // CTE, so both spellings are accepted and absorbed.
24484 if matches!(self.peek(), Token::Not) {
24485 self.advance(); // NOT
24486 if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
24487 if s.eq_ignore_ascii_case("materialized"))
24488 {
24489 self.advance();
24490 } else {
24491 return Err(self.err(format!(
24492 "expected MATERIALIZED after AS NOT, got {:?}",
24493 self.peek()
24494 )));
24495 }
24496 } else if matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s)
24497 if s.eq_ignore_ascii_case("materialized"))
24498 {
24499 self.advance();
24500 }
24501 if !matches!(self.peek(), Token::LParen) {
24502 return Err(self.err(format!(
24503 "expected '(' after AS in WITH clause, got {:?}",
24504 self.peek()
24505 )));
24506 }
24507 self.advance();
24508 // v7.37.43-T4.4 — accept INSERT / UPDATE / DELETE (with
24509 // RETURNING) as the CTE body in addition to SELECT.
24510 // PG writable CTE semantics. UPDATE / DELETE come in as
24511 // bare Idents (lexer keeps SELECT / INSERT as reserved
24512 // tokens but treats the rest of DML as case-insensitive
24513 // idents).
24514 let is_update_kw = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("update"));
24515 let is_delete_kw = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("delete"));
24516 let is_merge_kw = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("merge"));
24517 let body = match self.peek() {
24518 Token::Select => {
24519 let inner = self.parse_select_stmt()?;
24520 let Statement::Select(s) = inner else {
24521 unreachable!("parse_select_stmt returns Select");
24522 };
24523 crate::ast::CteBody::Select(s)
24524 }
24525 // v7.39 (round 869) — `WITH x AS (TABLE t)`. PG spells
24526 // `SELECT * FROM t` this way and accepts it wherever a
24527 // SELECT goes, so the CTE body dispatch needs its own
24528 // arm: this match is keyed on the FIRST token, and
24529 // `Token::Table` fell through to a tail that then
24530 // rejected what it got. `parse_table_shorthand` has
24531 // returned a desugared SelectStatement since the
24532 // shorthand landed — only the routing was missing.
24533 // Round 868 found this by putting the shorthand in a
24534 // subquery; every earlier check used a top-level form.
24535 // v7.39 (round 869) — `WITH x AS (TABLE t)`. PG spells
24536 // `SELECT * FROM t` this way and accepts it wherever a
24537 // SELECT goes, so the CTE body dispatch needs its own
24538 // arm: this match is keyed on the FIRST token, and
24539 // `Token::Table` fell through to a tail that rejected
24540 // what it got. `parse_table_shorthand` has returned a
24541 // desugared SelectStatement since the shorthand landed —
24542 // only the routing was missing, here and in the derived
24543 // table's second-token gate. Round 868 found both by
24544 // putting the shorthand in a subquery; every earlier
24545 // check had used a top-level form.
24546 Token::Table
24547 if matches!(
24548 self.tokens.get(self.pos + 1),
24549 Some(Token::Ident(_) | Token::QuotedIdent(_))
24550 ) =>
24551 {
24552 let mut head = self.parse_table_shorthand()?;
24553 self.parse_setop_chain_into(&mut head)?;
24554 self.parse_select_tail_into(&mut head)?;
24555 crate::ast::CteBody::Select(head)
24556 }
24557 // v7.37.17 (17.6 siblings) — VALUES as a CTE body:
24558 // WITH t(a) AS (VALUES (1), (2)) … lowers through
24559 // the shared rows helper onto a Select body.
24560 Token::Values => {
24561 self.advance(); // VALUES
24562 let mut head = self.parse_values_rows_body()?;
24563 // A VALUES seed can head a set-operation chain —
24564 // WITH RECURSIVE t(n) AS (VALUES(1) UNION ALL
24565 // SELECT n+1 FROM t …). Attach any trailing
24566 // UNION / INTERSECT / EXCEPT peers so the
24567 // recursive-CTE body parses like the SELECT seed.
24568 self.parse_setop_chain_into(&mut head)?;
24569 crate::ast::CteBody::Select(head)
24570 }
24571 Token::Insert => {
24572 let inner = self.parse_one_statement()?;
24573 let Statement::Insert(s) = inner else {
24574 unreachable!("Token::Insert routes to Insert");
24575 };
24576 crate::ast::CteBody::Insert(alloc::boxed::Box::new(s))
24577 }
24578 _ if is_update_kw => {
24579 let inner = self.parse_one_statement()?;
24580 let Statement::Update(s) = inner else {
24581 return Err(
24582 self.err(format!("expected UPDATE inside WITH (…), got {inner:?}"))
24583 );
24584 };
24585 crate::ast::CteBody::Update(alloc::boxed::Box::new(s))
24586 }
24587 _ if is_delete_kw => {
24588 let inner = self.parse_one_statement()?;
24589 let Statement::Delete(s) = inner else {
24590 return Err(
24591 self.err(format!("expected DELETE inside WITH (…), got {inner:?}"))
24592 );
24593 };
24594 crate::ast::CteBody::Delete(alloc::boxed::Box::new(s))
24595 }
24596 // v7.39 (round 149) — PG 17 allows MERGE as a
24597 // data-modifying CTE body.
24598 _ if is_merge_kw => {
24599 let inner = self.parse_one_statement()?;
24600 let Statement::Merge(s) = inner else {
24601 return Err(
24602 self.err(format!("expected MERGE inside WITH (…), got {inner:?}"))
24603 );
24604 };
24605 crate::ast::CteBody::Merge(alloc::boxed::Box::new(s))
24606 }
24607 // v7.39 (round 151) — a CTE body may itself be
24608 // WITH-headed (PG grammar: PreparableStmt carries its
24609 // own with_clause). The nested statement keeps its own
24610 // ctes; the modifying-CTE-at-top-level rule is enforced
24611 // at execution.
24612 Token::Ident(s) if s.eq_ignore_ascii_case("with") => {
24613 self.advance(); // WITH
24614 match self.parse_with_cte_then_select()? {
24615 Statement::Select(s) => crate::ast::CteBody::Select(s),
24616 Statement::Insert(s) => {
24617 crate::ast::CteBody::Insert(alloc::boxed::Box::new(s))
24618 }
24619 Statement::Update(s) => {
24620 crate::ast::CteBody::Update(alloc::boxed::Box::new(s))
24621 }
24622 Statement::Delete(s) => {
24623 crate::ast::CteBody::Delete(alloc::boxed::Box::new(s))
24624 }
24625 Statement::Merge(s) => {
24626 crate::ast::CteBody::Merge(alloc::boxed::Box::new(s))
24627 }
24628
24629 other => {
24630 return Err(self.err(format!(
24631 "WITH body must be SELECT / INSERT / UPDATE / DELETE / MERGE, got {other:?}"
24632 )));
24633 }
24634 }
24635 }
24636 other => {
24637 return Err(self.err(format!(
24638 "WITH body must be SELECT / INSERT / UPDATE / DELETE / MERGE, got {other:?}"
24639 )));
24640 }
24641 };
24642 if !matches!(self.peek(), Token::RParen) {
24643 return Err(self.err(format!(
24644 "expected ')' after CTE body, got {:?}",
24645 self.peek()
24646 )));
24647 }
24648 self.advance();
24649 // v7.38 (read01 U16) — optional SEARCH / CYCLE on a recursive
24650 // CTE, desugared into extra body columns by the engine.
24651 let search = self.parse_cte_search_clause()?;
24652 let cycle = self.parse_cte_cycle_clause()?;
24653 let mut cte = crate::ast::Cte {
24654 name,
24655 body,
24656 recursive,
24657 column_overrides,
24658 search,
24659 cycle,
24660 };
24661 self.validate_recursive_cte(&cte)?;
24662 self.desugar_cte_search_cycle(&mut cte)?;
24663 ctes.push(cte);
24664 if matches!(self.peek(), Token::Comma) {
24665 self.advance();
24666 continue;
24667 }
24668 break;
24669 }
24670 // v7.37.43-T4.4 — the outer body may be SELECT (classical),
24671 // or INSERT / UPDATE / DELETE (writable CTE outer). Attach
24672 // the parsed CTEs to whichever statement the body produces.
24673 let outer_is_update = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("update"));
24674 let outer_is_delete = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("delete"));
24675 let outer_is_merge = matches!(self.peek(), Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("merge"));
24676 match self.peek() {
24677 Token::Select => {
24678 let body_stmt = self.parse_select_stmt()?;
24679 let Statement::Select(mut body) = body_stmt else {
24680 unreachable!()
24681 };
24682 body.ctes = ctes;
24683 Ok(Statement::Select(body))
24684 }
24685 Token::Insert => {
24686 let body_stmt = self.parse_one_statement()?;
24687 let Statement::Insert(mut body) = body_stmt else {
24688 unreachable!()
24689 };
24690 body.ctes = ctes;
24691 Ok(Statement::Insert(body))
24692 }
24693 _ if outer_is_update => {
24694 let body_stmt = self.parse_one_statement()?;
24695 let Statement::Update(mut body) = body_stmt else {
24696 return Err(self.err(format!("expected UPDATE after WITH clause")));
24697 };
24698 body.ctes = ctes;
24699 Ok(Statement::Update(body))
24700 }
24701 _ if outer_is_delete => {
24702 let body_stmt = self.parse_one_statement()?;
24703 let Statement::Delete(mut body) = body_stmt else {
24704 return Err(self.err(format!("expected DELETE after WITH clause")));
24705 };
24706 body.ctes = ctes;
24707 Ok(Statement::Delete(body))
24708 }
24709 // v7.39 (round 149) — PG 15 allows a WITH clause on MERGE;
24710 // WITH RECURSIVE is rejected with PG's exact message
24711 // (parse analysis, transformWithClause).
24712 _ if outer_is_merge => {
24713 if recursive {
24714 return Err(self.err(String::from(
24715 "WITH RECURSIVE is not supported for MERGE statement",
24716 )));
24717 }
24718 let body_stmt = self.parse_one_statement()?;
24719 let Statement::Merge(mut body) = body_stmt else {
24720 return Err(self.err(format!("expected MERGE after WITH clause")));
24721 };
24722 body.ctes = ctes;
24723 Ok(Statement::Merge(body))
24724 }
24725 other => Err(self.err(format!(
24726 "expected SELECT / INSERT / UPDATE / DELETE / MERGE after WITH clause, got {other:?}"
24727 ))),
24728 }
24729 }
24730
24731 /// v4.10: parse `EXISTS (SELECT ...)`. Caller (`parse_atom`)
24732 /// already consumed the leading `EXISTS` ident via
24733 /// `self.advance()`.
24734 /// v7.13.0 — parse the rest of a `CASE … END` expression after
24735 /// the leading `CASE` ident has been consumed (mailrs round-5
24736 /// G9). Supports both the searched form
24737 /// (`CASE WHEN cond THEN val …`) and the simple form
24738 /// (`CASE operand WHEN val THEN val …`).
24739 fn parse_case_atom(&mut self) -> Result<Expr, ParseError> {
24740 // Disambiguate searched vs simple form: if the next token
24741 // is `WHEN`, we're in the searched form. Otherwise the
24742 // intervening expression is the operand.
24743 let operand = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("when")) {
24744 None
24745 } else {
24746 Some(Box::new(self.parse_expr(0)?))
24747 };
24748 let mut branches: Vec<(Expr, Expr)> = Vec::new();
24749 loop {
24750 match self.peek() {
24751 Token::Ident(s) if s.eq_ignore_ascii_case("when") => {
24752 self.advance();
24753 let cond = self.parse_expr(0)?;
24754 match self.peek() {
24755 Token::Ident(t) if t.eq_ignore_ascii_case("then") => {
24756 self.advance();
24757 }
24758 other => {
24759 return Err(self.err(alloc::format!(
24760 "expected THEN after CASE WHEN <expr>, got {other:?}"
24761 )));
24762 }
24763 }
24764 let value = self.parse_expr(0)?;
24765 branches.push((cond, value));
24766 }
24767 _ => break,
24768 }
24769 }
24770 if branches.is_empty() {
24771 return Err(self.err("CASE requires at least one WHEN … THEN … branch".into()));
24772 }
24773 let else_branch = if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("else"))
24774 {
24775 self.advance();
24776 Some(Box::new(self.parse_expr(0)?))
24777 } else {
24778 None
24779 };
24780 match self.peek() {
24781 Token::Ident(s) if s.eq_ignore_ascii_case("end") => {
24782 self.advance();
24783 }
24784 other => {
24785 return Err(self.err(alloc::format!(
24786 "expected END to close CASE expression, got {other:?}"
24787 )));
24788 }
24789 }
24790 Ok(Expr::Case {
24791 operand,
24792 branches,
24793 else_branch,
24794 })
24795 }
24796
24797 /// v7.39 (round 151) — nested `WITH … SELECT …` in a subquery /
24798 /// query-source position (EXISTS / IN / INSERT source / CTE body /
24799 /// view body). Caller consumed the WITH keyword. Only a SELECT
24800 /// outer is grammatical here; the data-modifying-CTE-at-top-level
24801 /// rule (PG 0A000) is enforced at execution, where the SQLSTATE
24802 /// maps correctly.
24803 fn parse_nested_with_select(&mut self) -> Result<crate::ast::SelectStatement, ParseError> {
24804 let inner = self.parse_with_cte_then_select()?;
24805 match inner {
24806 Statement::Select(s) => Ok(s),
24807 other => Err(self.err(format!(
24808 "expected SELECT after WITH in a subquery, got {other:?}"
24809 ))),
24810 }
24811 }
24812
24813 /// True when the next token is the (unquoted) WITH keyword. WITH is
24814 /// reserved in PG, so a bare `with` can never be a column reference
24815 /// in these positions; a quoted `"with"` stays an identifier.
24816 fn peek_is_with_kw(&self) -> bool {
24817 matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("with"))
24818 }
24819
24820 /// v7.39 (round 153) — the `ANY / ALL ( [WITH …] SELECT … )` body.
24821 /// `#[inline(never)]` keeps the large SelectStatement temporaries
24822 /// off parse_expr's recursive frame (the nesting-budget stack
24823 /// cliff — see the round-153 gate regression).
24824 #[inline(never)]
24825 fn parse_any_all_select_body(&mut self) -> Result<crate::ast::SelectStatement, ParseError> {
24826 if self.peek_is_with_kw() {
24827 self.advance();
24828 self.parse_nested_with_select()
24829 } else {
24830 match self.parse_select_stmt()? {
24831 Statement::Select(s) => Ok(s),
24832 other => Err(self.err(alloc::format!(
24833 "expected SELECT inside ANY/ALL, got {other:?}"
24834 ))),
24835 }
24836 }
24837 }
24838
24839 fn parse_exists_atom(&mut self, negated: bool) -> Result<Expr, ParseError> {
24840 if !matches!(self.peek(), Token::LParen) {
24841 return Err(self.err(format!("expected '(' after EXISTS, got {:?}", self.peek())));
24842 }
24843 self.advance();
24844 // v7.39 (round 151) — `EXISTS (WITH … SELECT …)` is legal PG.
24845 let s = if self.peek_is_with_kw() {
24846 self.advance();
24847 self.parse_nested_with_select()?
24848 } else {
24849 let inner = self.parse_select_stmt()?;
24850 let Statement::Select(s) = inner else {
24851 unreachable!("parse_select_stmt returns Select")
24852 };
24853 s
24854 };
24855 if !matches!(self.peek(), Token::RParen) {
24856 return Err(self.err(format!(
24857 "expected ')' after EXISTS-subquery, got {:?}",
24858 self.peek()
24859 )));
24860 }
24861 self.advance();
24862 Ok(Expr::Exists {
24863 subquery: Box::new(s),
24864 negated,
24865 })
24866 }
24867
24868 fn parse_in_tail(&mut self, expr: Expr, negated: bool) -> Result<Expr, ParseError> {
24869 self.advance(); // IN
24870 if !matches!(self.peek(), Token::LParen) {
24871 return Err(self.err(format!("expected '(' after IN, got {:?}", self.peek())));
24872 }
24873 self.advance();
24874 // v4.10: `IN (SELECT ...)` — subquery branch. v7.39 (round 151)
24875 // also accepts a WITH-headed subquery (`IN (WITH … SELECT …)`).
24876 if matches!(self.peek(), Token::Select) || self.peek_is_with_kw() {
24877 let s = if self.peek_is_with_kw() {
24878 self.advance();
24879 self.parse_nested_with_select()?
24880 } else {
24881 let inner = self.parse_select_stmt()?;
24882 let Statement::Select(s) = inner else {
24883 unreachable!("parse_select_stmt always returns Statement::Select")
24884 };
24885 s
24886 };
24887 if !matches!(self.peek(), Token::RParen) {
24888 return Err(self.err(format!(
24889 "expected ')' after IN-subquery, got {:?}",
24890 self.peek()
24891 )));
24892 }
24893 self.advance();
24894 return Ok(Expr::InSubquery {
24895 expr: Box::new(expr),
24896 subquery: Box::new(s),
24897 negated,
24898 });
24899 }
24900 let mut elements = Vec::new();
24901 if !matches!(self.peek(), Token::RParen) {
24902 loop {
24903 elements.push(self.parse_expr(0)?);
24904 match self.peek() {
24905 Token::Comma => {
24906 self.advance();
24907 }
24908 Token::RParen => break,
24909 other => {
24910 return Err(
24911 self.err(format!("expected ',' or ')' in IN list, got {other:?}"))
24912 );
24913 }
24914 }
24915 }
24916 }
24917 self.advance(); // ')'
24918 // v7.30.2 (mailrs round-25) — flat InList node instead of a
24919 // left-deep OR-Eq chain: chain depth scaled with the element
24920 // count and overflowed the stack (eval + drop are recursive).
24921 if elements.is_empty() {
24922 return Ok(maybe_not(Expr::Literal(Literal::Bool(false)), negated));
24923 }
24924 Ok(Expr::InList {
24925 expr: Box::new(expr),
24926 list: elements,
24927 negated,
24928 })
24929 }
24930
24931 /// Parse a pgvector array literal `[ x1, x2, ... ]`. The opening `[` is
24932 /// already consumed by the caller. Elements must be numeric literals
24933 /// (with optional unary `-`); any compound expression is rejected at
24934 /// parse time so the runtime never needs to evaluate inside a vector.
24935 /// `EXTRACT(<field> FROM <source>)`. The dispatching `parse_atom`
24936 /// has already consumed the `EXTRACT` token before calling us —
24937 /// we pick up at the opening `(`.
24938 /// v7.17.0 Phase 2.2 — MySQL `MATCH(col [, col ...]) AGAINST
24939 /// (expr [IN BOOLEAN MODE | IN NATURAL LANGUAGE MODE
24940 /// [WITH QUERY EXPANSION]])`. Rewritten in-place to a
24941 /// per-column OR-fold of
24942 /// `to_tsvector('simple', col) @@ plainto_tsquery('simple',
24943 /// term)` so the existing FTS evaluator handles semantics.
24944 ///
24945 /// The mode modifier is accepted-and-ignored at v7.17 — all
24946 /// modes map to the same `plainto_tsquery` rewrite. Boolean-
24947 /// mode operators (`+foo -bar`) would need their own parser
24948 /// (Phase 2.2c); customers who hit them today already get a
24949 /// correct lexeme-match against the bare term, only without
24950 /// the +/- precedence the customer asked for.
24951 fn parse_match_against_atom(&mut self) -> Result<Expr, ParseError> {
24952 // Already at `MATCH`-consumed position; the dispatcher
24953 // confirmed the next token is `(`.
24954 if !matches!(self.peek(), Token::LParen) {
24955 return Err(self.err(alloc::format!(
24956 "expected '(' after MATCH, got {:?}",
24957 self.peek()
24958 )));
24959 }
24960 self.advance();
24961 let mut cols: Vec<Expr> = Vec::new();
24962 loop {
24963 cols.push(self.parse_expr(0)?);
24964 match self.peek() {
24965 Token::Comma => {
24966 self.advance();
24967 }
24968 Token::RParen => break,
24969 other => {
24970 return Err(self.err(alloc::format!(
24971 "expected ',' or ')' in MATCH column list, got {other:?}"
24972 )));
24973 }
24974 }
24975 }
24976 self.advance(); // ')'
24977 // Expect AGAINST.
24978 match self.peek() {
24979 Token::Ident(s) | Token::QuotedIdent(s) if s.eq_ignore_ascii_case("against") => {
24980 self.advance();
24981 }
24982 other => {
24983 return Err(self.err(alloc::format!(
24984 "expected AGAINST after MATCH column list, got {other:?}"
24985 )));
24986 }
24987 }
24988 if !matches!(self.peek(), Token::LParen) {
24989 return Err(self.err(alloc::format!(
24990 "expected '(' after AGAINST, got {:?}",
24991 self.peek()
24992 )));
24993 }
24994 self.advance();
24995 // Read AGAINST's argument as a single primary token —
24996 // string literal, placeholder, or column-ref ident. We
24997 // can't call `parse_expr` / `parse_unary` here because
24998 // the postfix chain inside `parse_atom` would greedily
24999 // fold a trailing `IN BOOLEAN MODE` as `expr IN (...)`
25000 // and fail at "expected '(' after IN". Customers always
25001 // write a literal or bound parameter in AGAINST, so this
25002 // restriction is non-blocking; the error path explains
25003 // the limit if a more complex expression shows up.
25004 let term = match self.advance() {
25005 Token::String(s) => Expr::Literal(crate::ast::Literal::String(s)),
25006 Token::Placeholder(n) => Expr::Placeholder(n),
25007 Token::Ident(s) | Token::QuotedIdent(s) => Expr::Column(crate::ast::ColumnName {
25008 qualifier: None,
25009 name: s,
25010 }),
25011 other => {
25012 return Err(self.err(alloc::format!(
25013 "MATCH ... AGAINST(<term>) expects a string literal, \
25014 bound parameter, or column ref, got {other:?}"
25015 )));
25016 }
25017 };
25018 // Optional mode tail — accept-and-ignore at v7.17:
25019 // IN NATURAL LANGUAGE MODE [WITH QUERY EXPANSION]
25020 // IN BOOLEAN MODE
25021 // WITH QUERY EXPANSION
25022 loop {
25023 match self.peek() {
25024 // IN lexes as a reserved Token::In, not an ident,
25025 // so it gets its own arm.
25026 Token::In => {
25027 self.advance();
25028 }
25029 Token::Ident(s) | Token::QuotedIdent(s)
25030 if s.eq_ignore_ascii_case("natural")
25031 || s.eq_ignore_ascii_case("language")
25032 || s.eq_ignore_ascii_case("boolean")
25033 || s.eq_ignore_ascii_case("mode")
25034 || s.eq_ignore_ascii_case("with")
25035 || s.eq_ignore_ascii_case("query")
25036 || s.eq_ignore_ascii_case("expansion") =>
25037 {
25038 self.advance();
25039 }
25040 _ => break,
25041 }
25042 }
25043 if !matches!(self.peek(), Token::RParen) {
25044 return Err(self.err(alloc::format!(
25045 "expected ')' to close AGAINST, got {:?}",
25046 self.peek()
25047 )));
25048 }
25049 self.advance();
25050 // Build per-column `to_tsvector('simple', col) @@
25051 // plainto_tsquery('simple', term)` and OR-fold.
25052 let simple_lit = || Expr::Literal(crate::ast::Literal::String(String::from("simple")));
25053 let plainto = Expr::FunctionCall {
25054 name: String::from("plainto_tsquery"),
25055 args: alloc::vec![simple_lit(), term.clone()],
25056 };
25057 let mut folded: Option<Expr> = None;
25058 for col in cols {
25059 let to_tsv = Expr::FunctionCall {
25060 name: String::from("to_tsvector"),
25061 args: alloc::vec![simple_lit(), col],
25062 };
25063 let leaf = Expr::Binary {
25064 lhs: Box::new(to_tsv),
25065 op: crate::ast::BinOp::TsMatch,
25066 rhs: Box::new(plainto.clone()),
25067 };
25068 folded = Some(match folded {
25069 None => leaf,
25070 Some(prev) => Expr::Binary {
25071 lhs: Box::new(prev),
25072 op: crate::ast::BinOp::Or,
25073 rhs: Box::new(leaf),
25074 },
25075 });
25076 }
25077 match folded {
25078 Some(e) => Ok(e),
25079 None => Err(self.err(String::from(
25080 "MATCH(...) AGAINST(...) requires at least one column",
25081 ))),
25082 }
25083 }
25084
25085 fn parse_extract_atom(&mut self) -> Result<Expr, ParseError> {
25086 if !matches!(self.peek(), Token::LParen) {
25087 return Err(self.err(format!("expected '(' after EXTRACT, got {:?}", self.peek())));
25088 }
25089 self.advance();
25090 let field_name = self.expect_ident_like()?;
25091 let field = match field_name.to_ascii_lowercase().as_str() {
25092 // PG accepts the plural spellings (years/months/…/millenniums) as
25093 // aliases for the singular fields — its datetime unit table has both.
25094 // (quarter has no plural; dow/doy/isoyear/epoch/julian likewise.)
25095 "year" | "years" => ExtractField::Year,
25096 "month" | "months" => ExtractField::Month,
25097 "day" | "days" => ExtractField::Day,
25098 "hour" | "hours" => ExtractField::Hour,
25099 "minute" | "minutes" => ExtractField::Minute,
25100 "second" | "seconds" => ExtractField::Second,
25101 "microsecond" | "microseconds" => ExtractField::Microsecond,
25102 "epoch" => ExtractField::Epoch,
25103 "dow" => ExtractField::Dow,
25104 "isodow" => ExtractField::Isodow,
25105 "doy" => ExtractField::Doy,
25106 "week" | "weeks" => ExtractField::Week,
25107 "isoyear" => ExtractField::Isoyear,
25108 "quarter" => ExtractField::Quarter,
25109 "decade" | "decades" => ExtractField::Decade,
25110 "century" | "centuries" => ExtractField::Century,
25111 "millennium" | "millenniums" | "millennia" => ExtractField::Millennium,
25112 "julian" => ExtractField::Julian,
25113 "millisecond" | "milliseconds" => ExtractField::Millisecond,
25114 "timezone" => ExtractField::Timezone,
25115 "timezone_hour" => ExtractField::TimezoneHour,
25116 "timezone_minute" => ExtractField::TimezoneMinute,
25117 // v7.39 (round 253) — PG resolves EXTRACT fields at runtime and
25118 // reports an unknown one with the source type (22023); carry the
25119 // raw name so eval can word it.
25120 other => ExtractField::Other(alloc::string::String::from(other)),
25121 };
25122 if !matches!(self.peek(), Token::From) {
25123 return Err(self.err(format!(
25124 "expected FROM after EXTRACT field, got {:?}",
25125 self.peek()
25126 )));
25127 }
25128 self.advance();
25129 let source = self.parse_expr(0)?;
25130 if !matches!(self.peek(), Token::RParen) {
25131 return Err(self.err(format!(
25132 "expected ')' to close EXTRACT, got {:?}",
25133 self.peek()
25134 )));
25135 }
25136 self.advance();
25137 Ok(Expr::Extract {
25138 field,
25139 source: Box::new(source),
25140 })
25141 }
25142
25143 /// `INTERVAL '<n> <unit> [<n> <unit> ...]'` — the `INTERVAL` keyword
25144 /// is already consumed; we expect a single string literal next and
25145 /// resolve it into `Literal::Interval` at parse time so the engine
25146 /// never has to re-tokenise inside the string.
25147 /// The unquoted count of a MySQL `INTERVAL <n> <UNIT>`, when the
25148 /// tokens ahead really are one. A quoted count (`INTERVAL '2' DAY`)
25149 /// is the SQL-standard form and is left to the path below.
25150 fn peek_unquoted_interval_count(&self) -> Option<(alloc::string::String, usize)> {
25151 // A negative count lexes as `-` then the number (`INTERVAL -1 DAY`).
25152 let (offset, sign) = match self.peek() {
25153 Token::Minus => (1, "-"),
25154 _ => (0, ""),
25155 };
25156 let Some(Token::Integer(n)) = self.tokens.get(self.pos + offset) else {
25157 return None;
25158 };
25159 self.tokens
25160 .get(self.pos + offset + 1)
25161 .filter(|t| mysql_interval_unit(t).is_some())?;
25162 Some((alloc::format!("{sign}{n}"), offset + 1))
25163 }
25164
25165 /// v7.39 (round 422) — is the parenthesised group starting at the CURRENT
25166 /// `(` a single quantity followed by a time unit (`INTERVAL (1+1) DAY`),
25167 /// rather than the argument list of MySQL's `INTERVAL(N, N1, …)` function?
25168 ///
25169 /// Scans `self.tokens` by index and consumes NOTHING. Round 409 decided
25170 /// this by parsing the group and then restoring `self.pos` — which could
25171 /// never have worked, because `advance()` DESTROYS the token it returns
25172 /// (`mem::replace(.., Eof)`); the restore yielded a stream of Eof. It was
25173 /// inert only because both branches errored back then.
25174 fn interval_paren_is_quantity(&self) -> bool {
25175 let mut depth = 0usize;
25176 let mut saw_top_level_comma = false;
25177 let mut i = self.pos;
25178 while let Some(tok) = self.tokens.get(i) {
25179 match tok {
25180 Token::LParen => depth += 1,
25181 Token::RParen => {
25182 depth = depth.saturating_sub(1);
25183 if depth == 0 {
25184 return !saw_top_level_comma
25185 && mysql_interval_unit(self.tokens.get(i + 1).unwrap_or(&Token::Eof))
25186 .is_some();
25187 }
25188 }
25189 // A comma directly inside the outermost parens means the
25190 // argument list of the INTERVAL() function.
25191 Token::Comma if depth == 1 => saw_top_level_comma = true,
25192 Token::Eof => return false,
25193 _ => {}
25194 }
25195 i += 1;
25196 }
25197 false
25198 }
25199
25200 fn parse_interval_atom(&mut self) -> Result<Expr, ParseError> {
25201 // v7.39 (round 409) — MySQL's `INTERVAL(N, N1, N2, …)` function
25202 // (the index of the last Ni ≤ N), distinct from the interval literal.
25203 // `INTERVAL (` is ambiguous with `INTERVAL (expr) UNIT`, so the shape
25204 // is decided by a non-destructive lookahead (round 422) before either
25205 // branch consumes anything. MySQL only.
25206 if self.mysql_dialect
25207 && matches!(self.peek(), Token::LParen)
25208 && !self.interval_paren_is_quantity()
25209 {
25210 self.advance(); // (
25211 let mut args = Vec::new();
25212 if !matches!(self.peek(), Token::RParen) {
25213 loop {
25214 args.push(self.parse_expr(0)?);
25215 if matches!(self.peek(), Token::Comma) {
25216 self.advance();
25217 continue;
25218 }
25219 break;
25220 }
25221 }
25222 if !matches!(self.peek(), Token::RParen) {
25223 return Err(self.err(alloc::format!(
25224 "expected ')' after INTERVAL() arguments, got {:?}",
25225 self.peek()
25226 )));
25227 }
25228 self.advance(); // )
25229 return Ok(Expr::FunctionCall {
25230 name: alloc::string::String::from("interval"),
25231 args,
25232 });
25233 }
25234 // v7.39 (round 350, M7) — MySQL's `INTERVAL <n> <UNIT>`, with the
25235 // number UNQUOTED: `DATE_ADD(d, INTERVAL 1 MONTH)`,
25236 // `d + INTERVAL 90 MINUTE`, `INTERVAL -1 DAY`. It is how MySQL
25237 // writes every date arithmetic there is, and it did not parse at
25238 // all. PG rejects the unquoted form outright (`syntax error at or
25239 // near "1"`, measured), so it is taken only in the MySQL dialect —
25240 // PG's own `INTERVAL '1' DAY` is untouched below.
25241 if self.mysql_dialect
25242 && let Some((text, consume)) = self.peek_unquoted_interval_count()
25243 {
25244 for _ in 0..consume {
25245 self.advance(); // the optional `-` and the number
25246 }
25247 let Some(unit) = mysql_interval_unit(self.peek()) else {
25248 return Err(self.err(alloc::format!(
25249 "expected an interval unit after INTERVAL {text}, got {:?}",
25250 self.peek()
25251 )));
25252 };
25253 self.advance(); // the unit
25254 let (months, days, micros) = scale_mysql_interval(&text, unit)
25255 .ok_or_else(|| self.err(alloc::format!("cannot read INTERVAL {text} {unit}")))?;
25256 return Ok(Expr::Literal(Literal::Interval {
25257 months,
25258 days,
25259 micros,
25260 // The canonical rendering, so Display round-trips into a
25261 // form both dialects read back.
25262 text: alloc::format!("{text} {unit}"),
25263 }));
25264 }
25265 // v7.39 (round 422) — MySQL's interval QUANTITY may be any expression,
25266 // not just a literal: `DATE_ADD(d, INTERVAL n DAY)`,
25267 // `d + INTERVAL n*2 DAY`, `INTERVAL (1+1) DAY`, `INTERVAL ABS(-5) DAY`.
25268 // Those cannot fold into a compile-time `Literal::Interval`, so they
25269 // lower onto the existing `make_interval(y, mo, w, d, h, mi, s)`
25270 // builtin, which builds the value at run time (and yields NULL for a
25271 // NULL quantity, as MariaDB does). The literal path above still folds
25272 // the constant case — it is cheaper and round-trips through Display.
25273 //
25274 // Guarded off a String operand so PG's own `INTERVAL '1 day'` (and
25275 // MySQL's quoted spelling) keep the qualifier path below.
25276 if self.mysql_dialect && !matches!(self.peek(), Token::String(_)) {
25277 let qty = self.parse_expr(0)?;
25278 let Some(unit) = mysql_interval_unit(self.peek()) else {
25279 return Err(self.err(alloc::format!(
25280 "expected an interval unit after INTERVAL <expr>, got {:?}",
25281 self.peek()
25282 )));
25283 };
25284 self.advance(); // the unit
25285 return Ok(make_interval_call(qty, unit));
25286 }
25287 let tok = self.advance();
25288 let Token::String(text) = tok else {
25289 return Err(self.err(format!(
25290 "expected string literal after INTERVAL, got {tok:?}"
25291 )));
25292 };
25293 // v7.39 (read01 round 102) — SQL-standard trailing field qualifier
25294 // `<FIELD> [TO <FIELD>]` (`INTERVAL '2' YEAR`, `INTERVAL '1-6' YEAR TO
25295 // MONTH`, `INTERVAL '1 2:03:04' DAY TO SECOND`). It sets which field a
25296 // bare number means and the leading/trailing precision.
25297 let field1 = interval_field_of(self.peek());
25298 let qualifier = if let Some(f1) = field1 {
25299 self.advance();
25300 let f2 = if matches!(self.peek(), Token::To) {
25301 self.advance();
25302 let Some(f) = interval_field_of(self.peek()) else {
25303 return Err(self.err(format!(
25304 "expected an interval field after TO, got {:?}",
25305 self.peek()
25306 )));
25307 };
25308 self.advance();
25309 Some(f)
25310 } else {
25311 None
25312 };
25313 Some((f1, f2))
25314 } else {
25315 None
25316 };
25317 let (months, days, micros) = match qualifier {
25318 Some(q) => interpret_qualified_interval(&text, q),
25319 None => parse_interval_text(&text),
25320 }
25321 .ok_or_else(|| ParseError {
25322 message: format!(
25323 "cannot parse INTERVAL {text:?}; \
25324 expected `<n> <unit> [<n> <unit> ...]` with units \
25325 microsecond[s], millisecond[s], second[s], minute[s], \
25326 hour[s], day[s], week[s], month[s], year[s]"
25327 ),
25328 token_pos: self.consumed_pos(),
25329 })?;
25330 Ok(Expr::Literal(Literal::Interval {
25331 months,
25332 days,
25333 micros,
25334 text,
25335 }))
25336 }
25337
25338 /// v7.38 (read01, T10) — parse a bracketed sub-array `[e, e, …]` inside an
25339 /// `ARRAY[...]` constructor, recursing on further nested `[...]` so
25340 /// `ARRAY[[1,2],[3,4]]` (and deeper) becomes nested `Expr::Array` rather
25341 /// than a pgvector literal.
25342 fn parse_array_bracket_body(&mut self) -> Result<Expr, ParseError> {
25343 self.advance(); // consume `[`
25344 let mut items: Vec<Expr> = Vec::new();
25345 if !matches!(self.peek(), Token::RBracket) {
25346 loop {
25347 if matches!(self.peek(), Token::LBracket) {
25348 items.push(self.parse_array_bracket_body()?);
25349 } else {
25350 items.push(self.parse_expr(0)?);
25351 }
25352 match self.peek() {
25353 Token::Comma => {
25354 self.advance();
25355 }
25356 Token::RBracket => break,
25357 other => {
25358 return Err(self.err(alloc::format!(
25359 "expected ',' or ']' in array literal, got {other:?}"
25360 )));
25361 }
25362 }
25363 }
25364 }
25365 self.advance(); // consume `]`
25366 Ok(Expr::Array(items))
25367 }
25368
25369 fn parse_vector_literal_body(&mut self) -> Result<Expr, ParseError> {
25370 let mut elems = Vec::new();
25371 if matches!(self.peek(), Token::RBracket) {
25372 self.advance();
25373 return Ok(Expr::Literal(Literal::Vector(elems)));
25374 }
25375 loop {
25376 let e = self.parse_expr(0)?;
25377 let x = extract_numeric_literal(&e).ok_or_else(|| ParseError {
25378 message: format!("vector element must be a numeric literal, got {e:?}"),
25379 token_pos: self.pos,
25380 })?;
25381 elems.push(x);
25382 match self.peek() {
25383 Token::Comma => {
25384 self.advance();
25385 }
25386 Token::RBracket => {
25387 self.advance();
25388 break;
25389 }
25390 other => {
25391 return Err(self.err(format!("expected ',' or ']' in vector, got {other:?}")));
25392 }
25393 }
25394 }
25395 Ok(Expr::Literal(Literal::Vector(elems)))
25396 }
25397
25398 /// Atom that started with an identifier: could be `t.col`, `col`, or
25399 /// `func(arg, ...)`. Detect each shape by looking at the next token.
25400 /// v4.12: parse `(PARTITION BY expr, ... ORDER BY expr [DESC]
25401 /// [, ...])`. Caller has already consumed `OVER`. Either clause
25402 /// is optional; an empty `()` is also legal (PG semantics).
25403 /// v6.4.2 — consume an optional `IGNORE NULLS` / `RESPECT NULLS`
25404 /// modifier between `name(args)` and `OVER (...)`. Default is
25405 /// `Respect`. Unrecognised idents leave the stream unchanged.
25406 fn parse_null_treatment_modifier(&mut self) -> NullTreatment {
25407 let Token::Ident(s) = self.peek().clone() else {
25408 return NullTreatment::Respect;
25409 };
25410 let is_ignore = s.eq_ignore_ascii_case("ignore");
25411 let is_respect = s.eq_ignore_ascii_case("respect");
25412 if !is_ignore && !is_respect {
25413 return NullTreatment::Respect;
25414 }
25415 // Lookahead for NULLS — only consume both tokens together.
25416 // pos+1 must hold a "nulls" ident.
25417 if self.pos + 1 < self.tokens.len()
25418 && let Token::Ident(s2) = &self.tokens[self.pos + 1]
25419 && s2.eq_ignore_ascii_case("nulls")
25420 {
25421 self.advance();
25422 self.advance();
25423 return if is_ignore {
25424 NullTreatment::Ignore
25425 } else {
25426 NullTreatment::Respect
25427 };
25428 }
25429 NullTreatment::Respect
25430 }
25431
25432 /// v7.32 (mailrs round-29) — `agg(args) FILTER (WHERE cond)`.
25433 /// `FILTER` is an unreserved keyword, so it arrives as an `Ident`
25434 /// (same shape as the `OVER` tail). Consumes the whole clause and
25435 /// returns the predicate; returns `None` when no `FILTER` follows.
25436 fn parse_filter_clause(&mut self) -> Result<Option<Box<Expr>>, ParseError> {
25437 let (Token::Ident(s) | Token::QuotedIdent(s)) = self.peek() else {
25438 return Ok(None);
25439 };
25440 if !s.eq_ignore_ascii_case("filter") {
25441 return Ok(None);
25442 }
25443 self.advance(); // FILTER
25444 if !matches!(self.peek(), Token::LParen) {
25445 return Err(self.err(format!("expected '(' after FILTER, got {:?}", self.peek())));
25446 }
25447 self.advance(); // (
25448 if !matches!(self.peek(), Token::Where) {
25449 return Err(self.err(format!(
25450 "expected WHERE inside FILTER (...), got {:?}",
25451 self.peek()
25452 )));
25453 }
25454 self.advance(); // WHERE
25455 let cond = self.parse_expr(0)?;
25456 if !matches!(self.peek(), Token::RParen) {
25457 return Err(self.err(format!(
25458 "expected ')' to close FILTER (WHERE ...), got {:?}",
25459 self.peek()
25460 )));
25461 }
25462 self.advance(); // )
25463 Ok(Some(Box::new(cond)))
25464 }
25465
25466 /// v7.39 (round 354, M12) — consume a `SEPARATOR '<s>'` tail and push
25467 /// the separator as the aggregate's second argument, which is the
25468 /// shape `string_agg` already takes. Returns whether one was there.
25469 fn consume_group_concat_separator(&mut self, args: &mut Vec<Expr>) -> Result<bool, ParseError> {
25470 if !matches!(self.peek(), Token::Ident(k) if k.eq_ignore_ascii_case("separator")) {
25471 return Ok(false);
25472 }
25473 self.advance();
25474 let Token::String(sep) = self.peek().clone() else {
25475 return Err(self.err(alloc::format!(
25476 "expected a string literal after SEPARATOR, got {:?}",
25477 self.peek()
25478 )));
25479 };
25480 self.advance();
25481 args.push(Expr::Literal(Literal::String(sep)));
25482 Ok(true)
25483 }
25484
25485 /// v7.32 (round-29) — `WITHIN GROUP ( ORDER BY <sort_spec> )` tail
25486 /// for ordered-set aggregates. `WITHIN` is unreserved (arrives as an
25487 /// `Ident`); `GROUP` and `ORDER`/`BY` are keywords. Returns the sort
25488 /// keys, or an empty vec when no `WITHIN GROUP` follows.
25489 fn parse_within_group_clause(&mut self) -> Result<Vec<OrderBy>, ParseError> {
25490 let (Token::Ident(s) | Token::QuotedIdent(s)) = self.peek() else {
25491 return Ok(Vec::new());
25492 };
25493 if !s.eq_ignore_ascii_case("within") {
25494 return Ok(Vec::new());
25495 }
25496 self.advance(); // WITHIN
25497 if !matches!(self.peek(), Token::Group) {
25498 return Err(self.err(format!(
25499 "expected GROUP after WITHIN, got {:?}",
25500 self.peek()
25501 )));
25502 }
25503 self.advance(); // GROUP
25504 if !matches!(self.peek(), Token::LParen) {
25505 return Err(self.err(format!(
25506 "expected '(' after WITHIN GROUP, got {:?}",
25507 self.peek()
25508 )));
25509 }
25510 self.advance(); // (
25511 if !matches!(self.peek(), Token::Order) {
25512 return Err(self.err(format!(
25513 "expected ORDER BY inside WITHIN GROUP (...), got {:?}",
25514 self.peek()
25515 )));
25516 }
25517 self.advance(); // ORDER
25518 if !self.peek_is_by() {
25519 return Err(self.err(format!("expected BY after ORDER, got {:?}", self.peek())));
25520 }
25521 self.advance(); // BY
25522 let mut keys: Vec<OrderBy> = Vec::new();
25523 loop {
25524 // v7.39 (round 691) — save/restore, the discipline this parser
25525 // already uses around `pending_sample_preds`, so a subquery inside
25526 // a key neither inherits nor leaks the channel.
25527 let saved_flag = core::mem::replace(&mut self.in_order_by_key, true);
25528 let saved_coll = self.order_key_collation.take();
25529 let parsed = self.parse_expr(0);
25530 self.in_order_by_key = saved_flag;
25531 let collation = core::mem::replace(&mut self.order_key_collation, saved_coll);
25532 let expr = parsed?;
25533 let desc = if matches!(self.peek(), Token::Desc) {
25534 self.advance();
25535 true
25536 } else if matches!(self.peek(), Token::Asc) {
25537 self.advance();
25538 false
25539 } else {
25540 false
25541 };
25542 let nulls_first = self.parse_optional_nulls_placement()?;
25543 keys.push(OrderBy {
25544 expr,
25545 desc,
25546 nulls_first,
25547 collation,
25548 });
25549 if matches!(self.peek(), Token::Comma) {
25550 self.advance();
25551 } else {
25552 break;
25553 }
25554 }
25555 if !matches!(self.peek(), Token::RParen) {
25556 return Err(self.err(format!(
25557 "expected ')' to close WITHIN GROUP (ORDER BY ...), got {:?}",
25558 self.peek()
25559 )));
25560 }
25561 self.advance(); // )
25562 Ok(keys)
25563 }
25564
25565 /// No frame clause is supported.
25566 #[allow(clippy::type_complexity)] // (partitions, ordered-keys-with-desc) is the natural shape
25567 fn parse_over_clause(
25568 &mut self,
25569 ) -> Result<
25570 (
25571 Vec<Expr>,
25572 Vec<(Expr, bool, Option<bool>)>,
25573 Option<WindowFrame>,
25574 ),
25575 ParseError,
25576 > {
25577 // `OVER w` — a named-window reference. The WINDOW clause
25578 // parses after the select list, so the name rides out as a
25579 // marker in partition_by; parse_bare_select substitutes the
25580 // definition once the clause is known.
25581 if let Token::Ident(w) | Token::QuotedIdent(w) = self.peek() {
25582 let name = w.clone();
25583 self.advance();
25584 return Ok((
25585 alloc::vec![Expr::Column(crate::ast::ColumnName {
25586 qualifier: Some("__named_window__".to_string()),
25587 name,
25588 })],
25589 Vec::new(),
25590 None,
25591 ));
25592 }
25593 if !matches!(self.peek(), Token::LParen) {
25594 return Err(self.err(format!("expected '(' after OVER, got {:?}", self.peek())));
25595 }
25596 self.advance();
25597 let mut partition_by = Vec::new();
25598 let mut order_by = Vec::new();
25599 // v7.39 (round 229) — `OVER (w1 …)`: a *copy* of an existing named
25600 // window, refined in place. PG's rules (probed against 18.4) differ
25601 // from the bare `OVER w1` form, so the reference rides out under its
25602 // own marker and `substitute_named_windows` applies them. The base
25603 // name is any leading identifier that isn't a window-spec keyword.
25604 let base_window = match self.peek() {
25605 Token::Ident(s) | Token::QuotedIdent(s)
25606 if !s.eq_ignore_ascii_case("partition")
25607 && !s.eq_ignore_ascii_case("rows")
25608 && !s.eq_ignore_ascii_case("range")
25609 && !s.eq_ignore_ascii_case("groups") =>
25610 {
25611 let n = s.clone();
25612 self.advance();
25613 Some(n)
25614 }
25615 _ => None,
25616 };
25617 // PARTITION BY ?
25618 // v7.37.6-B promoted PARTITION to a reserved keyword
25619 // (Token::Partition); pre-7.37.6-B catalogs lexed it as
25620 // `Token::Ident("partition")`. Accept both so older sources
25621 // and the new lexer surface land on the same path.
25622 let is_partition_kw = match self.peek() {
25623 Token::Partition => true,
25624 Token::Ident(s) | Token::QuotedIdent(s) => s.eq_ignore_ascii_case("partition"),
25625 _ => false,
25626 };
25627 if is_partition_kw {
25628 self.advance();
25629 if !self.peek_is_by() {
25630 return Err(self.err(format!(
25631 "expected BY after PARTITION, got {:?}",
25632 self.peek()
25633 )));
25634 }
25635 self.advance();
25636 loop {
25637 partition_by.push(self.parse_expr(0)?);
25638 if matches!(self.peek(), Token::Comma) {
25639 self.advance();
25640 continue;
25641 }
25642 break;
25643 }
25644 }
25645 // ORDER BY ?
25646 if matches!(self.peek(), Token::Order) {
25647 self.advance();
25648 if !self.peek_is_by() {
25649 return Err(self.err(format!("expected BY after ORDER, got {:?}", self.peek())));
25650 }
25651 self.advance();
25652 loop {
25653 let e = self.parse_expr(0)?;
25654 let desc = if matches!(self.peek(), Token::Desc) {
25655 self.advance();
25656 true
25657 } else if matches!(self.peek(), Token::Asc) {
25658 self.advance();
25659 false
25660 } else {
25661 false
25662 };
25663 // v7.24.1 — NULLS FIRST/LAST inside OVER (…).
25664 let nulls_first = self.parse_optional_nulls_placement()?;
25665 order_by.push((e, desc, nulls_first));
25666 if matches!(self.peek(), Token::Comma) {
25667 self.advance();
25668 continue;
25669 }
25670 break;
25671 }
25672 }
25673 // v4.20: optional explicit frame, `ROWS ...` / `RANGE ...`.
25674 // Both keywords come through the lexer as identifiers; match
25675 // case-insensitively.
25676 let mut frame: Option<WindowFrame> = None;
25677 if let Token::Ident(s) | Token::QuotedIdent(s) = self.peek() {
25678 let kind = if s.eq_ignore_ascii_case("rows") {
25679 Some(FrameKind::Rows)
25680 } else if s.eq_ignore_ascii_case("range") {
25681 Some(FrameKind::Range)
25682 } else if s.eq_ignore_ascii_case("groups") {
25683 // v7.37.19 (19.11) — PG 11+ GROUPS frame mode.
25684 Some(FrameKind::Groups)
25685 } else {
25686 None
25687 };
25688 if let Some(kind) = kind {
25689 self.advance();
25690 frame = Some(self.parse_frame_tail(kind)?);
25691 }
25692 }
25693 if !matches!(self.peek(), Token::RParen) {
25694 return Err(self.err(format!(
25695 "expected ')' to close OVER clause, got {:?}",
25696 self.peek()
25697 )));
25698 }
25699 self.advance();
25700 if let Some(base) = base_window {
25701 // A copy may refine but never override the base's partitioning
25702 // (PG rejects it outright, before looking the name up).
25703 if !partition_by.is_empty() {
25704 return Err(self.err(alloc::format!(
25705 "cannot override PARTITION BY clause of window \"{base}\""
25706 )));
25707 }
25708 partition_by = alloc::vec![Expr::Column(crate::ast::ColumnName {
25709 qualifier: Some("__named_window_ref__".to_string()),
25710 name: base,
25711 })];
25712 }
25713 Ok((partition_by, order_by, frame))
25714 }
25715
25716 /// v4.20: parse the tail of an explicit frame, given the `ROWS`
25717 /// or `RANGE` keyword was just consumed. Accepts both
25718 /// `BETWEEN <bound> AND <bound>` and the single-bound shorthand
25719 /// (`ROWS UNBOUNDED PRECEDING`, `ROWS 5 PRECEDING`, etc.) which
25720 /// PG normalises to `BETWEEN <bound> AND CURRENT ROW`.
25721 fn parse_frame_tail(&mut self, kind: FrameKind) -> Result<WindowFrame, ParseError> {
25722 let (start, end) = if matches!(self.peek(), Token::Between) {
25723 self.advance();
25724 let start = self.parse_frame_bound()?;
25725 if !matches!(self.peek(), Token::And) {
25726 return Err(self.err(format!("expected AND in frame spec, got {:?}", self.peek())));
25727 }
25728 self.advance();
25729 let end = self.parse_frame_bound()?;
25730 (start, Some(end))
25731 } else {
25732 (self.parse_frame_bound()?, None)
25733 };
25734 let exclude = self.parse_frame_exclusion()?;
25735 Ok(WindowFrame {
25736 kind,
25737 start,
25738 end,
25739 exclude,
25740 })
25741 }
25742
25743 /// Optional `EXCLUDE {CURRENT ROW | GROUP | TIES | NO OTHERS}`
25744 /// after a frame spec. NO OTHERS is the default no-op.
25745 fn parse_frame_exclusion(&mut self) -> Result<FrameExclusion, ParseError> {
25746 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("exclude")) {
25747 return Ok(FrameExclusion::NoOthers);
25748 }
25749 self.advance(); // EXCLUDE
25750 match self.peek() {
25751 Token::Ident(s) if s.eq_ignore_ascii_case("current") => {
25752 self.advance();
25753 if !matches!(self.peek(), Token::Ident(r) if r.eq_ignore_ascii_case("row")) {
25754 return Err(self.err(format!(
25755 "expected ROW after EXCLUDE CURRENT, got {:?}",
25756 self.peek()
25757 )));
25758 }
25759 self.advance();
25760 Ok(FrameExclusion::CurrentRow)
25761 }
25762 // v7.39 (read01 round 109) — GROUP is a reserved keyword token, so
25763 // `EXCLUDE GROUP` arrives as `Token::Group`, not `Ident("group")`.
25764 // Without this arm it fell to the catch-all, whose message
25765 // self-contradictingly listed GROUP as expected.
25766 Token::Ident(s) if s.eq_ignore_ascii_case("group") => {
25767 self.advance();
25768 Ok(FrameExclusion::Group)
25769 }
25770 Token::Group => {
25771 self.advance();
25772 Ok(FrameExclusion::Group)
25773 }
25774 Token::Ident(s) if s.eq_ignore_ascii_case("ties") => {
25775 self.advance();
25776 Ok(FrameExclusion::Ties)
25777 }
25778 Token::Ident(s) if s.eq_ignore_ascii_case("no") => {
25779 self.advance();
25780 if !matches!(self.peek(), Token::Ident(r) if r.eq_ignore_ascii_case("others")) {
25781 return Err(self.err(format!(
25782 "expected OTHERS after EXCLUDE NO, got {:?}",
25783 self.peek()
25784 )));
25785 }
25786 self.advance();
25787 Ok(FrameExclusion::NoOthers)
25788 }
25789 other => Err(self.err(format!(
25790 "expected CURRENT ROW / GROUP / TIES / NO OTHERS after EXCLUDE, got {other:?}"
25791 ))),
25792 }
25793 }
25794
25795 /// Parse one frame bound: `UNBOUNDED PRECEDING`, `<n> PRECEDING`,
25796 /// `<interval> PRECEDING`, `CURRENT ROW`, `<n>/<interval> FOLLOWING`,
25797 /// `UNBOUNDED FOLLOWING`.
25798 fn parse_frame_bound(&mut self) -> Result<FrameBound, ParseError> {
25799 // Interval-typed offset for a value-based RANGE frame over a
25800 // DATE / TIMESTAMP ORDER BY column (PG time-series windows),
25801 // spelled `INTERVAL '1 day' PRECEDING` or `'1 day'::interval
25802 // PRECEDING`.
25803 if let Some((months, days, micros)) = self.try_take_interval_offset()? {
25804 let dir = self.expect_ident_like()?;
25805 return if dir.eq_ignore_ascii_case("preceding") {
25806 Ok(FrameBound::IntervalPreceding {
25807 months,
25808 days,
25809 micros,
25810 })
25811 } else if dir.eq_ignore_ascii_case("following") {
25812 Ok(FrameBound::IntervalFollowing {
25813 months,
25814 days,
25815 micros,
25816 })
25817 } else {
25818 Err(self.err(format!(
25819 "expected PRECEDING or FOLLOWING after interval offset, got {dir:?}"
25820 )))
25821 };
25822 }
25823 // Number-led: "<n> PRECEDING" / "<n> FOLLOWING".
25824 if let Token::Integer(n) = *self.peek() {
25825 self.advance();
25826 let n: u64 = u64::try_from(n).map_err(|_| {
25827 self.err(format!(
25828 "invalid frame offset {n} — expected non-negative integer"
25829 ))
25830 })?;
25831 let dir = self.expect_ident_like()?;
25832 return if dir.eq_ignore_ascii_case("preceding") {
25833 Ok(FrameBound::OffsetPreceding(n))
25834 } else if dir.eq_ignore_ascii_case("following") {
25835 Ok(FrameBound::OffsetFollowing(n))
25836 } else {
25837 Err(self.err(format!(
25838 "expected PRECEDING or FOLLOWING after offset, got {dir:?}"
25839 )))
25840 };
25841 }
25842 let first = self.expect_ident_like()?;
25843 if first.eq_ignore_ascii_case("unbounded") {
25844 let dir = self.expect_ident_like()?;
25845 return if dir.eq_ignore_ascii_case("preceding") {
25846 Ok(FrameBound::UnboundedPreceding)
25847 } else if dir.eq_ignore_ascii_case("following") {
25848 Ok(FrameBound::UnboundedFollowing)
25849 } else {
25850 Err(self.err(format!(
25851 "expected PRECEDING or FOLLOWING after UNBOUNDED, got {dir:?}"
25852 )))
25853 };
25854 }
25855 if first.eq_ignore_ascii_case("current") {
25856 let row = self.expect_ident_like()?;
25857 if !row.eq_ignore_ascii_case("row") {
25858 return Err(self.err(format!("expected ROW after CURRENT, got {row:?}")));
25859 }
25860 return Ok(FrameBound::CurrentRow);
25861 }
25862 Err(self.err(format!(
25863 "expected frame bound (UNBOUNDED/CURRENT/<n>), got {first:?}"
25864 )))
25865 }
25866
25867 /// Detect and consume a leading interval offset in a frame bound —
25868 /// `INTERVAL '1 day'` or `'1 day'::interval` — returning its folded
25869 /// `(months, days, micros)`. Leaves the cursor on the trailing
25870 /// PRECEDING / FOLLOWING keyword. Returns `None` (without advancing)
25871 /// when the next tokens are not an interval offset.
25872 fn try_take_interval_offset(&mut self) -> Result<Option<(i32, i32, i64)>, ParseError> {
25873 // Shape A — `INTERVAL '1 day'`.
25874 if matches!(self.peek(), Token::Interval) {
25875 self.advance(); // INTERVAL
25876 let atom = self.parse_interval_atom()?;
25877 if let Expr::Literal(Literal::Interval {
25878 months,
25879 days,
25880 micros,
25881 ..
25882 }) = atom
25883 {
25884 return Ok(Some((months, days, micros)));
25885 }
25886 return Err(self.err("expected an interval literal in frame offset".to_string()));
25887 }
25888 // Shape B — `'1 day'::interval`. Look ahead for the exact
25889 // string / `::` / interval-target triple before committing.
25890 if let Token::String(text) = self.peek() {
25891 let target_is_interval = match self.tokens.get(self.pos + 2) {
25892 Some(Token::Interval) => true,
25893 Some(Token::Ident(s)) => s.eq_ignore_ascii_case("interval"),
25894 _ => false,
25895 };
25896 let is_cast = matches!(self.tokens.get(self.pos + 1), Some(Token::DoubleColon))
25897 && target_is_interval;
25898 if is_cast {
25899 let text = text.clone();
25900 self.advance(); // string
25901 self.advance(); // ::
25902 self.advance(); // interval
25903 let parts = parse_interval_text(&text).ok_or_else(|| {
25904 self.err(format!("cannot parse INTERVAL {text:?} in frame offset"))
25905 })?;
25906 return Ok(Some(parts));
25907 }
25908 }
25909 Ok(None)
25910 }
25911
25912 fn finish_ident_atom(&mut self, first: String) -> Result<Expr, ParseError> {
25913 // v7.39.2 — MySQL's charset INTRODUCER: `_utf8mb4'x'`, `N'y'`,
25914 // `_binary'z'`. All three were `ERROR 1064 syntax error` here
25915 // and all three answer the literal on MySQL 9.7.2.
25916 //
25917 // It is not only syntax, which is why it waited for
25918 // `Expr::Collate`: measured, `_binary'A' = 'a'` is 0 on MySQL
25919 // because `_binary` makes the comparison byte-wise, while
25920 // `_utf8mb4'A' = _utf8mb4'a'` is 1. Accepting the syntax and
25921 // dropping the charset would have turned a hard error into a
25922 // silently wrong comparison — worse than the error it replaced.
25923 //
25924 // An UNKNOWN charset is NOT an introducer: MySQL answers
25925 // `Unknown column '_nosuch'`, because it parses as a column
25926 // reference followed by a string. So the table decides, and it
25927 // is the same table `SET NAMES` reads.
25928 //
25929 // A space is allowed between the two (`_utf8mb4 'x'`), which
25930 // falls out of asking the token stream rather than the bytes.
25931 if self.mysql_dialect
25932 && let Token::String(_) = self.peek()
25933 {
25934 let lower = first.to_ascii_lowercase();
25935 let charset = if lower == "n" {
25936 // `N'…'` is the national character set, which MySQL
25937 // documents as utf8 — utf8mb3 in 9.7.2's spelling.
25938 //
25939 // utf8mb3 and utf8mb4 both fold case in their default
25940 // collations, so nothing SPG can be asked distinguishes
25941 // the two here: an ablation swapping this to utf8mb4
25942 // reddens no pin. Recorded rather than implied — the
25943 // spelling follows MySQL's documentation, not a
25944 // measurement.
25945 Some("utf8mb3")
25946 } else {
25947 // No filter here: the lookup below IS the test for
25948 // "is this a charset". An ablation that removed a filter
25949 // in this spot reddened nothing, which is how the two
25950 // were found to be one check written twice.
25951 lower.strip_prefix('_')
25952 };
25953 if let Some(cs) = charset
25954 && let Some(collation) = crate::charset::charset_default_collation(cs)
25955 {
25956 let Token::String(body) = self.advance() else {
25957 unreachable!("peeked a string");
25958 };
25959 return Ok(Expr::Collate {
25960 expr: Box::new(Expr::Literal(Literal::String(body))),
25961 collation: String::from(collation),
25962 });
25963 }
25964 }
25965 if matches!(self.peek(), Token::Dot) {
25966 self.advance();
25967 let name = self.expect_ident_like()?;
25968 // v7.14.0 — schema-qualified function call
25969 // `<schema>.<fn>(args)`. PG dumps emit
25970 // `pg_catalog.set_config(...)` in the preamble. SPG
25971 // is single-namespace: drop the schema prefix and
25972 // route the dispatch on the bare function name.
25973 if matches!(self.peek(), Token::LParen) {
25974 return self.finish_ident_atom(name);
25975 }
25976 return Ok(Expr::Column(ColumnName {
25977 qualifier: Some(first),
25978 name,
25979 }));
25980 }
25981 if matches!(self.peek(), Token::LParen) {
25982 self.advance();
25983 // `COUNT(*)` — special-cased here because `*` isn't a normal
25984 // expression token. Lower-case match on `first` since the lexer
25985 // folds identifiers.
25986 if first.eq_ignore_ascii_case("count") && matches!(self.peek(), Token::Star) {
25987 self.advance();
25988 if !matches!(self.peek(), Token::RParen) {
25989 return Err(self.err(format!(
25990 "expected ')' after COUNT(*), got {:?}",
25991 self.peek()
25992 )));
25993 }
25994 self.advance();
25995 // v7.32 (round-29) — `COUNT(*) FILTER (WHERE …)`.
25996 let filter = self.parse_filter_clause()?;
25997 // v4.12: COUNT(*) OVER (...) — same window tail.
25998 let null_treatment = self.parse_null_treatment_modifier();
25999 if let Token::Ident(s) | Token::QuotedIdent(s) = self.peek()
26000 && s.eq_ignore_ascii_case("over")
26001 {
26002 self.advance();
26003 let (partition_by, order_by, frame) = self.parse_over_clause()?;
26004 return Ok(Expr::WindowFunction {
26005 name: "count_star".into(),
26006 args: Vec::new(),
26007 partition_by,
26008 order_by,
26009 frame,
26010 null_treatment,
26011 filter,
26012 });
26013 }
26014 if let Some(filter) = filter {
26015 return Ok(Expr::AggregateOrdered {
26016 call: Box::new(Expr::FunctionCall {
26017 name: "count_star".into(),
26018 args: Vec::new(),
26019 }),
26020 order_by: Vec::new(),
26021 distinct: false,
26022 filter: Some(filter),
26023 });
26024 }
26025 return Ok(Expr::FunctionCall {
26026 name: "count_star".into(),
26027 args: Vec::new(),
26028 });
26029 }
26030 // Function call. PG-style: zero-or-more comma-separated args.
26031 let mut args = Vec::new();
26032 // v7.38 (read01, T14) — named-argument notation `argname => value`.
26033 // Names are collected in lock-step with `args` and resolved to
26034 // positional order after the loop (the AST stays positional).
26035 let mut arg_names: Vec<Option<String>> = Vec::new();
26036 let mut agg_order_by: Vec<OrderBy> = Vec::new();
26037 // v7.39 (round 354, M12) — whether a `SEPARATOR '<s>'` tail was
26038 // seen, so the value arguments before it can be folded.
26039 let mut saw_separator = false;
26040 // v7.25 (round-17) — `COUNT(DISTINCT x)` and friends.
26041 // v7.32 (round-29) — accept the dual `ALL` quantifier too
26042 // (the default; ORMs emit `COUNT(ALL x)` / `SUM(ALL x)`).
26043 let agg_distinct = if matches!(self.peek(), Token::Distinct) {
26044 self.advance();
26045 true
26046 } else if matches!(self.peek(), Token::All) {
26047 self.advance();
26048 false
26049 } else {
26050 false
26051 };
26052 // v7.37.17 (17.6 siblings) — MySQL TIMESTAMPADD /
26053 // TIMESTAMPDIFF take a bare unit keyword as the first
26054 // argument (MINUTE, DAY, ...), and GET_FORMAT takes a
26055 // bare type keyword (DATE / TIME / DATETIME); lower them
26056 // onto string literals so the evaluator sees plain text.
26057 if ((first.eq_ignore_ascii_case("timestampadd")
26058 || first.eq_ignore_ascii_case("timestampdiff"))
26059 && matches!(self.peek(), Token::Ident(u) if matches!(
26060 u.to_ascii_lowercase().as_str(),
26061 "microsecond" | "second" | "minute" | "hour" | "day"
26062 | "week" | "month" | "quarter" | "year"
26063 )))
26064 || (first.eq_ignore_ascii_case("get_format")
26065 && matches!(self.peek(), Token::Ident(u) if matches!(
26066 u.to_ascii_lowercase().as_str(),
26067 "date" | "time" | "datetime" | "timestamp"
26068 )))
26069 {
26070 if let Token::Ident(u) = self.peek() {
26071 args.push(Expr::Literal(Literal::String(u.to_ascii_lowercase())));
26072 }
26073 self.advance();
26074 if matches!(self.peek(), Token::Comma) {
26075 self.advance();
26076 }
26077 }
26078 // `ROW(a, b, …)` keyword constructor. Followed by a
26079 // comparison operator or [NOT] IN it joins the paren
26080 // row-constructor machinery (fieldwise parse-time
26081 // expansion); bare, it stays a `row` call the evaluator
26082 // renders as PG record text.
26083 if first.eq_ignore_ascii_case("row") {
26084 let mut row_items = Vec::new();
26085 if !matches!(self.peek(), Token::RParen) {
26086 loop {
26087 row_items.push(self.parse_expr(0)?);
26088 match self.peek() {
26089 Token::Comma => {
26090 self.advance();
26091 }
26092 Token::RParen => break,
26093 other => {
26094 return Err(self.err(format!(
26095 "expected ',' or ')' in ROW(...), got {other:?}"
26096 )));
26097 }
26098 }
26099 }
26100 }
26101 self.advance(); // ')'
26102 let comparison_follows = matches!(
26103 self.peek(),
26104 Token::Eq
26105 | Token::NotEq
26106 | Token::Lt
26107 | Token::LtEq
26108 | Token::Gt
26109 | Token::GtEq
26110 | Token::In
26111 ) || (matches!(self.peek(), Token::Not)
26112 && matches!(self.tokens.get(self.pos + 1), Some(Token::In)))
26113 || matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("overlaps"));
26114 if comparison_follows && !row_items.is_empty() {
26115 return self.parse_row_comparison_tail(row_items);
26116 }
26117 return Ok(Expr::FunctionCall {
26118 name: String::from("row"),
26119 args: row_items,
26120 });
26121 }
26122 // v7.39 (read01 xml.c) — `XMLPARSE(DOCUMENT|CONTENT expr)`:
26123 // the parse-mode keyword introduces the source text. SPG
26124 // carries XML as text, so both modes lower to __xmlparse(expr)
26125 // which validates well-formedness and returns Value::Xml.
26126 if first.eq_ignore_ascii_case("xmlparse")
26127 && matches!(self.peek(), Token::Ident(kw)
26128 if kw.eq_ignore_ascii_case("document")
26129 || kw.eq_ignore_ascii_case("content"))
26130 {
26131 let mode = match self.advance() {
26132 Token::Ident(kw) => kw.to_ascii_lowercase(),
26133 _ => unreachable!("peeked an ident"),
26134 };
26135 let src = self.parse_expr(0)?;
26136 if !matches!(self.peek(), Token::RParen) {
26137 return Err(self.err(format!(
26138 "expected ')' to close XMLPARSE, got {:?}",
26139 self.peek()
26140 )));
26141 }
26142 self.advance();
26143 return Ok(Expr::FunctionCall {
26144 name: String::from("__xmlparse"),
26145 args: alloc::vec![src, Expr::Literal(Literal::String(mode))],
26146 });
26147 }
26148 // SQL/XML `XMLELEMENT(NAME ident [, content …])` — the NAME
26149 // keyword introduces the element name (a bare or quoted
26150 // identifier), then optional content expressions. Lower to a
26151 // plain `xmlelement(name_text, content …)` call.
26152 if first.eq_ignore_ascii_case("xmlelement")
26153 && matches!(self.peek(), Token::Ident(kw) if kw.eq_ignore_ascii_case("name"))
26154 {
26155 self.advance(); // consume NAME
26156 let elem_name = match self.peek().clone() {
26157 Token::Ident(n) | Token::QuotedIdent(n) => {
26158 self.advance();
26159 n
26160 }
26161 other => {
26162 return Err(self.err(format!(
26163 "expected element name after XMLELEMENT NAME, got {other:?}"
26164 )));
26165 }
26166 };
26167 let mut args = alloc::vec![Expr::Literal(Literal::String(elem_name))];
26168 while matches!(self.peek(), Token::Comma) {
26169 self.advance();
26170 args.push(self.parse_expr(0)?);
26171 }
26172 if !matches!(self.peek(), Token::RParen) {
26173 return Err(self.err(format!(
26174 "expected ')' to close XMLELEMENT, got {:?}",
26175 self.peek()
26176 )));
26177 }
26178 self.advance();
26179 return Ok(Expr::FunctionCall {
26180 name: String::from("xmlelement"),
26181 args,
26182 });
26183 }
26184 // SQL/XML `XMLFOREST(value [AS name], …)` — each `value AS name`
26185 // becomes a `<name>value</name>` element; a bare column infers its
26186 // own name. Lower to `xmlforest(name1, val1, name2, val2, …)`.
26187 // v7.39.2 — MySQL's two CONVERT forms, neither of which parsed.
26188 // `CONVERT(expr USING cs)` was a syntax error at USING, and
26189 // `CONVERT(expr, CHAR)` was read as PostgreSQL's three-argument
26190 // `convert(bytea, src, dest)` and answered `column "char" does
26191 // not exist`. Both are casts in MySQL: measured on 9.7.2,
26192 // `CONVERT(0x41 USING utf8mb4)` and `CONVERT(0x41, CHAR)` are
26193 // both 'A', and `CONVERT(123, CHAR)` is '123'.
26194 //
26195 // The charset is checked against the same table the introducers
26196 // use, so an unknown one is refused rather than quietly ignored.
26197 if self.mysql_dialect
26198 && first.eq_ignore_ascii_case("convert")
26199 && !matches!(self.peek(), Token::RParen)
26200 {
26201 let save = self.pos;
26202 let inner = self.parse_expr(0)?;
26203 if matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("using")) {
26204 self.advance();
26205 let cs = match self.peek().clone() {
26206 Token::Ident(n) | Token::QuotedIdent(n) => {
26207 self.advance();
26208 n
26209 }
26210 other => {
26211 return Err(self.err(alloc::format!(
26212 "expected a charset after USING, got {other:?}"
26213 )));
26214 }
26215 };
26216 let lc = cs.to_ascii_lowercase();
26217 if lc != "binary" && crate::charset::charset_default_collation(&lc).is_none() {
26218 return Err(self.err(alloc::format!("unknown character set: '{cs}'")));
26219 }
26220 if !matches!(self.peek(), Token::RParen) {
26221 return Err(self.err(alloc::format!(
26222 "expected ')' after CONVERT … USING, got {:?}",
26223 self.peek()
26224 )));
26225 }
26226 self.advance();
26227 let target = if lc == "binary" {
26228 CastTarget::Named("binary".to_string())
26229 } else {
26230 CastTarget::Text
26231 };
26232 return self.finish_postfix_casts(Expr::Cast {
26233 expr: alloc::boxed::Box::new(inner),
26234 target,
26235 });
26236 }
26237 if matches!(self.peek(), Token::Comma) {
26238 self.advance();
26239 // A type name here is MySQL's cast form; anything else
26240 // (three string arguments) is PostgreSQL's `convert`,
26241 // which keeps its own path.
26242 if let Ok(target) = self.parse_cast_target()
26243 && matches!(self.peek(), Token::RParen)
26244 {
26245 self.advance();
26246 return self.finish_postfix_casts(Expr::Cast {
26247 expr: alloc::boxed::Box::new(inner),
26248 target,
26249 });
26250 }
26251 }
26252 self.pos = save;
26253 }
26254 if first.eq_ignore_ascii_case("xmlforest") && !matches!(self.peek(), Token::RParen) {
26255 let mut args: Vec<Expr> = Vec::new();
26256 loop {
26257 let val = self.parse_expr(0)?;
26258 let name = if matches!(self.peek(), Token::As) {
26259 self.advance();
26260 match self.peek().clone() {
26261 Token::Ident(n) | Token::QuotedIdent(n) => {
26262 self.advance();
26263 n
26264 }
26265 other => {
26266 return Err(self.err(format!(
26267 "expected name after AS in XMLFOREST, got {other:?}"
26268 )));
26269 }
26270 }
26271 } else if let Expr::Column(c) = &val {
26272 c.name.clone()
26273 } else {
26274 return Err(
26275 self.err("XMLFOREST element without a column name needs AS".into())
26276 );
26277 };
26278 args.push(Expr::Literal(Literal::String(name)));
26279 args.push(val);
26280 if matches!(self.peek(), Token::Comma) {
26281 self.advance();
26282 } else {
26283 break;
26284 }
26285 }
26286 if !matches!(self.peek(), Token::RParen) {
26287 return Err(self.err(format!(
26288 "expected ')' to close XMLFOREST, got {:?}",
26289 self.peek()
26290 )));
26291 }
26292 self.advance();
26293 return Ok(Expr::FunctionCall {
26294 name: String::from("xmlforest"),
26295 args,
26296 });
26297 }
26298 // SQL-standard `POSITION(sub IN str)` — lowers onto
26299 // strpos(str, sub). IN is the argument separator here,
26300 // so the needle parses with the IN-tail suppressed.
26301 if first.eq_ignore_ascii_case("position") && !matches!(self.peek(), Token::RParen) {
26302 let saved = self.suppress_in_tail;
26303 self.suppress_in_tail = true;
26304 let needle = self.parse_expr(0);
26305 self.suppress_in_tail = saved;
26306 let needle = needle?;
26307 if matches!(self.peek(), Token::In) {
26308 self.advance();
26309 let haystack = self.parse_expr(0)?;
26310 if !matches!(self.peek(), Token::RParen) {
26311 return Err(self.err(format!(
26312 "expected ')' to close POSITION, got {:?}",
26313 self.peek()
26314 )));
26315 }
26316 self.advance();
26317 return Ok(Expr::FunctionCall {
26318 name: String::from("strpos"),
26319 args: alloc::vec![haystack, needle],
26320 });
26321 }
26322 // position(sub, str) comma form (incl. bytea) —
26323 // hand the parsed first arg to the generic list.
26324 args.push(needle);
26325 if matches!(self.peek(), Token::Comma) {
26326 self.advance();
26327 }
26328 }
26329 // SQL-standard `TRIM([BOTH|LEADING|TRAILING] [chars]
26330 // FROM str)` — lowers onto btrim / ltrim / rtrim. The
26331 // plain comma forms TRIM(str) / TRIM(str, chars) keep
26332 // riding the generic argument list below.
26333 if first.eq_ignore_ascii_case("trim") {
26334 let mode = match self.peek() {
26335 Token::Ident(k) if k.eq_ignore_ascii_case("both") => {
26336 self.advance();
26337 Some("btrim")
26338 }
26339 Token::Ident(k) if k.eq_ignore_ascii_case("leading") => {
26340 self.advance();
26341 Some("ltrim")
26342 }
26343 Token::Ident(k) if k.eq_ignore_ascii_case("trailing") => {
26344 self.advance();
26345 Some("rtrim")
26346 }
26347 _ => None,
26348 };
26349 if mode.is_some() || matches!(self.peek(), Token::From) {
26350 // TRIM([mode] FROM str) — no strip-chars.
26351 let chars = if matches!(self.peek(), Token::From) {
26352 None
26353 } else {
26354 Some(self.parse_expr(0)?)
26355 };
26356 if !matches!(self.peek(), Token::From) {
26357 return Err(self.err(format!(
26358 "expected FROM in TRIM([BOTH|LEADING|TRAILING] [chars] FROM str), got {:?}",
26359 self.peek()
26360 )));
26361 }
26362 self.advance();
26363 let target = self.parse_expr(0)?;
26364 if !matches!(self.peek(), Token::RParen) {
26365 return Err(
26366 self.err(format!("expected ')' to close TRIM, got {:?}", self.peek()))
26367 );
26368 }
26369 self.advance();
26370 let mut trim_args = alloc::vec![target];
26371 if let Some(c) = chars {
26372 trim_args.push(c);
26373 }
26374 return Ok(Expr::FunctionCall {
26375 name: String::from(mode.unwrap_or("btrim")),
26376 args: trim_args,
26377 });
26378 }
26379 }
26380 if !matches!(self.peek(), Token::RParen) {
26381 loop {
26382 // v7.38 (read01, T14) — `argname => value` names this arg.
26383 // v7.39 (read01 round 77) — `argname := value` is the same
26384 // thing, and it is the spelling PG's own docs lead with. It
26385 // was simply never lexed here, so every `f(x := 1)` died in
26386 // the parser regardless of what `f` was.
26387 let this_name = match (&self.tokens[self.pos], self.tokens.get(self.pos + 1)) {
26388 (
26389 Token::Ident(n) | Token::QuotedIdent(n),
26390 Some(Token::FatArrow | Token::ColonEq),
26391 ) => {
26392 let name = n.clone();
26393 self.advance(); // name
26394 self.advance(); // => / :=
26395 Some(name)
26396 }
26397 _ => None,
26398 };
26399 // v7.39 (read01 round 100) — `VARIADIC <array>` spreads an
26400 // array's elements into a variadic call's trailing args
26401 // (`concat_ws(',', VARIADIC ARRAY[…])`). VARIADIC isn't
26402 // reserved, so it arrives as a bare ident before the arg.
26403 let is_variadic = this_name.is_none()
26404 && matches!(&self.tokens[self.pos], Token::Ident(s) if s.eq_ignore_ascii_case("variadic"));
26405 if is_variadic {
26406 self.advance();
26407 }
26408 let arg = self.parse_expr(0)?;
26409 args.push(match &this_name {
26410 // The callee's parameter names decide the slot, and a
26411 // user function's live in the catalog. Carry the name
26412 // to eval rather than guessing here.
26413 Some(n) => Expr::NamedArg {
26414 name: n.clone(),
26415 expr: Box::new(arg),
26416 },
26417 None if is_variadic => Expr::Variadic(Box::new(arg)),
26418 None => arg,
26419 });
26420 arg_names.push(this_name);
26421 // v7.25 (round-17) — standard `CAST(expr AS type)`.
26422 // The `::` cast already worked; this lowers the
26423 // function form onto the same Expr::Cast node.
26424 if first.eq_ignore_ascii_case("cast")
26425 && args.len() == 1
26426 && matches!(self.peek(), Token::As)
26427 {
26428 self.advance();
26429 let target = self.parse_cast_target()?;
26430 if !matches!(self.peek(), Token::RParen) {
26431 return Err(self.err(format!(
26432 "expected ')' to close CAST, got {:?}",
26433 self.peek()
26434 )));
26435 }
26436 self.advance();
26437 return Ok(Expr::Cast {
26438 expr: Box::new(args.pop().expect("one arg")),
26439 target,
26440 });
26441 }
26442 // v7.38 (read01 P6.-) — `normalize(text, FORM)` where FORM is
26443 // a bare keyword NFC / NFD / NFKC / NFKD. PG parses these as
26444 // keywords; SPG's lexer makes them plain idents (so they'd be
26445 // read as column refs). Lower the keyword to the string form
26446 // the evaluator already accepts.
26447 if first.eq_ignore_ascii_case("normalize")
26448 && args.len() == 1
26449 && matches!(self.peek(), Token::Comma)
26450 {
26451 let form = match self.tokens.get(self.pos + 1) {
26452 Some(Token::Ident(f) | Token::QuotedIdent(f)) => {
26453 let up = f.to_ascii_uppercase();
26454 matches!(up.as_str(), "NFC" | "NFD" | "NFKC" | "NFKD").then_some(up)
26455 }
26456 _ => None,
26457 };
26458 if let Some(up) = form {
26459 self.advance(); // comma
26460 self.advance(); // form keyword
26461 args.push(Expr::Literal(Literal::String(up)));
26462 }
26463 }
26464 // v7.37.7 C.1.8 — PG `substring(str FROM pos FOR len)` syntactic
26465 // form. Desugars to the comma-list shape evaluator already
26466 // handles. Triggered after the first arg when the function
26467 // name is substring / substr and the next token is FROM
26468 // (a reserved keyword in PG; SPG also reserves it).
26469 // v7.39 (read01 regexp.c) — `substring(str SIMILAR pat
26470 // ESCAPE esc)` (SQL:1999 three-part form) desugars to the
26471 // internal __substring_similar(str, pat, esc) call.
26472 if (first.eq_ignore_ascii_case("substring")
26473 || first.eq_ignore_ascii_case("substr"))
26474 && args.len() == 1
26475 && matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("similar"))
26476 {
26477 self.advance(); // SIMILAR
26478 let pattern = self.parse_expr(0)?;
26479 if !matches!(self.peek(), Token::Ident(s) if s.eq_ignore_ascii_case("escape"))
26480 {
26481 return Err(self.err(format!(
26482 "expected ESCAPE in substring(... SIMILAR ...), got {:?}",
26483 self.peek()
26484 )));
26485 }
26486 self.advance(); // ESCAPE
26487 let esc = self.parse_expr(0)?;
26488 if !matches!(self.peek(), Token::RParen) {
26489 return Err(self.err(format!(
26490 "expected ')' to close substring(... SIMILAR ...), got {:?}",
26491 self.peek()
26492 )));
26493 }
26494 self.advance();
26495 args.push(pattern);
26496 args.push(esc);
26497 return Ok(Expr::FunctionCall {
26498 name: "__substring_similar".to_string(),
26499 args,
26500 });
26501 }
26502 if (first.eq_ignore_ascii_case("substring")
26503 || first.eq_ignore_ascii_case("substr"))
26504 && args.len() == 1
26505 && matches!(self.peek(), Token::From | Token::For)
26506 {
26507 // `substring(str FROM pos [FOR len])`, or the FOR-only
26508 // `substring(str FOR len)` which PG treats as FROM 1.
26509 if matches!(self.peek(), Token::From) {
26510 self.advance();
26511 let start = self.parse_expr(0)?;
26512 args.push(start);
26513 } else {
26514 args.push(Expr::Literal(Literal::Integer(1)));
26515 }
26516 if matches!(self.peek(), Token::For) {
26517 self.advance();
26518 let length = self.parse_expr(0)?;
26519 args.push(length);
26520 }
26521 if !matches!(self.peek(), Token::RParen) {
26522 return Err(self.err(format!(
26523 "expected ')' to close substring(... FROM ... [FOR ...]), got {:?}",
26524 self.peek()
26525 )));
26526 }
26527 self.advance();
26528 return Ok(Expr::FunctionCall {
26529 name: first.to_ascii_lowercase(),
26530 args,
26531 });
26532 }
26533 // PG `overlay(str PLACING repl FROM n [FOR len])`
26534 // syntactic form. Desugars to the `overlay(str,
26535 // repl, n[, len])` comma-list shape the evaluator
26536 // already implements. `PLACING` is not a reserved
26537 // token in SPG, so it arrives as a bare Ident.
26538 if first.eq_ignore_ascii_case("overlay")
26539 && args.len() == 1
26540 && matches!(self.peek(), Token::Ident(kw) if kw == "placing")
26541 {
26542 self.advance(); // consume PLACING
26543 args.push(self.parse_expr(0)?); // replacement
26544 if !matches!(self.peek(), Token::From) {
26545 return Err(self.err(format!(
26546 "expected FROM in overlay(... PLACING ... FROM ...), got {:?}",
26547 self.peek()
26548 )));
26549 }
26550 self.advance();
26551 args.push(self.parse_expr(0)?); // start position
26552 if matches!(self.peek(), Token::For) {
26553 self.advance();
26554 args.push(self.parse_expr(0)?); // length
26555 }
26556 if !matches!(self.peek(), Token::RParen) {
26557 return Err(self.err(format!(
26558 "expected ')' to close overlay(... PLACING ... FROM ... [FOR ...]), got {:?}",
26559 self.peek()
26560 )));
26561 }
26562 self.advance();
26563 return Ok(Expr::FunctionCall {
26564 name: String::from("overlay"),
26565 args,
26566 });
26567 }
26568 // `TRIM(chars FROM str)` — the keyword-less
26569 // spelling lands here after the chars parse
26570 // (the keyword forms return earlier).
26571 if first.eq_ignore_ascii_case("trim")
26572 && args.len() == 1
26573 && matches!(self.peek(), Token::From)
26574 {
26575 self.advance();
26576 let target = self.parse_expr(0)?;
26577 if !matches!(self.peek(), Token::RParen) {
26578 return Err(self.err(format!(
26579 "expected ')' to close TRIM(chars FROM str), got {:?}",
26580 self.peek()
26581 )));
26582 }
26583 self.advance();
26584 let chars = args.pop().expect("one arg");
26585 return Ok(Expr::FunctionCall {
26586 name: String::from("btrim"),
26587 args: alloc::vec![target, chars],
26588 });
26589 }
26590 // v7.24 (round-16 A) — aggregate-internal
26591 // ordering: `array_agg(x ORDER BY y DESC NULLS
26592 // LAST)`. Keys close the argument list.
26593 if matches!(self.peek(), Token::Order) {
26594 self.advance();
26595 if !self.peek_is_by() {
26596 return Err(self.err(format!(
26597 "expected BY after ORDER in aggregate args, got {:?}",
26598 self.peek()
26599 )));
26600 }
26601 self.advance();
26602 loop {
26603 // v7.39 (round 691) — save/restore, the discipline this parser
26604 // already uses around `pending_sample_preds`, so a subquery inside
26605 // a key neither inherits nor leaks the channel.
26606 let saved_flag = core::mem::replace(&mut self.in_order_by_key, true);
26607 let saved_coll = self.order_key_collation.take();
26608 let parsed = self.parse_expr(0);
26609 self.in_order_by_key = saved_flag;
26610 let collation =
26611 core::mem::replace(&mut self.order_key_collation, saved_coll);
26612 let expr = parsed?;
26613 let desc = if matches!(self.peek(), Token::Desc) {
26614 self.advance();
26615 true
26616 } else if matches!(self.peek(), Token::Asc) {
26617 self.advance();
26618 false
26619 } else {
26620 false
26621 };
26622 let nulls_first = self.parse_optional_nulls_placement()?;
26623 agg_order_by.push(OrderBy {
26624 expr,
26625 desc,
26626 nulls_first,
26627 collation,
26628 });
26629 if matches!(self.peek(), Token::Comma) {
26630 self.advance();
26631 } else {
26632 break;
26633 }
26634 }
26635 // v7.39 (round 354, M12) — `SEPARATOR '<s>'` may
26636 // follow the ORDER BY inside GROUP_CONCAT.
26637 if self.consume_group_concat_separator(&mut args)? {
26638 saw_separator = true;
26639 }
26640 if !matches!(self.peek(), Token::RParen) {
26641 return Err(self.err(format!(
26642 "expected ')' after aggregate ORDER BY, got {:?}",
26643 self.peek()
26644 )));
26645 }
26646 break;
26647 }
26648 // v7.39 (round 354, M12) — …or directly after the
26649 // arguments (`GROUP_CONCAT(t SEPARATOR '|')`). MySQL's
26650 // own spelling of what PG passes as string_agg's second
26651 // argument; it was a parse error, so every MySQL query
26652 // that names its own separator failed outright.
26653 if self.consume_group_concat_separator(&mut args)? {
26654 saw_separator = true;
26655 break;
26656 }
26657 match self.peek() {
26658 Token::Comma => {
26659 self.advance();
26660 }
26661 Token::RParen => break,
26662 other => {
26663 return Err(self.err(format!(
26664 "expected ',' or ')' in function args, got {other:?}"
26665 )));
26666 }
26667 }
26668 }
26669 }
26670 // v7.39 (round 354, M12) — MySQL's GROUP_CONCAT concatenates
26671 // its value arguments PER ROW: `GROUP_CONCAT(n, ':', t)` is
26672 // `3:c,1:a,…` (measured on MariaDB 11), NOT a second argument
26673 // meaning a separator — that is what the explicit SEPARATOR
26674 // tail is for. Fold them into one `concat(...)` so the
26675 // aggregate keeps its single value argument.
26676 if self.mysql_dialect && first.eq_ignore_ascii_case("group_concat") {
26677 let values = args.len() - usize::from(saw_separator);
26678 if values > 1 {
26679 let sep_arg = if saw_separator { args.pop() } else { None };
26680 let folded = Expr::FunctionCall {
26681 name: "concat".to_string(),
26682 args: core::mem::take(&mut args),
26683 };
26684 args.push(folded);
26685 if let Some(sep) = sep_arg {
26686 args.push(sep);
26687 }
26688 }
26689 }
26690 self.advance(); // consume ')'
26691 // v7.39 (read01 round 77) — named arguments are NOT reordered here
26692 // any more. The parser has no catalog, so it could only ever resolve
26693 // the handful of `make_*` builtins whose parameter names were baked
26694 // into a table right here — every user function got
26695 // "does not support named arguments", though the catalog has been
26696 // storing its parameter names all along. Reordering happens in eval,
26697 // in one place, for builtins and user functions alike.
26698 // v7.32 (round-29) — ordered-set aggregate tail
26699 // `name(direct_args) WITHIN GROUP (ORDER BY …)`
26700 // (percentile_cont / percentile_disc / mode). The sort spec
26701 // lands in the same `order_by` slot a decorated aggregate
26702 // uses; the executor dispatches on the function name. WITHIN
26703 // GROUP and an intra-argument ORDER BY are mutually
26704 // exclusive (PG rejects both).
26705 let within_group_order = self.parse_within_group_clause()?;
26706 if !within_group_order.is_empty() && !agg_order_by.is_empty() {
26707 return Err(self.err(
26708 "an aggregate may not carry both an in-argument ORDER BY and WITHIN GROUP"
26709 .into(),
26710 ));
26711 }
26712 let within_group_seen = !within_group_order.is_empty();
26713 let agg_order_by = if within_group_order.is_empty() {
26714 agg_order_by
26715 } else {
26716 within_group_order
26717 };
26718 // v7.32 (round-29) — `name(args) FILTER (WHERE …)`.
26719 let filter = self.parse_filter_clause()?;
26720 // v4.12: window-function tail — `name(args) OVER (...)`.
26721 // Promotes the just-parsed FunctionCall into a
26722 // WindowFunction node carrying partition + order.
26723 // v6.4.2: also accepts `name(args) IGNORE NULLS OVER (...)`
26724 // / `RESPECT NULLS OVER (...)` between the closing paren
26725 // and `OVER`.
26726 let null_treatment = self.parse_null_treatment_modifier();
26727 if let Token::Ident(s) | Token::QuotedIdent(s) = self.peek()
26728 && s.eq_ignore_ascii_case("over")
26729 {
26730 self.advance();
26731 // v7.39 (round 230) — PG implements neither modifier for a
26732 // windowed call and says so (0A000). Both used to be parsed
26733 // and then silently dropped here, so `count(DISTINCT v)
26734 // OVER (…)` quietly answered the non-distinct count.
26735 if agg_distinct {
26736 return Err(
26737 self.err("DISTINCT is not implemented for window functions".to_string())
26738 );
26739 }
26740 if !agg_order_by.is_empty() {
26741 // PG separates the two shapes that land here: a
26742 // WITHIN GROUP call is an ordered-set aggregate and gets
26743 // its own message naming the aggregate; a plain
26744 // `agg(x ORDER BY y)` gets the generic one.
26745 let msg = if within_group_seen {
26746 alloc::format!("OVER is not supported for ordered-set aggregate {first}")
26747 } else {
26748 "aggregate ORDER BY is not implemented for window functions".to_string()
26749 };
26750 return Err(self.err(msg));
26751 }
26752 let (partition_by, order_by, frame) = self.parse_over_clause()?;
26753 return Ok(Expr::WindowFunction {
26754 name: first,
26755 args,
26756 partition_by,
26757 order_by,
26758 frame,
26759 null_treatment,
26760 filter,
26761 });
26762 }
26763 if !agg_order_by.is_empty() || agg_distinct || filter.is_some() {
26764 return Ok(Expr::AggregateOrdered {
26765 call: Box::new(Expr::FunctionCall { name: first, args }),
26766 order_by: agg_order_by,
26767 distinct: agg_distinct,
26768 filter,
26769 });
26770 }
26771 // v7.39 (round 522) — PG declares `date_add` / `date_subtract`
26772 // over TIMESTAMPTZ and has no timestamp overload, so a
26773 // timestamp argument is coerced on the way in and the answer
26774 // is timestamptz — measured: `pg_typeof(date_add(TIMESTAMP
26775 // '2020-01-01', INTERVAL '1 hour'))` is `timestamp with time
26776 // zone`. SPG answered `timestamp without time zone`, dropping
26777 // the offset from every rendering.
26778 //
26779 // Writing the coercion PG performs makes the existing
26780 // argument-driven typing (the one `date_trunc` uses) reach the
26781 // right answer, rather than teaching the type layer a second
26782 // rule. MySQL's DATE_ADD is a different function that returns
26783 // DATE or DATETIME, so this is PG-dialect only.
26784 //
26785 // Out-of-line because this sits on the RECURSIVE descent
26786 // frame: an inline block with locals here costs every nesting
26787 // level, and the suite's deep-nesting sentinel overflowed the
26788 // 512 KiB parser stack the moment one was added (round 430's
26789 // lesson, in the same shape).
26790 if !self.mysql_dialect {
26791 lift_date_add_arg_to_timestamptz(&first, &mut args);
26792 }
26793 return Ok(Expr::FunctionCall { name: first, args });
26794 }
26795 // v7.9.20 — SQL-standard parenless keyword expressions
26796 // (PG treats these as functions called without parens).
26797 // Resolve to a synthetic FunctionCall so the engine's
26798 // eval path reuses the existing function-call routing.
26799 // mailrs G3.
26800 let lc = first.to_ascii_lowercase();
26801 if matches!(
26802 lc.as_str(),
26803 "current_date"
26804 | "current_time"
26805 | "current_timestamp"
26806 | "localtimestamp"
26807 | "localtime"
26808 // v7.37.17 (17.6 siblings) — session-identity SQL-
26809 // standard parenless keywords. current_user /
26810 // session_user / user were already caught by the
26811 // pgwire canned-response shortcut but bare-select
26812 // in the embedded engine went through Expr::Column
26813 // and errored. Adding them here so the parser
26814 // resolves to a synthetic FunctionCall that reuses
26815 // the existing eval/functions.rs dispatch.
26816 | "current_user"
26817 | "session_user"
26818 | "current_role"
26819 | "current_catalog"
26820 | "current_schema"
26821 | "current_database"
26822 // v7.39 (read01 round 51) — PG 16's system_user is parenless too.
26823 | "system_user"
26824 ) {
26825 return Ok(Expr::FunctionCall {
26826 name: lc,
26827 args: Vec::new(),
26828 });
26829 }
26830 Ok(Expr::Column(ColumnName {
26831 qualifier: None,
26832 name: first,
26833 }))
26834 }
26835}
26836
26837/// v7.39 (round 522) — write the coercion PG's `date_add` /
26838/// `date_subtract` signature performs.
26839///
26840/// PG declares both over TIMESTAMPTZ and has no timestamp overload, so a
26841/// timestamp argument is cast on the way in and the answer is
26842/// timestamptz — measured: `pg_typeof(date_add(TIMESTAMP '2020-01-01',
26843/// INTERVAL '1 hour'))` is `timestamp with time zone`. SPG answered
26844/// `timestamp without time zone`, dropping the offset from every
26845/// rendering of the result.
26846///
26847/// Writing the cast the signature implies lets the existing
26848/// argument-driven typing (the one `date_trunc` uses) reach the right
26849/// answer instead of teaching the type layer a second rule. MySQL's
26850/// DATE_ADD is a different function returning DATE or DATETIME, so the
26851/// caller applies this in PG dialect only.
26852///
26853/// A free function, and not a block at the call site, because the caller
26854/// is on the recursive-descent frame chain.
26855#[inline(never)]
26856fn lift_date_add_arg_to_timestamptz(name: &str, args: &mut alloc::vec::Vec<Expr>) {
26857 if args.len() != 2
26858 || !(name.eq_ignore_ascii_case("date_add") || name.eq_ignore_ascii_case("date_subtract"))
26859 {
26860 return;
26861 }
26862 let base = args.remove(0);
26863 args.insert(
26864 0,
26865 Expr::Cast {
26866 expr: Box::new(base),
26867 target: CastTarget::Timestamptz,
26868 },
26869 );
26870}
26871
26872/// v6.8.2 — walk an expression tree and return the first column
26873/// reference's bare name. Used by `parse_create_index_stmt_after_create`
26874/// to derive `CreateIndexStatement.column` from an expression
26875/// key (so downstream planner code resolving a primary column
26876/// position keeps working with expression indexes). Returns
26877/// `None` when the expression has no column ref at all — caller
26878/// surfaces that as a parse error.
26879fn extract_first_column(expr: &Expr) -> Option<String> {
26880 match expr {
26881 Expr::Column(cn) => Some(cn.name.clone()),
26882 Expr::FunctionCall { args, .. } => args.iter().find_map(extract_first_column),
26883 Expr::Binary { lhs, rhs, .. } => {
26884 extract_first_column(lhs).or_else(|| extract_first_column(rhs))
26885 }
26886 Expr::Unary { expr: e, .. } => extract_first_column(e),
26887 // v7.39 (read01 round 93) — a cast wraps its operand: a common
26888 // expression-index key is `lower(col::text)`, where the column
26889 // sits under the `::text` cast inside the function arg. Without
26890 // descending here the key was rejected as "references no column".
26891 Expr::Cast { expr: e, .. } => extract_first_column(e),
26892 // v7.39.2 — and a COLLATE wraps its operand the same way.
26893 // `CREATE INDEX rc ON t (c COLLATE "C" DESC)` stopped naming a
26894 // column the moment the clause became a node instead of being
26895 // absorbed, and the key was rejected as referencing none. This
26896 // is the shape the wildcard below silently produces, which is
26897 // why it is spelled out.
26898 Expr::Collate { expr: e, .. } => extract_first_column(e),
26899 _ => None,
26900 }
26901}
26902
26903fn maybe_not(expr: Expr, negated: bool) -> Expr {
26904 if negated {
26905 Expr::Unary {
26906 op: UnOp::Not,
26907 expr: Box::new(expr),
26908 }
26909 } else {
26910 expr
26911 }
26912}
26913
26914/// v7.39 (round 353, M9/M10) — three operator TOKENS mean different
26915/// things in the two dialects, and SPG read all three PG's way:
26916///
26917/// | token | PG (and SPG) | MySQL, measured |
26918/// |---|---|---|
26919/// | `\|\|` | string concatenation | **OR** — `1 \|\| 0` is 1, not '10' |
26920/// | `&&` | inet / array overlap | **AND** |
26921/// | `<=>` | pgvector cosine distance | **NULL-safe equal** |
26922///
26923/// `1 || 0` answering the string '10' on a MySQL session is a wrong
26924/// answer with no error, which is why they are routed here rather than
26925/// left to the shared table.
26926impl Parser {
26927 fn binop_here(&self, tok: &Token) -> Option<(BinOp, u8)> {
26928 if self.mysql_dialect {
26929 // v7.39 (round 353, M9) — `DIV` is MySQL's truncating integer
26930 // division (`5 DIV 2` is 2, `-7 DIV 2` is -3 — toward zero —
26931 // and `5 DIV 0` is NULL). It is a plain ident to the lexer.
26932 if let Token::Ident(w) = tok
26933 && w.eq_ignore_ascii_case("div")
26934 {
26935 return Some((BinOp::IntDiv, 8));
26936 }
26937 // v7.39 (round 394) — `MOD` is MySQL's modulo operator, a synonym
26938 // for `%` (`10 MOD 3` is 1, `5.5 MOD 2` is 1.5). A plain ident to
26939 // the lexer; the `MOD(x, y)` function form is unaffected (MOD
26940 // there sits in operand position, not infix).
26941 if let Token::Ident(w) = tok
26942 && w.eq_ignore_ascii_case("mod")
26943 {
26944 return Some((BinOp::Mod, 8));
26945 }
26946 // v7.39 (round 407) — `XOR` is MySQL's logical exclusive-or, a
26947 // plain ident to the lexer. Its precedence sits between OR (1)
26948 // and AND (3) — hence rung 2, the slot freed by moving AND up.
26949 if let Token::Ident(w) = tok
26950 && w.eq_ignore_ascii_case("xor")
26951 {
26952 return Some((BinOp::LogicalXor, 2));
26953 }
26954 match tok {
26955 Token::Concat => return Some((BinOp::Or, 1)),
26956 // MySQL's `&&` is logical AND, sharing AND's rung (3).
26957 Token::InetOverlap => return Some((BinOp::And, 3)),
26958 // MySQL's `<=>` is NULL-safe equal, at the comparison rung (5).
26959 Token::CosineDistance => return Some((BinOp::IsNotDistinctFrom, 5)),
26960 _ => {}
26961 }
26962 }
26963 binop_from(tok)
26964 }
26965}
26966
26967// v7.39 (round 407) — precedence ladder. To open a rung for MySQL's `XOR`
26968// (which sits strictly between OR and AND), every level from AND upward was
26969// shifted +1: the ladder is now OR=1, XOR=2, AND=3, IS=4, comparison=5,
26970// distance=6, additive/concat/bitwise=7, multiplicative/JSON=8, prefix=9.
26971// XOR only exists in the MySQL dialect (binop_here); PG never sees it, and
26972// the *relative* order of every PG operator is unchanged by the shift.
26973fn binop_from(tok: &Token) -> Option<(BinOp, u8)> {
26974 let pair = match tok {
26975 Token::Or => (BinOp::Or, 1),
26976 Token::And => (BinOp::And, 3),
26977 Token::Eq => (BinOp::Eq, 5),
26978 Token::NotEq => (BinOp::NotEq, 5),
26979 Token::Lt => (BinOp::Lt, 5),
26980 Token::LtEq => (BinOp::LtEq, 5),
26981 Token::Gt => (BinOp::Gt, 5),
26982 Token::GtEq => (BinOp::GtEq, 5),
26983 // pgvector distance ops all sit on the same rung — tighter than
26984 // comparisons (5) so `col <-> v < threshold` parses correctly.
26985 Token::L2Distance => (BinOp::L2Distance, 6),
26986 // v7.39 (read01 geo_ops.c) — geometric predicates ride the
26987 // comparison rung.
26988 Token::GeomParallel => (BinOp::GeomParallel, 5),
26989 // v7.39 (read01 rangetypes.c) — range `&<` / `&>` on the
26990 // comparison rung.
26991 Token::OverLeft => (BinOp::OverLeft, 5),
26992 Token::OverRight => (BinOp::OverRight, 5),
26993 Token::GeomPerp => (BinOp::GeomPerp, 5),
26994 Token::GeomSameAs => (BinOp::GeomSameAs, 5),
26995 Token::ClosestPoint => (BinOp::ClosestPoint, 6),
26996 Token::GeomHoriz => (BinOp::GeomHoriz, 5),
26997 Token::InnerProduct => (BinOp::InnerProduct, 6),
26998 Token::CosineDistance => (BinOp::CosineDistance, 6),
26999 Token::Plus => (BinOp::Add, 7),
27000 Token::Minus => (BinOp::Sub, 7),
27001 // v7.39 (round 760, F31-B1) — the generic-operator rung. PG
27002 // binds every "other" operator (`||`, `|`, `&`, `#`, the
27003 // pgvector distances above) BETWEEN additive (7) and the
27004 // comparisons (5): `'a' || 1 + 1` is `'a' || 2` → `a2`,
27005 // `a & b + 1` is `a & (b + 1)`, and `flags & $1 = 0` stays
27006 // `(flags & $1) = 0`. They shared rung 7 with `+ -` since v1
27007 // ("matches PG conceptually" — the round-753 audit measured it
27008 // false; the old rung errored on `'a' || 1 + 1` with
27009 // `text + integer`). Same-level chains left-fold, as PG does.
27010 Token::Concat => (BinOp::Concat, 6),
27011 Token::Pipe => (BinOp::BitOr, 6),
27012 Token::Amp => (BinOp::BitAnd, 6),
27013 Token::Star => (BinOp::Mul, 8),
27014 Token::Slash => (BinOp::Div, 8),
27015 Token::Percent => (BinOp::Mod, 8),
27016 // v4.14: JSON path ops bind tighter than comparisons (5)
27017 // and additive (7) so `doc->'k' = 'v'` parses correctly.
27018 // Same rung as the multiplicative ops.
27019 Token::JsonGet => (BinOp::JsonGet, 8),
27020 Token::JsonGetText => (BinOp::JsonGetText, 8),
27021 Token::JsonGetPath => (BinOp::JsonGetPath, 8),
27022 Token::JsonGetPathText => (BinOp::JsonGetPathText, 8),
27023 Token::JsonContains => (BinOp::JsonContains, 8),
27024 Token::JsonPathExists => (BinOp::JsonPathExists, 8),
27025 Token::JsonContainedBy => (BinOp::JsonContainedBy, 8),
27026 Token::JsonKeyExists => (BinOp::JsonKeyExists, 8),
27027 Token::JsonKeysAny => (BinOp::JsonKeysAny, 8),
27028 Token::JsonKeysAll => (BinOp::JsonKeysAll, 8),
27029 Token::JsonDeletePath => (BinOp::JsonDeletePath, 8),
27030 // v7.12.2 — `@@` binds at the comparison rung (looser than
27031 // arithmetic, tighter than AND / OR). PG places `@@` at
27032 // the same precedence as `=` / `<`, so we follow.
27033 Token::TsMatch => (BinOp::TsMatch, 5),
27034 // v7.17.0 Phase 3.P0-47 — PG INET / CIDR containment + overlap.
27035 // PG places these at the comparison rung (same level as `=`),
27036 // so we follow.
27037 Token::InetContainedBy => (BinOp::InetContainedBy, 5),
27038 Token::InetContainedByEq => (BinOp::InetContainedByEq, 5),
27039 Token::InetContains => (BinOp::InetContains, 5),
27040 Token::InetContainsEq => (BinOp::InetContainsEq, 5),
27041 Token::InetOverlap => (BinOp::InetOverlap, 5),
27042 // v7.39 (round 508) — the geometric and pattern-order predicates
27043 // ride the comparison rung, as every other predicate does.
27044 Token::Intersects => (BinOp::Intersects, 5),
27045 Token::IsBelow => (BinOp::IsBelow, 5),
27046 Token::IsAbove => (BinOp::IsAbove, 5),
27047 Token::PatternLt => (BinOp::PatternLt, 5),
27048 Token::PatternLtEq => (BinOp::PatternLtEq, 5),
27049 Token::PatternGt => (BinOp::PatternGt, 5),
27050 Token::PatternGtEq => (BinOp::PatternGtEq, 5),
27051 // `@@@` is the old spelling of `@@` and means exactly it.
27052 Token::TsMatchOld => (BinOp::TsMatch, 5),
27053 _ => return None,
27054 };
27055 Some(pair)
27056}
27057
27058#[allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
27059// `as f32` here is intentional: vector elements widen / narrow into f32 on
27060// purpose. i64 → f32 loses precision past 2^24, f64 → f32 loses precision
27061// past ~15 decimal digits — both are acceptable for a fixed-precision
27062// pgvector column.
27063/// v7.17.0 Phase 1.3 — words that would otherwise be eaten as an
27064/// implicit table alias and break trailing clauses. WITH lands
27065/// here so `… FROM t WITH NO DATA` doesn't consume WITH as the
27066/// alias for `t`; same for ON / WHERE / HAVING / GROUP / ORDER /
27067/// LIMIT / OFFSET / UNION / EXCEPT / INTERSECT / RETURNING / SET
27068/// / VALUES / FOR / LATERAL — all of which would otherwise be
27069/// silently swallowed by `parse_optional_alias`.
27070fn is_alias_stopword(s: &str) -> bool {
27071 matches!(
27072 s.to_ascii_lowercase().as_str(),
27073 "with"
27074 | "on"
27075 | "where"
27076 | "having"
27077 | "group"
27078 | "order"
27079 | "limit"
27080 | "offset"
27081 | "union"
27082 | "except"
27083 | "intersect"
27084 | "returning"
27085 | "set"
27086 | "values"
27087 | "for"
27088 | "window"
27089 | "tablesample"
27090 | "lateral"
27091 | "left"
27092 | "right"
27093 | "inner"
27094 | "outer"
27095 | "full"
27096 | "cross"
27097 | "join"
27098 | "natural"
27099 | "using"
27100 | "fetch"
27101 )
27102}
27103
27104fn extract_numeric_literal(e: &Expr) -> Option<f32> {
27105 match e {
27106 Expr::Literal(Literal::Integer(n)) => Some(*n as f32),
27107 Expr::Literal(Literal::Float(x)) => Some(*x as f32),
27108 // v7.38 (read01) — a dotted literal is now NUMERIC, so a vector element
27109 // like `2.5` arrives as Literal::Numeric; widen it into f32. (`no_std`,
27110 // so scale the divisor by hand instead of `f32::powi`.)
27111 Expr::Literal(Literal::Numeric { unscaled, scale }) => {
27112 let mut div = 1.0f32;
27113 for _ in 0..*scale {
27114 div *= 10.0;
27115 }
27116 Some(*unscaled as f32 / div)
27117 }
27118 Expr::Unary {
27119 op: UnOp::Neg,
27120 expr,
27121 } => extract_numeric_literal(expr).map(|x| -x),
27122 _ => None,
27123 }
27124}
27125
27126/// Parse the text inside `INTERVAL '...'` into `(months, micros)`. Accepts
27127/// one or more `<n> <unit>` pairs separated by whitespace. `<n>` may be
27128/// negative. Returns `None` if any pair fails to parse or no pair is found.
27129///
27130/// Recognised units (case-insensitive, optional trailing `s`):
27131/// `microsecond`, `millisecond`, `second`, `minute`, `hour`, `day`, `week`,
27132/// `month`, `year`. `week` widens to 7 days; `year` widens to 12 months.
27133/// v7.37.5 β — returns `(months, days, micros)`. `days` is preserved
27134/// as its own dimension so `INTERVAL '1 day'` ≠ `INTERVAL '24 hours'`
27135/// (PG-canonical: DST and month-boundary semantics depend on this).
27136/// `week` rolls into `days` (× 7). Sub-day units flow into `micros`.
27137/// ISO 8601 duration input for INTERVAL: `P1Y2M3DT4H5M6S`. Before the `T`,
27138/// `M` is months; after it, `M` is minutes. Returns `(months, days, micros)`.
27139#[allow(clippy::cast_possible_truncation)]
27140fn parse_iso8601_interval(rest: &str) -> Option<(i32, i32, i64)> {
27141 let mut months: i64 = 0;
27142 let mut days: i64 = 0;
27143 let mut micros: i64 = 0;
27144 let mut in_time = false;
27145 let mut num = alloc::string::String::new();
27146 for ch in rest.chars() {
27147 if ch.is_ascii_digit() || ch == '.' || ch == '-' || ch == '+' {
27148 num.push(ch);
27149 continue;
27150 }
27151 if ch == 'T' || ch == 't' {
27152 if !num.is_empty() {
27153 return None;
27154 }
27155 in_time = true;
27156 continue;
27157 }
27158 let n: f64 = num.parse().ok()?;
27159 num.clear();
27160 match (ch, in_time) {
27161 ('Y' | 'y', false) => months += (n * 12.0) as i64,
27162 ('M', false) => months += n as i64,
27163 ('W' | 'w', false) => days += (n * 7.0) as i64,
27164 ('D' | 'd', false) => days += n as i64,
27165 ('H' | 'h', true) => micros += (n * 3_600_000_000.0) as i64,
27166 ('M', true) => micros += (n * 60_000_000.0) as i64,
27167 ('S' | 's', true) => micros += (n * 1_000_000.0) as i64,
27168 _ => return None,
27169 }
27170 }
27171 if !num.is_empty() {
27172 return None;
27173 }
27174 Some((
27175 i32::try_from(months).ok()?,
27176 i32::try_from(days).ok()?,
27177 micros,
27178 ))
27179}
27180
27181/// PG year-month shorthand for INTERVAL: `1-2` = 1 year 2 mons (an optional
27182/// leading `-` negates the whole value). Rejects date-like strings.
27183fn parse_year_month_interval(s: &str) -> Option<(i32, i32, i64)> {
27184 let (neg, body) = match s.strip_prefix('-') {
27185 Some(b) => (true, b),
27186 None => (false, s),
27187 };
27188 let (y, m) = body.split_once('-')?;
27189 let years: i32 = y.parse().ok()?;
27190 let mons: i32 = m.parse().ok()?;
27191 if years < 0 || mons < 0 {
27192 return None;
27193 }
27194 let total = years.checked_mul(12)?.checked_add(mons)?;
27195 Some((if neg { -total } else { total }, 0, 0))
27196}
27197
27198/// Parse a clock-time interval token `HH:MM[:SS[.ffffff]]` (optionally signed)
27199/// into microseconds. Used for the `3 days 14:30:45` / bare `14:30:45` forms.
27200fn parse_interval_clock(tok: &str) -> Option<i64> {
27201 let (neg, body) = match tok.strip_prefix('-') {
27202 Some(r) => (true, r),
27203 None => (false, tok.strip_prefix('+').unwrap_or(tok)),
27204 };
27205 let mut it = body.split(':');
27206 let h: i64 = it.next()?.parse().ok()?;
27207 let m: i64 = it.next()?.parse().ok()?;
27208 let s_tok = it.next().unwrap_or("0");
27209 if it.next().is_some() {
27210 return None;
27211 }
27212 let sec_us: i64 = if let Some((sec, frac)) = s_tok.split_once('.') {
27213 let sec: i64 = sec.parse().ok()?;
27214 let mut f = alloc::string::String::from(frac);
27215 while f.len() < 6 {
27216 f.push('0');
27217 }
27218 f.truncate(6);
27219 let fus: i64 = f.parse().ok()?;
27220 sec.checked_mul(1_000_000)?.checked_add(fus)?
27221 } else {
27222 s_tok.parse::<i64>().ok()?.checked_mul(1_000_000)?
27223 };
27224 let total = h
27225 .checked_mul(3_600_000_000)?
27226 .checked_add(m.checked_mul(60_000_000)?)?
27227 .checked_add(sec_us)?;
27228 Some(if neg { -total } else { total })
27229}
27230
27231/// v7.39 (read01 round 77) — one canonical name per interval unit, covering
27232/// every spelling PG accepts (measured against live PG18.4, not guessed):
27233/// `min` / `mins` / `m` are minutes, `mon` / `mons` are months, `y` is years.
27234/// Before this, the unit table matched long names only, with an ad-hoc
27235/// `strip_suffix('s')` in front of it — so `'15 min'` (and `hrs`, `secs`,
27236/// `yrs`, every abbreviation anyone actually types) was "cannot parse as
27237/// INTERVAL", and it had also grown arms for the debris that stripping leaves
27238/// behind (`centurie`, `millenniu`). Two parallel unit matches (integer and
27239/// fractional) both read from this one table now.
27240fn canonical_interval_unit(raw: &str) -> Option<&'static str> {
27241 let u = raw.to_ascii_lowercase();
27242 Some(match u.as_str() {
27243 "microsecond" | "microseconds" | "us" | "usec" | "usecs" | "usecond" | "useconds" => {
27244 "microsecond"
27245 }
27246 "millisecond" | "milliseconds" | "ms" | "msec" | "msecs" | "msecond" | "mseconds" => {
27247 "millisecond"
27248 }
27249 "second" | "seconds" | "sec" | "secs" | "s" => "second",
27250 "minute" | "minutes" | "min" | "mins" | "m" => "minute",
27251 "hour" | "hours" | "hr" | "hrs" | "h" => "hour",
27252 "day" | "days" | "d" => "day",
27253 "week" | "weeks" | "w" => "week",
27254 "month" | "months" | "mon" | "mons" => "month",
27255 "year" | "years" | "yr" | "yrs" | "y" => "year",
27256 "decade" | "decades" | "dec" | "decs" => "decade",
27257 "century" | "centuries" | "cent" | "c" => "century",
27258 "millennium" | "millenniums" | "millennia" | "mil" | "mils" => "millennium",
27259 _ => return None,
27260 })
27261}
27262
27263/// v7.39 (read01 round 102) — the six SQL-standard interval fields that can
27264/// qualify an `INTERVAL '…' <FIELD> [TO <FIELD>]` literal.
27265#[derive(Debug, Clone, Copy, PartialEq, Eq)]
27266pub(crate) enum IntervalField {
27267 Year,
27268 Month,
27269 Day,
27270 Hour,
27271 Minute,
27272 Second,
27273}
27274
27275/// Recognise an interval field keyword (bare ident, case-insensitive). Plural
27276/// spellings aren't standard for the qualifier position, so only the singular
27277/// forms are accepted.
27278/// v7.39 (round 350, M7) — MySQL's interval units, measured against
27279/// MariaDB 11. QUARTER is three months and WEEK seven days; MICROSECOND
27280/// is the finest. (The compound spellings — `DAY_HOUR` and friends, which
27281/// take a `'1 2'` style literal — are not read here; they stay a parse
27282/// error rather than being silently misread.)
27283/// v7.39 (round 430) — lower a `@name` / `@@name` reference to its call.
27284///
27285/// ONE `@` is a MySQL USER variable: its own per-session namespace, nothing
27286/// to do with a `@@` engine setting, and an unset one reads NULL rather
27287/// than raising. (The parser used to strip every `@`, so `@x` and `@@x`
27288/// were the same node and `SELECT @x` answered "Unknown system variable".)
27289/// For `@@`, the `session.` / `global.` scope is KEPT: a global read must
27290/// not see a session override — measured, after `SET autocommit=0`,
27291/// `@@global.autocommit` is still 1.
27292///
27293/// Out-of-line and NOT a method: `parse_atom` is the giant recursive frame
27294/// the parser's nesting budget is tuned against, and building these
27295/// `String` + `Vec` locals inside it overflowed the guard's stack (the same
27296/// wall `parse_left_right_atom` and friends were factored out for).
27297#[inline(never)]
27298fn variable_ref_atom(raw: &str) -> Expr {
27299 let user_var = !raw.starts_with("@@");
27300 let bare = raw.trim_start_matches('@').to_ascii_lowercase();
27301 Expr::FunctionCall {
27302 name: String::from(if user_var {
27303 "__spg_user_var"
27304 } else {
27305 "__spg_session_var"
27306 }),
27307 args: alloc::vec![Expr::Literal(Literal::String(bare))],
27308 }
27309}
27310
27311fn mysql_interval_unit(tok: &Token) -> Option<&'static str> {
27312 let Token::Ident(s) = tok else { return None };
27313 Some(match () {
27314 () if s.eq_ignore_ascii_case("microsecond") => "microsecond",
27315 () if s.eq_ignore_ascii_case("second") => "second",
27316 () if s.eq_ignore_ascii_case("minute") => "minute",
27317 () if s.eq_ignore_ascii_case("hour") => "hour",
27318 () if s.eq_ignore_ascii_case("day") => "day",
27319 () if s.eq_ignore_ascii_case("week") => "week",
27320 () if s.eq_ignore_ascii_case("month") => "month",
27321 () if s.eq_ignore_ascii_case("quarter") => "quarter",
27322 () if s.eq_ignore_ascii_case("year") => "year",
27323 () => return None,
27324 })
27325}
27326
27327/// v7.39 (round 422) — lower `INTERVAL <expr> <unit>` onto the existing
27328/// `make_interval(years, months, weeks, days, hours, mins, secs)` builtin,
27329/// which constructs the value at run time. Only the slot the unit names
27330/// carries the quantity; QUARTER and MICROSECOND scale it into the nearest
27331/// slot the builtin has (months and fractional seconds respectively).
27332fn make_interval_call(qty: Expr, unit: &str) -> Expr {
27333 let zero = || Expr::Literal(Literal::Integer(0));
27334 let scaled = |op: crate::ast::BinOp, by: Expr| Expr::Binary {
27335 lhs: alloc::boxed::Box::new(qty.clone()),
27336 op,
27337 rhs: alloc::boxed::Box::new(by),
27338 };
27339 // (years, months, weeks, days, hours, mins, secs)
27340 let mut args = alloc::vec![zero(), zero(), zero(), zero(), zero(), zero(), zero()];
27341 match unit {
27342 "year" => args[0] = qty,
27343 "quarter" => {
27344 args[1] = scaled(crate::ast::BinOp::Mul, Expr::Literal(Literal::Integer(3)));
27345 }
27346 "month" => args[1] = qty,
27347 "week" => args[2] = qty,
27348 "day" => args[3] = qty,
27349 "hour" => args[4] = qty,
27350 "minute" => args[5] = qty,
27351 "second" => args[6] = qty,
27352 // The builtin's seconds slot takes a fraction, so microseconds ride
27353 // it scaled down; the divisor is a NUMERIC literal so the division
27354 // stays exact rather than going through a float.
27355 "microsecond" => {
27356 args[6] = scaled(
27357 crate::ast::BinOp::Div,
27358 Expr::Literal(Literal::Numeric {
27359 unscaled: 1_000_000,
27360 scale: 0,
27361 }),
27362 );
27363 }
27364 _ => args[3] = qty,
27365 }
27366 Expr::FunctionCall {
27367 name: alloc::string::String::from("make_interval"),
27368 args,
27369 }
27370}
27371
27372/// `(count, unit)` → `(months, days, micros)`.
27373fn scale_mysql_interval(count: &str, unit: &str) -> Option<(i32, i32, i64)> {
27374 let n: i64 = count.trim().parse().ok()?;
27375 Some(match unit {
27376 "microsecond" => (0, 0, n),
27377 "second" => (0, 0, n.checked_mul(1_000_000)?),
27378 "minute" => (0, 0, n.checked_mul(60_000_000)?),
27379 "hour" => (0, 0, n.checked_mul(3_600_000_000)?),
27380 "day" => (0, i32::try_from(n).ok()?, 0),
27381 "week" => (0, i32::try_from(n.checked_mul(7)?).ok()?, 0),
27382 "month" => (i32::try_from(n).ok()?, 0, 0),
27383 "quarter" => (i32::try_from(n.checked_mul(3)?).ok()?, 0, 0),
27384 "year" => (i32::try_from(n.checked_mul(12)?).ok()?, 0, 0),
27385 _ => return None,
27386 })
27387}
27388
27389fn interval_field_of(tok: &Token) -> Option<IntervalField> {
27390 let Token::Ident(s) = tok else { return None };
27391 Some(match () {
27392 () if s.eq_ignore_ascii_case("year") => IntervalField::Year,
27393 () if s.eq_ignore_ascii_case("month") => IntervalField::Month,
27394 () if s.eq_ignore_ascii_case("day") => IntervalField::Day,
27395 () if s.eq_ignore_ascii_case("hour") => IntervalField::Hour,
27396 () if s.eq_ignore_ascii_case("minute") => IntervalField::Minute,
27397 () if s.eq_ignore_ascii_case("second") => IntervalField::Second,
27398 () => return None,
27399 })
27400}
27401
27402/// v7.39 (read01 round 102) — interpret an interval literal under a field
27403/// qualifier. Returns `(months, days, micros)`.
27404///
27405/// * A single field applied to a bare number sets which unit the number means,
27406/// truncated to that field's precision (`INTERVAL '1.5' HOUR` → `01:00:00`);
27407/// SECOND keeps its fraction (`'90.5' SECOND` → `00:01:30.5`).
27408/// * `YEAR TO MONTH` reads the `Y-M` form (`'1-6'` → 1 year 6 months).
27409/// * Every other range, and any literal a single field can't read as a plain
27410/// number (`'2 days' DAY`), falls back to the unqualified parse — SPG's
27411/// interval-text parser already reads the `D H:MM:SS` / `H:MM` forms exactly
27412/// like PG, and the qualifier there only bounds precision.
27413fn interpret_qualified_interval(
27414 text: &str,
27415 (f1, f2): (IntervalField, Option<IntervalField>),
27416) -> Option<(i32, i32, i64)> {
27417 if let Some(f2) = f2 {
27418 if f1 == IntervalField::Year && f2 == IntervalField::Month {
27419 if let Some(m) = parse_year_month_literal(text) {
27420 return Some((m, 0, 0));
27421 }
27422 }
27423 return parse_interval_text(text);
27424 }
27425 // Single field: reinterpret a bare number; otherwise the default parse.
27426 let trimmed = text.trim();
27427 if let Ok(val) = trimmed.parse::<f64>() {
27428 // no_std: f64 has no trunc/round; cast toward zero + round-half-away.
27429 #[allow(clippy::cast_possible_truncation)]
27430 let whole = val as i64;
27431 #[allow(clippy::cast_possible_truncation)]
27432 let secs_micros = {
27433 let m = val * 1_000_000.0;
27434 if m >= 0.0 {
27435 (m + 0.5) as i64
27436 } else {
27437 (m - 0.5) as i64
27438 }
27439 };
27440 return Some(match f1 {
27441 IntervalField::Year => (i32::try_from(whole).ok()?.checked_mul(12)?, 0, 0),
27442 IntervalField::Month => (i32::try_from(whole).ok()?, 0, 0),
27443 IntervalField::Day => (0, i32::try_from(whole).ok()?, 0),
27444 IntervalField::Hour => (0, 0, whole.checked_mul(3_600_000_000)?),
27445 IntervalField::Minute => (0, 0, whole.checked_mul(60_000_000)?),
27446 IntervalField::Second => (0, 0, secs_micros),
27447 });
27448 }
27449 parse_interval_text(text)
27450}
27451
27452/// Parse the `Y-M` (optionally signed) year-to-month literal into total months.
27453fn parse_year_month_literal(text: &str) -> Option<i32> {
27454 let t = text.trim();
27455 let (neg, body) = match t.strip_prefix('-') {
27456 Some(r) => (true, r),
27457 None => (false, t.strip_prefix('+').unwrap_or(t)),
27458 };
27459 let mut it = body.split('-');
27460 let years: i32 = it.next()?.trim().parse().ok()?;
27461 let months: i32 = match it.next() {
27462 Some(m) => m.trim().parse().ok()?,
27463 None => 0,
27464 };
27465 if it.next().is_some() {
27466 return None;
27467 }
27468 let total = years.checked_mul(12)?.checked_add(months)?;
27469 Some(if neg { -total } else { total })
27470}
27471
27472pub fn parse_interval_text(s: &str) -> Option<(i32, i32, i64)> {
27473 // v7.38.19 — the two infinities, answered as the three extreme
27474 // fields PostgreSQL itself puts on the wire for them:
27475 //
27476 // COPY (SELECT 'infinity'::interval) TO STDOUT (FORMAT binary)
27477 // … 7fffffffffffffff 7fffffff 7fffffff
27478 //
27479 // So no caller has to know the spelling — every one of them already
27480 // reads the three numbers, and `IntervalKind::from_fields` names
27481 // what they mean.
27482 //
27483 // `inf` is NOT one of them, measured: `'inf'::interval` is *invalid
27484 // input syntax* on PostgreSQL 18.4 while `'inf'::float8` is
27485 // infinity. Interval takes the full word, in any case.
27486 {
27487 let word = s.trim();
27488 let word = word.strip_prefix('@').map_or(word, str::trim);
27489 let (neg, body) = match word.strip_prefix('-') {
27490 Some(rest) => (true, rest.trim_start()),
27491 None => (false, word.strip_prefix('+').map_or(word, str::trim_start)),
27492 };
27493 if body.eq_ignore_ascii_case("infinity") {
27494 return Some(if neg {
27495 (i32::MIN, i32::MIN, i64::MIN)
27496 } else {
27497 (i32::MAX, i32::MAX, i64::MAX)
27498 });
27499 }
27500 }
27501 // v7.39 (read01 timestamp.c) — PG's postgres_verbose forms: a leading
27502 // `@` is decorative; a trailing `ago` negates the whole interval.
27503 let mut trimmed = s.trim();
27504 trimmed = trimmed.strip_prefix('@').map_or(trimmed, str::trim);
27505 let mut negate = false;
27506 if let Some(rest) = trimmed
27507 .strip_suffix("ago")
27508 .filter(|r| r.ends_with(char::is_whitespace) || r.is_empty())
27509 {
27510 negate = true;
27511 trimmed = rest.trim();
27512 }
27513 let finish = |v: Option<(i32, i32, i64)>| -> Option<(i32, i32, i64)> {
27514 let (mo, d, us) = v?;
27515 if negate {
27516 Some((mo.checked_neg()?, d.checked_neg()?, us.checked_neg()?))
27517 } else {
27518 Some((mo, d, us))
27519 }
27520 };
27521 let s = trimmed;
27522 // ISO 8601 duration (`P1Y2M3DT4H`) and PG's year-month shorthand (`1-2`)
27523 // are single tokens, not the `<n> <unit>` pair form handled below.
27524 if let Some(rest) = trimmed.strip_prefix(['P', 'p']) {
27525 return finish(parse_iso8601_interval(rest));
27526 }
27527 if !trimmed.contains(char::is_whitespace) && trimmed.contains('-') {
27528 if let Some(iv) = parse_year_month_interval(trimmed) {
27529 return finish(Some(iv));
27530 }
27531 }
27532 // v7.39 (GUC knife 3, differential) — PG accepts a bare number as
27533 // SECONDS: `INTERVAL '0'` = 00:00:00, `INTERVAL '5'` = 00:00:05,
27534 // fractions kept to the microsecond (`'1.5'` = 00:00:01.5).
27535 if !trimmed.is_empty() && !trimmed.contains(char::is_whitespace) {
27536 if let Ok(n) = trimmed.parse::<i64>() {
27537 return finish(Some((0, 0, n.checked_mul(1_000_000)?)));
27538 }
27539 if let Ok(f) = trimmed.parse::<f64>() {
27540 if f.is_finite() {
27541 #[allow(clippy::cast_possible_truncation)]
27542 return finish(Some((0, 0, (f * 1_000_000.0) as i64)));
27543 }
27544 }
27545 }
27546 // v7.39 (round 243) — PG accepts the number and unit run together
27547 // (`'15h 2m 12s'`); split each token at the digit→letter boundary so
27548 // the `<n> <unit>` pair loop below sees them as two.
27549 let raw_parts: Vec<&str> = s.split_whitespace().collect();
27550 let mut parts: Vec<&str> = Vec::with_capacity(raw_parts.len());
27551 for p in raw_parts {
27552 let boundary = p
27553 .char_indices()
27554 .find(|(i, c)| {
27555 *i > 0
27556 && c.is_ascii_alphabetic()
27557 && p[..*i]
27558 .chars()
27559 .all(|d| d.is_ascii_digit() || matches!(d, '.' | '-' | '+'))
27560 && p[..*i].chars().any(|d| d.is_ascii_digit())
27561 })
27562 .map(|(i, _)| i);
27563 match boundary {
27564 Some(i) => {
27565 parts.push(&p[..i]);
27566 parts.push(&p[i..]);
27567 }
27568 None => parts.push(p),
27569 }
27570 }
27571 // A bare clock-time token `HH:MM[:SS[.ffffff]]` carries the time-of-day
27572 // part (PG: `3 days 14:30:45`, or `14:30:45` alone). Extract it; whatever
27573 // remains is the `<n> <unit>` pair form handled below.
27574 let mut clock_us: i64 = 0;
27575 let mut had_clock = false;
27576 if let Some(pos) = parts.iter().position(|p| p.contains(':')) {
27577 clock_us = parse_interval_clock(parts[pos])?;
27578 parts.remove(pos);
27579 had_clock = true;
27580 }
27581 // v7.39 (read01 timestamp.c) — a lone bare number alongside a clock
27582 // time is DAYS (PG: '3 4:05:06' = 3 days 04:05:06).
27583 let mut lone_days: i32 = 0;
27584 if had_clock && parts.len() == 1 {
27585 if let Ok(n) = parts[0].parse::<i64>() {
27586 lone_days = i32::try_from(n).ok()?;
27587 parts.clear();
27588 }
27589 }
27590 if !parts.len().is_multiple_of(2) || (parts.is_empty() && !had_clock && lone_days == 0) {
27591 return None;
27592 }
27593 let mut months: i32 = 0;
27594 let mut days: i32 = lone_days;
27595 let mut micros: i64 = clock_us;
27596 let mut i = 0;
27597 while i < parts.len() {
27598 let unit_stripped = canonical_interval_unit(parts[i + 1])?;
27599 if let Ok(n) = parts[i].parse::<i64>() {
27600 match unit_stripped {
27601 "microsecond" => micros = micros.checked_add(n)?,
27602 "millisecond" => micros = micros.checked_add(n.checked_mul(1_000)?)?,
27603 "second" => micros = micros.checked_add(n.checked_mul(1_000_000)?)?,
27604 "minute" => micros = micros.checked_add(n.checked_mul(60_000_000)?)?,
27605 "hour" => micros = micros.checked_add(n.checked_mul(3_600_000_000)?)?,
27606 "day" => {
27607 let n32 = i32::try_from(n).ok()?;
27608 days = days.checked_add(n32)?;
27609 }
27610 "week" => {
27611 let n32 = i32::try_from(n).ok()?;
27612 days = days.checked_add(n32.checked_mul(7)?)?;
27613 }
27614 "month" => {
27615 let n32 = i32::try_from(n).ok()?;
27616 months = months.checked_add(n32)?;
27617 }
27618 "year" => {
27619 let n32 = i32::try_from(n).ok()?;
27620 months = months.checked_add(n32.checked_mul(12)?)?;
27621 }
27622 // v7.39 (read01 timestamp.c) — the larger calendar units.
27623 "decade" => {
27624 let n32 = i32::try_from(n).ok()?;
27625 months = months.checked_add(n32.checked_mul(120)?)?;
27626 }
27627 "century" => {
27628 let n32 = i32::try_from(n).ok()?;
27629 months = months.checked_add(n32.checked_mul(1200)?)?;
27630 }
27631 "millennium" => {
27632 let n32 = i32::try_from(n).ok()?;
27633 months = months.checked_add(n32.checked_mul(12000)?)?;
27634 }
27635 _ => return None,
27636 }
27637 } else if let Ok(f) = parts[i].parse::<f64>() {
27638 // Fractional units cascade down to the next-finer field the way
27639 // PG does: `1.5 days` -> `1 day 12:00:00`, `1.5 months` ->
27640 // `1 mon 15 days` (30-day month), `1.5 years` -> `1 year 6 mons`.
27641 // no_std: f64 has no trunc/fract/round methods, so do them with
27642 // casts (toward-zero) + explicit round-half-away-from-zero.
27643 #[allow(clippy::cast_possible_truncation)]
27644 fn round_i64(x: f64) -> i64 {
27645 if x >= 0.0 {
27646 (x + 0.5) as i64
27647 } else {
27648 (x - 0.5) as i64
27649 }
27650 }
27651 #[allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
27652 fn add_days_frac(days: &mut i32, micros: &mut i64, d: f64) -> Option<()> {
27653 const DAY_US: f64 = 86_400_000_000.0;
27654 let whole = d as i64; // truncates toward zero
27655 let frac = d - whole as f64;
27656 *days = days.checked_add(i32::try_from(whole).ok()?)?;
27657 *micros = micros.checked_add(round_i64(frac * DAY_US))?;
27658 Some(())
27659 }
27660 #[allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
27661 match unit_stripped {
27662 "microsecond" => micros = micros.checked_add(round_i64(f))?,
27663 "millisecond" => micros = micros.checked_add(round_i64(f * 1_000.0))?,
27664 "second" => micros = micros.checked_add(round_i64(f * 1_000_000.0))?,
27665 "minute" => micros = micros.checked_add(round_i64(f * 60_000_000.0))?,
27666 "hour" => micros = micros.checked_add(round_i64(f * 3_600_000_000.0))?,
27667 "day" => add_days_frac(&mut days, &mut micros, f)?,
27668 "week" => add_days_frac(&mut days, &mut micros, f * 7.0)?,
27669 "month" => {
27670 let whole = f as i64;
27671 months = months.checked_add(i32::try_from(whole).ok()?)?;
27672 add_days_frac(&mut days, &mut micros, (f - whole as f64) * 30.0)?;
27673 }
27674 "year" => {
27675 let m = f * 12.0;
27676 let whole = m as i64;
27677 months = months.checked_add(i32::try_from(whole).ok()?)?;
27678 add_days_frac(&mut days, &mut micros, (m - whole as f64) * 30.0)?;
27679 }
27680 _ => return None,
27681 }
27682 } else {
27683 return None;
27684 }
27685 i += 2;
27686 }
27687 finish(Some((months, days, micros)))
27688}
27689
27690/// v7.37 — map a scalar type keyword to its [`CastTarget`] for the PG
27691/// `TYPE 'literal'` typed-literal syntax (`time '10:30'` == `'10:30'::time`).
27692/// `interval` is intentionally absent (handled by its own parser arm).
27693/// Returns `None` for names that aren't sensible as a bare typed literal, so
27694/// the caller falls back to treating the ident as a column reference.
27695fn typed_literal_cast_target(ident: &str) -> Option<CastTarget> {
27696 Some(match ident {
27697 "date" => CastTarget::Date,
27698 "timestamp" | "datetime" => CastTarget::Timestamp,
27699 "timestamptz" => CastTarget::Timestamptz,
27700 "bool" | "boolean" => CastTarget::Bool,
27701 "int" | "integer" | "int4" => CastTarget::Int,
27702 "bigint" | "int8" => CastTarget::BigInt,
27703 "float8" | "double precision" => CastTarget::Float,
27704 "uuid" => CastTarget::Uuid,
27705 "bytea" => CastTarget::Bytea,
27706 "json" => CastTarget::Json,
27707 "jsonb" => CastTarget::Jsonb,
27708 // Types without a dedicated CastTarget variant flow through the
27709 // generic Named path (engine resolves via column_type_to_data_type).
27710 "time" | "timetz" | "smallint" | "int2" | "numeric" | "decimal"
27711 | "real" | "float4" | "inet" | "cidr" | "macaddr" | "macaddr8"
27712 | "money" | "bit" | "varbit"
27713 // Geometric types accept the `TYPE 'literal'` prefix spelling too.
27714 | "point" | "line" | "lseg" | "box" | "path" | "polygon" | "circle"
27715 // Range / multirange types likewise.
27716 | "int4range" | "int8range" | "numrange" | "daterange" | "tsrange"
27717 | "tstzrange" | "int4multirange" | "int8multirange" | "nummultirange"
27718 | "datemultirange" | "tsmultirange" | "tstzmultirange"
27719 // v7.39 (read01 round 18) — oid / name / jsonpath literal prefixes.
27720 | "oid" | "name" | "jsonpath" | "pg_lsn" | "varchar" | "text" | "xid" | "xid8" => {
27721 CastTarget::Named(alloc::string::String::from(ident))
27722 }
27723 _ => return None,
27724 })
27725}
27726
27727/// v7.12.4 — map a bare type-name identifier (the form that
27728/// appears in a function arg list or RETURNS clause) to a
27729/// [`ColumnTypeName`]. Returns `None` for unknown / extension
27730/// types so the caller can preserve them as
27731/// [`FunctionArgType::Raw`] / [`FunctionReturn::Other`].
27732///
27733/// Subset of the full column-type grammar — we deliberately
27734/// don't parse parameterised forms (`VARCHAR(n)`, `NUMERIC(p,s)`)
27735/// here because function-arg types in v7.12.4 are mostly the
27736/// bare form (`text`, `int`, `bytea`, …).
27737/// v7.39 (round 315, V19) — does this whole phrase name a type, rather
27738/// than being `name TYPE`?
27739///
27740/// The multi-word spellings SQL allows for a bare argument type, each
27741/// verified accepted by PG 18.4 as `CREATE FUNCTION f(<phrase>)`.
27742///
27743/// NOTE this list also exists in `spg-storage`, which computes the
27744/// signature key from the rendered argument text and has to reach the
27745/// same verdict. The two crates are siblings — neither depends on the
27746/// other — and each already carries its own table of type spellings
27747/// (`map_type_ident_to_column_type_name` here, `normalize_type_name`
27748/// there), so this follows the structure rather than inventing new
27749/// duplication. Recorded as V49.
27750pub fn is_multiword_type_phrase(phrase: &str) -> bool {
27751 let t = phrase.trim().to_ascii_lowercase();
27752 let base = t.split_once('(').map_or(t.as_str(), |(h, _)| h).trim();
27753 matches!(
27754 base,
27755 "double precision"
27756 | "character varying"
27757 | "bit varying"
27758 | "timestamp with time zone"
27759 | "timestamp without time zone"
27760 | "time with time zone"
27761 | "time without time zone"
27762 | "national character"
27763 | "national character varying"
27764 )
27765}
27766
27767fn map_type_ident_to_column_type_name(ident: &str) -> Option<ColumnTypeName> {
27768 Some(match ident.to_ascii_lowercase().as_str() {
27769 "smallint" | "tinyint" => ColumnTypeName::SmallInt,
27770 "int" | "integer" | "mediumint" => ColumnTypeName::Int,
27771 "bigint" => ColumnTypeName::BigInt,
27772 "float" | "double" => ColumnTypeName::Float,
27773 // v7.39 (round 269) — real is 32-bit.
27774 "real" | "float4" => ColumnTypeName::Real,
27775 "text" => ColumnTypeName::Text,
27776 "bool" | "boolean" => ColumnTypeName::Bool,
27777 "date" => ColumnTypeName::Date,
27778 "timestamp" | "datetime" => ColumnTypeName::Timestamp,
27779 "timestamptz" => ColumnTypeName::Timestamptz,
27780 "json" => ColumnTypeName::Json,
27781 "jsonb" => ColumnTypeName::Jsonb,
27782 "bytea" | "bytes" => ColumnTypeName::Bytes,
27783 "tsvector" => ColumnTypeName::TsVector,
27784 "tsquery" => ColumnTypeName::TsQuery,
27785 "uuid" => ColumnTypeName::Uuid,
27786 "interval" => ColumnTypeName::Interval,
27787 "time" => ColumnTypeName::Time,
27788 "year" => ColumnTypeName::Year,
27789 "timetz" => ColumnTypeName::TimeTz,
27790 "money" => ColumnTypeName::Money,
27791 _ => return None,
27792 })
27793}
27794
27795/// v7.12.4 — parse a PL/pgSQL function body (the bytes between
27796/// `$$ ... $$`). Returns the parsed `BEGIN ... END;` block.
27797///
27798/// v7.12.4 grammar (strict subset — IF / LOOP / DECLARE / RAISE
27799/// / embedded SQL land in v7.12.5+):
27800///
27801/// ```text
27802/// body := [ws] block [ws]
27803/// block := BEGIN stmt ( ; stmt )* [ ; ] END [ ; ]
27804/// stmt := assign | return
27805/// assign := assign_target := expr
27806/// assign_target := ( NEW | OLD ) . ident | ident
27807/// return := RETURN ( NEW | OLD | NULL | expr )
27808/// ```
27809///
27810/// `expr` is parsed by recursing into the regular `Parser` — so a
27811/// PL/pgSQL `NEW.search_vector := to_tsvector('english',
27812/// NEW.subject || ' ' || NEW.sender)` body shape works without
27813/// the body parser knowing what `to_tsvector` is.
27814///
27815/// Errors here cause the caller to fall back to
27816/// `FunctionBody::Raw` — keeping the CREATE FUNCTION DDL itself
27817/// successful, but the executor will refuse to invoke the
27818/// function with an "unparseable body" error.
27819/// v7.12.4 — public alias for [`parse_plpgsql_body`] re-exported
27820/// from the crate root as `spg_sql::parse_function_body`.
27821pub fn parse_function_body(body: &str) -> Result<PlPgSqlBlock, ParseError> {
27822 parse_plpgsql_body(body)
27823}
27824
27825fn parse_plpgsql_body(body: &str) -> Result<PlPgSqlBlock, ParseError> {
27826 // Use the regular lexer on the body text. The trailing
27827 // `END;` may or may not have a semicolon; the lexer treats
27828 // both forms identically.
27829 let tokens = lexer::tokenize(body).map_err(|e| ParseError {
27830 message: alloc::format!("plpgsql body lex error: {e}"),
27831 token_pos: 0,
27832 })?;
27833 let mut parser = Parser::new(tokens);
27834 parser.parse_plpgsql_block()
27835}
27836
27837/// v7.39 (GUC) — the textual body of a SET value, for list joining.
27838fn set_value_text(v: &crate::ast::SetValue) -> alloc::string::String {
27839 match v {
27840 crate::ast::SetValue::String(s)
27841 | crate::ast::SetValue::Ident(s)
27842 | crate::ast::SetValue::Number(s) => s.clone(),
27843 crate::ast::SetValue::Default => "DEFAULT".into(),
27844 }
27845}
27846
27847/// v7.39 (round 145, parse_cte.c / parse_agg.c) — true when an expression
27848/// contains an aggregate call at ITS OWN query level (recursion stops at
27849/// sublink boundaries — a sublink's aggregates belong to the sublink).
27850/// Backs the "aggregate functions are not allowed in a recursive query's
27851/// recursive term" well-formedness check.
27852fn expr_has_toplevel_aggregate(e: &Expr) -> bool {
27853 const AGG_NAMES: &[&str] = &[
27854 "count",
27855 "sum",
27856 "min",
27857 "max",
27858 "avg",
27859 "string_agg",
27860 "array_agg",
27861 "bool_and",
27862 "bool_or",
27863 "every",
27864 "any_value",
27865 "json_agg",
27866 "jsonb_agg",
27867 "json_object_agg",
27868 "jsonb_object_agg",
27869 "bit_and",
27870 "bit_or",
27871 "bit_xor",
27872 "var_pop",
27873 "var_samp",
27874 "variance",
27875 "stddev",
27876 "stddev_pop",
27877 "stddev_samp",
27878 "range_agg",
27879 "range_intersect_agg",
27880 "percentile_cont",
27881 "percentile_disc",
27882 "mode",
27883 "corr",
27884 "covar_pop",
27885 "covar_samp",
27886 ];
27887 match e {
27888 Expr::AggregateOrdered { .. } => true,
27889 Expr::FunctionCall { name, args } => {
27890 AGG_NAMES.contains(&name.to_ascii_lowercase().as_str())
27891 || args.iter().any(expr_has_toplevel_aggregate)
27892 }
27893 Expr::NamedArg { expr, .. }
27894 | Expr::Variadic(expr)
27895 | Expr::Unary { expr, .. }
27896 | Expr::Cast { expr, .. }
27897 | Expr::IsNull { expr, .. }
27898 | Expr::FieldAccess { base: expr, .. }
27899 | Expr::Extract { source: expr, .. } => expr_has_toplevel_aggregate(expr),
27900 Expr::Binary { lhs, rhs, .. } => {
27901 expr_has_toplevel_aggregate(lhs) || expr_has_toplevel_aggregate(rhs)
27902 }
27903 Expr::Like { expr, pattern, .. } => {
27904 expr_has_toplevel_aggregate(expr) || expr_has_toplevel_aggregate(pattern)
27905 }
27906 Expr::Array(items) => items.iter().any(expr_has_toplevel_aggregate),
27907 Expr::InList { expr, list, .. } => {
27908 expr_has_toplevel_aggregate(expr) || list.iter().any(expr_has_toplevel_aggregate)
27909 }
27910 Expr::ArraySubscript { target, index } => {
27911 expr_has_toplevel_aggregate(target) || expr_has_toplevel_aggregate(index)
27912 }
27913 Expr::ArraySlice { target, lo, hi } => {
27914 expr_has_toplevel_aggregate(target)
27915 || lo.as_deref().is_some_and(expr_has_toplevel_aggregate)
27916 || hi.as_deref().is_some_and(expr_has_toplevel_aggregate)
27917 }
27918 Expr::AnyAll { expr, array, .. } => {
27919 expr_has_toplevel_aggregate(expr) || expr_has_toplevel_aggregate(array)
27920 }
27921 Expr::Case {
27922 operand,
27923 branches,
27924 else_branch,
27925 } => {
27926 operand.as_deref().is_some_and(expr_has_toplevel_aggregate)
27927 || branches
27928 .iter()
27929 .any(|(w, t)| expr_has_toplevel_aggregate(w) || expr_has_toplevel_aggregate(t))
27930 || else_branch
27931 .as_deref()
27932 .is_some_and(expr_has_toplevel_aggregate)
27933 }
27934 // The outer-level operands of a sublink can aggregate; the sublink's
27935 // own body cannot leak its aggregates up here.
27936 Expr::InSubquery { expr, .. } => expr_has_toplevel_aggregate(expr),
27937 Expr::RowInSubquery { row, .. } | Expr::RowCmpSubquery { row, .. } => {
27938 row.iter().any(expr_has_toplevel_aggregate)
27939 }
27940 _ => false,
27941 }
27942}
27943
27944/// v7.39 (round 145, parse_cte.c) — true when any sublink expression
27945/// (EXISTS / IN / scalar subquery) inside this SELECT term references the
27946/// named table anywhere in its subtree. A plain FROM derived table is NOT a
27947/// sublink and is legal in a recursive term, so it is not walked here.
27948fn select_has_self_ref_in_sublink(s: &crate::ast::SelectStatement, name: &str) -> bool {
27949 let mut exprs: Vec<&Expr> = Vec::new();
27950 for it in &s.items {
27951 if let crate::ast::SelectItem::Expr { expr, .. } = it {
27952 exprs.push(expr);
27953 }
27954 }
27955 if let Some(w) = &s.where_ {
27956 exprs.push(w);
27957 }
27958 if let Some(h) = &s.having {
27959 exprs.push(h);
27960 }
27961 if let Some(g) = &s.group_by {
27962 exprs.extend(g.iter());
27963 }
27964 if let Some(from) = &s.from {
27965 for j in &from.joins {
27966 if let Some(on) = &j.on {
27967 exprs.push(on);
27968 }
27969 }
27970 }
27971 exprs.into_iter().any(|e| expr_sublink_mentions(e, name))
27972}
27973
27974/// Does this expression contain a sublink whose subquery mentions `name`?
27975fn expr_sublink_mentions(e: &Expr, name: &str) -> bool {
27976 match e {
27977 Expr::ScalarSubquery(sub) => select_mentions_table(sub, name),
27978 Expr::Exists { subquery, .. } => select_mentions_table(subquery, name),
27979 Expr::InSubquery { expr, subquery, .. } => {
27980 expr_sublink_mentions(expr, name) || select_mentions_table(subquery, name)
27981 }
27982 Expr::RowInSubquery { row, subquery, .. } => {
27983 row.iter().any(|x| expr_sublink_mentions(x, name))
27984 || select_mentions_table(subquery, name)
27985 }
27986 Expr::RowCmpSubquery { row, subquery, .. } => {
27987 row.iter().any(|x| expr_sublink_mentions(x, name))
27988 || select_mentions_table(subquery, name)
27989 }
27990 Expr::NamedArg { expr, .. }
27991 | Expr::Variadic(expr)
27992 | Expr::Unary { expr, .. }
27993 | Expr::Cast { expr, .. }
27994 | Expr::IsNull { expr, .. }
27995 | Expr::FieldAccess { base: expr, .. }
27996 | Expr::Extract { source: expr, .. } => expr_sublink_mentions(expr, name),
27997 Expr::Binary { lhs, rhs, .. } => {
27998 expr_sublink_mentions(lhs, name) || expr_sublink_mentions(rhs, name)
27999 }
28000 Expr::Like { expr, pattern, .. } => {
28001 expr_sublink_mentions(expr, name) || expr_sublink_mentions(pattern, name)
28002 }
28003 Expr::FunctionCall { args, .. } | Expr::Array(args) => {
28004 args.iter().any(|x| expr_sublink_mentions(x, name))
28005 }
28006 Expr::InList { expr, list, .. } => {
28007 expr_sublink_mentions(expr, name) || list.iter().any(|x| expr_sublink_mentions(x, name))
28008 }
28009 Expr::ArraySubscript { target, index } => {
28010 expr_sublink_mentions(target, name) || expr_sublink_mentions(index, name)
28011 }
28012 Expr::ArraySlice { target, lo, hi } => {
28013 expr_sublink_mentions(target, name)
28014 || lo
28015 .as_deref()
28016 .is_some_and(|x| expr_sublink_mentions(x, name))
28017 || hi
28018 .as_deref()
28019 .is_some_and(|x| expr_sublink_mentions(x, name))
28020 }
28021 Expr::AnyAll { expr, array, .. } => {
28022 expr_sublink_mentions(expr, name) || expr_sublink_mentions(array, name)
28023 }
28024 Expr::Case {
28025 operand,
28026 branches,
28027 else_branch,
28028 } => {
28029 operand
28030 .as_deref()
28031 .is_some_and(|x| expr_sublink_mentions(x, name))
28032 || branches
28033 .iter()
28034 .any(|(w, t)| expr_sublink_mentions(w, name) || expr_sublink_mentions(t, name))
28035 || else_branch
28036 .as_deref()
28037 .is_some_and(|x| expr_sublink_mentions(x, name))
28038 }
28039 _ => false,
28040 }
28041}
28042
28043/// Does this SELECT (in full — FROM tables, derived tables, its own
28044/// sublinks, and union arms) mention the named table?
28045fn select_mentions_table(s: &crate::ast::SelectStatement, name: &str) -> bool {
28046 if let Some(from) = &s.from {
28047 if from.primary.name.eq_ignore_ascii_case(name) {
28048 return true;
28049 }
28050 if let Some(sub) = &from.primary.lateral_subquery
28051 && select_mentions_table(sub, name)
28052 {
28053 return true;
28054 }
28055 for j in &from.joins {
28056 if j.table.name.eq_ignore_ascii_case(name) {
28057 return true;
28058 }
28059 if let Some(sub) = &j.table.lateral_subquery
28060 && select_mentions_table(sub, name)
28061 {
28062 return true;
28063 }
28064 }
28065 }
28066 if select_has_self_ref_in_sublink(s, name) {
28067 return true;
28068 }
28069 s.unions.iter().any(|(_, u)| select_mentions_table(u, name))
28070}
28071
28072/// v7.39 (round 284) — fold a constant `LIMIT` / `OFFSET` expression to a
28073/// row count, the way PG evaluates one before applying it.
28074///
28075/// `None` = not a constant (a column, a subquery, a function call).
28076/// `Some(Err(msg))` = PG rejects it, and the message is PG's; `{L}` in the
28077/// message stands in for LIMIT / OFFSET, which the caller substitutes.
28078/// All wordings were read off live PG 18.4.
28079fn fold_limit_constant(e: &crate::ast::Expr) -> Option<Result<i128, alloc::string::String>> {
28080 use crate::ast::{BinOp, Expr, Literal, UnOp};
28081 match e {
28082 Expr::Literal(Literal::Integer(n)) => Some(Ok(i128::from(*n))),
28083 Expr::Literal(Literal::Numeric { unscaled, scale }) => {
28084 Some(Ok(round_scaled_half_away(*unscaled, *scale)))
28085 }
28086 // PG coerces a string by its CONTENT, and fails on the value.
28087 Expr::Literal(Literal::String(t)) => Some(t.trim().parse::<i64>().map_or_else(
28088 |_| {
28089 Err(alloc::format!(
28090 "invalid input syntax for type bigint: \"{t}\""
28091 ))
28092 },
28093 |n| Ok(i128::from(n)),
28094 )),
28095 Expr::Literal(Literal::Bool(_)) => Some(Err(
28096 "argument of {L} must be type bigint, not type boolean".into(),
28097 )),
28098 Expr::Unary {
28099 op: UnOp::Neg,
28100 expr,
28101 } => match fold_limit_constant(expr)? {
28102 Ok(v) => Some(Ok(-v)),
28103 e @ Err(_) => Some(e),
28104 },
28105 Expr::Binary { lhs, op, rhs } => {
28106 let a = match fold_limit_constant(lhs)? {
28107 Ok(v) => v,
28108 e @ Err(_) => return Some(e),
28109 };
28110 let b = match fold_limit_constant(rhs)? {
28111 Ok(v) => v,
28112 e @ Err(_) => return Some(e),
28113 };
28114 let out = match op {
28115 BinOp::Add => a.checked_add(b),
28116 BinOp::Sub => a.checked_sub(b),
28117 BinOp::Mul => a.checked_mul(b),
28118 BinOp::Div if b != 0 => a.checked_div(b),
28119 BinOp::Div => return Some(Err("division by zero".into())),
28120 BinOp::Mod if b != 0 => a.checked_rem(b),
28121 BinOp::Mod => return Some(Err("division by zero".into())),
28122 _ => return None,
28123 };
28124 // PG evaluates the arithmetic in the operand's own type, so an
28125 // int-by-int product that leaves int range fails there — before
28126 // the row count is ever looked at.
28127 match out {
28128 Some(v) if v > i128::from(i32::MAX) || v < i128::from(i32::MIN) => {
28129 Some(Err("integer out of range".into()))
28130 }
28131 Some(v) => Some(Ok(v)),
28132 None => Some(Err("integer out of range".into())),
28133 }
28134 }
28135 _ => None,
28136 }
28137}
28138
28139/// Round `unscaled / 10^scale` half away from zero — PG's numeric→bigint
28140/// cast, which is what makes `LIMIT 2.5` keep three rows.
28141fn round_scaled_half_away(unscaled: i128, scale: u16) -> i128 {
28142 if scale == 0 {
28143 return unscaled;
28144 }
28145 let Some(div) = 10i128.checked_pow(u32::from(scale)) else {
28146 return 0;
28147 };
28148 let neg = unscaled < 0;
28149 let mag = unscaled.unsigned_abs() as i128;
28150 let rounded = (mag + div / 2) / div;
28151 if neg { -rounded } else { rounded }
28152}
28153
28154#[cfg(test)]
28155mod tests {
28156 use super::*;
28157 use alloc::string::ToString;
28158
28159 fn parse(s: &str) -> Statement {
28160 parse_statement(s).expect("parse ok")
28161 }
28162
28163 // v7.37.43-T4 sentori cutover acceptance — `release`, `index`,
28164 // `tables`, `partition`, etc. are unreserved keywords per PG's
28165 // `pg_get_keywords()` and MUST be usable as column / table /
28166 // alias names. Pre-T4 every drop-in user whose schema had one
28167 // of these as a column name (sentori events.release, mailrs
28168 // messages.index in some forks) blew the parser up at CREATE
28169 // TABLE time with "expected identifier, got Release". The
28170 // generalisation lives in `unreserved_keyword_text` + the
28171 // `expect_ident_like` and `parse_atom` arms that consult it.
28172 #[test]
28173 fn release_usable_as_column_name_in_create_table() {
28174 let stmt =
28175 parse("CREATE TABLE events (id INT PRIMARY KEY, release TEXT NOT NULL, payload TEXT)");
28176 if let Statement::CreateTable(t) = stmt {
28177 let names: alloc::vec::Vec<&str> = t.columns.iter().map(|c| c.name.as_str()).collect();
28178 assert_eq!(names, alloc::vec!["id", "release", "payload"]);
28179 } else {
28180 panic!("expected CreateTable");
28181 }
28182 }
28183
28184 #[test]
28185 fn release_usable_as_column_ref_in_select_projection() {
28186 // The sentori `0003_partition_events.sql` INSERT-SELECT
28187 // walk references `release` in both column lists; the
28188 // projection-side use exercises `parse_atom`'s relaxed
28189 // identifier set.
28190 parse("SELECT id, release, payload FROM events WHERE id = 1");
28191 }
28192
28193 #[test]
28194 fn release_usable_as_column_ref_in_insert_column_list() {
28195 // INSERT INTO t (id, release, payload) VALUES (…)
28196 parse("INSERT INTO events (id, release, payload) VALUES (1, '1.0.0', 'data')");
28197 }
28198
28199 #[test]
28200 fn alter_column_drop_not_null_uses_keyword_drop_token() {
28201 // Sentori `0013_audit_tombstone.sql` issues
28202 // `ALTER TABLE … ALTER COLUMN x DROP NOT NULL`. The lexer
28203 // emits Token::Drop (not Ident("drop")); the parser must
28204 // accept both in the ALTER COLUMN sub-dispatch.
28205 parse("ALTER TABLE audit_logs ALTER COLUMN org_id DROP NOT NULL");
28206 }
28207
28208 #[test]
28209 fn create_index_accepts_parenthesised_expression_key() {
28210 // sentori `0040_events_bundle_idx.sql` shape — JSONB
28211 // expression index. Pre-T4 the parser bailed at the
28212 // inner `(` with "expected column ident or expression,
28213 // got LParen". The Token::LParen arm in CREATE INDEX
28214 // routes through the expression parser instead.
28215 parse(
28216 "CREATE INDEX IF NOT EXISTS events_bundle_id_idx \
28217 ON events ((payload->'bundle'->>'id'))",
28218 );
28219 }
28220
28221 // v7.30.2 (mailrs round-25 ask 2) — nesting / chain budgets must
28222 // surface as parse errors, never stack overflows (embed hosts
28223 // abort on overflow).
28224 /// The nesting budget is a COUNT; what it has to fit inside is a
28225 /// number of BYTES, and only one of those two is stable across
28226 /// compiler versions. Round 847 measured 30,336 bytes per level
28227 /// after a toolchain move, which puts 64 levels at 1.94 MB and
28228 /// overflows a 2 MiB thread — `nesting_budget_errors_cleanly`
28229 /// aborted instead of erroring, which is precisely the outcome it
28230 /// exists to rule out.
28231 ///
28232 /// So the budget is metered rather than assumed. The ceiling leaves
28233 /// the depth SPG advertises fitting in a default 2 MiB thread with
28234 /// room to spare, in the debug build, where frames are widest.
28235 #[test]
28236 fn nesting_frame_cost_stays_under_ceiling() {
28237 // Room for MAX_NEST_DEPTH levels inside 1.2 MB, so a 2 MiB
28238 // thread keeps a margin for whatever called the parser.
28239 const CEILING: usize = 1_200_000 / MAX_NEST_DEPTH;
28240
28241 frame_meter::reset();
28242 let depth = frame_meter::SAMPLE_HI + 8;
28243 let sql = format!("SELECT {}1{}", "(".repeat(depth), ")".repeat(depth));
28244 parse(&sql);
28245
28246 let per_level = frame_meter::bytes_per_level();
28247 {
28248 extern crate std;
28249 std::eprintln!("nesting frame: {per_level} bytes/level, ceiling {CEILING}");
28250 }
28251 assert!(
28252 per_level <= CEILING,
28253 "{per_level} bytes per nesting level exceeds {CEILING}; \
28254 {MAX_NEST_DEPTH} levels would want {} bytes. Out-line arms \
28255 in parse_expr_inner / parse_unary rather than lowering the \
28256 depth or widening the stack.",
28257 per_level * MAX_NEST_DEPTH
28258 );
28259 }
28260
28261 #[test]
28262 fn nesting_budget_errors_cleanly() {
28263 let depth = MAX_NEST_DEPTH + 50;
28264 let sql = format!("SELECT {}1{}", "(".repeat(depth), ")".repeat(depth));
28265 let err = parse_statement(&sql).expect_err("must reject");
28266 assert!(err.message.contains("nests deeper"), "{err:?}");
28267 // Within budget still parses.
28268 let sql = format!("SELECT {}1{}", "(".repeat(48), ")".repeat(48));
28269 parse(&sql);
28270 }
28271
28272 #[test]
28273 fn binary_chain_budget_errors_cleanly() {
28274 let sql = format!("SELECT 1{}", " + 1".repeat(MAX_BINARY_CHAIN + 50));
28275 let err = parse_statement(&sql).expect_err("must reject");
28276 assert!(err.message.contains("chained binary"), "{err:?}");
28277 // Within budget still parses (chain depth ≤ budget is safe
28278 // for recursive eval/drop on 2 MiB stacks).
28279 let sql = format!("SELECT 1{}", " + 1".repeat(200));
28280 parse(&sql);
28281 }
28282
28283 #[test]
28284 fn in_list_unaffected_by_chain_budget() {
28285 // Flat InList: 20k elements parse fine and stay flat.
28286 let items: alloc::vec::Vec<String> = (0..20_000).map(|k| k.to_string()).collect();
28287 let sql = format!("SELECT 1 WHERE 5 IN ({})", items.join(","));
28288 let Statement::Select(s) = parse(&sql) else {
28289 panic!("expected select")
28290 };
28291 let Some(Expr::InList { list, negated, .. }) = s.where_ else {
28292 panic!("expected flat InList, got {:?}", s.where_)
28293 };
28294 assert_eq!(list.len(), 20_000);
28295 assert!(!negated);
28296 }
28297
28298 fn lit_int(n: i64) -> Expr {
28299 Expr::Literal(Literal::Integer(n))
28300 }
28301
28302 fn col(name: &str) -> Expr {
28303 Expr::Column(ColumnName {
28304 qualifier: None,
28305 name: name.into(),
28306 })
28307 }
28308
28309 #[test]
28310 fn select_single_integer() {
28311 let s = parse("SELECT 1");
28312 let Statement::Select(s) = s else {
28313 panic!("expected SELECT")
28314 };
28315 assert_eq!(s.items.len(), 1);
28316 assert!(s.from.is_none());
28317 assert!(s.where_.is_none());
28318 }
28319
28320 #[test]
28321 fn select_multiple_literal_kinds() {
28322 let s = parse("SELECT 1, 'hi', NULL, TRUE, 1.5");
28323 let Statement::Select(s) = s else {
28324 panic!("expected SELECT")
28325 };
28326 assert_eq!(s.items.len(), 5);
28327 }
28328
28329 #[test]
28330 fn select_wildcard_from_table() {
28331 let s = parse("SELECT * FROM users");
28332 let Statement::Select(s) = s else {
28333 panic!("expected SELECT")
28334 };
28335 assert!(matches!(s.items[..], [SelectItem::Wildcard]));
28336 assert_eq!(s.from.as_ref().unwrap().primary.name, "users");
28337 }
28338
28339 #[test]
28340 fn select_with_table_alias() {
28341 let s = parse("SELECT * FROM users AS u");
28342 let Statement::Select(s) = s else {
28343 panic!("expected SELECT")
28344 };
28345 let t = &s.from.as_ref().unwrap().primary;
28346 assert_eq!(t.name, "users");
28347 assert_eq!(t.alias.as_deref(), Some("u"));
28348 }
28349
28350 #[test]
28351 fn select_with_where_eq() {
28352 let s = parse("SELECT a FROM t WHERE a = 1");
28353 let Statement::Select(s) = s else {
28354 panic!("expected SELECT")
28355 };
28356 let w = s.where_.unwrap();
28357 assert_eq!(
28358 w,
28359 Expr::Binary {
28360 lhs: Box::new(col("a")),
28361 op: BinOp::Eq,
28362 rhs: Box::new(lit_int(1)),
28363 }
28364 );
28365 }
28366
28367 #[test]
28368 fn arithmetic_precedence() {
28369 let s = parse("SELECT 1 + 2 * 3");
28370 let Statement::Select(s) = s else {
28371 panic!("expected SELECT")
28372 };
28373 let SelectItem::Expr { expr, .. } = &s.items[0] else {
28374 panic!("wildcard?")
28375 };
28376 assert_eq!(
28377 expr,
28378 &Expr::Binary {
28379 lhs: Box::new(lit_int(1)),
28380 op: BinOp::Add,
28381 rhs: Box::new(Expr::Binary {
28382 lhs: Box::new(lit_int(2)),
28383 op: BinOp::Mul,
28384 rhs: Box::new(lit_int(3)),
28385 }),
28386 }
28387 );
28388 }
28389
28390 #[test]
28391 fn parentheses_override_precedence() {
28392 let s = parse("SELECT (1 + 2) * 3");
28393 let Statement::Select(s) = s else {
28394 panic!("expected SELECT")
28395 };
28396 let SelectItem::Expr { expr, .. } = &s.items[0] else {
28397 panic!()
28398 };
28399 assert_eq!(
28400 expr,
28401 &Expr::Binary {
28402 lhs: Box::new(Expr::Binary {
28403 lhs: Box::new(lit_int(1)),
28404 op: BinOp::Add,
28405 rhs: Box::new(lit_int(2)),
28406 }),
28407 op: BinOp::Mul,
28408 rhs: Box::new(lit_int(3)),
28409 }
28410 );
28411 }
28412
28413 #[test]
28414 fn not_binds_below_comparison() {
28415 // `NOT a = 1` should parse as `NOT (a = 1)`.
28416 let s = parse("SELECT NOT a = 1 FROM t");
28417 let Statement::Select(s) = s else {
28418 panic!("expected SELECT")
28419 };
28420 let SelectItem::Expr { expr, .. } = &s.items[0] else {
28421 panic!()
28422 };
28423 assert_eq!(
28424 expr,
28425 &Expr::Unary {
28426 op: UnOp::Not,
28427 expr: Box::new(Expr::Binary {
28428 lhs: Box::new(col("a")),
28429 op: BinOp::Eq,
28430 rhs: Box::new(lit_int(1)),
28431 }),
28432 }
28433 );
28434 }
28435
28436 #[test]
28437 fn unary_minus_binds_above_multiplication() {
28438 // `-a * 2` should be `(-a) * 2`.
28439 let s = parse("SELECT -a * 2 FROM t");
28440 let Statement::Select(s) = s else {
28441 panic!("expected SELECT")
28442 };
28443 let SelectItem::Expr { expr, .. } = &s.items[0] else {
28444 panic!()
28445 };
28446 assert_eq!(
28447 expr,
28448 &Expr::Binary {
28449 lhs: Box::new(Expr::Unary {
28450 op: UnOp::Neg,
28451 expr: Box::new(col("a")),
28452 }),
28453 op: BinOp::Mul,
28454 rhs: Box::new(lit_int(2)),
28455 }
28456 );
28457 }
28458
28459 #[test]
28460 fn qualified_column() {
28461 let s = parse("SELECT t.col FROM t");
28462 let Statement::Select(s) = s else {
28463 panic!("expected SELECT")
28464 };
28465 let SelectItem::Expr { expr, .. } = &s.items[0] else {
28466 panic!()
28467 };
28468 assert_eq!(
28469 expr,
28470 &Expr::Column(ColumnName {
28471 qualifier: Some("t".into()),
28472 name: "col".into()
28473 })
28474 );
28475 }
28476
28477 #[test]
28478 fn select_item_alias_with_as() {
28479 let s = parse("SELECT a AS y FROM t");
28480 let Statement::Select(s) = s else {
28481 panic!("expected SELECT")
28482 };
28483 let SelectItem::Expr { alias, .. } = &s.items[0] else {
28484 panic!()
28485 };
28486 assert_eq!(alias.as_deref(), Some("y"));
28487 }
28488
28489 #[test]
28490 fn trailing_semicolon_accepted() {
28491 let s = parse("SELECT 1;");
28492 let Statement::Select(s) = s else {
28493 panic!("expected SELECT")
28494 };
28495 assert_eq!(s.items.len(), 1);
28496 }
28497
28498 #[test]
28499 fn boolean_chain_with_and_or_not() {
28500 // (NOT a) OR (b AND (NOT c))
28501 let s = parse("SELECT NOT a OR b AND NOT c FROM t");
28502 let Statement::Select(s) = s else {
28503 panic!("expected SELECT")
28504 };
28505 let SelectItem::Expr { expr, .. } = &s.items[0] else {
28506 panic!()
28507 };
28508 let expected = Expr::Binary {
28509 lhs: Box::new(Expr::Unary {
28510 op: UnOp::Not,
28511 expr: Box::new(col("a")),
28512 }),
28513 op: BinOp::Or,
28514 rhs: Box::new(Expr::Binary {
28515 lhs: Box::new(col("b")),
28516 op: BinOp::And,
28517 rhs: Box::new(Expr::Unary {
28518 op: UnOp::Not,
28519 expr: Box::new(col("c")),
28520 }),
28521 }),
28522 };
28523 assert_eq!(expr, &expected);
28524 }
28525
28526 #[test]
28527 fn empty_input_errors() {
28528 // v7.14.0 — pg_dump preambles emit several comment-only
28529 // / blank-line statements that collapse to Statement::
28530 // Empty rather than a parse error. The old "SELECT in
28531 // message" assertion is stale; verify the new contract:
28532 // empty / whitespace / comment-only input parses to
28533 // Statement::Empty.
28534 assert!(matches!(parse_statement("").unwrap(), Statement::Empty));
28535 assert!(matches!(
28536 parse_statement(" \n\t ").unwrap(),
28537 Statement::Empty
28538 ));
28539 // Sanity: malformed-but-non-empty still errors.
28540 assert!(parse_statement("SELECT FROM WHERE").is_err());
28541 }
28542
28543 #[test]
28544 fn unmatched_paren_errors() {
28545 assert!(parse_statement("SELECT (1 + 2").is_err());
28546 }
28547
28548 #[test]
28549 fn display_round_trip_simple_select() {
28550 let original = parse("SELECT a + 1 FROM t WHERE a > 0");
28551 let text = original.to_string();
28552 let again = parse_statement(&text).expect("re-parse");
28553 assert_eq!(original, again);
28554 }
28555
28556 // --- CREATE TABLE & INSERT (v0.3) ---------------------------------------
28557
28558 #[test]
28559 fn create_table_single_column() {
28560 let s = parse("CREATE TABLE foo (a INT)");
28561 let Statement::CreateTable(c) = s else {
28562 panic!("expected CreateTable")
28563 };
28564 assert_eq!(c.name, "foo");
28565 assert_eq!(c.columns.len(), 1);
28566 assert_eq!(c.columns[0].name, "a");
28567 assert_eq!(c.columns[0].ty, ColumnTypeName::Int);
28568 assert!(c.columns[0].nullable);
28569 }
28570
28571 #[test]
28572 fn create_table_multi_column_with_not_null_mix() {
28573 let s = parse("CREATE TABLE u (id INT NOT NULL, name TEXT, score FLOAT NOT NULL, ok BOOL)");
28574 let Statement::CreateTable(c) = s else {
28575 panic!()
28576 };
28577 assert_eq!(c.columns.len(), 4);
28578 assert_eq!(c.columns[0].ty, ColumnTypeName::Int);
28579 assert!(!c.columns[0].nullable);
28580 assert_eq!(c.columns[1].ty, ColumnTypeName::Text);
28581 assert!(c.columns[1].nullable);
28582 assert_eq!(c.columns[2].ty, ColumnTypeName::Float);
28583 assert!(!c.columns[2].nullable);
28584 assert_eq!(c.columns[3].ty, ColumnTypeName::Bool);
28585 }
28586
28587 #[test]
28588 fn create_table_bigint_supported() {
28589 let s = parse("CREATE TABLE accounts (id BIGINT NOT NULL)");
28590 let Statement::CreateTable(c) = s else {
28591 panic!()
28592 };
28593 assert_eq!(c.columns[0].ty, ColumnTypeName::BigInt);
28594 }
28595
28596 #[test]
28597 fn create_table_vector_default_is_f32() {
28598 let s = parse("CREATE TABLE t (v VECTOR(128))");
28599 let Statement::CreateTable(c) = s else {
28600 panic!()
28601 };
28602 assert_eq!(
28603 c.columns[0].ty,
28604 ColumnTypeName::Vector {
28605 dim: 128,
28606 encoding: VecEncoding::F32,
28607 },
28608 );
28609 }
28610
28611 #[test]
28612 fn create_table_vector_using_sq8() {
28613 // v6.0.1: `USING SQ8` selects scalar-quantised encoding.
28614 // Case-insensitive on both `USING` and the encoding name.
28615 for sql in [
28616 "CREATE TABLE t (v VECTOR(128) USING SQ8)",
28617 "CREATE TABLE t (v VECTOR(128) using sq8)",
28618 ] {
28619 let s = parse(sql);
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::Sq8,
28628 },
28629 "{sql}",
28630 );
28631 }
28632 }
28633
28634 #[test]
28635 fn create_table_vector_using_unknown_errors() {
28636 // v7.16.1 — the inline `USING <encoding>` shape on
28637 // CREATE TABLE column defs was withdrawn before
28638 // v7.14.0 in favour of `CREATE INDEX … USING hnsw
28639 // (col vector_<metric>_ops)`; the parser now rejects
28640 // USING at column-list position with a clearer
28641 // "expected ',' or ')'" message. Test asserts the
28642 // current rejection, not the old "unknown vector
28643 // encoding" string.
28644 let err = parse_statement("CREATE TABLE t (v VECTOR(8) USING PQ8)").unwrap_err();
28645 assert!(
28646 err.message.contains("USING")
28647 || err.message.contains("using")
28648 || err.message.contains("')'")
28649 || err.message.contains("','"),
28650 "expected USING/column-list rejection, got: {}",
28651 err.message
28652 );
28653 }
28654
28655 #[test]
28656 fn vector_using_sq8_display_roundtrips() {
28657 // The Display impl must produce text that re-parses to the
28658 // same AST. Guard for the v6.0.1 `USING SQ8` suffix.
28659 let s = parse("CREATE TABLE t (v VECTOR(64) USING SQ8)");
28660 let Statement::CreateTable(c) = s else {
28661 panic!()
28662 };
28663 assert_eq!(c.columns[0].ty.to_string(), "VECTOR(64) USING SQ8");
28664 }
28665
28666 #[test]
28667 fn parser_recognises_placeholders() {
28668 use crate::ast::{Expr, SelectItem, Statement};
28669 // $N in expression position parses as Expr::Placeholder(N).
28670 let s = parse("SELECT $1, $2 + 1 FROM t WHERE x = $3");
28671 let Statement::Select(sel) = s else { panic!() };
28672 assert!(matches!(
28673 sel.items[0],
28674 SelectItem::Expr {
28675 expr: Expr::Placeholder(1),
28676 alias: None
28677 }
28678 ));
28679 // $2 + 1
28680 let SelectItem::Expr {
28681 expr: Expr::Binary { lhs, rhs, .. },
28682 ..
28683 } = &sel.items[1]
28684 else {
28685 panic!()
28686 };
28687 assert!(matches!(**lhs, Expr::Placeholder(2)));
28688 assert!(matches!(**rhs, Expr::Literal(Literal::Integer(1))));
28689 // WHERE x = $3
28690 let Some(Expr::Binary { rhs, .. }) = sel.where_.as_ref() else {
28691 panic!()
28692 };
28693 assert!(matches!(**rhs, Expr::Placeholder(3)));
28694 }
28695
28696 #[test]
28697 fn parser_rejects_dollar_zero() {
28698 // $0 is not valid in PG; the lexer rejects it.
28699 assert!(parse_statement("SELECT $0").is_err());
28700 }
28701
28702 #[test]
28703 fn placeholder_display_roundtrips() {
28704 // The Display impl must produce text that re-lexes to the
28705 // same Placeholder token.
28706 let s = parse("SELECT $42 FROM t");
28707 let printed = s.to_string();
28708 assert!(printed.contains("$42"));
28709 let again = parse(&printed);
28710 assert_eq!(s, again);
28711 }
28712
28713 #[test]
28714 fn alter_index_rebuild_bare() {
28715 use crate::ast::{AlterIndexTarget, Statement};
28716 let s = parse("ALTER INDEX my_idx REBUILD");
28717 let Statement::AlterIndex(a) = s else {
28718 panic!("expected AlterIndex, got {s:?}")
28719 };
28720 assert_eq!(a.name, "my_idx");
28721 assert_eq!(a.target, AlterIndexTarget::Rebuild { encoding: None });
28722 }
28723
28724 #[test]
28725 fn alter_index_rebuild_with_encoding() {
28726 use crate::ast::{AlterIndexTarget, Statement};
28727 for (sql, want) in [
28728 (
28729 "ALTER INDEX my_idx REBUILD WITH (encoding = F32)",
28730 VecEncoding::F32,
28731 ),
28732 (
28733 "ALTER INDEX my_idx REBUILD WITH (encoding = sq8)",
28734 VecEncoding::Sq8,
28735 ),
28736 (
28737 "ALTER INDEX my_idx REBUILD WITH (encoding = HALF)",
28738 VecEncoding::F16,
28739 ),
28740 ] {
28741 let s = parse(sql);
28742 let Statement::AlterIndex(a) = s else {
28743 panic!("{sql}: expected AlterIndex")
28744 };
28745 assert_eq!(a.name, "my_idx");
28746 assert_eq!(
28747 a.target,
28748 AlterIndexTarget::Rebuild {
28749 encoding: Some(want)
28750 },
28751 "{sql}"
28752 );
28753 }
28754 }
28755
28756 #[test]
28757 fn alter_index_rebuild_unknown_encoding_errors() {
28758 let err = parse_statement("ALTER INDEX my_idx REBUILD WITH (encoding = PQ8)").unwrap_err();
28759 assert!(
28760 err.message.contains("unknown vector encoding"),
28761 "got: {}",
28762 err.message
28763 );
28764 }
28765
28766 #[test]
28767 fn alter_index_rebuild_display_roundtrips() {
28768 for (input, want) in [
28769 ("ALTER INDEX my_idx REBUILD", "ALTER INDEX my_idx REBUILD"),
28770 (
28771 "ALTER INDEX my_idx REBUILD WITH (encoding = SQ8)",
28772 "ALTER INDEX my_idx REBUILD WITH (encoding = SQ8)",
28773 ),
28774 (
28775 "ALTER INDEX my_idx REBUILD WITH (encoding = HALF)",
28776 "ALTER INDEX my_idx REBUILD WITH (encoding = HALF)",
28777 ),
28778 ] {
28779 let s = parse(input);
28780 assert_eq!(s.to_string(), want);
28781 }
28782 }
28783
28784 #[test]
28785 fn create_table_unknown_type_defers_to_engine() {
28786 // v4.9 picked XML as a parse-time "unsupported column
28787 // type" probe. v7.17.0 Phase 1.4 changed the contract:
28788 // an unknown type ident parses as Text + `user_type_ref`
28789 // so CREATE TABLE can resolve user-defined enum / domain
28790 // types — rejection of truly-unknown types moved to the
28791 // engine's catalog lookup. v7.37.5 ζ-A then promoted XML
28792 // to a first-class built-in, so this probe switched to a
28793 // synthetic name nothing in the lexer will ever recognise.
28794 let stmt = parse_statement("CREATE TABLE x (a my_user_type)").unwrap();
28795 let Statement::CreateTable(t) = stmt else {
28796 panic!("expected CreateTable");
28797 };
28798 assert_eq!(t.columns[0].user_type_ref.as_deref(), Some("my_user_type"));
28799 }
28800
28801 #[test]
28802 fn create_table_missing_table_keyword_errors() {
28803 assert!(parse_statement("CREATE x (a INT)").is_err());
28804 }
28805
28806 // v7.37.6-B(sentori Epic 2 P0)— `PARTITION BY RANGE` parent +
28807 // `PARTITION OF parent <bounds>` child parse + Display round-trip.
28808
28809 #[test]
28810 fn parse_create_table_partition_by_range() {
28811 use crate::ast::{PartitionBySpec, PartitionKindAst};
28812 let stmt = parse_statement(
28813 "CREATE TABLE events_partitioned (id BIGINT NOT NULL, ts TIMESTAMPTZ NOT NULL, \
28814 payload JSONB) PARTITION BY RANGE (ts)",
28815 )
28816 .unwrap();
28817 let Statement::CreateTable(t) = stmt else {
28818 panic!("expected CreateTable");
28819 };
28820 assert!(t.partition_of.is_none(), "parent has no partition_of");
28821 assert_eq!(t.columns.len(), 3);
28822 let by = t.partition_by.as_ref().expect("expected PARTITION BY");
28823 assert_eq!(
28824 by,
28825 &PartitionBySpec {
28826 kind: PartitionKindAst::Range,
28827 key_columns: alloc::vec!["ts".to_string()],
28828 }
28829 );
28830 // Display round-trip preserves the suffix. `quote_ident`
28831 // only adds double quotes when the ident needs escaping, so
28832 // a plain `ts` survives bare here.
28833 assert!(
28834 t.to_string().contains("PARTITION BY RANGE (ts)"),
28835 "Display lost PARTITION BY suffix: {t}"
28836 );
28837 }
28838
28839 #[test]
28840 fn parse_create_table_partition_of_range() {
28841 use crate::ast::{PartitionOfBoundsAst, PartitionOfSpec};
28842 let stmt = parse_statement(
28843 "CREATE TABLE events_2026_06 PARTITION OF events_partitioned \
28844 FOR VALUES FROM ('2026-06-01 00:00:00+00') TO ('2026-07-01 00:00:00+00')",
28845 )
28846 .unwrap();
28847 let Statement::CreateTable(t) = stmt else {
28848 panic!("expected CreateTable");
28849 };
28850 assert!(t.columns.is_empty(), "child inherits columns from parent");
28851 assert!(t.partition_by.is_none());
28852 let of = t.partition_of.as_ref().expect("expected PARTITION OF");
28853 assert_eq!(of.parent_name, "events_partitioned");
28854 let PartitionOfSpec { bounds, .. } = of.clone();
28855 match bounds {
28856 PartitionOfBoundsAst::Range { lower, upper } => {
28857 assert!(lower.to_string().contains("2026-06-01"));
28858 assert!(upper.to_string().contains("2026-07-01"));
28859 }
28860 other => panic!("expected Range, got {other:?}"),
28861 }
28862 // Display round-trip emits the FOR VALUES tail. `quote_ident`
28863 // skips quotes when not required, so the parent name appears
28864 // bare here.
28865 let s = t.to_string();
28866 assert!(
28867 s.contains("PARTITION OF events_partitioned"),
28868 "Display lost PARTITION OF: {s}"
28869 );
28870 assert!(s.contains("FOR VALUES FROM"), "Display lost FROM: {s}");
28871 assert!(s.contains(") TO ("), "Display lost TO: {s}");
28872 }
28873
28874 #[test]
28875 fn parse_create_table_partition_of_default() {
28876 use crate::ast::PartitionOfBoundsAst;
28877 let stmt =
28878 parse_statement("CREATE TABLE events_default PARTITION OF events_partitioned DEFAULT")
28879 .unwrap();
28880 let Statement::CreateTable(t) = stmt else {
28881 panic!("expected CreateTable");
28882 };
28883 let of = t.partition_of.as_ref().expect("expected PARTITION OF");
28884 assert_eq!(of.parent_name, "events_partitioned");
28885 assert!(matches!(of.bounds, PartitionOfBoundsAst::Default));
28886 assert!(
28887 t.to_string()
28888 .contains("PARTITION OF events_partitioned DEFAULT"),
28889 "Display lost DEFAULT: {t}"
28890 );
28891 }
28892
28893 #[test]
28894 fn parse_create_table_partition_by_list() {
28895 // v7.37.16 (16.1) — `PARTITION BY LIST (key)` parent + a
28896 // child with `FOR VALUES IN (lit, lit, …)`.
28897 use crate::ast::{PartitionBySpec, PartitionKindAst, PartitionOfBoundsAst};
28898 let parent =
28899 parse_statement("CREATE TABLE events_listed (region TEXT) PARTITION BY LIST (region)")
28900 .unwrap();
28901 let Statement::CreateTable(t) = parent else {
28902 panic!("expected CreateTable");
28903 };
28904 let Some(PartitionBySpec {
28905 kind,
28906 ref key_columns,
28907 }) = t.partition_by
28908 else {
28909 panic!("expected PARTITION BY");
28910 };
28911 assert_eq!(kind, PartitionKindAst::List);
28912 assert_eq!(*key_columns, vec!["region".to_string()]);
28913 assert!(t.to_string().contains("PARTITION BY LIST (region)"));
28914
28915 let child = parse_statement(
28916 "CREATE TABLE events_apac PARTITION OF events_listed \
28917 FOR VALUES IN ('jp', 'kr', 'tw')",
28918 )
28919 .unwrap();
28920 let Statement::CreateTable(c) = child else {
28921 panic!("expected CreateTable");
28922 };
28923 let of = c.partition_of.as_ref().expect("expected PARTITION OF");
28924 let PartitionOfBoundsAst::List { values } = &of.bounds else {
28925 panic!("expected List bounds, got {:?}", of.bounds);
28926 };
28927 assert_eq!(values.len(), 3);
28928 let disp = c.to_string();
28929 assert!(disp.contains("FOR VALUES IN ("), "Display lost IN: {disp}");
28930 }
28931
28932 #[test]
28933 fn parse_create_table_partition_by_hash() {
28934 // v7.37.16 (16.2) — `PARTITION BY HASH (key)` parent + a
28935 // child with `FOR VALUES WITH (MODULUS m, REMAINDER r)`.
28936 use crate::ast::{PartitionBySpec, PartitionKindAst, PartitionOfBoundsAst};
28937 let parent =
28938 parse_statement("CREATE TABLE orders_h (id BIGINT) PARTITION BY HASH (id)").unwrap();
28939 let Statement::CreateTable(t) = parent else {
28940 panic!("expected CreateTable");
28941 };
28942 let Some(PartitionBySpec {
28943 kind,
28944 ref key_columns,
28945 }) = t.partition_by
28946 else {
28947 panic!("expected PARTITION BY");
28948 };
28949 assert_eq!(kind, PartitionKindAst::Hash);
28950 assert_eq!(*key_columns, vec!["id".to_string()]);
28951 assert!(t.to_string().contains("PARTITION BY HASH (id)"));
28952
28953 let child = parse_statement(
28954 "CREATE TABLE orders_h_0 PARTITION OF orders_h \
28955 FOR VALUES WITH (MODULUS 4, REMAINDER 0)",
28956 )
28957 .unwrap();
28958 let Statement::CreateTable(c) = child else {
28959 panic!("expected CreateTable");
28960 };
28961 let of = c.partition_of.as_ref().expect("expected PARTITION OF");
28962 let PartitionOfBoundsAst::Hash { modulus, remainder } = of.bounds else {
28963 panic!("expected Hash bounds");
28964 };
28965 assert_eq!(modulus, 4);
28966 assert_eq!(remainder, 0);
28967 let disp = c.to_string();
28968 assert!(
28969 disp.contains("FOR VALUES WITH (MODULUS 4, REMAINDER 0)"),
28970 "Display lost HASH bounds: {disp}"
28971 );
28972
28973 // Validation: REMAINDER ≥ MODULUS is rejected at parse time.
28974 let bad = parse_statement(
28975 "CREATE TABLE orders_h_bad PARTITION OF orders_h \
28976 FOR VALUES WITH (MODULUS 4, REMAINDER 4)",
28977 );
28978 let msg = format!("{}", bad.unwrap_err());
28979 assert!(
28980 msg.contains("REMAINDER") && msg.contains("MODULUS"),
28981 "expected REMAINDER/MODULUS validation error: {msg}"
28982 );
28983 }
28984
28985 #[test]
28986 fn parse_create_table_partition_of_rejects_columns() {
28987 // v7.37.6-B contract: PARTITION OF children inherit columns
28988 // from the parent; an explicit list MUST surface as a parse
28989 // error rather than getting silently ignored.
28990 let err = parse_statement(
28991 "CREATE TABLE events_2026_06 PARTITION OF events_partitioned (id BIGINT) \
28992 FOR VALUES FROM ('a') TO ('b')",
28993 );
28994 assert!(err.is_err(), "expected parse error for explicit columns");
28995 let msg = format!("{}", err.unwrap_err());
28996 assert!(
28997 msg.contains("PARTITION OF") && msg.contains("column"),
28998 "error should mention PARTITION OF + columns: {msg}"
28999 );
29000 }
29001
29002 #[test]
29003 fn insert_single_value() {
29004 let s = parse("INSERT INTO foo VALUES (42)");
29005 let Statement::Insert(i) = s else {
29006 panic!("expected Insert")
29007 };
29008 assert_eq!(i.table, "foo");
29009 assert_eq!(i.rows.len(), 1);
29010 assert_eq!(i.rows[0].len(), 1);
29011 assert!(matches!(i.rows[0][0], Expr::Literal(Literal::Integer(42))));
29012 }
29013
29014 #[test]
29015 fn insert_multi_value_with_mixed_literals() {
29016 let s = parse("INSERT INTO foo VALUES (1, 'hi', 3.14, TRUE, NULL)");
29017 let Statement::Insert(i) = s else { panic!() };
29018 assert_eq!(i.rows.len(), 1);
29019 assert_eq!(i.rows[0].len(), 5);
29020 }
29021
29022 #[test]
29023 fn insert_missing_into_errors() {
29024 assert!(parse_statement("INSERT foo VALUES (1)").is_err());
29025 }
29026
29027 #[test]
29028 fn create_table_round_trip() {
29029 let original =
29030 parse("CREATE TABLE foo (id BIGINT NOT NULL, label TEXT, score FLOAT NOT NULL)");
29031 let text = original.to_string();
29032 let again = parse_statement(&text).expect("re-parse");
29033 assert_eq!(original, again);
29034 }
29035
29036 #[test]
29037 fn insert_round_trip_with_negation_and_string() {
29038 let original = parse("INSERT INTO t VALUES (-1, 'it''s', NULL)");
29039 let text = original.to_string();
29040 let again = parse_statement(&text).expect("re-parse");
29041 assert_eq!(original, again);
29042 }
29043
29044 #[test]
29045 fn unknown_keyword_at_statement_start_errors() {
29046 // v4.4: UPDATE is real SQL now. Use a fabricated keyword so
29047 // the top-level dispatch still has no branch to take.
29048 let err = parse_statement("FROBNICATE foo SET x = 1").unwrap_err();
29049 assert_eq!(err.message, "syntax error at or near \"FROBNICATE\"");
29050 }
29051
29052 // --- v0.8 CREATE INDEX --------------------------------------------------
29053
29054 #[test]
29055 fn create_index_basic() {
29056 let s = parse("CREATE INDEX idx_id ON users (id)");
29057 let Statement::CreateIndex(c) = s else {
29058 panic!("expected CreateIndex")
29059 };
29060 assert_eq!(c.name, "idx_id");
29061 assert_eq!(c.table, "users");
29062 assert_eq!(c.column, "id");
29063 }
29064
29065 #[test]
29066 fn create_index_missing_on_errors() {
29067 assert!(parse_statement("CREATE INDEX foo users (id)").is_err());
29068 }
29069
29070 #[test]
29071 fn create_index_missing_paren_errors() {
29072 assert!(parse_statement("CREATE INDEX foo ON users id").is_err());
29073 }
29074
29075 #[test]
29076 fn create_index_round_trip() {
29077 let original = parse("CREATE INDEX by_name ON users (name)");
29078 let again = parse_statement(&original.to_string()).unwrap();
29079 assert_eq!(original, again);
29080 }
29081
29082 // --- v7.9.29 CREATE UNIQUE INDEX [WHERE pred] (mailrs K1) -------------
29083
29084 #[test]
29085 fn create_unique_index_basic() {
29086 let s = parse("CREATE UNIQUE INDEX uq_x ON t (a)");
29087 let Statement::CreateIndex(c) = s else {
29088 panic!("expected CreateIndex");
29089 };
29090 assert!(c.is_unique);
29091 assert_eq!(c.column, "a");
29092 assert!(c.partial_predicate.is_none());
29093 }
29094
29095 #[test]
29096 fn create_unique_index_partial() {
29097 // mailrs's email_templates "one default per user" shape.
29098 let s = parse(
29099 "CREATE UNIQUE INDEX idx_email_templates_user_default \
29100 ON email_templates (user_address) WHERE is_default = true",
29101 );
29102 let Statement::CreateIndex(c) = s else {
29103 panic!("expected CreateIndex");
29104 };
29105 assert!(c.is_unique);
29106 assert_eq!(c.table, "email_templates");
29107 assert_eq!(c.column, "user_address");
29108 assert!(c.partial_predicate.is_some());
29109 }
29110
29111 #[test]
29112 fn create_unique_index_composite_with_predicate() {
29113 // mailrs's calendar_events instance: composite columns.
29114 let s = parse(
29115 "CREATE UNIQUE INDEX uq_calendar_events_instance \
29116 ON calendar_events (calendar_id, uid, recurrence_id) \
29117 WHERE recurrence_id IS NOT NULL",
29118 );
29119 let Statement::CreateIndex(c) = s else {
29120 panic!("expected CreateIndex");
29121 };
29122 assert!(c.is_unique);
29123 assert_eq!(c.column, "calendar_id");
29124 assert_eq!(
29125 c.extra_columns,
29126 vec!["uid".to_string(), "recurrence_id".to_string()]
29127 );
29128 assert!(c.partial_predicate.is_some());
29129 }
29130
29131 #[test]
29132 fn create_unique_index_using_btree_ok() {
29133 let s = parse("CREATE UNIQUE INDEX uq_x ON t USING btree (a)");
29134 assert!(matches!(s, Statement::CreateIndex(ref c) if c.is_unique));
29135 }
29136
29137 #[test]
29138 fn create_unique_index_using_hnsw_rejected() {
29139 let err =
29140 parse_statement("CREATE UNIQUE INDEX uq_v ON t USING hnsw (embedding)").unwrap_err();
29141 assert!(err.message.contains("UNIQUE"), "{}", err.message);
29142 }
29143
29144 #[test]
29145 fn create_unique_index_round_trip() {
29146 let original = parse(
29147 "CREATE UNIQUE INDEX uq_calendar_events_master \
29148 ON calendar_events (calendar_id, uid) WHERE recurrence_id IS NULL",
29149 );
29150 let again = parse_statement(&original.to_string()).unwrap();
29151 assert_eq!(original, again);
29152 }
29153
29154 #[test]
29155 fn create_unique_without_index_errors() {
29156 let err = parse_statement("CREATE UNIQUE TABLE t (a INT)").unwrap_err();
29157 // v7.39 (round 340, V56) — PG 18.4, verbatim.
29158 assert_eq!(err.message, "syntax error at or near \"TABLE\"");
29159 }
29160
29161 // --- v7.10.4 BYTES / BYTEA column type (Epic 1) ----------------------
29162
29163 #[test]
29164 fn create_table_bytea_column() {
29165 let s = parse("CREATE TABLE t (id INT NOT NULL, payload BYTEA NOT NULL)");
29166 let Statement::CreateTable(c) = s else {
29167 panic!("expected CreateTable");
29168 };
29169 assert_eq!(c.columns.len(), 2);
29170 assert_eq!(c.columns[1].ty, ColumnTypeName::Bytes);
29171 assert!(!c.columns[1].nullable);
29172 }
29173
29174 #[test]
29175 fn create_table_bytes_alias_column() {
29176 let s = parse("CREATE TABLE t (blob BYTES)");
29177 let Statement::CreateTable(c) = s else {
29178 panic!("expected CreateTable");
29179 };
29180 assert_eq!(c.columns[0].ty, ColumnTypeName::Bytes);
29181 }
29182
29183 #[test]
29184 fn bytea_round_trip_display() {
29185 let original = parse("CREATE TABLE t (a BYTEA NOT NULL)");
29186 let again = parse_statement(&original.to_string()).unwrap();
29187 assert_eq!(original, again);
29188 }
29189
29190 // --- v0.9 transactions -------------------------------------------------
29191
29192 #[test]
29193 fn begin_commit_rollback_parse_as_unit_variants() {
29194 let plain = crate::ast::TransactionModes::default();
29195 assert_eq!(parse("BEGIN"), Statement::Begin(plain));
29196 assert_eq!(parse("COMMIT"), Statement::Commit);
29197 // r1066 — PG synonyms pgbench's tpcb script relies on.
29198 assert_eq!(parse("END"), Statement::Commit);
29199 assert_eq!(parse("END TRANSACTION"), Statement::Commit);
29200 assert_eq!(parse("COMMIT WORK"), Statement::Commit);
29201 assert_eq!(parse("ROLLBACK"), Statement::Rollback);
29202 // Trailing semicolons accepted too.
29203 assert_eq!(parse("BEGIN;"), Statement::Begin(plain));
29204 // v7.39 (read01 round 118, B3) — an explicit ISOLATION LEVEL rides the
29205 // statement (with or without the WORK/TRANSACTION noise word).
29206 assert_eq!(
29207 parse("BEGIN ISOLATION LEVEL REPEATABLE READ"),
29208 Statement::Begin(crate::ast::TransactionModes {
29209 isolation: Some(IsolationLevel::RepeatableRead),
29210 read_only: None,
29211 })
29212 );
29213 assert_eq!(
29214 parse("START TRANSACTION ISOLATION LEVEL SERIALIZABLE"),
29215 Statement::Begin(crate::ast::TransactionModes {
29216 isolation: Some(IsolationLevel::Serializable),
29217 read_only: None,
29218 })
29219 );
29220 // v7.39 — this line used to read
29221 //
29222 // // A non-isolation mode keeps the session default (None).
29223 // assert_eq!(parse("BEGIN READ ONLY"), Statement::Begin(None));
29224 //
29225 // which pinned the defect rather than catching it: the READ ONLY
29226 // was thrown away, so the statement opened an ordinary read-write
29227 // transaction and every write inside it was accepted. The
29228 // isolation level is still absent here, because this statement
29229 // does not name one — that part was right.
29230 assert_eq!(
29231 parse("BEGIN READ ONLY"),
29232 Statement::Begin(crate::ast::TransactionModes {
29233 isolation: None,
29234 read_only: Some(true),
29235 })
29236 );
29237 assert_eq!(
29238 parse("START TRANSACTION READ WRITE"),
29239 Statement::Begin(crate::ast::TransactionModes {
29240 isolation: None,
29241 read_only: Some(false),
29242 })
29243 );
29244 assert_eq!(
29245 parse("BEGIN ISOLATION LEVEL SERIALIZABLE, READ ONLY"),
29246 Statement::Begin(crate::ast::TransactionModes {
29247 isolation: Some(IsolationLevel::Serializable),
29248 read_only: Some(true),
29249 })
29250 );
29251 }
29252
29253 // --- v1.2: pgvector distance ops + ::vector cast --------------------
29254
29255 #[test]
29256 fn inner_product_binop_parses() {
29257 let s = parse("SELECT v <#> [1.0, 2.0] FROM t");
29258 let Statement::Select(s) = s else { panic!() };
29259 let SelectItem::Expr { expr, .. } = &s.items[0] else {
29260 panic!()
29261 };
29262 assert!(matches!(
29263 expr,
29264 Expr::Binary {
29265 op: BinOp::InnerProduct,
29266 ..
29267 }
29268 ));
29269 }
29270
29271 #[test]
29272 fn cosine_distance_binop_parses() {
29273 let s = parse("SELECT v <=> [1.0, 2.0] FROM t");
29274 let Statement::Select(s) = s else { panic!() };
29275 let SelectItem::Expr { expr, .. } = &s.items[0] else {
29276 panic!()
29277 };
29278 assert!(matches!(
29279 expr,
29280 Expr::Binary {
29281 op: BinOp::CosineDistance,
29282 ..
29283 }
29284 ));
29285 }
29286
29287 #[test]
29288 fn vector_cast_postfix_wraps_string_literal() {
29289 let s = parse("SELECT '[1,2,3]'::vector FROM t");
29290 let Statement::Select(s) = s else { panic!() };
29291 let SelectItem::Expr { expr, .. } = &s.items[0] else {
29292 panic!()
29293 };
29294 assert!(matches!(
29295 expr,
29296 Expr::Cast {
29297 target: CastTarget::Vector,
29298 ..
29299 }
29300 ));
29301 }
29302
29303 #[test]
29304 fn unsupported_cast_target_errors() {
29305 // v7.37.5 ship triage promoted the parser to accept every
29306 // ident as a `CastTarget::Named(canonical)`; the engine
29307 // surfaces the "unsupported cast target" error at eval
29308 // time when `type_name_to_data_type` can't resolve it.
29309 // Parser-side error now requires a NON-ident after `::`
29310 // (e.g. a punctuation token).
29311 let err = parse_statement("SELECT 1::, FROM t").unwrap_err();
29312 assert_eq!(err.message, "syntax error at or near \",\"");
29313 }
29314
29315 #[test]
29316 fn tx_statements_round_trip() {
29317 for q in ["BEGIN", "COMMIT", "ROLLBACK"] {
29318 let original = parse(q);
29319 let again = parse_statement(&original.to_string()).unwrap();
29320 assert_eq!(original, again);
29321 }
29322 }
29323
29324 #[test]
29325 fn interval_text_parsing_units() {
29326 // v7.37.5 β — three-field shape `(months, days, micros)` so
29327 // `'1 day'` and `'24 hours'` no longer collide (PG parity).
29328 // Single unit.
29329 assert_eq!(parse_interval_text("1 day"), Some((0, 1, 0)));
29330 assert_eq!(
29331 parse_interval_text("24 hours"),
29332 Some((0, 0, 86_400_000_000))
29333 );
29334 assert_eq!(parse_interval_text("1 second"), Some((0, 0, 1_000_000)));
29335 assert_eq!(parse_interval_text("1 month"), Some((1, 0, 0)));
29336 assert_eq!(parse_interval_text("2 years"), Some((24, 0, 0)));
29337 assert_eq!(parse_interval_text("1 week"), Some((0, 7, 0)));
29338 // Compound spans accumulate per-dimension.
29339 assert_eq!(parse_interval_text("1 year 6 months"), Some((18, 0, 0)));
29340 assert_eq!(
29341 parse_interval_text("1 day 2 hours"),
29342 Some((0, 1, 7_200_000_000))
29343 );
29344 // Negative numbers carry through per-dimension.
29345 assert_eq!(parse_interval_text("-1 day"), Some((0, -1, 0)));
29346 // Bad shapes return None.
29347 assert_eq!(parse_interval_text(""), None);
29348 assert_eq!(parse_interval_text("garbage"), None);
29349 assert_eq!(parse_interval_text("1 fortnight"), None);
29350 // v7.39 (GUC knife 3) — PG reads a bare number as SECONDS
29351 // (`INTERVAL '1'` = 00:00:01), verified against the oracle.
29352 assert_eq!(parse_interval_text("1"), Some((0, 0, 1_000_000)));
29353 assert_eq!(parse_interval_text("0"), Some((0, 0, 0)));
29354 assert_eq!(parse_interval_text("1.5"), Some((0, 0, 1_500_000)));
29355 }
29356
29357 #[test]
29358 fn interval_literal_roundtrips_via_display() {
29359 let parsed = parse("SELECT INTERVAL '1 day 2 hours'");
29360 let s = parsed.to_string();
29361 // Display preserves the original text verbatim.
29362 assert!(s.contains("INTERVAL '1 day 2 hours'"), "got: {s}");
29363 // And re-parsing yields a structurally equal statement.
29364 let again = parse_statement(&s).unwrap();
29365 assert_eq!(parsed, again);
29366 }
29367
29368 // ── v6.1.2: CREATE / DROP PUBLICATION ────────────────────
29369
29370 #[test]
29371 fn parser_recognises_create_publication_bare() {
29372 let s = parse("CREATE PUBLICATION pub_a");
29373 let Statement::CreatePublication(p) = s else {
29374 panic!("expected CreatePublication, got {s:?}")
29375 };
29376 assert_eq!(p.name, "pub_a");
29377 assert_eq!(p.scope, PublicationScope::AllTables);
29378 }
29379
29380 #[test]
29381 fn parser_recognises_create_publication_for_all_tables() {
29382 let s = parse("CREATE PUBLICATION pub_a FOR ALL TABLES");
29383 let Statement::CreatePublication(p) = s else {
29384 panic!("expected CreatePublication, got {s:?}")
29385 };
29386 assert_eq!(p.name, "pub_a");
29387 assert_eq!(p.scope, PublicationScope::AllTables);
29388 }
29389
29390 #[test]
29391 fn parser_recognises_drop_publication() {
29392 let s = parse("DROP PUBLICATION pub_a");
29393 let Statement::DropPublication { name, .. } = s else {
29394 panic!("expected DropPublication, got {s:?}")
29395 };
29396 assert_eq!(name, "pub_a");
29397 }
29398
29399 #[test]
29400 fn parser_recognises_for_table_list() {
29401 let s = parse("CREATE PUBLICATION pub_a FOR TABLE t1, t2, t3");
29402 let Statement::CreatePublication(p) = s else {
29403 panic!("expected CreatePublication, got {s:?}")
29404 };
29405 assert_eq!(p.name, "pub_a");
29406 let PublicationScope::ForTables(ts) = p.scope else {
29407 panic!("expected ForTables scope")
29408 };
29409 assert_eq!(ts, alloc::vec!["t1", "t2", "t3"]);
29410 }
29411
29412 #[test]
29413 fn parser_rejects_bare_for_tables_and_takes_in_schema() {
29414 // v7.39 (round 754, F31-B5) — PG18-measured: the bare plural
29415 // is rejected (`invalid publication object list`; the old
29416 // test pinned an unverifiable "PG 19 accepts both" claim);
29417 // TABLES pairs with IN SCHEMA.
29418 let err = parse_statement("CREATE PUBLICATION pub_a FOR TABLES t1, t2")
29419 .expect_err("bare FOR TABLES must reject");
29420 assert!(
29421 alloc::format!("{err}").contains("invalid publication object list"),
29422 "got: {err}"
29423 );
29424 let s = parse("CREATE PUBLICATION pub_a FOR TABLES IN SCHEMA public");
29425 let Statement::CreatePublication(p) = s else {
29426 panic!("expected CreatePublication, got {s:?}")
29427 };
29428 let PublicationScope::TablesInSchema(schema) = p.scope else {
29429 panic!("expected TablesInSchema")
29430 };
29431 assert_eq!(schema, "public");
29432 }
29433
29434 #[test]
29435 fn parser_recognises_for_all_tables_except_list() {
29436 let s = parse("CREATE PUBLICATION p FOR ALL TABLES EXCEPT t1, t2");
29437 let Statement::CreatePublication(p) = s else {
29438 panic!()
29439 };
29440 let PublicationScope::AllTablesExcept(ts) = p.scope else {
29441 panic!("expected AllTablesExcept")
29442 };
29443 assert_eq!(ts, alloc::vec!["t1", "t2"]);
29444 }
29445
29446 #[test]
29447 fn parser_rejects_for_table_with_empty_list() {
29448 // `FOR TABLE` with nothing after is a parse error.
29449 let err = parse_statement("CREATE PUBLICATION p FOR TABLE")
29450 .expect_err("must error on empty list");
29451 // No specific message asserted — the call falls through to
29452 // expect_ident_like which yields "expected identifier, got …".
29453 assert!(!err.message.is_empty());
29454 }
29455
29456 #[test]
29457 fn parser_recognises_show_publications() {
29458 // v6.1.3 — SHOW PUBLICATIONS lands here. PUBLICATIONS is a
29459 // bare ident in this position, NOT a reserved keyword.
29460 let s = parse("SHOW PUBLICATIONS");
29461 assert!(matches!(s, Statement::ShowPublications));
29462 }
29463
29464 // ── v6.1.4: CREATE / DROP SUBSCRIPTION + SHOW SUBSCRIPTIONS ─
29465
29466 #[test]
29467 fn parser_recognises_create_subscription_single_publication() {
29468 let s = parse(
29469 "CREATE SUBSCRIPTION sub_a CONNECTION 'host=127.0.0.1 port=20002' PUBLICATION pub_a",
29470 );
29471 let Statement::CreateSubscription(c) = s else {
29472 panic!("expected CreateSubscription, got {s:?}")
29473 };
29474 assert_eq!(c.name, "sub_a");
29475 assert_eq!(c.conn_str, "host=127.0.0.1 port=20002");
29476 assert_eq!(c.publications, alloc::vec!["pub_a"]);
29477 }
29478
29479 #[test]
29480 fn parser_recognises_create_subscription_multi_publication() {
29481 let s = parse("CREATE SUBSCRIPTION sub_a CONNECTION 'host=h' PUBLICATION p1, p2, p3");
29482 let Statement::CreateSubscription(c) = s else {
29483 panic!()
29484 };
29485 assert_eq!(c.publications, alloc::vec!["p1", "p2", "p3"]);
29486 }
29487
29488 #[test]
29489 fn parser_rejects_create_subscription_missing_connection() {
29490 let err = parse_statement("CREATE SUBSCRIPTION s PUBLICATION p")
29491 .expect_err("must error on missing CONNECTION");
29492 assert_eq!(err.message, "syntax error at or near \"PUBLICATION\"");
29493 }
29494
29495 #[test]
29496 fn parser_rejects_create_subscription_missing_publication() {
29497 let err = parse_statement("CREATE SUBSCRIPTION s CONNECTION 'host=x'")
29498 .expect_err("must error on missing PUBLICATION");
29499 assert_eq!(err.message, "syntax error at end of input");
29500 }
29501
29502 #[test]
29503 fn parser_recognises_drop_subscription() {
29504 let s = parse("DROP SUBSCRIPTION sub_a");
29505 let Statement::DropSubscription { name, .. } = s else {
29506 panic!("expected DropSubscription, got {s:?}")
29507 };
29508 assert_eq!(name, "sub_a");
29509 }
29510
29511 #[test]
29512 fn parser_recognises_show_subscriptions() {
29513 let s = parse("SHOW SUBSCRIPTIONS");
29514 assert!(matches!(s, Statement::ShowSubscriptions));
29515 }
29516
29517 #[test]
29518 fn parser_recognises_wait_for_wal_position_no_timeout() {
29519 let s = parse("WAIT FOR WAL POSITION 12345");
29520 let Statement::WaitForWalPosition { pos, timeout_ms } = s else {
29521 panic!("expected WaitForWalPosition, got {s:?}")
29522 };
29523 assert_eq!(pos, 12345);
29524 assert!(timeout_ms.is_none());
29525 }
29526
29527 #[test]
29528 fn parser_recognises_wait_for_wal_position_with_timeout() {
29529 let s = parse("WAIT FOR WAL POSITION 67890 WITH TIMEOUT 5000");
29530 let Statement::WaitForWalPosition { pos, timeout_ms } = s else {
29531 panic!()
29532 };
29533 assert_eq!(pos, 67890);
29534 assert_eq!(timeout_ms, Some(5000));
29535 }
29536
29537 #[test]
29538 fn parser_rejects_wait_with_negative_position() {
29539 // The lexer treats `-` as a token; `expect_u64_literal`
29540 // only sees the Integer that follows, so the negative
29541 // arrives as a unary-minus expression at higher levels.
29542 // Bare `WAIT FOR WAL POSITION -1` thus surfaces as a
29543 // parse error one way or another.
29544 let err = parse_statement("WAIT FOR WAL POSITION -1").unwrap_err();
29545 assert!(!err.message.is_empty());
29546 }
29547
29548 #[test]
29549 fn parser_recognises_bare_analyze() {
29550 let s = parse("ANALYZE");
29551 assert!(matches!(s, Statement::Analyze(None)));
29552 }
29553
29554 #[test]
29555 fn parser_recognises_analyze_with_table() {
29556 let s = parse("ANALYZE users");
29557 let Statement::Analyze(Some(name)) = s else {
29558 panic!("expected Analyze, got {s:?}")
29559 };
29560 assert_eq!(name, "users");
29561 }
29562
29563 #[test]
29564 fn parser_recognises_analyze_with_quoted_table() {
29565 let s = parse("ANALYZE \"Mixed Case\"");
29566 let Statement::Analyze(Some(name)) = s else {
29567 panic!()
29568 };
29569 assert_eq!(name, "Mixed Case");
29570 }
29571
29572 #[test]
29573 fn parser_rejects_analyze_with_garbage_token() {
29574 let err = parse_statement("ANALYZE 42").expect_err("must error");
29575 assert!(!err.message.is_empty());
29576 }
29577
29578 #[test]
29579 fn analyze_display_roundtrips() {
29580 for sql in ["ANALYZE", "ANALYZE users"] {
29581 let s = parse(sql);
29582 let printed = s.to_string();
29583 let again = parse_statement(&printed)
29584 .unwrap_or_else(|e| panic!("re-parse failed for {printed:?}: {e}"));
29585 assert_eq!(s, again);
29586 }
29587 }
29588
29589 #[test]
29590 fn wait_for_display_roundtrips() {
29591 for sql in [
29592 "WAIT FOR WAL POSITION 12345",
29593 "WAIT FOR WAL POSITION 67890 WITH TIMEOUT 5000",
29594 ] {
29595 let s = parse(sql);
29596 let printed = s.to_string();
29597 let again = parse_statement(&printed)
29598 .unwrap_or_else(|e| panic!("re-parse failed for {printed:?}: {e}"));
29599 assert_eq!(s, again, "round-trip mismatch for {sql:?}");
29600 }
29601 }
29602
29603 #[test]
29604 fn subscription_ddl_display_roundtrips() {
29605 for sql in [
29606 "CREATE SUBSCRIPTION sub_a CONNECTION 'host=h port=20002' PUBLICATION pub_a",
29607 "CREATE SUBSCRIPTION sub_b CONNECTION 'host=h' PUBLICATION p1, p2",
29608 "DROP SUBSCRIPTION sub_a",
29609 "SHOW SUBSCRIPTIONS",
29610 ] {
29611 let s = parse(sql);
29612 let printed = s.to_string();
29613 let again = parse_statement(&printed)
29614 .unwrap_or_else(|e| panic!("re-parse failed for {printed:?}: {e}"));
29615 assert_eq!(s, again, "round-trip mismatch for {sql:?}");
29616 }
29617 }
29618
29619 #[test]
29620 fn parser_drop_dispatches_user_vs_publication() {
29621 // Pre-v6.1.2 DROP USER took the bare-ident path; v6.1.2
29622 // tokenises DROP. Both targets must still parse.
29623 let s = parse("DROP USER 'alice'");
29624 let Statement::DropUser { name, .. } = s else {
29625 panic!("expected DropUser, got {s:?}")
29626 };
29627 assert_eq!(name, "alice");
29628 // And DROP PUBLICATION lands the new variant.
29629 let s = parse("DROP PUBLICATION p1");
29630 assert!(matches!(s, Statement::DropPublication { .. }));
29631 }
29632
29633 #[test]
29634 fn publication_ddl_display_roundtrips() {
29635 // Every CREATE PUBLICATION variant must Display → parse →
29636 // same AST. v6.1.3 covers all three scope shapes.
29637 for sql in [
29638 "CREATE PUBLICATION pub_a",
29639 "CREATE PUBLICATION pub_a FOR ALL TABLES",
29640 "CREATE PUBLICATION pub_a FOR TABLE t1, t2",
29641 "CREATE PUBLICATION pub_a FOR ALL TABLES EXCEPT t1",
29642 "DROP PUBLICATION pub_a",
29643 "SHOW PUBLICATIONS",
29644 ] {
29645 let s = parse(sql);
29646 let printed = s.to_string();
29647 let again = parse_statement(&printed)
29648 .unwrap_or_else(|e| panic!("re-parse failed for {printed:?}: {e}"));
29649 assert_eq!(s, again, "round-trip mismatch for {sql:?}");
29650 }
29651 }
29652
29653 // --- v7.12.4: CREATE FUNCTION + CREATE TRIGGER + PL/pgSQL ---
29654
29655 #[test]
29656 fn create_function_returns_trigger_plpgsql_minimal() {
29657 let sql = "CREATE FUNCTION noop() RETURNS TRIGGER LANGUAGE plpgsql AS $$ BEGIN RETURN NEW; END; $$";
29658 let s = parse(sql);
29659 let Statement::CreateFunction(f) = s else {
29660 panic!("expected CreateFunction");
29661 };
29662 assert_eq!(f.name, "noop");
29663 assert!(!f.or_replace);
29664 assert!(f.args.is_empty());
29665 assert!(matches!(f.returns, FunctionReturn::Trigger));
29666 assert_eq!(f.language, "plpgsql");
29667 let FunctionBody::PlPgSql(block) = f.body else {
29668 panic!("expected PlPgSql body");
29669 };
29670 assert_eq!(block.statements.len(), 1);
29671 assert!(matches!(
29672 block.statements[0],
29673 PlPgSqlStmt::Return(ReturnTarget::New)
29674 ));
29675 }
29676
29677 #[test]
29678 fn create_function_or_replace_with_assignment() {
29679 // mailrs-shape trigger function: NEW.col := to_tsvector(...);
29680 // RETURN NEW.
29681 let sql = "CREATE OR REPLACE FUNCTION update_sv() RETURNS TRIGGER LANGUAGE plpgsql AS $$
29682BEGIN
29683 NEW.search_vector := to_tsvector('english', NEW.subject);
29684 RETURN NEW;
29685END;
29686$$";
29687 let s = parse(sql);
29688 let Statement::CreateFunction(f) = s else {
29689 panic!("expected CreateFunction");
29690 };
29691 assert!(f.or_replace);
29692 let FunctionBody::PlPgSql(block) = &f.body else {
29693 panic!("expected PlPgSql body");
29694 };
29695 assert_eq!(block.statements.len(), 2);
29696 // First statement: NEW.search_vector := to_tsvector(...)
29697 let PlPgSqlStmt::Assign { target, .. } = &block.statements[0] else {
29698 panic!("expected Assign as first stmt");
29699 };
29700 match target {
29701 AssignTarget::NewColumn(c) => assert_eq!(c, "search_vector"),
29702 other => panic!("expected NEW.col, got {other:?}"),
29703 }
29704 // Second statement: RETURN NEW
29705 assert!(matches!(
29706 block.statements[1],
29707 PlPgSqlStmt::Return(ReturnTarget::New)
29708 ));
29709 }
29710
29711 #[test]
29712 fn create_trigger_after_insert_or_update() {
29713 let sql = "CREATE TRIGGER tg AFTER INSERT OR UPDATE ON messages FOR EACH ROW EXECUTE FUNCTION update_sv()";
29714 let s = parse(sql);
29715 let Statement::CreateTrigger(t) = s else {
29716 panic!("expected CreateTrigger");
29717 };
29718 assert_eq!(t.name, "tg");
29719 assert_eq!(t.table, "messages");
29720 assert_eq!(t.timing, TriggerTiming::After);
29721 assert_eq!(t.events, vec![TriggerEvent::Insert, TriggerEvent::Update]);
29722 assert_eq!(t.for_each, TriggerForEach::Row);
29723 assert_eq!(t.function, "update_sv");
29724 }
29725
29726 #[test]
29727 fn create_trigger_before_delete_execute_procedure_alias() {
29728 // PG also accepts the legacy `EXECUTE PROCEDURE` spelling.
29729 let sql =
29730 "CREATE TRIGGER guard BEFORE DELETE ON t FOR EACH ROW EXECUTE PROCEDURE block_delete()";
29731 let s = parse(sql);
29732 let Statement::CreateTrigger(t) = s else {
29733 panic!("expected CreateTrigger");
29734 };
29735 assert_eq!(t.timing, TriggerTiming::Before);
29736 assert_eq!(t.events, vec![TriggerEvent::Delete]);
29737 }
29738
29739 #[test]
29740 fn drop_trigger_if_exists_round_trips() {
29741 // No parser support for DROP TRIGGER yet — added in v7.12.5
29742 // alongside the broader DROP …{IF EXISTS} cleanup. The
29743 // AST + Display impls are in place so we round-trip via
29744 // construction:
29745 let s = Statement::DropTrigger {
29746 name: "tg".into(),
29747 table: "messages".into(),
29748 if_exists: true,
29749 };
29750 assert_eq!(s.to_string(), "DROP TRIGGER IF EXISTS tg ON messages");
29751 }
29752
29753 #[test]
29754 fn trigger_ddl_display_roundtrips_through_parser() {
29755 // CREATE TRIGGER + its referenced CREATE FUNCTION must
29756 // Display → parse → same AST (modulo PL/pgSQL body
29757 // formatting which is parser-canonicalised).
29758 for sql in [
29759 "CREATE TRIGGER tg AFTER INSERT ON t FOR EACH ROW EXECUTE FUNCTION f()",
29760 "CREATE TRIGGER tg2 BEFORE UPDATE OR DELETE ON t FOR EACH ROW EXECUTE FUNCTION g()",
29761 ] {
29762 let s = parse(sql);
29763 let printed = s.to_string();
29764 let again = parse_statement(&printed)
29765 .unwrap_or_else(|e| panic!("re-parse failed for {printed:?}: {e}"));
29766 assert_eq!(s, again, "round-trip mismatch for {sql:?}");
29767 }
29768 }
29769}