fatou_parser/parser/reparse.rs
1//! Incremental reparse: turn a text edit plus the previous parse into a new
2//! tree without re-lexing and re-parsing the whole file.
3//!
4//! Modelled on rust-analyzer's `reparsing.rs` and arity's `reparse.rs`: try
5//! the cheapest strategy first and fall back to progressively more work, with
6//! a full [`parse`](crate::parser::parse) as the always-correct last resort.
7//! The staged plan lives in `TODO.md` (`### Incremental`). Two tiers are in:
8//! the token tier ([`reparse_token`]) relexes an edit confined to a single
9//! `Ident`, comment, or whitespace leaf in isolation and splices it into the
10//! previous green tree ([`try_splice_plain_token`]), or — for a `StringContent`
11//! leaf, whose lexing depends on the enclosing delimiter's mode — relexes the
12//! whole enclosing string literal in isolation and requires it to reproduce its
13//! own token run ([`try_splice_string_content`]); the top-level statement tier
14//! ([`reparse_toplevel`]) reparses the run of `ROOT` children touching the edit
15//! with the public [`parse`](crate::parser::parse) and splices the resulting
16//! children and diagnostics in place, provided the boundary guards prove the
17//! region decoupled from its neighbors.
18//!
19//! Two entry points sit above those tiers. [`reparse_edits`] takes the precise
20//! chain the language server staged from a `didChange` batch and replays it one
21//! edit at a time; [`reparse`] takes a single edit, which is what the host's
22//! `fatou::incremental` recovers with [`diff_edit`] from a pair of whole texts
23//! when no chain is available. The chain is tried first: a collapsed diff of
24//! scattered edits spans everything between them, and the region that produces
25//! is wide enough that [`region_is_too_wide`] declines it outright, where
26//! replaying the chain splices each edit on its own. The Tenet-4 oracle
27//! ([`assert_matches_full_parse`] plus `tests/incremental_reparse.rs`) checks
28//! every splice against a full parse.
29//!
30//! **Correctness invariant (Tenet 4):** a successful reparse must yield a
31//! green tree *and* diagnostics byte-identical to a full parse of the edited
32//! text. A reparse is an output-pure performance hint: the salsa layer only
33//! ever sees text changes, and a miss is a full parse, never an error.
34
35use rowan::{GreenNode, GreenToken, TextRange, TextSize, TokenAtOffset};
36
37use crate::parser::ParseDiagnostic;
38use crate::parser::diagnostics::DIAGNOSTIC_STREAMS;
39use crate::parser::lexer::lex;
40use crate::parser::tree_builder::syntax_kind_for;
41use crate::syntax::{SyntaxKind, SyntaxNode, SyntaxToken};
42
43/// The edit currency, defined in the sibling `edit` module and re-exported
44/// here so `parser::Edit` stays the path the parser layer uses.
45pub use super::edit::{Edit, apply_edits, diff_edit, try_apply_edits};
46
47/// Structural fingerprint of a tree: one line per descendant element with
48/// `kind@range` plus the token text (empty for nodes). Two trees with equal
49/// fingerprints are byte-identical. Oracle/debug support: the Tenet-4 assert
50/// in [`reparse`] and the `tests/incremental_reparse.rs` harness share this
51/// definition so they can never diverge.
52pub fn fingerprint(node: &SyntaxNode) -> String {
53 use std::fmt::Write;
54
55 let mut out = String::new();
56 for el in node.descendants_with_tokens() {
57 let text = el
58 .as_token()
59 .map(|t| t.text().to_string())
60 .unwrap_or_default();
61 let _ = writeln!(out, "{:?}@{:?} {:?}", el.kind(), el.text_range(), text);
62 }
63 out
64}
65
66/// Which strategy produced a [`Reparsed`]. Surfaced for tests and benchmarks.
67/// Ordered cheapest-first, so a chain of reparses can report the most
68/// expensive tier it had to reach with [`Ord::max`].
69#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
70pub enum ReparseTier {
71 /// A single leaf token was relexed in isolation and spliced in place.
72 Token,
73 /// A run of top-level statements was reparsed and spliced under `ROOT`.
74 TopLevel,
75}
76
77/// The result of a successful incremental reparse: the new whole-file green
78/// tree and its parse diagnostics (absolute offsets in the new text).
79#[derive(Debug, Clone)]
80pub struct Reparsed {
81 pub green: GreenNode,
82 pub diagnostics: Vec<ParseDiagnostic>,
83 pub tier: ReparseTier,
84}
85
86/// Attempt an incremental reparse of the previous parse (`prev_green`, parsed
87/// from `prev_text` with `prev_diags`) under `edit`, which transforms
88/// `prev_text` into `new_text`. Returns `None` when no incremental strategy
89/// applies — the caller must then do a full parse.
90///
91/// In debug builds every successful reparse is checked against a full parse
92/// of `new_text` (Tenet 4): tree fingerprint and diagnostics vector must be
93/// identical.
94///
95/// `edit` is not trusted, for the same reason [`reparse_edits`]'s chain is
96/// not: one that does not fit `prev_text` — a stale range, a buffer that moved
97/// underneath it — is rejected here rather than applied at offsets that would
98/// panic on a slice.
99pub fn reparse(
100 prev_text: &str,
101 prev_green: &GreenNode,
102 prev_diags: &[ParseDiagnostic],
103 edit: &Edit,
104 new_text: &str,
105) -> Option<Reparsed> {
106 if !fits(prev_text, edit) {
107 return None;
108 }
109 let result = reparse_impl(prev_text, prev_green, prev_diags, edit, new_text)?;
110 assert_matches_full_parse(&result, new_text);
111 Some(result)
112}
113
114/// Attempt an incremental reparse across a *chain* of edits, each expressed
115/// against the text its predecessors produced — the shape an LSP `didChange`
116/// batch arrives in (see `fatou::text::apply_content_changes` host-side).
117///
118/// This is for the case [`diff_edit`] cannot express: when the edits are
119/// scattered, collapsing them into one spanning edit covers everything between
120/// the first and the last change and blows both tiers' guards, while replaying
121/// them one at a time keeps each splice small. Try this *before* the collapsed
122/// edit: a wide span is the one shape the collapsed edit can never answer, so
123/// offering it first is the difference between a hit and a
124/// `region_is_too_wide` bail (`benches/reparse.rs`).
125///
126/// Fewer than two edits returns `None` immediately: a single edit's span
127/// always contains [`diff_edit`]'s for the same transform, so there is nothing
128/// to gain over the caller's next step.
129///
130/// `edits` is not trusted: a slice that is not the exact transform from
131/// `prev_text` to `new_text` (a stale chain, a buffer that moved underneath
132/// it) is rejected rather than applied at meaningless offsets. Each step is
133/// proven to `fits` the text its predecessors produced *before* that text is
134/// sliced, and the composed result is proven to be `new_text` before anything
135/// is returned — so a stale chain costs at most the splices it got through,
136/// bounded by the caller's chain budget (`incremental::MAX_CHAIN_EDITS`).
137///
138/// Validating step by step rather than pre-folding the whole chain with
139/// [`try_apply_edits`] keeps the replay to one application: the fold would have
140/// to redo every `replace_range` the loop below already does, which on a
141/// 131 KB buffer is the same order as the splice it is guarding
142/// (`benches/reparse.rs`, `precise_chain`).
143pub fn reparse_edits(
144 prev_text: &str,
145 prev_green: &GreenNode,
146 prev_diags: &[ParseDiagnostic],
147 edits: &[Edit],
148 new_text: &str,
149) -> Option<Reparsed> {
150 if edits.len() < 2 {
151 return None;
152 }
153
154 let mut text = prev_text.to_string();
155 let mut green = prev_green.clone();
156 let mut diagnostics = prev_diags.to_vec();
157 let mut tier = ReparseTier::Token;
158
159 for edit in edits {
160 if !fits(&text, edit) {
161 return None;
162 }
163 let next = edit.apply(&text);
164 let step = reparse_impl(&text, &green, &diagnostics, edit, &next)?;
165 tier = tier.max(step.tier);
166 text = next;
167 green = step.green;
168 diagnostics = step.diagnostics;
169 }
170
171 // The chain applied cleanly, but that alone does not make it the transform
172 // the caller is asking about: a stale chain can fit and still land
173 // somewhere else entirely. This check is what the Tenet-4 assert below
174 // relies on to be comparing the same two texts.
175 if text != new_text {
176 return None;
177 }
178
179 let result = Reparsed {
180 green,
181 diagnostics,
182 tier,
183 };
184 assert_matches_full_parse(&result, new_text);
185 Some(result)
186}
187
188/// Tenet 4: a successful reparse yields a green tree *and* diagnostics
189/// byte-identical to a full parse of the edited text.
190///
191/// Debug builds only — this runs a whole `parse` plus two whole-tree
192/// [`fingerprint`] builds, so it is far too expensive to leave on. A chain
193/// asserts once on its composed result rather than once per step, which would
194/// otherwise make a debug-build language server unusable on a large file.
195#[cfg(debug_assertions)]
196fn assert_matches_full_parse(result: &Reparsed, new_text: &str) {
197 let full = crate::parser::parse(new_text);
198 debug_assert_eq!(
199 fingerprint(&SyntaxNode::new_root(result.green.clone())),
200 fingerprint(&full.cst),
201 "Tenet 4: reparse ({:?}) tree differs from full parse",
202 result.tier,
203 );
204 debug_assert_eq!(
205 result.diagnostics, full.diagnostics,
206 "Tenet 4: reparse ({:?}) diagnostics differ from full parse",
207 result.tier,
208 );
209}
210
211#[cfg(not(debug_assertions))]
212fn assert_matches_full_parse(_result: &Reparsed, _new_text: &str) {}
213
214/// Whether `edit`'s range is a well-formed byte range of `text`. Both tiers
215/// slice `prev_text` at the edit's offsets, so an edit that does not fit is a
216/// panic rather than a miss.
217fn fits(text: &str, edit: &Edit) -> bool {
218 edit.range.start <= edit.range.end
219 && edit.range.end <= text.len()
220 && text.is_char_boundary(edit.range.start)
221 && text.is_char_boundary(edit.range.end)
222}
223
224/// Tier dispatch: cheapest first.
225fn reparse_impl(
226 prev_text: &str,
227 prev_green: &GreenNode,
228 prev_diags: &[ParseDiagnostic],
229 edit: &Edit,
230 new_text: &str,
231) -> Option<Reparsed> {
232 reparse_token(prev_text, prev_green, prev_diags, edit)
233 .or_else(|| reparse_toplevel(prev_text, prev_green, prev_diags, edit, new_text))
234}
235
236/// Leaf kinds proved by the isolated single-token relex plus the two join
237/// probes ([`try_splice_plain_token`]). Their text never spans a newline and
238/// their lexing is local enough for those probes to prove a splice sound.
239/// Deliberately absent: `NEWLINE` (statement structure), `CHAR` and the
240/// numeric kinds (quote-context and juxtaposition traps outweigh the win), and
241/// the string prefix/suffix kinds (glued to their delimiters).
242/// `STRING_CONTENT` is handled too, but by a different proof — its lexing
243/// depends on the enclosing delimiter's mode, so it takes the whole-literal
244/// relex in [`try_splice_string_content`] instead.
245const TOKEN_REPARSE_KINDS: &[SyntaxKind] = &[
246 SyntaxKind::IDENT,
247 SyntaxKind::COMMENT,
248 SyntaxKind::BLOCK_COMMENT,
249 SyntaxKind::WHITESPACE,
250];
251
252/// Literal node kinds a `STRING_CONTENT` leaf may sit directly under. Error
253/// recovery can leave one under an `ERROR` node instead, which
254/// [`try_splice_string_content`] rejects: the proof needs the delimiters.
255const STRING_LITERAL_KINDS: &[SyntaxKind] = &[
256 SyntaxKind::STRING_LITERAL,
257 SyntaxKind::CMD_LITERAL,
258 // `var"…"`, a non-standard *identifier*. Nothing keys on its content text
259 // either, so a content edit is as safe here as in a plain string.
260 SyntaxKind::NONSTANDARD_IDENTIFIER,
261];
262
263/// Identifier texts the parser treats structurally even though they lex as
264/// plain `Ident`, so a same-kind relex proves nothing about the tree shape.
265/// True keywords (`where`, `mutable`, …) need no entry: they lex as their own
266/// kinds, and the same-kind guard rejects them.
267///
268/// This list is the one guard here with no compile-time link to what it
269/// describes, and the Tenet-4 oracle only catches an omission if a seeded edit
270/// happens to spell the missing word — so
271/// [`every_contextual_ident_is_listed`](tests::every_contextual_ident_is_listed)
272/// scans the grammar for `Ident`-text comparisons and fails if one names a word
273/// that is not here.
274const CONTEXTUAL_IDENTS: &[&str] = &[
275 "as", // import alias (structural.rs)
276 "abstract", // `abstract type` (expr.rs)
277 "primitive", // `primitive type` (expr.rs)
278 "type", // the `type` in `abstract`/`primitive` type decls (expr.rs)
279 "typegroup", // `typegroup` declaration (expr.rs)
280 "public", // top-level `public` statement (expr.rs)
281 "var", // `var"..."` quoting prefix (expr.rs)
282 "in", // iteration spec / infix operator (expr.rs)
283 "∈", // Unicode `in` (expr.rs)
284 "isa", // infix operator (expr.rs)
285 "doc", // `doc"..."` / `@doc` marker (expr.rs)
286 "outer", // future `for outer i` support
287];
288
289/// The token tier: when the edit is confined to a single eligible leaf, relex
290/// that leaf in isolation and splice it via [`SyntaxToken::replace_with`].
291///
292/// A pure insertion at a token boundary is ambiguous (it could extend either
293/// neighbor), so both candidates are tried, left first; whichever passes the
294/// guards yields the same new file text, and the guards plus the Tenet-4
295/// oracle arbitrate tree equality. Any guard failure returns `None`, which
296/// the caller answers with a full parse, so every bail here is safe.
297fn reparse_token(
298 prev_text: &str,
299 prev_green: &GreenNode,
300 prev_diags: &[ParseDiagnostic],
301 edit: &Edit,
302) -> Option<Reparsed> {
303 let root = SyntaxNode::new_root(prev_green.clone());
304 let (s, e) = (edit.range.start, edit.range.end);
305 let mut candidates: Vec<SyntaxToken> = Vec::with_capacity(2);
306 if s == e {
307 match root.token_at_offset(TextSize::new(s as u32)) {
308 TokenAtOffset::None => return None,
309 TokenAtOffset::Single(t) => candidates.push(t),
310 TokenAtOffset::Between(l, r) => candidates.extend([l, r]),
311 }
312 } else {
313 // A range spanning a token boundary covers a node, not a token, and
314 // bails here.
315 let range = TextRange::new(TextSize::new(s as u32), TextSize::new(e as u32));
316 candidates.push(root.covering_element(range).into_token()?);
317 }
318
319 candidates
320 .into_iter()
321 .find_map(|token| try_splice_token(prev_text, prev_diags, edit, &token))
322}
323
324/// Dispatch one candidate leaf to the proof its kind takes.
325fn try_splice_token(
326 prev_text: &str,
327 prev_diags: &[ParseDiagnostic],
328 edit: &Edit,
329 token: &SyntaxToken,
330) -> Option<Reparsed> {
331 match token.kind() {
332 SyntaxKind::STRING_CONTENT => try_splice_string_content(prev_text, prev_diags, edit, token),
333 kind if TOKEN_REPARSE_KINDS.contains(&kind) => {
334 try_splice_plain_token(prev_text, prev_diags, edit, token)
335 }
336 _ => None,
337 }
338}
339
340/// Run the stage-2 guards against one candidate leaf and splice on success.
341fn try_splice_plain_token(
342 prev_text: &str,
343 prev_diags: &[ParseDiagnostic],
344 edit: &Edit,
345 token: &SyntaxToken,
346) -> Option<Reparsed> {
347 // Newline ban: these kinds either cannot contain a newline or (block
348 // comments) rarely see one edited, and a newline moves statement
349 // boundaries, so bail early. The single-token relex guard below would
350 // catch most of these anyway; this is the cheap early-out. It does not
351 // apply to `STRING_CONTENT`, whose run legitimately spans newlines and
352 // whose proof does not depend on line structure.
353 if edit.insert.contains(['\n', '\r']) || prev_text[edit.range.clone()].contains(['\n', '\r']) {
354 return None;
355 }
356
357 let tr = token.text_range();
358 let (t0, t1) = (usize::from(tr.start()), usize::from(tr.end()));
359
360 // The leaf's new text, with the edit applied at its relative offset. A
361 // whole-leaf deletion leaves it empty, which the relex guard rejects.
362 let mut new_leaf = token.text().to_string();
363 new_leaf.replace_range((edit.range.start - t0)..(edit.range.end - t0), &edit.insert);
364
365 // Isolated relex: still exactly one token, same kind, spanning all of it.
366 // This alone rejects kind flips (ident ⇒ keyword, whitespace ⇒ newline)
367 // and splits (`=#` typed into a block comment).
368 let relexed = lex(&new_leaf);
369 let [only] = relexed.as_slice() else {
370 return None;
371 };
372 if only.start != 0 || only.end != new_leaf.len() || syntax_kind_for(only.kind) != token.kind() {
373 return None;
374 }
375
376 // Contextual identifiers parse structurally, so text changes into or out
377 // of one can reshape the tree under an unchanged token kind.
378 if token.kind() == SyntaxKind::IDENT
379 && (CONTEXTUAL_IDENTS.contains(&token.text())
380 || CONTEXTUAL_IDENTS.contains(&new_leaf.as_str()))
381 {
382 return None;
383 }
384
385 // Forward join: the next source character must not extend the token
386 // (`r` + `"…"` fusing into a prefixed string, an unterminated block
387 // comment swallowing its neighbor). At EOF probe with `\n`, which no
388 // eligible kind can absorb, so a leaf left unterminated at EOF (`#=` typed
389 // into the last block comment) still fails here.
390 let next_char = prev_text[t1..].chars().next().unwrap_or('\n');
391 let mut probe = new_leaf.clone();
392 probe.push(next_char);
393 let first = lex(&probe).into_iter().next()?;
394 if first.end != new_leaf.len() || syntax_kind_for(first.kind) != token.kind() {
395 return None;
396 }
397
398 // Backward join: the previous leaf plus the new text must relex to the
399 // same two tokens. Catches juxtaposition fusing (`2` + `e10` ⇒ one
400 // `Float`), and conservatively bails whenever the previous leaf's
401 // isolated lex diverges from its in-context one (e.g. a closing string
402 // delimiter reopening as a string).
403 if let Some(prev) = token.prev_token() {
404 let mut probe = prev.text().to_string();
405 let boundary = probe.len();
406 probe.push_str(&new_leaf);
407 let relexed = lex(&probe);
408 let [a, b] = relexed.as_slice() else {
409 return None;
410 };
411 if a.end != boundary
412 || syntax_kind_for(a.kind) != prev.kind()
413 || b.end != probe.len()
414 || syntax_kind_for(b.kind) != token.kind()
415 {
416 return None;
417 }
418 }
419
420 // Keep diagnostic remapping trivial: bail when any diagnostic touches the
421 // leaf, boundaries included. A clean same-kind relex introduces none of
422 // its own, so the rest keep their meaning and only need shifting.
423 if prev_diags.iter().any(|d| d.start <= t1 && d.end >= t0) {
424 return None;
425 }
426 let delta = edit.delta();
427 let shift = |pos: usize| (pos as isize + delta) as usize;
428 let diagnostics = prev_diags
429 .iter()
430 .map(|d| {
431 let mut d = d.clone();
432 if d.start >= t1 {
433 d.start = shift(d.start);
434 d.end = shift(d.end);
435 }
436 d
437 })
438 .collect();
439
440 let green = token.replace_with(GreenToken::new(token.kind().into(), &new_leaf));
441 Some(Reparsed {
442 green,
443 diagnostics,
444 tier: ReparseTier::Token,
445 })
446}
447
448/// `text` — the slice of the previous buffer starting at absolute offset
449/// `base` — with `edit` applied at its relative offset.
450fn edited_slice(text: &str, base: usize, edit: &Edit) -> String {
451 let mut out = text.to_string();
452 out.replace_range(
453 (edit.range.start - base)..(edit.range.end - base),
454 &edit.insert,
455 );
456 out
457}
458
459/// Every token under `node` as `(kind, end)`, with `end` relative to the node
460/// start. Tokens tile the node's text contiguously from `0` and the tree
461/// synthesizes none of its own (`tree_builder`), so this *is* the lexer's
462/// token run for that span and the ends alone pin every range.
463fn literal_token_ends(node: &SyntaxNode) -> Vec<(SyntaxKind, usize)> {
464 let base = usize::from(node.text_range().start());
465 node.descendants_with_tokens()
466 .filter_map(|el| el.into_token())
467 .map(|t| (t.kind(), usize::from(t.text_range().end()) - base))
468 .collect()
469}
470
471/// Whether an isolated [`lex`] of `text` reproduces `expected`: the same kinds
472/// in the same order, with the same ends, covering all of `text`. `grown`
473/// names the index whose end — and every end after it — is expected to have
474/// moved by that delta.
475fn relex_matches(
476 text: &str,
477 expected: &[(SyntaxKind, usize)],
478 grown: Option<(usize, isize)>,
479) -> bool {
480 let tokens = lex(text);
481 if tokens.len() != expected.len() {
482 return false;
483 }
484 for (i, (got, &(kind, end))) in tokens.iter().zip(expected).enumerate() {
485 let shift = match grown {
486 Some((from, delta)) if i >= from => delta,
487 _ => 0,
488 };
489 // A shift that runs an end past zero cannot describe any token, so it
490 // simply fails to match rather than wrapping into a huge `usize`.
491 let Ok(want) = usize::try_from(end as isize + shift) else {
492 return false;
493 };
494 if syntax_kind_for(got.kind) != kind || got.end != want {
495 return false;
496 }
497 }
498 tokens.last().is_some_and(|t| t.end == text.len())
499}
500
501/// The token tier's `STRING_CONTENT` path: relex the *whole enclosing string
502/// literal* in isolation and require it to reproduce its own token run with
503/// only the edited content leaf resized.
504///
505/// A content leaf cannot take [`try_splice_plain_token`]'s proof, because both
506/// of its probes assume a leaf that lexes the same standing alone: `abc` in
507/// isolation is an `Ident`, never `StringContent`, and a leaf following an
508/// interpolation (`"a$x b"`) has an `Ident` before it, which blows the
509/// two-token backward probe. Taking the delimiters along instead puts the
510/// isolated lexer into the right mode for free, so triple/raw/prefixed need no
511/// hand-written character table.
512///
513/// Why this is sound. Nothing outside the lexer reads `StringContent` *text*:
514/// the parser only pushes the token, docstring eligibility tests the first
515/// token's kind, and the flag passes are kind-based — so preserving the node's
516/// token-kind sequence preserves both the tree and the diagnostics. (Nor does
517/// triple-quoted dedenting intrude: that lives in the test-only JuliaSyntax
518/// projector, which rebuilds the string at projection time.) Given that:
519///
520/// * The bytes before the node are unchanged and the lexer is a deterministic
521/// left-to-right machine, so its mode stack at the node start is identical.
522/// That subsumes the backward join probe.
523/// * The relex proves the kinds are unchanged and that only the edited leaf's
524/// end, and the ends after it, moved by the delta. The bytes after the leaf
525/// *within* the node are therefore byte-identical — including the close
526/// delimiter and any suffix — so the mode stack is back to its starting
527/// depth at the node end and the rest of the file lexes identically, merely
528/// shifted. That subsumes the forward join probe, which could not have
529/// expressed a delimiter-plus-suffix tail with its one probe character
530/// anyway.
531///
532/// The second bullet needs the literal to be *terminated*, which gets its own
533/// check rather than leaning on an `UnterminatedLiteral` diagnostic's anchor.
534///
535/// Newlines are allowed here, unlike in [`try_splice_plain_token`]: a content
536/// run spans them freely (a triple-quoted body, and a single-quoted one too —
537/// no newline stops the run, which is why an unterminated literal reaches EOF),
538/// and a newline *inside* one emits no `NEWLINE` token, so no statement
539/// boundary moves. Diagnostics are byte offsets and do not care about lines.
540/// That is what puts Enter in a docstring on this tier.
541///
542/// The faithfulness relex (the unedited node text must reproduce the node's own
543/// tokens) is what makes the second bullet an induction rather than an
544/// assumption: an isolated [`lex`] starts from an empty mode stack and no token
545/// history, while in context the node may sit under a `Str`/`Interp` frame and
546/// `prev_ends_value` consults the preceding token. Proving faithfulness for the
547/// old text, plus an identical kind sequence, carries it to the new text.
548fn try_splice_string_content(
549 prev_text: &str,
550 prev_diags: &[ParseDiagnostic],
551 edit: &Edit,
552 token: &SyntaxToken,
553) -> Option<Reparsed> {
554 let node = token.parent()?;
555 if !STRING_LITERAL_KINDS.contains(&node.kind()) {
556 return None;
557 }
558 let nr = node.text_range();
559 let (n0, n1) = (usize::from(nr.start()), usize::from(nr.end()));
560
561 // Bail when any diagnostic touches the literal, boundaries included. The
562 // relex proves the token sequence unchanged, so the four string-related
563 // kinds would in fact survive a shift — but "no diagnostic may touch the
564 // literal we splice" is an invariant that keeps holding as diagnostics are
565 // added, and what it gives up (`x = "abc`, `"a"b`, `var"x"y`) is the error
566 // path, not the keystroke path.
567 //
568 // This runs *first*, before either relex: an unterminated literal runs to
569 // EOF, so its node is the rest of the file, and every keystroke inside one
570 // would otherwise pay two whole-file lexes to learn what a diagnostic
571 // already says (`benches/reparse.rs`, `rejected_attempt`).
572 if prev_diags.iter().any(|d| d.start <= n1 && d.end >= n0) {
573 return None;
574 }
575
576 let expected = literal_token_ends(&node);
577 let tr = token.text_range();
578 let (t0, t1) = (usize::from(tr.start()), usize::from(tr.end()));
579 // Ends are unique (tokens tile with positive width), so this identifies the
580 // leaf by index — which the edited relex then pins as the one that resized.
581 let leaf = expected
582 .iter()
583 .position(|&(kind, end)| kind == SyntaxKind::STRING_CONTENT && end == t1 - n0)?;
584
585 // Terminated: a matching close delimiter must follow the leaf. A prefixed
586 // literal's suffix is lexed as part of the same node, so nothing but the
587 // close can end the run.
588 let close = if node.kind() == SyntaxKind::CMD_LITERAL {
589 SyntaxKind::CMD_DELIM_CLOSE
590 } else {
591 SyntaxKind::STRING_DELIM_CLOSE
592 };
593 if !expected[leaf + 1..].iter().any(|&(kind, _)| kind == close) {
594 return None;
595 }
596
597 let old_node_text = &prev_text[n0..n1];
598 if !relex_matches(old_node_text, &expected, None) {
599 return None;
600 }
601 let delta = edit.delta();
602 let edited = edited_slice(old_node_text, n0, edit);
603 if !relex_matches(&edited, &expected, Some((leaf, delta))) {
604 return None;
605 }
606
607 let new_leaf = &edited[(t0 - n0)..((t1 - n0) as isize + delta) as usize];
608
609 // Every surviving diagnostic lies wholly outside the literal, so the ones
610 // past it shift and the rest keep their positions.
611 let shift = |pos: usize| (pos as isize + delta) as usize;
612 let diagnostics = prev_diags
613 .iter()
614 .map(|d| {
615 let mut d = d.clone();
616 if d.start >= n1 {
617 d.start = shift(d.start);
618 d.end = shift(d.end);
619 }
620 d
621 })
622 .collect();
623
624 let green = token.replace_with(GreenToken::new(SyntaxKind::STRING_CONTENT.into(), new_leaf));
625 Some(Reparsed {
626 green,
627 diagnostics,
628 tier: ReparseTier::Token,
629 })
630}
631
632/// Whether a token kind is trivia. Mirrors `TokKind::is_trivia` on the CST
633/// side.
634fn is_trivia_kind(kind: SyntaxKind) -> bool {
635 matches!(
636 kind,
637 SyntaxKind::WHITESPACE
638 | SyntaxKind::NEWLINE
639 | SyntaxKind::COMMENT
640 | SyntaxKind::BLOCK_COMMENT
641 )
642}
643
644/// Whether a `ROOT` child carries statement-level content: a positive-width
645/// node, or a loose non-trivia token (the unparseable-token recovery emits
646/// bare tokens at the root, without an `ERROR` wrapper). Zero-width nodes and
647/// trivia are gap material.
648fn is_significant_child(el: &crate::syntax::SyntaxElement) -> bool {
649 match el {
650 rowan::NodeOrToken::Node(n) => !n.text_range().is_empty(),
651 rowan::NodeOrToken::Token(t) => !is_trivia_kind(t.kind()),
652 }
653}
654
655/// Below this the tier always tries, whatever fraction of the file the region
656/// is: on a small file every parse in play is small, and the fraction test
657/// alone would refuse the common case of a file with one or two statements.
658const REGION_ALWAYS_TRY_BYTES: usize = 4 * 1024;
659
660/// Above this fraction of the file, a region is wide enough that answering it
661/// costs more than the full parse it is trying to avoid.
662const REGION_MAX_FRACTION: usize = 4;
663
664/// Whether a region is too wide to be worth attempting. The tier answers a
665/// region with a fragment parse of it *plus* up to three neighbor-sized
666/// boundary parses, so the attempt scales with the region — and a miss pays
667/// the full parse on top. A collapsed [`diff_edit`] of scattered edits spans
668/// everything between the first and the last, which is exactly how a region
669/// grows to most of the file; on ~131 KB that attempt cost about 12 ms against
670/// a 6.3 ms full parse (`benches/reparse.rs`, `scattered_via_diff_edit`).
671/// Bailing on the cheap evidence — the region's byte span, known before any
672/// parse — caps the loss at the full parse the caller was going to pay anyway.
673///
674/// This is a performance guard only. Like every other bail here it returns the
675/// caller to a full parse, so it cannot affect what a reparse yields.
676fn region_is_too_wide(region_len: usize, text_len: usize) -> bool {
677 region_len > REGION_ALWAYS_TRY_BYTES && region_len > text_len / REGION_MAX_FRACTION
678}
679
680/// The reparse region at the top-level tier: a contiguous run of `ROOT`
681/// children (statement nodes and loose trivia tokens both), identified by
682/// child index and old-text byte range. The boundary child kinds
683/// disambiguate zero-width diagnostics anchored exactly on a boundary.
684struct Region {
685 first: usize,
686 last: usize,
687 start: usize,
688 end: usize,
689 first_is_node: bool,
690 last_is_node: bool,
691}
692
693/// Select the contiguous run of `ROOT` children whose closed range touches
694/// the edit. Closed-interval touch makes boundary insertions extend into the
695/// neighbor they abut (blank-line typing selects the surrounding newline
696/// tokens; deleting a newline pulls in both neighbor statements). A
697/// zero-width child abutting the region start is absorbed: stray-closer
698/// recovery pairs a zero-width `ERROR` with its byte-bearing run at the same
699/// offset, and the fragment reparse regenerates both, so leaving it outside
700/// would duplicate it. `None` when the run covers every child (the fragment
701/// reparse would be the full parse) or `ROOT` is empty.
702fn select_region(root: &SyntaxNode, edit: &Edit) -> Option<Region> {
703 let (s, e) = (edit.range.start, edit.range.end);
704 let children: Vec<_> = root.children_with_tokens().collect();
705 let touches = |el: &crate::syntax::SyntaxElement| {
706 let r = el.text_range();
707 usize::from(r.start()) <= e && usize::from(r.end()) >= s
708 };
709 let mut first = children.iter().position(touches)?;
710 let last = children.len() - 1 - children.iter().rev().position(touches)?;
711
712 let start = usize::from(children[first].text_range().start());
713 while first > 0
714 && children[first - 1].text_range() == TextRange::empty(TextSize::new(start as u32))
715 {
716 first -= 1;
717 }
718
719 if first == 0 && last == children.len() - 1 {
720 return None;
721 }
722 Some(Region {
723 first,
724 last,
725 start,
726 end: usize::from(children[last].text_range().end()),
727 first_is_node: children[first].as_node().is_some(),
728 last_is_node: children[last].as_node().is_some(),
729 })
730}
731
732/// The nearest positive-width `ROOT` child nodes outside the region. Trivia
733/// tokens are gap material, and a zero-width `ERROR` node contributes no text
734/// (taking one as the forward neighbor would make the boundary guard
735/// vacuous), so both are skipped.
736fn sibling_nodes(root: &SyntaxNode, region: &Region) -> (Option<SyntaxNode>, Option<SyntaxNode>) {
737 let positive_node =
738 |el: crate::syntax::SyntaxElement| el.into_node().filter(|n| !n.text_range().is_empty());
739 let prev = root
740 .children_with_tokens()
741 .take(region.first)
742 .filter_map(positive_node)
743 .last();
744 let next = root
745 .children_with_tokens()
746 .skip(region.last + 1)
747 .find_map(positive_node);
748 (prev, next)
749}
750
751/// Whether no `ROOT` child (node or loose token) of `cst` straddles `seam`.
752/// Children tile the text, so no-straddle is equivalent to a boundary
753/// existing exactly at the seam: the two sides parse independently. A
754/// straddling node is a cross-boundary statement (trailing-operator
755/// continuation, a docstring fold, an unterminated block absorbing its
756/// neighbor); a straddling token is a lexical fusion (`#=` swallowing
757/// trailing trivia, an unterminated string).
758fn no_straddle(cst: &SyntaxNode, seam: usize) -> bool {
759 let seam = TextSize::new(seam as u32);
760 cst.children_with_tokens().all(|el| {
761 let r = el.text_range();
762 !(r.start() < seam && seam < r.end())
763 })
764}
765
766/// Splice the fragment's diagnostics into the previous ones, stream by
767/// stream ([`DiagnosticKind::stream`](crate::parser::diagnostics::DiagnosticKind::stream)):
768/// keep those before the region, drop those the fragment reparse
769/// regenerates, and shift those after by the edit delta. Emission order
770/// within a stream is positional across statements (the drive loop is
771/// per-line sequential, the flag passes walk in document order), so
772/// `before ++ fragment ++ after` per stream, concatenated in stream order,
773/// reproduces a full parse's vector.
774///
775/// A zero-width diagnostic exactly on a region boundary is ambiguous by
776/// position alone and is disambiguated by the boundary child's kind: a
777/// statement node owns a start-anchored marker at `start` (`const x` at the
778/// region's front) and an end-anchored one at `end`, while a trivia boundary
779/// child owns nothing, so the marker belongs to the statement just outside.
780/// This is sound only because the decoupling guards already rejected
781/// zero-gap node adjacency at the boundaries. A diagnostic genuinely
782/// straddling a boundary aborts the splice (`None`); regions are whole
783/// children and diagnostics never outgrow their statement, so this arm
784/// should be unreachable and is purely defensive.
785///
786/// A stream index outside [`DIAGNOSTIC_STREAMS`] aborts the splice too, rather
787/// than indexing out of bounds. [`DiagnosticKind::stream`] is an exhaustive
788/// match today, so this cannot fire — but a bail returns the caller to a full
789/// parse, whereas a panic here takes down an analysis query, and this module
790/// buys its whole safety argument from every failure being a bail.
791fn splice_diagnostics(
792 prev_diags: &[ParseDiagnostic],
793 frag_diags: &[ParseDiagnostic],
794 region: &Region,
795 delta: isize,
796) -> Option<Vec<ParseDiagnostic>> {
797 enum Class {
798 Before,
799 Drop,
800 After,
801 }
802 let classify = |d: &ParseDiagnostic| -> Option<Class> {
803 let (rs, re) = (region.start, region.end);
804 if d.start == d.end && d.start == rs {
805 return Some(if region.first_is_node {
806 Class::Drop
807 } else {
808 Class::Before
809 });
810 }
811 if d.start == d.end && d.start == re {
812 return Some(if region.last_is_node {
813 Class::Drop
814 } else {
815 Class::After
816 });
817 }
818 if d.end <= rs {
819 return Some(Class::Before);
820 }
821 if d.start >= re {
822 return Some(Class::After);
823 }
824 if d.start >= rs && d.end <= re {
825 return Some(Class::Drop);
826 }
827 None
828 };
829
830 let mut before: [Vec<ParseDiagnostic>; DIAGNOSTIC_STREAMS] = Default::default();
831 let mut after: [Vec<ParseDiagnostic>; DIAGNOSTIC_STREAMS] = Default::default();
832 for d in prev_diags {
833 let class = classify(d)?;
834 let stream = d.kind.stream();
835 if stream >= DIAGNOSTIC_STREAMS {
836 debug_assert!(false, "diagnostic stream {stream} is out of range");
837 return None;
838 }
839 match class {
840 Class::Before => before[stream].push(d.clone()),
841 Class::Drop => {}
842 Class::After => {
843 let mut d = d.clone();
844 d.start = (d.start as isize + delta) as usize;
845 d.end = (d.end as isize + delta) as usize;
846 after[stream].push(d);
847 }
848 }
849 }
850
851 let mut out = Vec::with_capacity(prev_diags.len() + frag_diags.len());
852 for stream in 0..DIAGNOSTIC_STREAMS {
853 out.append(&mut before[stream]);
854 out.extend(
855 frag_diags
856 .iter()
857 .filter(|d| d.kind.stream() == stream)
858 .map(|d| {
859 let mut d = d.clone();
860 d.start += region.start;
861 d.end += region.start;
862 d
863 }),
864 );
865 out.append(&mut after[stream]);
866 }
867 Some(out)
868}
869
870/// The top-level statement tier: reparse the run of `ROOT` children touching
871/// the edit with the public [`parse`](crate::parser::parse) (which reruns
872/// `fold_docstrings` and the flag passes on the fragment) and splice the
873/// resulting children and diagnostics in place.
874///
875/// The guards, in order: decoupling scans require a newline between each
876/// untouched sibling statement and the fragment's non-trivia content, because
877/// the drive loop's per-line junk recovery couples same-line siblings; the
878/// boundary parses then prove the seam survives — the neighbor plus the
879/// actual gap bytes plus the fragment must parse with a boundary exactly at
880/// the old one, which rejects trailing-operator continuation, docstring
881/// folds, forward absorption by an unterminated block, and lexical fusion.
882/// Diagnostics splice last (see [`splice_diagnostics`]).
883fn reparse_toplevel(
884 prev_text: &str,
885 prev_green: &GreenNode,
886 prev_diags: &[ParseDiagnostic],
887 edit: &Edit,
888 new_text: &str,
889) -> Option<Reparsed> {
890 let root = SyntaxNode::new_root(prev_green.clone());
891 let region = select_region(&root, edit)?;
892 if region_is_too_wide(region.end - region.start, prev_text.len()) {
893 return None;
894 }
895 let delta = edit.delta();
896 let frag_end = (region.end as isize + delta) as usize;
897 let fragment = &new_text[region.start..frag_end];
898
899 let frag = crate::parser::parse(fragment);
900 let (prev_node, next_node) = sibling_nodes(&root, ®ion);
901
902 let positive_nodes = |node: &SyntaxNode| {
903 node.children_with_tokens()
904 .filter_map(|el| el.into_node())
905 .filter(|n| !n.text_range().is_empty())
906 };
907
908 // Backward decoupling: a newline must separate the last significant
909 // content before the region (a statement node or a loose junk token, not
910 // just `prev_node`) from the fragment's first significant content (or
911 // all of a trivia-only fragment). The drive loop's per-line junk
912 // recovery couples same-line siblings, and the newline often lives in
913 // the unchanged gap outside the region, so scan `new_text`.
914 let prev_significant_end = root
915 .children_with_tokens()
916 .take(region.first)
917 .filter(is_significant_child)
918 .last()
919 .map(|el| usize::from(el.text_range().end()));
920 if let Some(from) = prev_significant_end {
921 let anchor = frag
922 .cst
923 .children_with_tokens()
924 .find(is_significant_child)
925 .map(|el| usize::from(el.text_range().start()))
926 .unwrap_or(fragment.len());
927 if !new_text[from..region.start + anchor].contains('\n') {
928 return None;
929 }
930 }
931
932 // Forward decoupling, in both texts: the next statement is kept, not
933 // reparsed, so its line must have been decoupled from the old region
934 // content too (deleting `x` from `x y` puts the kept junk `y` at line
935 // start, where a full parse reads it as a plain statement). The new-text
936 // scan anchors at the fragment's last statement *node* end, not its last
937 // token: an unterminated bracket consumes the separating newline into
938 // the statement, and a consumed newline decouples nothing.
939 if let Some(next) = &next_node {
940 let anchor = frag
941 .cst
942 .children_with_tokens()
943 .filter(is_significant_child)
944 .last()
945 .map(|el| usize::from(el.text_range().end()))
946 .unwrap_or(0);
947 let to = (usize::from(next.text_range().start()) as isize + delta) as usize;
948 if !new_text[region.start + anchor..to].contains('\n') {
949 return None;
950 }
951
952 let old_anchor = root
953 .children_with_tokens()
954 .skip(region.first)
955 .take(region.last + 1 - region.first)
956 .filter(is_significant_child)
957 .last()
958 .map(|el| usize::from(el.text_range().end()))
959 .or(prev_significant_end)
960 .unwrap_or(0);
961 if !prev_text[old_anchor..usize::from(next.text_range().start())].contains('\n') {
962 return None;
963 }
964 }
965
966 // Cross-region coupling: when the fragment carries no significant content,
967 // the two neighbors sit next to each other across it and can couple in a
968 // way neither boundary parse can see, because each concatenation contains
969 // only one of them. Deleting the `]` from `"doc" ]\nf() = 1` is the shape:
970 // the stray closer was all that kept the string from being the statement
971 // below it, and once it goes `fold_docstrings` folds the pair into a `DOC`
972 // spanning both seams. Parse the neighbors together, through the real gap
973 // bytes, and require both seams to survive. Only reachable when the region
974 // is reduced to trivia, so the extra parse is rare and neighbor-sized.
975 if !frag
976 .cst
977 .children_with_tokens()
978 .any(|el| is_significant_child(&el))
979 && let (Some(prev), Some(next)) = (&prev_node, &next_node)
980 {
981 let back_gap = &prev_text[usize::from(prev.text_range().end())..region.start];
982 let fwd_gap = &prev_text[region.end..usize::from(next.text_range().start())];
983 let guard = format!(
984 "{}{}{}{}{}",
985 prev.text(),
986 back_gap,
987 fragment,
988 fwd_gap,
989 next.text()
990 );
991 let back_seam = usize::from(prev.text_range().len()) + back_gap.len();
992 let fwd_seam = back_seam + fragment.len();
993 let parsed = crate::parser::parse(&guard);
994 if !no_straddle(&parsed.cst, back_seam) || !no_straddle(&parsed.cst, fwd_seam) {
995 return None;
996 }
997 }
998
999 // A zero-width diagnostic exactly on a guard seam is disambiguated by
1000 // the fragment's own boundary child, mirroring `splice_diagnostics`: a
1001 // statement node owns a start-anchored marker at its start and an
1002 // end-anchored one at its end; a trivia boundary child owns nothing.
1003 let frag_starts_with_node = frag
1004 .cst
1005 .children_with_tokens()
1006 .next()
1007 .is_some_and(|el| el.as_node().is_some());
1008 let frag_ends_with_node = frag
1009 .cst
1010 .children_with_tokens()
1011 .last()
1012 .is_some_and(|el| el.as_node().is_some());
1013
1014 // Backward boundary parse: previous statement + gap trivia + fragment
1015 // must keep a boundary exactly at the old one, and the fragment must
1016 // parse in context with exactly the diagnostics it produced in
1017 // isolation.
1018 if let Some(prev) = &prev_node {
1019 let back_gap = &prev_text[usize::from(prev.text_range().end())..region.start];
1020 let guard = format!("{}{}{}", prev.text(), back_gap, fragment);
1021 let seam = guard.len() - fragment.len();
1022 let parsed = crate::parser::parse(&guard);
1023 if !no_straddle(&parsed.cst, seam) {
1024 return None;
1025 }
1026 let guard_frag: Vec<ParseDiagnostic> = parsed
1027 .diagnostics
1028 .iter()
1029 .filter(|d| {
1030 d.start > seam
1031 || (d.start == seam && d.end > seam)
1032 || (d.start == seam && d.end == seam && frag_starts_with_node)
1033 })
1034 .map(|d| {
1035 let mut d = d.clone();
1036 d.start -= seam;
1037 d.end -= seam;
1038 d
1039 })
1040 .collect();
1041 if guard_frag != frag.diagnostics {
1042 return None;
1043 }
1044 }
1045
1046 // Forward boundary parse: fragment + gap trivia + next statement (or the
1047 // trailing trivia to EOF when no statement follows). Three checks
1048 // against one parse of the concatenation: no element straddles the seam;
1049 // the fragment reparses exactly as it did in isolation (an EOF-adjacent
1050 // junk run wraps differently once text follows it); and the kept next
1051 // statement reappears untouched — same start, kind, and length — since a
1052 // fragment ending in an unterminated bracket can consume the separating
1053 // newline and demote the next line to flat junk without anything
1054 // straddling the seam.
1055 let tail = match &next_node {
1056 Some(next) => {
1057 let gap = &prev_text[region.end..usize::from(next.text_range().start())];
1058 format!("{}{}", gap, next.text())
1059 }
1060 None => prev_text[region.end..].to_string(),
1061 };
1062 if !tail.is_empty() {
1063 let guard = format!("{fragment}{tail}");
1064 let seam = fragment.len();
1065 let parsed = crate::parser::parse(&guard);
1066 if !no_straddle(&parsed.cst, seam) {
1067 return None;
1068 }
1069 let signature = |el: crate::syntax::SyntaxElement| {
1070 (
1071 usize::from(el.text_range().start()),
1072 el.as_node().is_some(),
1073 el.kind(),
1074 usize::from(el.text_range().len()),
1075 )
1076 };
1077 let frag_sig: Vec<_> = frag.cst.children_with_tokens().map(signature).collect();
1078 let pre_seam: Vec<_> = parsed
1079 .cst
1080 .children_with_tokens()
1081 .filter(|el| {
1082 let r = el.text_range();
1083 // A zero-width child exactly at the seam pairs with a junk
1084 // run on the far side; it belongs to the tail.
1085 usize::from(r.end()) <= seam && !(r.is_empty() && usize::from(r.start()) == seam)
1086 })
1087 .map(signature)
1088 .collect();
1089 if pre_seam != frag_sig {
1090 return None;
1091 }
1092 // The fragment's in-context diagnostics must also match: a trailing
1093 // token at the fragment's EOF can emit differently once text follows
1094 // it (`:` with a newline after it flags `QuoteColonWhitespace`; at
1095 // EOF it does not).
1096 let guard_frag: Vec<ParseDiagnostic> = parsed
1097 .diagnostics
1098 .iter()
1099 .filter(|d| {
1100 d.end < seam
1101 || (d.end == seam && d.start < seam)
1102 || (d.start == seam && d.end == seam && frag_ends_with_node)
1103 })
1104 .cloned()
1105 .collect();
1106 if guard_frag != frag.diagnostics {
1107 return None;
1108 }
1109 if let Some(next) = &next_node {
1110 let gap_len = usize::from(next.text_range().start()) - region.end;
1111 let reparsed_next =
1112 positive_nodes(&parsed.cst).find(|n| usize::from(n.text_range().start()) >= seam);
1113 let intact = reparsed_next.is_some_and(|n| {
1114 usize::from(n.text_range().start()) == seam + gap_len
1115 && n.kind() == next.kind()
1116 && n.text_range().len() == next.text_range().len()
1117 });
1118 if !intact {
1119 return None;
1120 }
1121 }
1122 }
1123
1124 let diagnostics = splice_diagnostics(prev_diags, &frag.diagnostics, ®ion, delta)?;
1125
1126 let frag_green = frag.cst.green();
1127 let green = prev_green.splice_children(
1128 region.first..=region.last,
1129 frag_green.children().map(|c| c.to_owned()),
1130 );
1131 Some(Reparsed {
1132 green,
1133 diagnostics,
1134 tier: ReparseTier::TopLevel,
1135 })
1136}
1137
1138#[cfg(test)]
1139mod tests {
1140 use super::*;
1141
1142 fn edit(range: std::ops::Range<usize>, insert: &str) -> Edit {
1143 Edit {
1144 range,
1145 insert: insert.to_string(),
1146 }
1147 }
1148
1149 /// The grammar decides some tree shapes by an identifier's *text*, and the
1150 /// token tier's same-kind relex cannot see such a change — so every such
1151 /// word has to be in [`CONTEXTUAL_IDENTS`]. Nothing in the type system says
1152 /// so, and the oracle only surfaces an omission by luck, so scan the
1153 /// grammar for the two spellings those comparisons take and check the words
1154 /// they name.
1155 ///
1156 /// The known words are asserted to have been *found*, not just listed: if a
1157 /// comparison is ever rewritten into a shape this scan does not recognize,
1158 /// the test fails loudly rather than quietly covering nothing.
1159 #[test]
1160 fn every_contextual_ident_is_listed() {
1161 // `t.kind == TokKind::Ident && t.text == "word"` and the `matches!`
1162 // form `matches!(t.text.as_ref(), "word" | "other")`. Both are keyed on
1163 // `t.text`, which is what makes them text-structural.
1164 fn words_after(haystack: &str, marker: &str) -> Vec<String> {
1165 let mut out = Vec::new();
1166 for tail in haystack.split(marker).skip(1) {
1167 // Take the literals in the comparison that follows, stopping at
1168 // the first token that ends it.
1169 let stop = tail.find([';', '{', '}']).unwrap_or(tail.len());
1170 let mut rest = &tail[..stop];
1171 while let Some(open) = rest.find('"') {
1172 let Some(close) = rest[open + 1..].find('"') else {
1173 break;
1174 };
1175 out.push(rest[open + 1..open + 1 + close].to_string());
1176 rest = &rest[open + 1 + close + 1..];
1177 }
1178 }
1179 out
1180 }
1181
1182 // The grammar files, read from source: this is a lint over the code,
1183 // not over anything the crate exposes at runtime.
1184 let dir = std::path::Path::new(env!("CARGO_MANIFEST_DIR")).join("src/parser");
1185 let mut found: Vec<String> = Vec::new();
1186 for name in ["expr.rs", "structural.rs", "core.rs", "recovery.rs"] {
1187 let src = std::fs::read_to_string(dir.join(name))
1188 .unwrap_or_else(|e| panic!("read {name}: {e}"));
1189 found.extend(words_after(&src, "t.text =="));
1190 found.extend(words_after(&src, "t.text.as_ref(),"));
1191 found.extend(words_after(&src, "next.text =="));
1192 }
1193 assert!(
1194 !found.is_empty(),
1195 "the scan matched nothing — the grammar's `Ident`-text comparisons \
1196 have been rewritten, and this guard is no longer guarding anything"
1197 );
1198
1199 for word in &found {
1200 assert!(
1201 CONTEXTUAL_IDENTS.contains(&word.as_str()),
1202 "the grammar branches on the identifier text {word:?}, so an \
1203 edit into or out of it can reshape the tree under an unchanged \
1204 token kind. Add it to CONTEXTUAL_IDENTS.",
1205 );
1206 }
1207
1208 // Every word the scan is known to reach today. A rewrite that hides one
1209 // from the scan trips this rather than silently narrowing the guard.
1210 for expected in ["as", "abstract", "primitive", "type", "typegroup", "public"] {
1211 assert!(
1212 found.iter().any(|w| w == expected),
1213 "the scan no longer finds the {expected:?} comparison",
1214 );
1215 }
1216 }
1217
1218 #[test]
1219 fn fingerprint_is_stable_and_discriminating() {
1220 let a1 = fingerprint(&crate::parser::parse("x = 1\n").cst);
1221 let a2 = fingerprint(&crate::parser::parse("x = 1\n").cst);
1222 let b = fingerprint(&crate::parser::parse("x = 2\n").cst);
1223 assert_eq!(a1, a2);
1224 assert_ne!(a1, b);
1225 }
1226
1227 /// A numeric literal is deliberately outside `TOKEN_REPARSE_KINDS`, and a
1228 /// lone `x = 1` statement is the whole `ROOT`, so the top-level tier finds
1229 /// no decoupled region either.
1230 #[test]
1231 fn a_number_edit_in_a_lone_statement_falls_back() {
1232 let parsed = crate::parser::parse("x = 1\n");
1233 let green = parsed.cst.green().to_owned();
1234 let e = diff_edit("x = 1\n", "x = 2\n");
1235 assert!(reparse("x = 1\n", &green, &parsed.diagnostics, &e, "x = 2\n").is_none());
1236 }
1237
1238 /// The point of the `STRING_CONTENT` path: a docstring body edit stays at
1239 /// the token tier. `fold_docstrings` folds a docstring with the definition
1240 /// it documents into one `ROOT` child, so the statement tier would answer
1241 /// this by reparsing the docstring *and* the whole function under it.
1242 #[test]
1243 fn a_docstring_body_edit_stays_at_the_token_tier() {
1244 let old = "\"\"\"\ndoc\n\"\"\"\nfunction f(x)\n x + 1\nend\n";
1245 let parsed = crate::parser::parse(old);
1246 let green = parsed.cst.green().to_owned();
1247 let e = edit(5..5, "s");
1248 let new = e.apply(old);
1249
1250 let out = reparse(old, &green, &parsed.diagnostics, &e, &new)
1251 .expect("a docstring body edit is spliceable");
1252 assert_eq!(out.tier, ReparseTier::Token);
1253 }
1254
1255 /// An unterminated literal runs to EOF, so its node is the rest of the
1256 /// file. The diagnostic bail must reject a content edit inside one — and
1257 /// must do so before either relex, which is why it is hoisted to the top
1258 /// of `try_splice_string_content`.
1259 #[test]
1260 fn a_content_edit_in_an_unterminated_literal_falls_back() {
1261 let old = "x = 1\ny = \"abc\nz = 2\n";
1262 let parsed = crate::parser::parse(old);
1263 let green = parsed.cst.green().to_owned();
1264 // Offset 12 is inside the content run, which here swallows the rest of
1265 // the file; `UnterminatedLiteral` is anchored at the literal's start.
1266 let token = SyntaxNode::new_root(green.clone())
1267 .token_at_offset(TextSize::new(12))
1268 .right_biased()
1269 .expect("a leaf at offset 12");
1270 assert_eq!(token.kind(), SyntaxKind::STRING_CONTENT);
1271
1272 let e = edit(12..13, "X");
1273 let new = e.apply(old);
1274 assert!(try_splice_string_content(old, &parsed.diagnostics, &e, &token).is_none());
1275 // The statement tier may still answer it (the in-crate Tenet-4 assert
1276 // validates that splice); what matters is that the token tier declined.
1277 let tier = reparse(old, &green, &parsed.diagnostics, &e, &new).map(|r| r.tier);
1278 assert_ne!(tier, Some(ReparseTier::Token));
1279 }
1280
1281 /// An edit that does not fit `prev_text` is rejected rather than applied at
1282 /// offsets that would panic on a slice. `reparse_edits` proves its chain
1283 /// with `try_apply_edits`; the single-edit entry point has to check too.
1284 #[test]
1285 fn reparse_rejects_an_edit_that_does_not_fit() {
1286 let old = "α = 1\n";
1287 let parsed = crate::parser::parse(old);
1288 let green = parsed.cst.green().to_owned();
1289 // `new_text` is deliberately not the edit's result: a misfitting edit
1290 // has none, which is the whole point.
1291 let run = |e: Edit| reparse(old, &green, &parsed.diagnostics, &e, old);
1292
1293 // Past the end of the text.
1294 assert!(run(edit(99..99, "x")).is_none());
1295 // Off a char boundary (α is two bytes).
1296 assert!(run(edit(1..1, "x")).is_none());
1297 // Inverted range (built by hand: a literal `4..2` is a clippy error).
1298 let inverted = Edit {
1299 range: std::ops::Range { start: 4, end: 2 },
1300 insert: "x".to_string(),
1301 };
1302 assert!(run(inverted).is_none());
1303 }
1304
1305 /// A region wide enough that answering it costs more than the full parse it
1306 /// avoids is declined on the cheap evidence, before any fragment parse. The
1307 /// shape is a collapsed `diff_edit` of scattered edits, which spans
1308 /// everything between the first and the last.
1309 #[test]
1310 fn a_region_covering_most_of_a_large_file_is_declined() {
1311 let mut old = String::new();
1312 for i in 0..2000 {
1313 old.push_str(&format!("var_{i} = {i}\n"));
1314 }
1315 let parsed = crate::parser::parse(&old);
1316 let green = parsed.cst.green().to_owned();
1317
1318 // A narrow edit near the end still splices: the guard keys on the
1319 // region, not on the file size.
1320 let narrow = edit(old.len()..old.len(), "tail = 1\n");
1321 assert!(
1322 reparse(
1323 &old,
1324 &green,
1325 &parsed.diagnostics,
1326 &narrow,
1327 &narrow.apply(&old)
1328 )
1329 .is_some()
1330 );
1331
1332 // The same net change seen as one collapsed edit spanning most of the
1333 // file: `select_region` returns nearly every child, and the tier bails.
1334 // Both sites are inside an identifier, and the second is expressed
1335 // against the text the first produced (one byte longer), so the chain
1336 // below stays at the token tier. The last line is `var_1999 = 1999\n`,
1337 // 16 bytes, whose name runs from its start.
1338 let scattered = vec![edit(4..4, "X"), edit(old.len() - 11..old.len() - 11, "X")];
1339 let new = apply_edits(&old, &scattered);
1340 let collapsed = diff_edit(&old, &new);
1341 assert!(collapsed.range.len() > old.len() / 2, "{collapsed:?}");
1342 assert!(reparse(&old, &green, &parsed.diagnostics, &collapsed, &new).is_none());
1343
1344 // Replaying the chain still answers it at the token tier — the guard
1345 // costs the precise-edit path nothing.
1346 let out = reparse_edits(&old, &green, &parsed.diagnostics, &scattered, &new)
1347 .expect("the precise chain is unaffected");
1348 assert_eq!(out.tier, ReparseTier::Token);
1349 }
1350
1351 /// The width guard is a fraction *and* a floor: a small file's region is
1352 /// always worth trying, however much of the file it covers.
1353 #[test]
1354 fn the_region_width_guard_leaves_small_files_alone() {
1355 assert!(!region_is_too_wide(4 * 1024, 4 * 1024));
1356 assert!(!region_is_too_wide(64 * 1024, 1024 * 1024));
1357 assert!(region_is_too_wide(64 * 1024, 128 * 1024));
1358 }
1359
1360 /// The shape `reparse_edits` exists for: two identifier edits far apart,
1361 /// which `diff_edit` would collapse into one span covering both statements.
1362 #[test]
1363 fn reparse_edits_chains_scattered_edits() {
1364 let old = "alpha = 1\nfiller = 2\nomega = 3\n";
1365 let parsed = crate::parser::parse(old);
1366 let green = parsed.cst.green().to_owned();
1367 // `alpha` -> `alphaX`, then `omega` -> `omegaX` at its post-first-edit
1368 // offset (one byte further along).
1369 let edits = vec![edit(5..5, "X"), edit(27..27, "X")];
1370 let new = apply_edits(old, &edits);
1371 assert_eq!(new, "alphaX = 1\nfiller = 2\nomegaX = 3\n");
1372
1373 // The collapsed diff spans from `alpha` all the way to `omega`, so the
1374 // cheap token tier cannot touch it and the whole run of statements has
1375 // to be reparsed. Replaying the chain keeps both splices at the token
1376 // tier. On three statements that is a wash; on a real file the
1377 // difference is two orders of magnitude (`benches/reparse.rs`), which
1378 // is why the chain is tried first.
1379 let collapsed = diff_edit(old, &new);
1380 assert!(collapsed.range.len() > 20, "{collapsed:?} should be coarse");
1381 let via_diff = reparse(old, &green, &parsed.diagnostics, &collapsed, &new)
1382 .expect("the whole-file region is still spliceable at this size");
1383 assert_eq!(via_diff.tier, ReparseTier::TopLevel);
1384
1385 let out = reparse_edits(old, &green, &parsed.diagnostics, &edits, &new)
1386 .expect("two identifier edits should chain");
1387 assert_eq!(out.tier, ReparseTier::Token);
1388 assert_eq!(
1389 fingerprint(&SyntaxNode::new_root(out.green)),
1390 fingerprint(&crate::parser::parse(&new).cst),
1391 );
1392 }
1393
1394 /// The chain reports the most expensive tier it had to reach.
1395 #[test]
1396 fn reparse_edits_reports_the_max_tier() {
1397 let old = "alpha = 1\nfiller = 2\nomega = 3\n";
1398 let parsed = crate::parser::parse(old);
1399 let green = parsed.cst.green().to_owned();
1400 // An ident edit (token tier) plus a whole added statement (top level),
1401 // appended at the post-first-edit end of the buffer.
1402 let edits = vec![edit(5..5, "X"), edit(32..32, "beta = 4\n")];
1403 let new = apply_edits(old, &edits);
1404
1405 let out = reparse_edits(old, &green, &parsed.diagnostics, &edits, &new)
1406 .expect("ident edit plus a new statement should chain");
1407 assert_eq!(out.tier, ReparseTier::TopLevel);
1408 }
1409
1410 /// Fewer than two edits is always `diff_edit`'s job: its span for the same
1411 /// transform is contained in the single edit's, and the caller tried it.
1412 #[test]
1413 fn reparse_edits_declines_short_chains() {
1414 let old = "alpha = 1\nomega = 3\n";
1415 let parsed = crate::parser::parse(old);
1416 let green = parsed.cst.green().to_owned();
1417
1418 assert!(reparse_edits(old, &green, &parsed.diagnostics, &[], old).is_none());
1419 let one = vec![edit(5..5, "X")];
1420 let new = apply_edits(old, &one);
1421 assert!(reparse_edits(old, &green, &parsed.diagnostics, &one, &new).is_none());
1422 }
1423
1424 /// A chain that is not the exact transform between the two texts is
1425 /// rejected before any offset is trusted — both when it simply does not
1426 /// reproduce the target, and when it does not fit the source at all.
1427 #[test]
1428 fn reparse_edits_rejects_a_stale_chain() {
1429 let old = "alpha = 1\nfiller = 2\nomega = 3\n";
1430 let parsed = crate::parser::parse(old);
1431 let green = parsed.cst.green().to_owned();
1432
1433 // Applies cleanly, but the buffer moved on underneath it.
1434 let edits = vec![edit(5..5, "X"), edit(26..26, "X")];
1435 let elsewhere = "alphaX = 1\nfiller = 2\nomegaX = 3\nextra = 4\n";
1436 assert!(reparse_edits(old, &green, &parsed.diagnostics, &edits, elsewhere).is_none());
1437
1438 // Does not even fit the source text.
1439 let past_end = vec![edit(500..500, "X"), edit(501..501, "X")];
1440 assert!(reparse_edits(old, &green, &parsed.diagnostics, &past_end, old).is_none());
1441 }
1442
1443 /// One unhandleable step aborts the whole chain — the caller full-parses.
1444 #[test]
1445 fn reparse_edits_aborts_on_an_unhandleable_step() {
1446 let old = "alpha = 1\nfiller = 2\nomega = 3\n";
1447 let parsed = crate::parser::parse(old);
1448 let green = parsed.cst.green().to_owned();
1449 // The second edit opens an unterminated string, which no tier splices.
1450 let edits = vec![edit(5..5, "X"), edit(26..26, "\"")];
1451 let new = apply_edits(old, &edits);
1452 assert!(reparse_edits(old, &green, &parsed.diagnostics, &edits, &new).is_none());
1453 }
1454}