1 //! This is an NFA-based parser, which calls out to the main Rust parser for named non-terminals
2 //! (which it commits to fully when it hits one in a grammar). There's a set of current NFA threads
3 //! and a set of next ones. Instead of NTs, we have a special case for Kleene star. The big-O, in
4 //! pathological cases, is worse than traditional use of NFA or Earley parsing, but it's an easier
5 //! fit for Macro-by-Example-style rules.
7 //! (In order to prevent the pathological case, we'd need to lazily construct the resulting
8 //! `NamedMatch`es at the very end. It'd be a pain, and require more memory to keep around old
9 //! matcher positions, but it would also save overhead)
11 //! We don't say this parser uses the Earley algorithm, because it's unnecessarily inaccurate.
12 //! The macro parser restricts itself to the features of finite state automata. Earley parsers
13 //! can be described as an extension of NFAs with completion rules, prediction rules, and recursion.
15 //! Quick intro to how the parser works:
17 //! A "matcher position" (a.k.a. "position" or "mp") is a dot in the middle of a matcher, usually
18 //! written as a `·`. For example `· a $( a )* a b` is one, as is `a $( · a )* a b`.
20 //! The parser walks through the input a token at a time, maintaining a list
21 //! of threads consistent with the current position in the input string: `cur_mps`.
23 //! As it processes them, it fills up `eof_mps` with threads that would be valid if
24 //! the macro invocation is now over, `bb_mps` with threads that are waiting on
25 //! a Rust non-terminal like `$e:expr`, and `next_mps` with threads that are waiting
26 //! on a particular token. Most of the logic concerns moving the · through the
27 //! repetitions indicated by Kleene stars. The rules for moving the · without
28 //! consuming any input are called epsilon transitions. It only advances or calls
29 //! out to the real Rust parser when no `cur_mps` threads remain.
34 //! Start parsing a a a a b against [· a $( a )* a b].
36 //! Remaining input: a a a a b
37 //! next: [· a $( a )* a b]
39 //! - - - Advance over an a. - - -
41 //! Remaining input: a a a b
42 //! cur: [a · $( a )* a b]
43 //! Descend/Skip (first position).
44 //! next: [a $( · a )* a b] [a $( a )* · a b].
46 //! - - - Advance over an a. - - -
48 //! Remaining input: a a b
49 //! cur: [a $( a · )* a b] [a $( a )* a · b]
50 //! Follow epsilon transition: Finish/Repeat (first position)
51 //! next: [a $( a )* · a b] [a $( · a )* a b] [a $( a )* a · b]
53 //! - - - Advance over an a. - - - (this looks exactly like the last step)
55 //! Remaining input: a b
56 //! cur: [a $( a · )* a b] [a $( a )* a · b]
57 //! Follow epsilon transition: Finish/Repeat (first position)
58 //! next: [a $( a )* · a b] [a $( · a )* a b] [a $( a )* a · b]
60 //! - - - Advance over an a. - - - (this looks exactly like the last step)
62 //! Remaining input: b
63 //! cur: [a $( a · )* a b] [a $( a )* a · b]
64 //! Follow epsilon transition: Finish/Repeat (first position)
65 //! next: [a $( a )* · a b] [a $( · a )* a b] [a $( a )* a · b]
67 //! - - - Advance over a b. - - -
69 //! Remaining input: ''
70 //! eof: [a $( a )* a b ·]
73 crate use NamedMatch::*;
74 crate use ParseResult::*;
76 use crate::mbe::{KleeneOp, TokenTree};
78 use rustc_ast::token::{self, DocComment, Nonterminal, NonterminalKind, Token};
79 use rustc_parse::parser::{NtOrTt, Parser};
80 use rustc_session::parse::ParseSess;
81 use rustc_span::symbol::MacroRulesNormalizedIdent;
84 use smallvec::{smallvec, SmallVec};
86 use rustc_data_structures::fx::FxHashMap;
87 use rustc_data_structures::sync::Lrc;
88 use rustc_span::symbol::Ident;
90 use std::collections::hash_map::Entry::{Occupied, Vacant};
92 // One element is enough to cover 95-99% of vectors for most benchmarks. Also, vectors longer than
93 // one frequently have many elements, not just two or three.
94 type NamedMatchVec = SmallVec<[NamedMatch; 1]>;
96 // This type is used a lot. Make sure it doesn't unintentionally get bigger.
97 #[cfg(all(target_arch = "x86_64", target_pointer_width = "64"))]
98 rustc_data_structures::static_assert_size!(NamedMatchVec, 48);
100 /// A unit within a matcher that a `MatcherPos` can refer to. Similar to (and derived from)
101 /// `mbe::TokenTree`, but designed specifically for fast and easy traversal during matching.
102 /// Notable differences to `mbe::TokenTree`:
103 /// - It is non-recursive, i.e. there is no nesting.
104 /// - The end pieces of each sequence (the separator, if present, and the Kleene op) are
105 /// represented explicitly, as is the very end of the matcher.
107 /// This means a matcher can be represented by `&[MatcherLoc]`, and traversal mostly involves
108 /// simply incrementing the current matcher position index by one.
109 pub(super) enum MatcherLoc {
116 num_metavar_decls: usize,
117 idx_first_after: usize,
121 SequenceKleeneOpNoSep {
128 SequenceKleeneOpAfterSep {
134 kind: Option<NonterminalKind>,
141 pub(super) fn compute_locs(sess: &ParseSess, matcher: &[TokenTree]) -> Vec<MatcherLoc> {
145 locs: &mut Vec<MatcherLoc>,
146 next_metavar: &mut usize,
151 TokenTree::Token(token) => {
152 locs.push(MatcherLoc::Token { token: token.clone() });
154 TokenTree::Delimited(_, delimited) => {
155 locs.push(MatcherLoc::Delimited);
156 inner(sess, &delimited.all_tts, locs, next_metavar, seq_depth);
158 TokenTree::Sequence(_, seq) => {
159 // We can't determine `idx_first_after` and construct the final
160 // `MatcherLoc::Sequence` until after `inner()` is called and the sequence end
161 // pieces are processed. So we push a dummy value (`Eof` is cheapest to
162 // construct) now, and overwrite it with the proper value below.
163 let dummy = MatcherLoc::Eof;
166 let next_metavar_orig = *next_metavar;
167 let op = seq.kleene.op;
168 let idx_first = locs.len();
169 let idx_seq = idx_first - 1;
170 inner(sess, &seq.tts, locs, next_metavar, seq_depth + 1);
172 if let Some(separator) = &seq.separator {
173 locs.push(MatcherLoc::SequenceSep { separator: separator.clone() });
174 locs.push(MatcherLoc::SequenceKleeneOpAfterSep { idx_first });
176 locs.push(MatcherLoc::SequenceKleeneOpNoSep { op, idx_first });
179 // Overwrite the dummy value pushed above with the proper value.
180 locs[idx_seq] = MatcherLoc::Sequence {
182 num_metavar_decls: seq.num_captures,
183 idx_first_after: locs.len(),
184 next_metavar: next_metavar_orig,
188 &TokenTree::MetaVarDecl(span, bind, kind) => {
189 locs.push(MatcherLoc::MetaVarDecl {
193 next_metavar: *next_metavar,
198 TokenTree::MetaVar(..) | TokenTree::MetaVarExpr(..) => unreachable!(),
203 let mut locs = vec![];
204 let mut next_metavar = 0;
205 inner(sess, matcher, &mut locs, &mut next_metavar, /* seq_depth */ 0);
207 // A final entry is needed for eof.
208 locs.push(MatcherLoc::Eof);
213 /// A single matcher position, representing the state of matching.
215 /// The index into `TtParser::locs`, which represents the "dot".
218 /// The matches made against metavar decls so far. On a successful match, this vector ends up
219 /// with one element per metavar decl in the matcher. Each element records token trees matched
220 /// against the relevant metavar by the black box parser. An element will be a `MatchedSeq` if
221 /// the corresponding metavar decl is within a sequence.
222 matches: Lrc<NamedMatchVec>,
225 // This type is used a lot. Make sure it doesn't unintentionally get bigger.
226 #[cfg(all(target_arch = "x86_64", target_pointer_width = "64"))]
227 rustc_data_structures::static_assert_size!(MatcherPos, 16);
230 /// Adds `m` as a named match for the `metavar_idx`-th metavar. There are only two call sites,
231 /// and both are hot enough to be always worth inlining.
233 fn push_match(&mut self, metavar_idx: usize, seq_depth: usize, m: NamedMatch) {
234 let matches = Lrc::make_mut(&mut self.matches);
237 // We are not within a sequence. Just append `m`.
238 assert_eq!(metavar_idx, matches.len());
242 // We are within a sequence. Find the final `MatchedSeq` at the appropriate depth
243 // and append `m` to its vector.
244 let mut curr = &mut matches[metavar_idx];
245 for _ in 0..seq_depth - 1 {
248 let seq = Lrc::make_mut(seq);
249 curr = seq.last_mut().unwrap();
256 let seq = Lrc::make_mut(seq);
266 enum EofMatcherPositions {
272 /// Represents the possible results of an attempted parse.
273 crate enum ParseResult<T> {
274 /// Parsed successfully.
276 /// Arm failed to match. If the second parameter is `token::Eof`, it indicates an unexpected
277 /// end of macro invocation. Otherwise, it indicates that no rules expected the given token.
278 Failure(Token, &'static str),
279 /// Fatal error (malformed macro?). Abort compilation.
280 Error(rustc_span::Span, String),
284 /// A `ParseResult` where the `Success` variant contains a mapping of
285 /// `MacroRulesNormalizedIdent`s to `NamedMatch`es. This represents the mapping
286 /// of metavars to the token trees they bind to.
287 crate type NamedParseResult = ParseResult<FxHashMap<MacroRulesNormalizedIdent, NamedMatch>>;
289 /// Count how many metavars declarations are in `matcher`.
290 pub(super) fn count_metavar_decls(matcher: &[TokenTree]) -> usize {
294 TokenTree::MetaVarDecl(..) => 1,
295 TokenTree::Sequence(_, seq) => seq.num_captures,
296 TokenTree::Delimited(_, delim) => count_metavar_decls(delim.inner_tts()),
297 TokenTree::Token(..) => 0,
298 TokenTree::MetaVar(..) | TokenTree::MetaVarExpr(..) => unreachable!(),
303 /// `NamedMatch` is a pattern-match result for a single metavar. All
304 /// `MatchedNonterminal`s in the `NamedMatch` have the same non-terminal type
305 /// (expr, item, etc).
307 /// The in-memory structure of a particular `NamedMatch` represents the match
308 /// that occurred when a particular subset of a matcher was applied to a
309 /// particular token tree.
311 /// The width of each `MatchedSeq` in the `NamedMatch`, and the identity of
312 /// the `MatchedNtNonTts`s, will depend on the token tree it was applied
313 /// to: each `MatchedSeq` corresponds to a single repetition in the originating
314 /// token tree. The depth of the `NamedMatch` structure will therefore depend
315 /// only on the nesting depth of repetitions in the originating token tree it
316 /// was derived from.
318 /// In layman's terms: `NamedMatch` will form a tree representing nested matches of a particular
319 /// meta variable. For example, if we are matching the following macro against the following
323 /// macro_rules! foo {
324 /// ($($($x:ident),+);+) => {}
327 /// foo!(a, b, c, d; a, b, c, d, e);
330 /// Then, the tree will have the following shape:
335 /// MatchedNonterminal(a),
336 /// MatchedNonterminal(b),
337 /// MatchedNonterminal(c),
338 /// MatchedNonterminal(d),
341 /// MatchedNonterminal(a),
342 /// MatchedNonterminal(b),
343 /// MatchedNonterminal(c),
344 /// MatchedNonterminal(d),
345 /// MatchedNonterminal(e),
349 #[derive(Debug, Clone)]
350 crate enum NamedMatch {
351 MatchedSeq(Lrc<NamedMatchVec>),
353 // A metavar match of type `tt`.
354 MatchedTokenTree(rustc_ast::tokenstream::TokenTree),
356 // A metavar match of any type other than `tt`.
357 MatchedNonterminal(Lrc<Nonterminal>),
360 /// Performs a token equality check, ignoring syntax context (that is, an unhygienic comparison)
361 fn token_name_eq(t1: &Token, t2: &Token) -> bool {
362 if let (Some((ident1, is_raw1)), Some((ident2, is_raw2))) = (t1.ident(), t2.ident()) {
363 ident1.name == ident2.name && is_raw1 == is_raw2
364 } else if let (Some(ident1), Some(ident2)) = (t1.lifetime(), t2.lifetime()) {
365 ident1.name == ident2.name
371 // Note: the vectors could be created and dropped within `parse_tt`, but to avoid excess
372 // allocations we have a single vector fo each kind that is cleared and reused repeatedly.
373 pub struct TtParser {
376 /// The set of current mps to be processed. This should be empty by the end of a successful
377 /// execution of `parse_tt_inner`.
378 cur_mps: Vec<MatcherPos>,
380 /// The set of newly generated mps. These are used to replenish `cur_mps` in the function
382 next_mps: Vec<MatcherPos>,
384 /// The set of mps that are waiting for the black-box parser.
385 bb_mps: Vec<MatcherPos>,
387 /// Pre-allocate an empty match array, so it can be cloned cheaply for macros with many rules
388 /// that have no metavars.
389 empty_matches: Lrc<NamedMatchVec>,
393 pub(super) fn new(macro_name: Ident) -> TtParser {
399 empty_matches: Lrc::new(smallvec![]),
403 /// Process the matcher positions of `cur_mps` until it is empty. In the process, this will
404 /// produce more mps in `next_mps` and `bb_mps`.
408 /// `Some(result)` if everything is finished, `None` otherwise. Note that matches are kept
409 /// track of through the mps generated.
413 matcher: &[MatcherLoc],
415 ) -> Option<NamedParseResult> {
416 // Matcher positions that would be valid if the macro invocation was over now. Only
417 // modified if `token == Eof`.
418 let mut eof_mps = EofMatcherPositions::None;
420 while let Some(mut mp) = self.cur_mps.pop() {
421 match &matcher[mp.idx] {
422 MatcherLoc::Token { token: t } => {
423 // If it's a doc comment, we just ignore it and move on to the next tt in the
424 // matcher. This is a bug, but #95267 showed that existing programs rely on
425 // this behaviour, and changing it would require some care and a transition
428 // If the token matches, we can just advance the parser.
430 // Otherwise, this match has failed, there is nothing to do, and hopefully
431 // another mp in `cur_mps` will match.
432 if matches!(t, Token { kind: DocComment(..), .. }) {
434 self.cur_mps.push(mp);
435 } else if token_name_eq(&t, token) {
437 self.next_mps.push(mp);
440 MatcherLoc::Delimited => {
441 // Entering the delimeter is trivial.
443 self.cur_mps.push(mp);
445 &MatcherLoc::Sequence {
452 // Install an empty vec for each metavar within the sequence.
453 for metavar_idx in next_metavar..next_metavar + num_metavar_decls {
457 MatchedSeq(self.empty_matches.clone()),
461 if op == KleeneOp::ZeroOrMore || op == KleeneOp::ZeroOrOne {
462 // Try zero matches of this sequence, by skipping over it.
463 self.cur_mps.push(MatcherPos {
464 idx: idx_first_after,
465 matches: mp.matches.clone(), // a cheap clone
469 // Try one or more matches of this sequence, by entering it.
471 self.cur_mps.push(mp);
473 &MatcherLoc::SequenceKleeneOpNoSep { op, idx_first } => {
474 // We are past the end of a sequence with no separator. Try ending the
475 // sequence. If that's not possible, `ending_mp` will fail quietly when it is
476 // processed next time around the loop.
477 let ending_mp = MatcherPos {
478 idx: mp.idx + 1, // +1 skips the Kleene op
479 matches: mp.matches.clone(), // a cheap clone
481 self.cur_mps.push(ending_mp);
483 if op != KleeneOp::ZeroOrOne {
484 // Try another repetition.
486 self.cur_mps.push(mp);
489 MatcherLoc::SequenceSep { separator } => {
490 // We are past the end of a sequence with a separator but we haven't seen the
491 // separator yet. Try ending the sequence. If that's not possible, `ending_mp`
492 // will fail quietly when it is processed next time around the loop.
493 let ending_mp = MatcherPos {
494 idx: mp.idx + 2, // +2 skips the separator and the Kleene op
495 matches: mp.matches.clone(), // a cheap clone
497 self.cur_mps.push(ending_mp);
499 if token_name_eq(token, separator) {
500 // The separator matches the current token. Advance past it.
502 self.next_mps.push(mp);
505 &MatcherLoc::SequenceKleeneOpAfterSep { idx_first } => {
506 // We are past the sequence separator. This can't be a `?` Kleene op, because
507 // they don't permit separators. Try another repetition.
509 self.cur_mps.push(mp);
511 &MatcherLoc::MetaVarDecl { span, kind, .. } => {
512 // Built-in nonterminals never start with these tokens, so we can eliminate
513 // them from consideration. We use the span of the metavariable declaration
514 // to determine any edition-specific matching behavior for non-terminals.
515 if let Some(kind) = kind {
516 if Parser::nonterminal_may_begin_with(kind, token) {
517 self.bb_mps.push(mp);
520 // Both this check and the one in `nameize` are necessary, surprisingly.
521 if sess.missing_fragment_specifiers.borrow_mut().remove(&span).is_some() {
522 // E.g. `$e` instead of `$e:expr`.
523 return Some(Error(span, "missing fragment specifier".to_string()));
528 // We are past the matcher's end, and not in a sequence. Try to end things.
529 debug_assert_eq!(mp.idx, matcher.len() - 1);
530 if *token == token::Eof {
531 eof_mps = match eof_mps {
532 EofMatcherPositions::None => EofMatcherPositions::One(mp),
533 EofMatcherPositions::One(_) | EofMatcherPositions::Multiple => {
534 EofMatcherPositions::Multiple
542 // If we reached the end of input, check that there is EXACTLY ONE possible matcher.
543 // Otherwise, either the parse is ambiguous (which is an error) or there is a syntax error.
544 if *token == token::Eof {
546 EofMatcherPositions::One(mut eof_mp) => {
547 // Need to take ownership of the matches from within the `Lrc`.
548 Lrc::make_mut(&mut eof_mp.matches);
549 let matches = Lrc::try_unwrap(eof_mp.matches).unwrap().into_iter();
550 self.nameize(sess, matcher, matches)
552 EofMatcherPositions::Multiple => {
553 Error(token.span, "ambiguity: multiple successful parses".to_string())
555 EofMatcherPositions::None => Failure(
558 if token.span.is_dummy() { token.span } else { token.span.shrink_to_hi() },
560 "missing tokens in macro arguments",
568 /// Match the token stream from `parser` against `matcher`.
569 pub(super) fn parse_tt(
571 parser: &mut Cow<'_, Parser<'_>>,
572 matcher: &[MatcherLoc],
573 ) -> NamedParseResult {
574 // A queue of possible matcher positions. We initialize it with the matcher position in
575 // which the "dot" is before the first token of the first token tree in `matcher`.
576 // `parse_tt_inner` then processes all of these possible matcher positions and produces
577 // possible next positions into `next_mps`. After some post-processing, the contents of
578 // `next_mps` replenish `cur_mps` and we start over again.
579 self.cur_mps.clear();
580 self.cur_mps.push(MatcherPos { idx: 0, matches: self.empty_matches.clone() });
583 self.next_mps.clear();
586 // Process `cur_mps` until either we have finished the input or we need to get some
587 // parsing from the black-box parser done.
588 if let Some(res) = self.parse_tt_inner(&parser.sess, matcher, &parser.token) {
592 // `parse_tt_inner` handled all of `cur_mps`, so it's empty.
593 assert!(self.cur_mps.is_empty());
595 // Error messages here could be improved with links to original rules.
596 match (self.next_mps.len(), self.bb_mps.len()) {
598 // There are no possible next positions AND we aren't waiting for the black-box
599 // parser: syntax error.
601 parser.token.clone(),
602 "no rules expected this token in macro call",
607 // Dump all possible `next_mps` into `cur_mps` for the next iteration. Then
608 // process the next token.
609 self.cur_mps.extend(self.next_mps.drain(..));
610 parser.to_mut().bump();
614 // We need to call the black-box parser to get some nonterminal.
615 let mut mp = self.bb_mps.pop().unwrap();
616 let loc = &matcher[mp.idx];
617 if let &MatcherLoc::MetaVarDecl {
625 // We use the span of the metavariable declaration to determine any
626 // edition-specific matching behavior for non-terminals.
627 let nt = match parser.to_mut().parse_nonterminal(kind) {
632 "while parsing argument for this `{kind}` macro fragment"
636 return ErrorReported;
641 NtOrTt::Nt(nt) => MatchedNonterminal(Lrc::new(nt)),
642 NtOrTt::Tt(tt) => MatchedTokenTree(tt),
644 mp.push_match(next_metavar, seq_depth, m);
649 self.cur_mps.push(mp);
653 // Too many possibilities!
654 return self.ambiguity_error(matcher, parser.token.span);
658 assert!(!self.cur_mps.is_empty());
664 matcher: &[MatcherLoc],
665 token_span: rustc_span::Span,
666 ) -> NamedParseResult {
670 .map(|mp| match &matcher[mp.idx] {
671 MatcherLoc::MetaVarDecl { bind, kind: Some(kind), .. } => {
672 format!("{} ('{}')", kind, bind)
676 .collect::<Vec<String>>()
682 "local ambiguity when calling macro `{}`: multiple parsing options: {}",
684 match self.next_mps.len() {
685 0 => format!("built-in NTs {}.", nts),
686 1 => format!("built-in NTs {} or 1 other option.", nts),
687 n => format!("built-in NTs {} or {} other options.", nts, n),
693 fn nameize<I: Iterator<Item = NamedMatch>>(
696 matcher: &[MatcherLoc],
698 ) -> NamedParseResult {
699 // Make that each metavar has _exactly one_ binding. If so, insert the binding into the
700 // `NamedParseResult`. Otherwise, it's an error.
701 let mut ret_val = FxHashMap::default();
703 if let &MatcherLoc::MetaVarDecl { span, bind, kind, .. } = loc {
705 match ret_val.entry(MacroRulesNormalizedIdent::new(bind)) {
706 Vacant(spot) => spot.insert(res.next().unwrap()),
708 return Error(span, format!("duplicated bind name: {}", bind));
712 // Both this check and the one in `parse_tt_inner` are necessary, surprisingly.
713 if sess.missing_fragment_specifiers.borrow_mut().remove(&span).is_some() {
714 // E.g. `$e` instead of `$e:expr`.
715 return Error(span, "missing fragment specifier".to_string());