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 //! items, 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 'position' is a dot in the middle of a matcher, usually represented as a
18 //! dot. For example `· a $( a )* a b` is a position, as is `a $( · a )* a b`.
20 //! The parser walks through the input a character at a time, maintaining a list
21 //! of threads consistent with the current position in the input string: `cur_items`.
23 //! As it processes them, it fills up `eof_items` with threads that would be valid if
24 //! the macro invocation is now over, `bb_items` with threads that are waiting on
25 //! a Rust non-terminal like `$e:expr`, and `next_items` 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_items` 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 item).
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 item)
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 item)
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 item)
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::*;
75 use TokenTreeOrTokenTreeSlice::*;
77 use crate::mbe::{self, TokenTree};
79 use rustc_ast::token::{self, DocComment, Nonterminal, Token};
80 use rustc_parse::parser::{OrPatNonterminalMode, Parser};
81 use rustc_session::parse::ParseSess;
82 use rustc_span::{edition::Edition, symbol::MacroRulesNormalizedIdent};
84 use smallvec::{smallvec, SmallVec};
86 use rustc_data_structures::fx::FxHashMap;
87 use rustc_data_structures::sync::Lrc;
89 use std::collections::hash_map::Entry::{Occupied, Vacant};
91 use std::ops::{Deref, DerefMut};
93 // To avoid costly uniqueness checks, we require that `MatchSeq` always has a nonempty body.
95 /// Either a sequence of token trees or a single one. This is used as the representation of the
96 /// sequence of tokens that make up a matcher.
98 enum TokenTreeOrTokenTreeSlice<'tt> {
100 TtSeq(&'tt [TokenTree]),
103 impl<'tt> TokenTreeOrTokenTreeSlice<'tt> {
104 /// Returns the number of constituent top-level token trees of `self` (top-level in that it
105 /// will not recursively descend into subtrees).
106 fn len(&self) -> usize {
108 TtSeq(ref v) => v.len(),
109 Tt(ref tt) => tt.len(),
113 /// The `index`-th token tree of `self`.
114 fn get_tt(&self, index: usize) -> TokenTree {
116 TtSeq(ref v) => v[index].clone(),
117 Tt(ref tt) => tt.get_tt(index),
122 /// An unzipping of `TokenTree`s... see the `stack` field of `MatcherPos`.
124 /// This is used by `inner_parse_loop` to keep track of delimited submatchers that we have
127 struct MatcherTtFrame<'tt> {
128 /// The "parent" matcher that we are descending into.
129 elts: TokenTreeOrTokenTreeSlice<'tt>,
130 /// The position of the "dot" in `elts` at the time we descended.
134 type NamedMatchVec = SmallVec<[NamedMatch; 4]>;
136 /// Represents a single "position" (aka "matcher position", aka "item"), as
137 /// described in the module documentation.
141 /// - `'root` represents the lifetime of the stack slot that holds the root
142 /// `MatcherPos`. As described in `MatcherPosHandle`, the root `MatcherPos`
143 /// structure is stored on the stack, but subsequent instances are put into
145 /// - `'tt` represents the lifetime of the token trees that this matcher
146 /// position refers to.
148 /// It is important to distinguish these two lifetimes because we have a
149 /// `SmallVec<TokenTreeOrTokenTreeSlice<'tt>>` below, and the destructor of
150 /// that is considered to possibly access the data from its elements (it lacks
151 /// a `#[may_dangle]` attribute). As a result, the compiler needs to know that
152 /// all the elements in that `SmallVec` strictly outlive the root stack slot
153 /// lifetime. By separating `'tt` from `'root`, we can show that.
155 struct MatcherPos<'root, 'tt> {
156 /// The token or sequence of tokens that make up the matcher
157 top_elts: TokenTreeOrTokenTreeSlice<'tt>,
159 /// The position of the "dot" in this matcher
162 /// For each named metavar in the matcher, we keep track of token trees matched against the
163 /// metavar by the black box parser. In particular, there may be more than one match per
164 /// metavar if we are in a repetition (each repetition matches each of the variables).
165 /// Moreover, matchers and repetitions can be nested; the `matches` field is shared (hence the
166 /// `Rc`) among all "nested" matchers. `match_lo`, `match_cur`, and `match_hi` keep track of
167 /// the current position of the `self` matcher position in the shared `matches` list.
169 /// Also, note that while we are descending into a sequence, matchers are given their own
170 /// `matches` vector. Only once we reach the end of a full repetition of the sequence do we add
171 /// all bound matches from the submatcher into the shared top-level `matches` vector. If `sep`
172 /// and `up` are `Some`, then `matches` is _not_ the shared top-level list. Instead, if one
173 /// wants the shared `matches`, one should use `up.matches`.
174 matches: Box<[Lrc<NamedMatchVec>]>,
175 /// The position in `matches` corresponding to the first metavar in this matcher's sequence of
176 /// token trees. In other words, the first metavar in the first token of `top_elts` corresponds
177 /// to `matches[match_lo]`.
179 /// The position in `matches` corresponding to the metavar we are currently trying to match
180 /// against the source token stream. `match_lo <= match_cur <= match_hi`.
182 /// Similar to `match_lo` except `match_hi` is the position in `matches` of the _last_ metavar
186 // The following fields are used if we are matching a repetition. If we aren't, they should be
188 /// The KleeneOp of this sequence if we are in a repetition.
189 seq_op: Option<mbe::KleeneOp>,
191 /// The separator if we are in a repetition.
194 /// The "parent" matcher position if we are in a repetition. That is, the matcher position just
195 /// before we enter the sequence.
196 up: Option<MatcherPosHandle<'root, 'tt>>,
198 /// Specifically used to "unzip" token trees. By "unzip", we mean to unwrap the delimiters from
199 /// a delimited token tree (e.g., something wrapped in `(` `)`) or to get the contents of a doc
202 /// When matching against matchers with nested delimited submatchers (e.g., `pat ( pat ( .. )
203 /// pat ) pat`), we need to keep track of the matchers we are descending into. This stack does
204 /// that where the bottom of the stack is the outermost matcher.
205 /// Also, throughout the comments, this "descent" is often referred to as "unzipping"...
206 stack: SmallVec<[MatcherTtFrame<'tt>; 1]>,
209 impl<'root, 'tt> MatcherPos<'root, 'tt> {
210 /// Adds `m` as a named match for the `idx`-th metavar.
211 fn push_match(&mut self, idx: usize, m: NamedMatch) {
212 let matches = Lrc::make_mut(&mut self.matches[idx]);
217 // Lots of MatcherPos instances are created at runtime. Allocating them on the
218 // heap is slow. Furthermore, using SmallVec<MatcherPos> to allocate them all
219 // on the stack is also slow, because MatcherPos is quite a large type and
220 // instances get moved around a lot between vectors, which requires lots of
221 // slow memcpy calls.
223 // Therefore, the initial MatcherPos is always allocated on the stack,
224 // subsequent ones (of which there aren't that many) are allocated on the heap,
225 // and this type is used to encapsulate both cases.
226 enum MatcherPosHandle<'root, 'tt> {
227 Ref(&'root mut MatcherPos<'root, 'tt>),
228 Box(Box<MatcherPos<'root, 'tt>>),
231 impl<'root, 'tt> Clone for MatcherPosHandle<'root, 'tt> {
232 // This always produces a new Box.
233 fn clone(&self) -> Self {
234 MatcherPosHandle::Box(match *self {
235 MatcherPosHandle::Ref(ref r) => Box::new((**r).clone()),
236 MatcherPosHandle::Box(ref b) => b.clone(),
241 impl<'root, 'tt> Deref for MatcherPosHandle<'root, 'tt> {
242 type Target = MatcherPos<'root, 'tt>;
243 fn deref(&self) -> &Self::Target {
245 MatcherPosHandle::Ref(ref r) => r,
246 MatcherPosHandle::Box(ref b) => b,
251 impl<'root, 'tt> DerefMut for MatcherPosHandle<'root, 'tt> {
252 fn deref_mut(&mut self) -> &mut MatcherPos<'root, 'tt> {
254 MatcherPosHandle::Ref(ref mut r) => r,
255 MatcherPosHandle::Box(ref mut b) => b,
260 /// Represents the possible results of an attempted parse.
261 crate enum ParseResult<T> {
262 /// Parsed successfully.
264 /// Arm failed to match. If the second parameter is `token::Eof`, it indicates an unexpected
265 /// end of macro invocation. Otherwise, it indicates that no rules expected the given token.
266 Failure(Token, &'static str),
267 /// Fatal error (malformed macro?). Abort compilation.
268 Error(rustc_span::Span, String),
272 /// A `ParseResult` where the `Success` variant contains a mapping of
273 /// `MacroRulesNormalizedIdent`s to `NamedMatch`es. This represents the mapping
274 /// of metavars to the token trees they bind to.
275 crate type NamedParseResult = ParseResult<FxHashMap<MacroRulesNormalizedIdent, NamedMatch>>;
277 /// Count how many metavars are named in the given matcher `ms`.
278 pub(super) fn count_names(ms: &[TokenTree]) -> usize {
279 ms.iter().fold(0, |count, elt| {
282 TokenTree::Sequence(_, ref seq) => seq.num_captures,
283 TokenTree::Delimited(_, ref delim) => count_names(&delim.tts),
284 TokenTree::MetaVar(..) => 0,
285 TokenTree::MetaVarDecl(..) => 1,
286 TokenTree::Token(..) => 0,
291 /// `len` `Vec`s (initially shared and empty) that will store matches of metavars.
292 fn create_matches(len: usize) -> Box<[Lrc<NamedMatchVec>]> {
296 let empty_matches = Lrc::new(SmallVec::new());
297 vec![empty_matches; len]
302 /// Generates the top-level matcher position in which the "dot" is before the first token of the
304 fn initial_matcher_pos<'root, 'tt>(ms: &'tt [TokenTree]) -> MatcherPos<'root, 'tt> {
305 let match_idx_hi = count_names(ms);
306 let matches = create_matches(match_idx_hi);
308 // Start with the top level matcher given to us
309 top_elts: TtSeq(ms), // "elts" is an abbr. for "elements"
310 // The "dot" is before the first token of the matcher
313 // Initialize `matches` to a bunch of empty `Vec`s -- one for each metavar in `top_elts`.
314 // `match_lo` for `top_elts` is 0 and `match_hi` is `matches.len()`. `match_cur` is 0 since
315 // we haven't actually matched anything yet.
319 match_hi: match_idx_hi,
321 // Haven't descended into any delimiters, so empty stack
324 // Haven't descended into any sequences, so both of these are `None`.
331 /// `NamedMatch` is a pattern-match result for a single `token::MATCH_NONTERMINAL`:
332 /// so it is associated with a single ident in a parse, and all
333 /// `MatchedNonterminal`s in the `NamedMatch` have the same non-terminal type
334 /// (expr, item, etc). Each leaf in a single `NamedMatch` corresponds to a
335 /// single `token::MATCH_NONTERMINAL` in the `TokenTree` that produced it.
337 /// The in-memory structure of a particular `NamedMatch` represents the match
338 /// that occurred when a particular subset of a matcher was applied to a
339 /// particular token tree.
341 /// The width of each `MatchedSeq` in the `NamedMatch`, and the identity of
342 /// the `MatchedNonterminal`s, will depend on the token tree it was applied
343 /// to: each `MatchedSeq` corresponds to a single `TTSeq` in the originating
344 /// token tree. The depth of the `NamedMatch` structure will therefore depend
345 /// only on the nesting depth of `ast::TTSeq`s in the originating
346 /// token tree it was derived from.
347 #[derive(Debug, Clone)]
348 crate enum NamedMatch {
349 MatchedSeq(Lrc<NamedMatchVec>),
350 MatchedNonterminal(Lrc<Nonterminal>),
353 /// Takes a sequence of token trees `ms` representing a matcher which successfully matched input
354 /// and an iterator of items that matched input and produces a `NamedParseResult`.
355 fn nameize<I: Iterator<Item = NamedMatch>>(
359 ) -> NamedParseResult {
360 // Recursively descend into each type of matcher (e.g., sequences, delimited, metavars) and make
361 // sure that each metavar has _exactly one_ binding. If a metavar does not have exactly one
362 // binding, then there is an error. If it does, then we insert the binding into the
363 // `NamedParseResult`.
364 fn n_rec<I: Iterator<Item = NamedMatch>>(
368 ret_val: &mut FxHashMap<MacroRulesNormalizedIdent, NamedMatch>,
369 ) -> Result<(), (rustc_span::Span, String)> {
371 TokenTree::Sequence(_, ref seq) => {
372 for next_m in &seq.tts {
373 n_rec(sess, next_m, res.by_ref(), ret_val)?
376 TokenTree::Delimited(_, ref delim) => {
377 for next_m in &delim.tts {
378 n_rec(sess, next_m, res.by_ref(), ret_val)?;
381 TokenTree::MetaVarDecl(sp, bind_name, _) => match ret_val
382 .entry(MacroRulesNormalizedIdent::new(bind_name))
385 spot.insert(res.next().unwrap());
387 Occupied(..) => return Err((sp, format!("duplicated bind name: {}", bind_name))),
389 TokenTree::MetaVar(..) | TokenTree::Token(..) => (),
395 let mut ret_val = FxHashMap::default();
397 match n_rec(sess, m, res.by_ref(), &mut ret_val) {
399 Err((sp, msg)) => return Error(sp, msg),
406 /// Performs a token equality check, ignoring syntax context (that is, an unhygienic comparison)
407 fn token_name_eq(t1: &Token, t2: &Token) -> bool {
408 if let (Some((ident1, is_raw1)), Some((ident2, is_raw2))) = (t1.ident(), t2.ident()) {
409 ident1.name == ident2.name && is_raw1 == is_raw2
410 } else if let (Some(ident1), Some(ident2)) = (t1.lifetime(), t2.lifetime()) {
411 ident1.name == ident2.name
417 /// In edition 2015/18, `:pat` can only match `pat<no_top_alt>` because otherwise, we have
418 /// breakage. As of edition 2021, `:pat` matches `top_pat`.
420 /// See <https://github.com/rust-lang/rust/issues/54883> for more info.
421 fn or_pat_mode(edition: Edition) -> OrPatNonterminalMode {
423 Edition::Edition2015 | Edition::Edition2018 => OrPatNonterminalMode::NoTopAlt,
424 // FIXME(mark-i-m): uncomment this when edition 2021 machinery is added.
425 // Edition::Edition2021 => OrPatNonterminalMode::TopPat,
429 /// Process the matcher positions of `cur_items` until it is empty. In the process, this will
430 /// produce more items in `next_items`, `eof_items`, and `bb_items`.
432 /// For more info about the how this happens, see the module-level doc comments and the inline
433 /// comments of this function.
437 /// - `cur_items`: the set of current items to be processed. This should be empty by the end of a
438 /// successful execution of this function.
439 /// - `next_items`: the set of newly generated items. These are used to replenish `cur_items` in
440 /// the function `parse`.
441 /// - `eof_items`: the set of items that would be valid if this was the EOF.
442 /// - `bb_items`: the set of items that are waiting for the black-box parser.
443 /// - `token`: the current token of the parser.
447 /// A `ParseResult`. Note that matches are kept track of through the items generated.
448 fn inner_parse_loop<'root, 'tt>(
449 cur_items: &mut SmallVec<[MatcherPosHandle<'root, 'tt>; 1]>,
450 next_items: &mut Vec<MatcherPosHandle<'root, 'tt>>,
451 eof_items: &mut SmallVec<[MatcherPosHandle<'root, 'tt>; 1]>,
452 bb_items: &mut SmallVec<[MatcherPosHandle<'root, 'tt>; 1]>,
454 ) -> ParseResult<()> {
455 // Pop items from `cur_items` until it is empty.
456 while let Some(mut item) = cur_items.pop() {
457 // When unzipped trees end, remove them. This corresponds to backtracking out of a
458 // delimited submatcher into which we already descended. In backtracking out again, we need
459 // to advance the "dot" past the delimiters in the outer matcher.
460 while item.idx >= item.top_elts.len() {
461 match item.stack.pop() {
462 Some(MatcherTtFrame { elts, idx }) => {
463 item.top_elts = elts;
470 // Get the current position of the "dot" (`idx`) in `item` and the number of token trees in
471 // the matcher (`len`).
473 let len = item.top_elts.len();
475 // If `idx >= len`, then we are at or past the end of the matcher of `item`.
477 // We are repeating iff there is a parent. If the matcher is inside of a repetition,
478 // then we could be at the end of a sequence or at the beginning of the next
480 if item.up.is_some() {
481 // At this point, regardless of whether there is a separator, we should add all
482 // matches from the complete repetition of the sequence to the shared, top-level
483 // `matches` list (actually, `up.matches`, which could itself not be the top-level,
484 // but anyway...). Moreover, we add another item to `cur_items` in which the "dot"
485 // is at the end of the `up` matcher. This ensures that the "dot" in the `up`
486 // matcher is also advanced sufficiently.
488 // NOTE: removing the condition `idx == len` allows trailing separators.
490 // Get the `up` matcher
491 let mut new_pos = item.up.clone().unwrap();
493 // Add matches from this repetition to the `matches` of `up`
494 for idx in item.match_lo..item.match_hi {
495 let sub = item.matches[idx].clone();
496 new_pos.push_match(idx, MatchedSeq(sub));
499 // Move the "dot" past the repetition in `up`
500 new_pos.match_cur = item.match_hi;
502 cur_items.push(new_pos);
505 // Check if we need a separator.
506 if idx == len && item.sep.is_some() {
507 // We have a separator, and it is the current token. We can advance past the
509 if item.sep.as_ref().map(|sep| token_name_eq(token, sep)).unwrap_or(false) {
511 next_items.push(item);
514 // We don't need a separator. Move the "dot" back to the beginning of the matcher
515 // and try to match again UNLESS we are only allowed to have _one_ repetition.
516 else if item.seq_op != Some(mbe::KleeneOp::ZeroOrOne) {
517 item.match_cur = item.match_lo;
519 cur_items.push(item);
522 // If we are not in a repetition, then being at the end of a matcher means that we have
523 // reached the potential end of the input.
525 eof_items.push(item);
528 // We are in the middle of a matcher.
530 // Look at what token in the matcher we are trying to match the current token (`token`)
531 // against. Depending on that, we may generate new items.
532 match item.top_elts.get_tt(idx) {
533 // Need to descend into a sequence
534 TokenTree::Sequence(sp, seq) => {
535 // Examine the case where there are 0 matches of this sequence. We are
536 // implicitly disallowing OneOrMore from having 0 matches here. Thus, that will
537 // result in a "no rules expected token" error by virtue of this matcher not
539 if seq.kleene.op == mbe::KleeneOp::ZeroOrMore
540 || seq.kleene.op == mbe::KleeneOp::ZeroOrOne
542 let mut new_item = item.clone();
543 new_item.match_cur += seq.num_captures;
545 for idx in item.match_cur..item.match_cur + seq.num_captures {
546 new_item.push_match(idx, MatchedSeq(Lrc::new(smallvec![])));
548 cur_items.push(new_item);
551 let matches = create_matches(item.matches.len());
552 cur_items.push(MatcherPosHandle::Box(Box::new(MatcherPos {
554 sep: seq.separator.clone(),
555 seq_op: Some(seq.kleene.op),
558 match_lo: item.match_cur,
559 match_cur: item.match_cur,
560 match_hi: item.match_cur + seq.num_captures,
562 top_elts: Tt(TokenTree::Sequence(sp, seq)),
566 // We need to match a metavar with a valid ident... call out to the black-box
567 // parser by adding an item to `bb_items`.
568 TokenTree::MetaVarDecl(span, _, kind) => {
569 // Built-in nonterminals never start with these tokens, so we can eliminate
570 // them from consideration.
572 // We use the span of the metavariable declaration to determine any
573 // edition-specific matching behavior for non-terminals.
574 if Parser::nonterminal_may_begin_with(kind, token, or_pat_mode(span.edition()))
580 // We need to descend into a delimited submatcher or a doc comment. To do this, we
581 // push the current matcher onto a stack and push a new item containing the
582 // submatcher onto `cur_items`.
584 // At the beginning of the loop, if we reach the end of the delimited submatcher,
585 // we pop the stack to backtrack out of the descent.
588 (TokenTree::Delimited(..)
589 | TokenTree::Token(Token { kind: DocComment(..), .. })) => {
590 let lower_elts = mem::replace(&mut item.top_elts, Tt(seq));
592 item.stack.push(MatcherTtFrame { elts: lower_elts, idx });
594 cur_items.push(item);
597 // We just matched a normal token. We can just advance the parser.
598 TokenTree::Token(t) if token_name_eq(&t, token) => {
600 next_items.push(item);
603 // There was another token that was not `token`... This means we can't add any
604 // rules. NOTE that this is not necessarily an error unless _all_ items in
605 // `cur_items` end up doing this. There may still be some other matchers that do
606 // end up working out.
607 TokenTree::Token(..) | TokenTree::MetaVar(..) => {}
612 // Yay a successful parse (so far)!
616 /// Use the given sequence of token trees (`ms`) as a matcher. Match the token
617 /// stream from the given `parser` against it and return the match.
618 pub(super) fn parse_tt(parser: &mut Cow<'_, Parser<'_>>, ms: &[TokenTree]) -> NamedParseResult {
619 // A queue of possible matcher positions. We initialize it with the matcher position in which
620 // the "dot" is before the first token of the first token tree in `ms`. `inner_parse_loop` then
621 // processes all of these possible matcher positions and produces possible next positions into
622 // `next_items`. After some post-processing, the contents of `next_items` replenish `cur_items`
623 // and we start over again.
625 // This MatcherPos instance is allocated on the stack. All others -- and
626 // there are frequently *no* others! -- are allocated on the heap.
627 let mut initial = initial_matcher_pos(ms);
628 let mut cur_items = smallvec![MatcherPosHandle::Ref(&mut initial)];
629 let mut next_items = Vec::new();
632 // Matcher positions black-box parsed by parser.rs (`parser`)
633 let mut bb_items = SmallVec::new();
635 // Matcher positions that would be valid if the macro invocation was over now
636 let mut eof_items = SmallVec::new();
637 assert!(next_items.is_empty());
639 // Process `cur_items` until either we have finished the input or we need to get some
640 // parsing from the black-box parser done. The result is that `next_items` will contain a
641 // bunch of possible next matcher positions in `next_items`.
642 match inner_parse_loop(
650 Failure(token, msg) => return Failure(token, msg),
651 Error(sp, msg) => return Error(sp, msg),
652 ErrorReported => return ErrorReported,
655 // inner parse loop handled all cur_items, so it's empty
656 assert!(cur_items.is_empty());
658 // We need to do some post processing after the `inner_parser_loop`.
660 // Error messages here could be improved with links to original rules.
662 // If we reached the EOF, check that there is EXACTLY ONE possible matcher. Otherwise,
663 // either the parse is ambiguous (which should never happen) or there is a syntax error.
664 if parser.token == token::Eof {
665 if eof_items.len() == 1 {
667 eof_items[0].matches.iter_mut().map(|dv| Lrc::make_mut(dv).pop().unwrap());
668 return nameize(parser.sess, ms, matches);
669 } else if eof_items.len() > 1 {
672 "ambiguity: multiple successful parses".to_string(),
678 if parser.token.span.is_dummy() {
681 parser.token.span.shrink_to_hi()
684 "missing tokens in macro arguments",
688 // Performance hack: eof_items may share matchers via Rc with other things that we want
689 // to modify. Dropping eof_items now may drop these refcounts to 1, preventing an
690 // unnecessary implicit clone later in Rc::make_mut.
693 // If there are no possible next positions AND we aren't waiting for the black-box parser,
694 // then there is a syntax error.
695 if bb_items.is_empty() && next_items.is_empty() {
696 return Failure(parser.token.clone(), "no rules expected this token in macro call");
698 // Another possibility is that we need to call out to parse some rust nonterminal
699 // (black-box) parser. However, if there is not EXACTLY ONE of these, something is wrong.
700 else if (!bb_items.is_empty() && !next_items.is_empty()) || bb_items.len() > 1 {
703 .map(|item| match item.top_elts.get_tt(item.idx) {
704 TokenTree::MetaVarDecl(_, bind, kind) => format!("{} ('{}')", kind, bind),
707 .collect::<Vec<String>>()
713 "local ambiguity: multiple parsing options: {}",
714 match next_items.len() {
715 0 => format!("built-in NTs {}.", nts),
716 1 => format!("built-in NTs {} or 1 other option.", nts),
717 n => format!("built-in NTs {} or {} other options.", nts, n),
722 // Dump all possible `next_items` into `cur_items` for the next iteration.
723 else if !next_items.is_empty() {
724 // Now process the next token
725 cur_items.extend(next_items.drain(..));
726 parser.to_mut().bump();
728 // Finally, we have the case where we need to call the black-box parser to get some
731 assert_eq!(bb_items.len(), 1);
733 let mut item = bb_items.pop().unwrap();
734 if let TokenTree::MetaVarDecl(span, _, kind) = item.top_elts.get_tt(item.idx) {
735 let match_cur = item.match_cur;
736 // We use the span of the metavariable declaration to determine any
737 // edition-specific matching behavior for non-terminals.
738 let nt = match parser.to_mut().parse_nonterminal(kind, or_pat_mode(span.edition()))
743 format!("while parsing argument for this `{}` macro fragment", kind),
746 return ErrorReported;
750 item.push_match(match_cur, MatchedNonterminal(Lrc::new(nt)));
756 cur_items.push(item);
759 assert!(!cur_items.is_empty());