1 use crate::base::{DummyResult, ExtCtxt, MacResult, TTMacroExpander};
2 use crate::base::{SyntaxExtension, SyntaxExtensionKind};
3 use crate::expand::{ensure_complete_parse, parse_ast_fragment, AstFragment, AstFragmentKind};
5 use crate::mbe::diagnostics::{annotate_doc_comment, parse_failure_msg};
6 use crate::mbe::macro_check;
7 use crate::mbe::macro_parser::{Error, ErrorReported, Failure, Success, TtParser};
8 use crate::mbe::macro_parser::{MatchedSeq, MatchedTokenTree, MatcherLoc};
9 use crate::mbe::transcribe::transcribe;
12 use rustc_ast::token::{self, Delimiter, NonterminalKind, Token, TokenKind, TokenKind::*};
13 use rustc_ast::tokenstream::{DelimSpan, TokenStream};
14 use rustc_ast::{NodeId, DUMMY_NODE_ID};
15 use rustc_ast_pretty::pprust;
16 use rustc_attr::{self as attr, TransparencyError};
17 use rustc_data_structures::fx::{FxHashMap, FxIndexMap};
18 use rustc_errors::{Applicability, ErrorGuaranteed};
19 use rustc_feature::Features;
20 use rustc_lint_defs::builtin::{
21 RUST_2021_INCOMPATIBLE_OR_PATTERNS, SEMICOLON_IN_EXPRESSIONS_FROM_MACROS,
23 use rustc_lint_defs::BuiltinLintDiagnostics;
24 use rustc_parse::parser::{Parser, Recovery};
25 use rustc_session::parse::ParseSess;
26 use rustc_session::Session;
27 use rustc_span::edition::Edition;
28 use rustc_span::hygiene::Transparency;
29 use rustc_span::symbol::{kw, sym, Ident, MacroRulesNormalizedIdent};
33 use std::collections::hash_map::Entry;
34 use std::{mem, slice};
36 use super::diagnostics;
37 use super::macro_parser::{NamedMatches, NamedParseResult};
39 pub(crate) struct ParserAnyMacro<'a> {
42 /// Span of the expansion site of the macro this parser is for
44 /// The ident of the macro we're parsing
47 is_trailing_mac: bool,
49 /// Whether or not this macro is defined in the current crate
53 impl<'a> ParserAnyMacro<'a> {
54 pub(crate) fn make(mut self: Box<ParserAnyMacro<'a>>, kind: AstFragmentKind) -> AstFragment {
64 let snapshot = &mut parser.create_snapshot_for_diagnostic();
65 let fragment = match parse_ast_fragment(parser, kind) {
68 diagnostics::emit_frag_parse_err(err, parser, snapshot, site_span, arm_span, kind);
69 return kind.dummy(site_span);
73 // We allow semicolons at the end of expressions -- e.g., the semicolon in
74 // `macro_rules! m { () => { panic!(); } }` isn't parsed by `.parse_expr()`,
75 // but `m!()` is allowed in expression positions (cf. issue #34706).
76 if kind == AstFragmentKind::Expr && parser.token == token::Semi {
78 parser.sess.buffer_lint_with_diagnostic(
79 SEMICOLON_IN_EXPRESSIONS_FROM_MACROS,
82 "trailing semicolon in macro used in expression position",
83 BuiltinLintDiagnostics::TrailingMacro(is_trailing_mac, macro_ident),
89 // Make sure we don't have any tokens left to parse so we don't silently drop anything.
90 let path = ast::Path::from_ident(macro_ident.with_span_pos(site_span));
91 ensure_complete_parse(parser, &path, kind.name(), site_span);
96 struct MacroRulesMacroExpander {
100 transparency: Transparency,
101 lhses: Vec<Vec<MatcherLoc>>,
102 rhses: Vec<mbe::TokenTree>,
106 impl TTMacroExpander for MacroRulesMacroExpander {
109 cx: &'cx mut ExtCtxt<'_>,
112 ) -> Box<dyn MacResult + 'cx> {
114 return DummyResult::any(sp);
130 fn macro_rules_dummy_expander<'cx>(
131 _: &'cx mut ExtCtxt<'_>,
134 ) -> Box<dyn MacResult + 'cx> {
135 DummyResult::any(span)
138 fn trace_macros_note(cx_expansions: &mut FxIndexMap<Span, Vec<String>>, sp: Span, message: String) {
139 let sp = sp.macro_backtrace().last().map_or(sp, |trace| trace.call_site);
140 cx_expansions.entry(sp).or_default().push(message);
143 pub(super) trait Tracker<'matcher> {
144 /// The contents of `ParseResult::Failure`.
147 /// Arm failed to match. If the token is `token::Eof`, it indicates an unexpected
148 /// end of macro invocation. Otherwise, it indicates that no rules expected the given token.
149 /// The usize is the approximate position of the token in the input token stream.
150 fn build_failure(tok: Token, position: usize, msg: &'static str) -> Self::Failure;
152 /// This is called before trying to match next MatcherLoc on the current token.
153 fn before_match_loc(&mut self, _parser: &TtParser, _matcher: &'matcher MatcherLoc) {}
155 /// This is called after an arm has been parsed, either successfully or unsuccessfully. When this is called,
156 /// `before_match_loc` was called at least once (with a `MatcherLoc::Eof`).
157 fn after_arm(&mut self, _result: &NamedParseResult<Self::Failure>) {}
160 fn description() -> &'static str;
162 fn recovery() -> Recovery {
167 /// A noop tracker that is used in the hot path of the expansion, has zero overhead thanks to monomorphization.
168 pub(super) struct NoopTracker;
170 impl<'matcher> Tracker<'matcher> for NoopTracker {
173 fn build_failure(_tok: Token, _position: usize, _msg: &'static str) -> Self::Failure {}
175 fn description() -> &'static str {
180 /// Expands the rules based macro defined by `lhses` and `rhses` for a given
182 #[instrument(skip(cx, transparency, arg, lhses, rhses))]
183 fn expand_macro<'cx>(
184 cx: &'cx mut ExtCtxt<'_>,
189 transparency: Transparency,
191 lhses: &[Vec<MatcherLoc>],
192 rhses: &[mbe::TokenTree],
193 ) -> Box<dyn MacResult + 'cx> {
194 let sess = &cx.sess.parse_sess;
195 // Macros defined in the current crate have a real node id,
196 // whereas macros from an external crate have a dummy id.
197 let is_local = node_id != DUMMY_NODE_ID;
199 if cx.trace_macros() {
200 let msg = format!("expanding `{}! {{ {} }}`", name, pprust::tts_to_string(&arg));
201 trace_macros_note(&mut cx.expansions, sp, msg);
204 // Track nothing for the best performance.
205 let try_success_result = try_match_macro(sess, name, &arg, lhses, &mut NoopTracker);
207 match try_success_result {
208 Ok((i, named_matches)) => {
209 let (rhs, rhs_span): (&mbe::Delimited, DelimSpan) = match &rhses[i] {
210 mbe::TokenTree::Delimited(span, delimited) => (&delimited, *span),
211 _ => cx.span_bug(sp, "malformed macro rhs"),
213 let arm_span = rhses[i].span();
215 let rhs_spans = rhs.tts.iter().map(|t| t.span()).collect::<Vec<_>>();
216 // rhs has holes ( `$id` and `$(...)` that need filled)
217 let mut tts = match transcribe(cx, &named_matches, &rhs, rhs_span, transparency) {
221 return DummyResult::any(arm_span);
225 // Replace all the tokens for the corresponding positions in the macro, to maintain
226 // proper positions in error reporting, while maintaining the macro_backtrace.
227 if rhs_spans.len() == tts.len() {
228 tts = tts.map_enumerated(|i, tt| {
229 let mut tt = tt.clone();
230 let mut sp = rhs_spans[i];
231 sp = sp.with_ctxt(tt.span().ctxt());
237 if cx.trace_macros() {
238 let msg = format!("to `{}`", pprust::tts_to_string(&tts));
239 trace_macros_note(&mut cx.expansions, sp, msg);
242 let mut p = Parser::new(sess, tts, false, None);
243 p.last_type_ascription = cx.current_expansion.prior_type_ascription;
246 cx.resolver.record_macro_rule_usage(node_id, i);
249 // Let the context choose how to interpret the result.
250 // Weird, but useful for X-macros.
251 return Box::new(ParserAnyMacro {
254 // Pass along the original expansion site and the name of the macro
255 // so we can print a useful error message if the parse of the expanded
256 // macro leaves unparsed tokens.
259 lint_node_id: cx.current_expansion.lint_node_id,
260 is_trailing_mac: cx.current_expansion.is_trailing_mac,
265 Err(CanRetry::No(_)) => {
266 debug!("Will not retry matching as an error was emitted already");
267 return DummyResult::any(sp);
269 Err(CanRetry::Yes) => {
270 // Retry and emit a better error below.
274 diagnostics::failed_to_match_macro(cx, sp, def_span, name, arg, lhses)
277 pub(super) enum CanRetry {
279 /// We are not allowed to retry macro expansion as a fatal error has been emitted already.
283 /// Try expanding the macro. Returns the index of the successful arm and its named_matches if it was successful,
284 /// and nothing if it failed. On failure, it's the callers job to use `track` accordingly to record all errors
286 #[instrument(level = "debug", skip(sess, arg, lhses, track), fields(tracking = %T::description()))]
287 pub(super) fn try_match_macro<'matcher, T: Tracker<'matcher>>(
291 lhses: &'matcher [Vec<MatcherLoc>],
293 ) -> Result<(usize, NamedMatches), CanRetry> {
294 // We create a base parser that can be used for the "black box" parts.
295 // Every iteration needs a fresh copy of that parser. However, the parser
296 // is not mutated on many of the iterations, particularly when dealing with
299 // macro_rules! foo {
303 // // ... etc. (maybe hundreds more)
306 // as seen in the `html5ever` benchmark. We use a `Cow` so that the base
307 // parser is only cloned when necessary (upon mutation). Furthermore, we
308 // reinitialize the `Cow` with the base parser at the start of every
309 // iteration, so that any mutated parsers are not reused. This is all quite
310 // hacky, but speeds up the `html5ever` benchmark significantly. (Issue
311 // 68836 suggests a more comprehensive but more complex change to deal with
313 let parser = parser_from_cx(sess, arg.clone(), T::recovery());
314 // Try each arm's matchers.
315 let mut tt_parser = TtParser::new(name);
316 for (i, lhs) in lhses.iter().enumerate() {
317 let _tracing_span = trace_span!("Matching arm", %i);
319 // Take a snapshot of the state of pre-expansion gating at this point.
320 // This is used so that if a matcher is not `Success(..)`ful,
321 // then the spans which became gated when parsing the unsuccessful matcher
322 // are not recorded. On the first `Success(..)`ful matcher, the spans are merged.
323 let mut gated_spans_snapshot = mem::take(&mut *sess.gated_spans.spans.borrow_mut());
325 let result = tt_parser.parse_tt(&mut Cow::Borrowed(&parser), lhs, track);
327 track.after_arm(&result);
330 Success(named_matches) => {
331 debug!("Parsed arm successfully");
332 // The matcher was `Success(..)`ful.
333 // Merge the gated spans from parsing the matcher with the pre-existing ones.
334 sess.gated_spans.merge(gated_spans_snapshot);
336 return Ok((i, named_matches));
339 trace!("Failed to match arm, trying the next one");
343 debug!("Fatal error occurred during matching");
344 // We haven't emitted an error yet, so we can retry.
345 return Err(CanRetry::Yes);
347 ErrorReported(guarantee) => {
348 debug!("Fatal error occurred and was reported during matching");
349 // An error has been reported already, we cannot retry as that would cause duplicate errors.
350 return Err(CanRetry::No(guarantee));
354 // The matcher was not `Success(..)`ful.
355 // Restore to the state before snapshotting and maybe try again.
356 mem::swap(&mut gated_spans_snapshot, &mut sess.gated_spans.spans.borrow_mut());
362 // Note that macro-by-example's input is also matched against a token tree:
363 // $( $lhs:tt => $rhs:tt );+
365 // Holy self-referential!
367 /// Converts a macro item into a syntax extension.
368 pub fn compile_declarative_macro(
373 ) -> (SyntaxExtension, Vec<(usize, Span)>) {
374 debug!("compile_declarative_macro: {:?}", def);
375 let mk_syn_ext = |expander| {
376 SyntaxExtension::new(
378 SyntaxExtensionKind::LegacyBang(expander),
386 let dummy_syn_ext = || (mk_syn_ext(Box::new(macro_rules_dummy_expander)), Vec::new());
388 let diag = &sess.parse_sess.span_diagnostic;
389 let lhs_nm = Ident::new(sym::lhs, def.span);
390 let rhs_nm = Ident::new(sym::rhs, def.span);
391 let tt_spec = Some(NonterminalKind::TT);
393 let macro_def = match &def.kind {
394 ast::ItemKind::MacroDef(def) => def,
397 let macro_rules = macro_def.macro_rules;
399 // Parse the macro_rules! invocation
401 // The pattern that macro_rules matches.
402 // The grammar for macro_rules! is:
403 // $( $lhs:tt => $rhs:tt );+
404 // ...quasiquoting this would be nice.
405 // These spans won't matter, anyways
406 let argument_gram = vec![
407 mbe::TokenTree::Sequence(
409 mbe::SequenceRepetition {
411 mbe::TokenTree::MetaVarDecl(def.span, lhs_nm, tt_spec),
412 mbe::TokenTree::token(token::FatArrow, def.span),
413 mbe::TokenTree::MetaVarDecl(def.span, rhs_nm, tt_spec),
415 separator: Some(Token::new(
416 if macro_rules { token::Semi } else { token::Comma },
419 kleene: mbe::KleeneToken::new(mbe::KleeneOp::OneOrMore, def.span),
423 // to phase into semicolon-termination instead of semicolon-separation
424 mbe::TokenTree::Sequence(
426 mbe::SequenceRepetition {
427 tts: vec![mbe::TokenTree::token(
428 if macro_rules { token::Semi } else { token::Comma },
432 kleene: mbe::KleeneToken::new(mbe::KleeneOp::ZeroOrMore, def.span),
437 // Convert it into `MatcherLoc` form.
438 let argument_gram = mbe::macro_parser::compute_locs(&argument_gram);
440 let create_parser = || {
441 let body = macro_def.body.tokens.clone();
442 Parser::new(&sess.parse_sess, body, true, rustc_parse::MACRO_ARGUMENTS)
445 let parser = create_parser();
447 TtParser::new(Ident::with_dummy_span(if macro_rules { kw::MacroRules } else { kw::Macro }));
449 match tt_parser.parse_tt(&mut Cow::Owned(parser), &argument_gram, &mut NoopTracker) {
452 // The fast `NoopTracker` doesn't have any info on failure, so we need to retry it with another one
453 // that gives us the information we need.
454 // For this we need to reclone the macro body as the previous parser consumed it.
455 let retry_parser = create_parser();
457 let parse_result = tt_parser.parse_tt(
458 &mut Cow::Owned(retry_parser),
460 &mut diagnostics::FailureForwarder,
462 let Failure((token, _, msg)) = parse_result else {
463 unreachable!("matcher returned something other than Failure after retry");
466 let s = parse_failure_msg(&token);
467 let sp = token.span.substitute_dummy(def.span);
468 let mut err = sess.parse_sess.span_diagnostic.struct_span_err(sp, &s);
469 err.span_label(sp, msg);
470 annotate_doc_comment(&mut err, sess.source_map(), sp);
472 return dummy_syn_ext();
477 .struct_span_err(sp.substitute_dummy(def.span), &msg)
479 return dummy_syn_ext();
481 ErrorReported(_) => {
482 return dummy_syn_ext();
486 let mut valid = true;
488 // Extract the arguments:
489 let lhses = match argument_map[&MacroRulesNormalizedIdent::new(lhs_nm)] {
490 MatchedSeq(ref s) => s
493 if let MatchedTokenTree(ref tt) = *m {
494 let tt = mbe::quoted::parse(
495 TokenStream::new(vec![tt.clone()]),
504 valid &= check_lhs_nt_follows(&sess.parse_sess, &def, &tt);
507 sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs")
509 .collect::<Vec<mbe::TokenTree>>(),
510 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs"),
513 let rhses = match argument_map[&MacroRulesNormalizedIdent::new(rhs_nm)] {
514 MatchedSeq(ref s) => s
517 if let MatchedTokenTree(ref tt) = *m {
518 return mbe::quoted::parse(
519 TokenStream::new(vec![tt.clone()]),
529 sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs")
531 .collect::<Vec<mbe::TokenTree>>(),
532 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured rhs"),
536 valid &= check_rhs(&sess.parse_sess, rhs);
539 // don't abort iteration early, so that errors for multiple lhses can be reported
541 valid &= check_lhs_no_empty_seq(&sess.parse_sess, slice::from_ref(lhs));
544 valid &= macro_check::check_meta_variables(&sess.parse_sess, def.id, def.span, &lhses, &rhses);
546 let (transparency, transparency_error) = attr::find_transparency(&def.attrs, macro_rules);
547 match transparency_error {
548 Some(TransparencyError::UnknownTransparency(value, span)) => {
549 diag.span_err(span, &format!("unknown macro transparency: `{}`", value));
551 Some(TransparencyError::MultipleTransparencyAttrs(old_span, new_span)) => {
552 diag.span_err(vec![old_span, new_span], "multiple macro transparency attributes");
557 // Compute the spans of the macro rules for unused rule linting.
558 // To avoid warning noise, only consider the rules of this
559 // macro for the lint, if all rules are valid.
560 // Also, we are only interested in non-foreign macros.
561 let rule_spans = if valid && def.id != DUMMY_NODE_ID {
566 // If the rhs contains an invocation like compile_error!,
567 // don't consider the rule for the unused rule lint.
568 .filter(|(_idx, (_lhs, rhs))| !has_compile_error_macro(rhs))
569 // We only take the span of the lhs here,
570 // so that the spans of created warnings are smaller.
571 .map(|(idx, (lhs, _rhs))| (idx, lhs.span()))
577 // Convert the lhses into `MatcherLoc` form, which is better for doing the
578 // actual matching. Unless the matcher is invalid.
579 let lhses = if valid {
583 // Ignore the delimiters around the matcher.
585 mbe::TokenTree::Delimited(_, delimited) => {
586 mbe::macro_parser::compute_locs(&delimited.tts)
588 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "malformed macro lhs"),
596 let expander = Box::new(MacroRulesMacroExpander {
605 (mk_syn_ext(expander), rule_spans)
608 fn check_lhs_nt_follows(sess: &ParseSess, def: &ast::Item, lhs: &mbe::TokenTree) -> bool {
609 // lhs is going to be like TokenTree::Delimited(...), where the
610 // entire lhs is those tts. Or, it can be a "bare sequence", not wrapped in parens.
611 if let mbe::TokenTree::Delimited(_, delimited) = lhs {
612 check_matcher(sess, def, &delimited.tts)
614 let msg = "invalid macro matcher; matchers must be contained in balanced delimiters";
615 sess.span_diagnostic.span_err(lhs.span(), msg);
618 // we don't abort on errors on rejection, the driver will do that for us
619 // after parsing/expansion. we can report every error in every macro this way.
622 /// Checks that the lhs contains no repetition which could match an empty token
623 /// tree, because then the matcher would hang indefinitely.
624 fn check_lhs_no_empty_seq(sess: &ParseSess, tts: &[mbe::TokenTree]) -> bool {
629 | TokenTree::MetaVar(..)
630 | TokenTree::MetaVarDecl(..)
631 | TokenTree::MetaVarExpr(..) => (),
632 TokenTree::Delimited(_, ref del) => {
633 if !check_lhs_no_empty_seq(sess, &del.tts) {
637 TokenTree::Sequence(span, ref seq) => {
638 if seq.separator.is_none()
639 && seq.tts.iter().all(|seq_tt| match *seq_tt {
640 TokenTree::MetaVarDecl(_, _, Some(NonterminalKind::Vis)) => true,
641 TokenTree::Sequence(_, ref sub_seq) => {
642 sub_seq.kleene.op == mbe::KleeneOp::ZeroOrMore
643 || sub_seq.kleene.op == mbe::KleeneOp::ZeroOrOne
648 let sp = span.entire();
649 sess.span_diagnostic.span_err(sp, "repetition matches empty token tree");
652 if !check_lhs_no_empty_seq(sess, &seq.tts) {
662 fn check_rhs(sess: &ParseSess, rhs: &mbe::TokenTree) -> bool {
664 mbe::TokenTree::Delimited(..) => return true,
666 sess.span_diagnostic.span_err(rhs.span(), "macro rhs must be delimited");
672 fn check_matcher(sess: &ParseSess, def: &ast::Item, matcher: &[mbe::TokenTree]) -> bool {
673 let first_sets = FirstSets::new(matcher);
674 let empty_suffix = TokenSet::empty();
675 let err = sess.span_diagnostic.err_count();
676 check_matcher_core(sess, def, &first_sets, matcher, &empty_suffix);
677 err == sess.span_diagnostic.err_count()
680 fn has_compile_error_macro(rhs: &mbe::TokenTree) -> bool {
682 mbe::TokenTree::Delimited(_sp, d) => {
683 let has_compile_error = d.tts.array_windows::<3>().any(|[ident, bang, args]| {
684 if let mbe::TokenTree::Token(ident) = ident &&
685 let TokenKind::Ident(ident, _) = ident.kind &&
686 ident == sym::compile_error &&
687 let mbe::TokenTree::Token(bang) = bang &&
688 let TokenKind::Not = bang.kind &&
689 let mbe::TokenTree::Delimited(_, del) = args &&
690 del.delim != Delimiter::Invisible
697 if has_compile_error { true } else { d.tts.iter().any(has_compile_error_macro) }
703 // `The FirstSets` for a matcher is a mapping from subsequences in the
704 // matcher to the FIRST set for that subsequence.
706 // This mapping is partially precomputed via a backwards scan over the
707 // token trees of the matcher, which provides a mapping from each
708 // repetition sequence to its *first* set.
710 // (Hypothetically, sequences should be uniquely identifiable via their
711 // spans, though perhaps that is false, e.g., for macro-generated macros
712 // that do not try to inject artificial span information. My plan is
713 // to try to catch such cases ahead of time and not include them in
714 // the precomputed mapping.)
715 struct FirstSets<'tt> {
716 // this maps each TokenTree::Sequence `$(tt ...) SEP OP` that is uniquely identified by its
717 // span in the original matcher to the First set for the inner sequence `tt ...`.
719 // If two sequences have the same span in a matcher, then map that
720 // span to None (invalidating the mapping here and forcing the code to
722 first: FxHashMap<Span, Option<TokenSet<'tt>>>,
725 impl<'tt> FirstSets<'tt> {
726 fn new(tts: &'tt [mbe::TokenTree]) -> FirstSets<'tt> {
729 let mut sets = FirstSets { first: FxHashMap::default() };
730 build_recur(&mut sets, tts);
733 // walks backward over `tts`, returning the FIRST for `tts`
734 // and updating `sets` at the same time for all sequence
735 // substructure we find within `tts`.
736 fn build_recur<'tt>(sets: &mut FirstSets<'tt>, tts: &'tt [TokenTree]) -> TokenSet<'tt> {
737 let mut first = TokenSet::empty();
738 for tt in tts.iter().rev() {
741 | TokenTree::MetaVar(..)
742 | TokenTree::MetaVarDecl(..)
743 | TokenTree::MetaVarExpr(..) => {
744 first.replace_with(TtHandle::TtRef(tt));
746 TokenTree::Delimited(span, ref delimited) => {
747 build_recur(sets, &delimited.tts);
748 first.replace_with(TtHandle::from_token_kind(
749 token::OpenDelim(delimited.delim),
753 TokenTree::Sequence(sp, ref seq_rep) => {
754 let subfirst = build_recur(sets, &seq_rep.tts);
756 match sets.first.entry(sp.entire()) {
757 Entry::Vacant(vac) => {
758 vac.insert(Some(subfirst.clone()));
760 Entry::Occupied(mut occ) => {
761 // if there is already an entry, then a span must have collided.
762 // This should not happen with typical macro_rules macros,
763 // but syntax extensions need not maintain distinct spans,
764 // so distinct syntax trees can be assigned the same span.
765 // In such a case, the map cannot be trusted; so mark this
766 // entry as unusable.
771 // If the sequence contents can be empty, then the first
772 // token could be the separator token itself.
774 if let (Some(sep), true) = (&seq_rep.separator, subfirst.maybe_empty) {
775 first.add_one_maybe(TtHandle::from_token(sep.clone()));
778 // Reverse scan: Sequence comes before `first`.
779 if subfirst.maybe_empty
780 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrMore
781 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrOne
783 // If sequence is potentially empty, then
784 // union them (preserving first emptiness).
785 first.add_all(&TokenSet { maybe_empty: true, ..subfirst });
787 // Otherwise, sequence guaranteed
788 // non-empty; replace first.
799 // walks forward over `tts` until all potential FIRST tokens are
801 fn first(&self, tts: &'tt [mbe::TokenTree]) -> TokenSet<'tt> {
804 let mut first = TokenSet::empty();
805 for tt in tts.iter() {
806 assert!(first.maybe_empty);
809 | TokenTree::MetaVar(..)
810 | TokenTree::MetaVarDecl(..)
811 | TokenTree::MetaVarExpr(..) => {
812 first.add_one(TtHandle::TtRef(tt));
815 TokenTree::Delimited(span, ref delimited) => {
816 first.add_one(TtHandle::from_token_kind(
817 token::OpenDelim(delimited.delim),
822 TokenTree::Sequence(sp, ref seq_rep) => {
824 let subfirst = match self.first.get(&sp.entire()) {
825 Some(Some(subfirst)) => subfirst,
827 subfirst_owned = self.first(&seq_rep.tts);
831 panic!("We missed a sequence during FirstSets construction");
835 // If the sequence contents can be empty, then the first
836 // token could be the separator token itself.
837 if let (Some(sep), true) = (&seq_rep.separator, subfirst.maybe_empty) {
838 first.add_one_maybe(TtHandle::from_token(sep.clone()));
841 assert!(first.maybe_empty);
842 first.add_all(subfirst);
843 if subfirst.maybe_empty
844 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrMore
845 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrOne
847 // Continue scanning for more first
848 // tokens, but also make sure we
849 // restore empty-tracking state.
850 first.maybe_empty = true;
859 // we only exit the loop if `tts` was empty or if every
860 // element of `tts` matches the empty sequence.
861 assert!(first.maybe_empty);
866 // Most `mbe::TokenTree`s are pre-existing in the matcher, but some are defined
867 // implicitly, such as opening/closing delimiters and sequence repetition ops.
868 // This type encapsulates both kinds. It implements `Clone` while avoiding the
869 // need for `mbe::TokenTree` to implement `Clone`.
872 /// This is used in most cases.
873 TtRef(&'tt mbe::TokenTree),
875 /// This is only used for implicit token trees. The `mbe::TokenTree` *must*
876 /// be `mbe::TokenTree::Token`. No other variants are allowed. We store an
877 /// `mbe::TokenTree` rather than a `Token` so that `get()` can return a
878 /// `&mbe::TokenTree`.
879 Token(mbe::TokenTree),
882 impl<'tt> TtHandle<'tt> {
883 fn from_token(tok: Token) -> Self {
884 TtHandle::Token(mbe::TokenTree::Token(tok))
887 fn from_token_kind(kind: TokenKind, span: Span) -> Self {
888 TtHandle::from_token(Token::new(kind, span))
891 // Get a reference to a token tree.
892 fn get(&'tt self) -> &'tt mbe::TokenTree {
894 TtHandle::TtRef(tt) => tt,
895 TtHandle::Token(token_tt) => &token_tt,
900 impl<'tt> PartialEq for TtHandle<'tt> {
901 fn eq(&self, other: &TtHandle<'tt>) -> bool {
902 self.get() == other.get()
906 impl<'tt> Clone for TtHandle<'tt> {
907 fn clone(&self) -> Self {
909 TtHandle::TtRef(tt) => TtHandle::TtRef(tt),
911 // This variant *must* contain a `mbe::TokenTree::Token`, and not
912 // any other variant of `mbe::TokenTree`.
913 TtHandle::Token(mbe::TokenTree::Token(tok)) => {
914 TtHandle::Token(mbe::TokenTree::Token(tok.clone()))
922 // A set of `mbe::TokenTree`s, which may include `TokenTree::Match`s
923 // (for macro-by-example syntactic variables). It also carries the
924 // `maybe_empty` flag; that is true if and only if the matcher can
925 // match an empty token sequence.
927 // The First set is computed on submatchers like `$($a:expr b),* $(c)* d`,
928 // which has corresponding FIRST = {$a:expr, c, d}.
929 // Likewise, `$($a:expr b),* $(c)+ d` has FIRST = {$a:expr, c}.
931 // (Notably, we must allow for *-op to occur zero times.)
932 #[derive(Clone, Debug)]
933 struct TokenSet<'tt> {
934 tokens: Vec<TtHandle<'tt>>,
938 impl<'tt> TokenSet<'tt> {
939 // Returns a set for the empty sequence.
941 TokenSet { tokens: Vec::new(), maybe_empty: true }
944 // Returns the set `{ tok }` for the single-token (and thus
945 // non-empty) sequence [tok].
946 fn singleton(tt: TtHandle<'tt>) -> Self {
947 TokenSet { tokens: vec![tt], maybe_empty: false }
950 // Changes self to be the set `{ tok }`.
951 // Since `tok` is always present, marks self as non-empty.
952 fn replace_with(&mut self, tt: TtHandle<'tt>) {
954 self.tokens.push(tt);
955 self.maybe_empty = false;
958 // Changes self to be the empty set `{}`; meant for use when
959 // the particular token does not matter, but we want to
960 // record that it occurs.
961 fn replace_with_irrelevant(&mut self) {
963 self.maybe_empty = false;
966 // Adds `tok` to the set for `self`, marking sequence as non-empty.
967 fn add_one(&mut self, tt: TtHandle<'tt>) {
968 if !self.tokens.contains(&tt) {
969 self.tokens.push(tt);
971 self.maybe_empty = false;
974 // Adds `tok` to the set for `self`. (Leaves `maybe_empty` flag alone.)
975 fn add_one_maybe(&mut self, tt: TtHandle<'tt>) {
976 if !self.tokens.contains(&tt) {
977 self.tokens.push(tt);
981 // Adds all elements of `other` to this.
983 // (Since this is a set, we filter out duplicates.)
985 // If `other` is potentially empty, then preserves the previous
986 // setting of the empty flag of `self`. If `other` is guaranteed
987 // non-empty, then `self` is marked non-empty.
988 fn add_all(&mut self, other: &Self) {
989 for tt in &other.tokens {
990 if !self.tokens.contains(tt) {
991 self.tokens.push(tt.clone());
994 if !other.maybe_empty {
995 self.maybe_empty = false;
1000 // Checks that `matcher` is internally consistent and that it
1001 // can legally be followed by a token `N`, for all `N` in `follow`.
1002 // (If `follow` is empty, then it imposes no constraint on
1005 // Returns the set of NT tokens that could possibly come last in
1006 // `matcher`. (If `matcher` matches the empty sequence, then
1007 // `maybe_empty` will be set to true.)
1009 // Requires that `first_sets` is pre-computed for `matcher`;
1010 // see `FirstSets::new`.
1011 fn check_matcher_core<'tt>(
1014 first_sets: &FirstSets<'tt>,
1015 matcher: &'tt [mbe::TokenTree],
1016 follow: &TokenSet<'tt>,
1017 ) -> TokenSet<'tt> {
1020 let mut last = TokenSet::empty();
1022 // 2. For each token and suffix [T, SUFFIX] in M:
1023 // ensure that T can be followed by SUFFIX, and if SUFFIX may be empty,
1024 // then ensure T can also be followed by any element of FOLLOW.
1025 'each_token: for i in 0..matcher.len() {
1026 let token = &matcher[i];
1027 let suffix = &matcher[i + 1..];
1029 let build_suffix_first = || {
1030 let mut s = first_sets.first(suffix);
1037 // (we build `suffix_first` on demand below; you can tell
1038 // which cases are supposed to fall through by looking for the
1039 // initialization of this variable.)
1042 // First, update `last` so that it corresponds to the set
1043 // of NT tokens that might end the sequence `... token`.
1045 TokenTree::Token(..)
1046 | TokenTree::MetaVar(..)
1047 | TokenTree::MetaVarDecl(..)
1048 | TokenTree::MetaVarExpr(..) => {
1049 if token_can_be_followed_by_any(token) {
1050 // don't need to track tokens that work with any,
1051 last.replace_with_irrelevant();
1052 // ... and don't need to check tokens that can be
1053 // followed by anything against SUFFIX.
1054 continue 'each_token;
1056 last.replace_with(TtHandle::TtRef(token));
1057 suffix_first = build_suffix_first();
1060 TokenTree::Delimited(span, ref d) => {
1061 let my_suffix = TokenSet::singleton(TtHandle::from_token_kind(
1062 token::CloseDelim(d.delim),
1065 check_matcher_core(sess, def, first_sets, &d.tts, &my_suffix);
1066 // don't track non NT tokens
1067 last.replace_with_irrelevant();
1069 // also, we don't need to check delimited sequences
1071 continue 'each_token;
1073 TokenTree::Sequence(_, ref seq_rep) => {
1074 suffix_first = build_suffix_first();
1075 // The trick here: when we check the interior, we want
1076 // to include the separator (if any) as a potential
1077 // (but not guaranteed) element of FOLLOW. So in that
1078 // case, we make a temp copy of suffix and stuff
1079 // delimiter in there.
1081 // FIXME: Should I first scan suffix_first to see if
1082 // delimiter is already in it before I go through the
1083 // work of cloning it? But then again, this way I may
1084 // get a "tighter" span?
1086 let my_suffix = if let Some(sep) = &seq_rep.separator {
1087 new = suffix_first.clone();
1088 new.add_one_maybe(TtHandle::from_token(sep.clone()));
1094 // At this point, `suffix_first` is built, and
1095 // `my_suffix` is some TokenSet that we can use
1096 // for checking the interior of `seq_rep`.
1097 let next = check_matcher_core(sess, def, first_sets, &seq_rep.tts, my_suffix);
1098 if next.maybe_empty {
1099 last.add_all(&next);
1104 // the recursive call to check_matcher_core already ran the 'each_last
1105 // check below, so we can just keep going forward here.
1106 continue 'each_token;
1110 // (`suffix_first` guaranteed initialized once reaching here.)
1112 // Now `last` holds the complete set of NT tokens that could
1113 // end the sequence before SUFFIX. Check that every one works with `suffix`.
1114 for tt in &last.tokens {
1115 if let &TokenTree::MetaVarDecl(span, name, Some(kind)) = tt.get() {
1116 for next_token in &suffix_first.tokens {
1117 let next_token = next_token.get();
1119 // Check if the old pat is used and the next token is `|`
1120 // to warn about incompatibility with Rust 2021.
1121 // We only emit this lint if we're parsing the original
1122 // definition of this macro_rules, not while (re)parsing
1123 // the macro when compiling another crate that is using the
1124 // macro. (See #86567.)
1125 // Macros defined in the current crate have a real node id,
1126 // whereas macros from an external crate have a dummy id.
1127 if def.id != DUMMY_NODE_ID
1128 && matches!(kind, NonterminalKind::PatParam { inferred: true })
1129 && matches!(next_token, TokenTree::Token(token) if token.kind == BinOp(token::BinOpToken::Or))
1131 // It is suggestion to use pat_param, for example: $x:pat -> $x:pat_param.
1132 let suggestion = quoted_tt_to_string(&TokenTree::MetaVarDecl(
1135 Some(NonterminalKind::PatParam { inferred: false }),
1137 sess.buffer_lint_with_diagnostic(
1138 &RUST_2021_INCOMPATIBLE_OR_PATTERNS,
1141 "the meaning of the `pat` fragment specifier is changing in Rust 2021, which may affect this macro",
1142 BuiltinLintDiagnostics::OrPatternsBackCompat(span, suggestion),
1145 match is_in_follow(next_token, kind) {
1146 IsInFollow::Yes => {}
1147 IsInFollow::No(possible) => {
1148 let may_be = if last.tokens.len() == 1 && suffix_first.tokens.len() == 1
1155 let sp = next_token.span();
1156 let mut err = sess.span_diagnostic.struct_span_err(
1159 "`${name}:{frag}` {may_be} followed by `{next}`, which \
1160 is not allowed for `{frag}` fragments",
1163 next = quoted_tt_to_string(next_token),
1167 err.span_label(sp, format!("not allowed after `{}` fragments", kind));
1169 if kind == NonterminalKind::PatWithOr
1170 && sess.edition.rust_2021()
1171 && next_token.is_token(&BinOp(token::BinOpToken::Or))
1173 let suggestion = quoted_tt_to_string(&TokenTree::MetaVarDecl(
1176 Some(NonterminalKind::PatParam { inferred: false }),
1178 err.span_suggestion(
1180 "try a `pat_param` fragment specifier instead",
1182 Applicability::MaybeIncorrect,
1186 let msg = "allowed there are: ";
1191 "only {} is allowed after `{}` fragments",
1199 ts[..ts.len() - 1].to_vec().join(", "),
1214 fn token_can_be_followed_by_any(tok: &mbe::TokenTree) -> bool {
1215 if let mbe::TokenTree::MetaVarDecl(_, _, Some(kind)) = *tok {
1216 frag_can_be_followed_by_any(kind)
1218 // (Non NT's can always be followed by anything in matchers.)
1223 /// Returns `true` if a fragment of type `frag` can be followed by any sort of
1224 /// token. We use this (among other things) as a useful approximation
1225 /// for when `frag` can be followed by a repetition like `$(...)*` or
1226 /// `$(...)+`. In general, these can be a bit tricky to reason about,
1227 /// so we adopt a conservative position that says that any fragment
1228 /// specifier which consumes at most one token tree can be followed by
1229 /// a fragment specifier (indeed, these fragments can be followed by
1230 /// ANYTHING without fear of future compatibility hazards).
1231 fn frag_can_be_followed_by_any(kind: NonterminalKind) -> bool {
1234 NonterminalKind::Item // always terminated by `}` or `;`
1235 | NonterminalKind::Block // exactly one token tree
1236 | NonterminalKind::Ident // exactly one token tree
1237 | NonterminalKind::Literal // exactly one token tree
1238 | NonterminalKind::Meta // exactly one token tree
1239 | NonterminalKind::Lifetime // exactly one token tree
1240 | NonterminalKind::TT // exactly one token tree
1246 No(&'static [&'static str]),
1249 /// Returns `true` if `frag` can legally be followed by the token `tok`. For
1250 /// fragments that can consume an unbounded number of tokens, `tok`
1251 /// must be within a well-defined follow set. This is intended to
1252 /// guarantee future compatibility: for example, without this rule, if
1253 /// we expanded `expr` to include a new binary operator, we might
1254 /// break macros that were relying on that binary operator as a
1256 // when changing this do not forget to update doc/book/macros.md!
1257 fn is_in_follow(tok: &mbe::TokenTree, kind: NonterminalKind) -> IsInFollow {
1260 if let TokenTree::Token(Token { kind: token::CloseDelim(_), .. }) = *tok {
1261 // closing a token tree can never be matched by any fragment;
1262 // iow, we always require that `(` and `)` match, etc.
1266 NonterminalKind::Item => {
1267 // since items *must* be followed by either a `;` or a `}`, we can
1268 // accept anything after them
1271 NonterminalKind::Block => {
1272 // anything can follow block, the braces provide an easy boundary to
1276 NonterminalKind::Stmt | NonterminalKind::Expr => {
1277 const TOKENS: &[&str] = &["`=>`", "`,`", "`;`"];
1279 TokenTree::Token(token) => match token.kind {
1280 FatArrow | Comma | Semi => IsInFollow::Yes,
1281 _ => IsInFollow::No(TOKENS),
1283 _ => IsInFollow::No(TOKENS),
1286 NonterminalKind::PatParam { .. } => {
1287 const TOKENS: &[&str] = &["`=>`", "`,`", "`=`", "`|`", "`if`", "`in`"];
1289 TokenTree::Token(token) => match token.kind {
1290 FatArrow | Comma | Eq | BinOp(token::Or) => IsInFollow::Yes,
1291 Ident(name, false) if name == kw::If || name == kw::In => IsInFollow::Yes,
1292 _ => IsInFollow::No(TOKENS),
1294 _ => IsInFollow::No(TOKENS),
1297 NonterminalKind::PatWithOr { .. } => {
1298 const TOKENS: &[&str] = &["`=>`", "`,`", "`=`", "`if`", "`in`"];
1300 TokenTree::Token(token) => match token.kind {
1301 FatArrow | Comma | Eq => IsInFollow::Yes,
1302 Ident(name, false) if name == kw::If || name == kw::In => IsInFollow::Yes,
1303 _ => IsInFollow::No(TOKENS),
1305 _ => IsInFollow::No(TOKENS),
1308 NonterminalKind::Path | NonterminalKind::Ty => {
1309 const TOKENS: &[&str] = &[
1310 "`{`", "`[`", "`=>`", "`,`", "`>`", "`=`", "`:`", "`;`", "`|`", "`as`",
1314 TokenTree::Token(token) => match token.kind {
1315 OpenDelim(Delimiter::Brace)
1316 | OpenDelim(Delimiter::Bracket)
1324 | BinOp(token::Or) => IsInFollow::Yes,
1325 Ident(name, false) if name == kw::As || name == kw::Where => {
1328 _ => IsInFollow::No(TOKENS),
1330 TokenTree::MetaVarDecl(_, _, Some(NonterminalKind::Block)) => IsInFollow::Yes,
1331 _ => IsInFollow::No(TOKENS),
1334 NonterminalKind::Ident | NonterminalKind::Lifetime => {
1335 // being a single token, idents and lifetimes are harmless
1338 NonterminalKind::Literal => {
1339 // literals may be of a single token, or two tokens (negative numbers)
1342 NonterminalKind::Meta | NonterminalKind::TT => {
1343 // being either a single token or a delimited sequence, tt is
1347 NonterminalKind::Vis => {
1348 // Explicitly disallow `priv`, on the off chance it comes back.
1349 const TOKENS: &[&str] = &["`,`", "an ident", "a type"];
1351 TokenTree::Token(token) => match token.kind {
1352 Comma => IsInFollow::Yes,
1353 Ident(name, is_raw) if is_raw || name != kw::Priv => IsInFollow::Yes,
1355 if token.can_begin_type() {
1358 IsInFollow::No(TOKENS)
1362 TokenTree::MetaVarDecl(
1365 Some(NonterminalKind::Ident | NonterminalKind::Ty | NonterminalKind::Path),
1366 ) => IsInFollow::Yes,
1367 _ => IsInFollow::No(TOKENS),
1374 fn quoted_tt_to_string(tt: &mbe::TokenTree) -> String {
1376 mbe::TokenTree::Token(ref token) => pprust::token_to_string(&token).into(),
1377 mbe::TokenTree::MetaVar(_, name) => format!("${}", name),
1378 mbe::TokenTree::MetaVarDecl(_, name, Some(kind)) => format!("${}:{}", name, kind),
1379 mbe::TokenTree::MetaVarDecl(_, name, None) => format!("${}:", name),
1382 "unexpected mbe::TokenTree::{Sequence or Delimited} \
1383 in follow set checker"
1388 pub(super) fn parser_from_cx(sess: &ParseSess, tts: TokenStream, recovery: Recovery) -> Parser<'_> {
1389 Parser::new(sess, tts, true, rustc_parse::MACRO_ARGUMENTS).recovery(recovery)