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::macro_check;
6 use crate::mbe::macro_parser::{Error, ErrorReported, Failure, Success, TtParser};
7 use crate::mbe::macro_parser::{MatchedSeq, MatchedTokenTree, MatcherLoc};
8 use crate::mbe::transcribe::transcribe;
11 use rustc_ast::token::{self, Delimiter, NonterminalKind, Token, TokenKind, TokenKind::*};
12 use rustc_ast::tokenstream::{DelimSpan, TokenStream};
13 use rustc_ast::{NodeId, DUMMY_NODE_ID};
14 use rustc_ast_pretty::pprust;
15 use rustc_attr::{self as attr, TransparencyError};
16 use rustc_data_structures::fx::FxHashMap;
17 use rustc_errors::{Applicability, Diagnostic, DiagnosticBuilder};
18 use rustc_feature::Features;
19 use rustc_lint_defs::builtin::{
20 RUST_2021_INCOMPATIBLE_OR_PATTERNS, SEMICOLON_IN_EXPRESSIONS_FROM_MACROS,
22 use rustc_lint_defs::BuiltinLintDiagnostics;
23 use rustc_parse::parser::Parser;
24 use rustc_session::parse::ParseSess;
25 use rustc_session::Session;
26 use rustc_span::edition::Edition;
27 use rustc_span::hygiene::Transparency;
28 use rustc_span::symbol::{kw, sym, Ident, MacroRulesNormalizedIdent};
32 use std::collections::hash_map::Entry;
33 use std::{mem, slice};
36 pub(crate) struct ParserAnyMacro<'a> {
39 /// Span of the expansion site of the macro this parser is for
41 /// The ident of the macro we're parsing
44 is_trailing_mac: bool,
46 /// Whether or not this macro is defined in the current crate
50 pub(crate) fn annotate_err_with_kind(err: &mut Diagnostic, kind: AstFragmentKind, span: Span) {
52 AstFragmentKind::Ty => {
53 err.span_label(span, "this macro call doesn't expand to a type");
55 AstFragmentKind::Pat => {
56 err.span_label(span, "this macro call doesn't expand to a pattern");
62 fn emit_frag_parse_err(
63 mut e: DiagnosticBuilder<'_, rustc_errors::ErrorGuaranteed>,
65 orig_parser: &mut Parser<'_>,
68 kind: AstFragmentKind,
70 // FIXME(davidtwco): avoid depending on the error message text
71 if parser.token == token::Eof && e.message[0].0.expect_str().ends_with(", found `<eof>`") {
72 if !e.span.is_dummy() {
73 // early end of macro arm (#52866)
74 e.replace_span_with(parser.sess.source_map().next_point(parser.token.span));
76 let msg = &e.message[0];
78 rustc_errors::DiagnosticMessage::Str(format!(
79 "macro expansion ends with an incomplete expression: {}",
80 msg.0.expect_str().replace(", found `<eof>`", ""),
85 if e.span.is_dummy() {
86 // Get around lack of span in error (#30128)
87 e.replace_span_with(site_span);
88 if !parser.sess.source_map().is_imported(arm_span) {
89 e.span_label(arm_span, "in this macro arm");
91 } else if parser.sess.source_map().is_imported(parser.token.span) {
92 e.span_label(site_span, "in this macro invocation");
95 // Try a statement if an expression is wanted but failed and suggest adding `;` to call.
96 AstFragmentKind::Expr => match parse_ast_fragment(orig_parser, AstFragmentKind::Stmts) {
97 Err(err) => err.cancel(),
100 "the macro call doesn't expand to an expression, but it can expand to a statement",
102 e.span_suggestion_verbose(
103 site_span.shrink_to_hi(),
104 "add `;` to interpret the expansion as a statement",
106 Applicability::MaybeIncorrect,
110 _ => annotate_err_with_kind(&mut e, kind, site_span),
115 impl<'a> ParserAnyMacro<'a> {
116 pub(crate) fn make(mut self: Box<ParserAnyMacro<'a>>, kind: AstFragmentKind) -> AstFragment {
126 let snapshot = &mut parser.clone();
127 let fragment = match parse_ast_fragment(parser, kind) {
130 emit_frag_parse_err(err, parser, snapshot, site_span, arm_span, kind);
131 return kind.dummy(site_span);
135 // We allow semicolons at the end of expressions -- e.g., the semicolon in
136 // `macro_rules! m { () => { panic!(); } }` isn't parsed by `.parse_expr()`,
137 // but `m!()` is allowed in expression positions (cf. issue #34706).
138 if kind == AstFragmentKind::Expr && parser.token == token::Semi {
140 parser.sess.buffer_lint_with_diagnostic(
141 SEMICOLON_IN_EXPRESSIONS_FROM_MACROS,
144 "trailing semicolon in macro used in expression position",
145 BuiltinLintDiagnostics::TrailingMacro(is_trailing_mac, macro_ident),
151 // Make sure we don't have any tokens left to parse so we don't silently drop anything.
152 let path = ast::Path::from_ident(macro_ident.with_span_pos(site_span));
153 ensure_complete_parse(parser, &path, kind.name(), site_span);
158 struct MacroRulesMacroExpander {
162 transparency: Transparency,
163 lhses: Vec<Vec<MatcherLoc>>,
164 rhses: Vec<mbe::TokenTree>,
168 impl TTMacroExpander for MacroRulesMacroExpander {
171 cx: &'cx mut ExtCtxt<'_>,
174 ) -> Box<dyn MacResult + 'cx> {
176 return DummyResult::any(sp);
192 fn macro_rules_dummy_expander<'cx>(
193 _: &'cx mut ExtCtxt<'_>,
196 ) -> Box<dyn MacResult + 'cx> {
197 DummyResult::any(span)
200 fn trace_macros_note(cx_expansions: &mut FxHashMap<Span, Vec<String>>, sp: Span, message: String) {
201 let sp = sp.macro_backtrace().last().map_or(sp, |trace| trace.call_site);
202 cx_expansions.entry(sp).or_default().push(message);
205 /// Expands the rules based macro defined by `lhses` and `rhses` for a given
207 fn expand_macro<'cx>(
208 cx: &'cx mut ExtCtxt<'_>,
213 transparency: Transparency,
215 lhses: &[Vec<MatcherLoc>],
216 rhses: &[mbe::TokenTree],
217 ) -> Box<dyn MacResult + 'cx> {
218 let sess = &cx.sess.parse_sess;
219 // Macros defined in the current crate have a real node id,
220 // whereas macros from an external crate have a dummy id.
221 let is_local = node_id != DUMMY_NODE_ID;
223 if cx.trace_macros() {
224 let msg = format!("expanding `{}! {{ {} }}`", name, pprust::tts_to_string(&arg));
225 trace_macros_note(&mut cx.expansions, sp, msg);
228 // Which arm's failure should we report? (the one furthest along)
229 let mut best_failure: Option<(Token, &str)> = None;
231 // We create a base parser that can be used for the "black box" parts.
232 // Every iteration needs a fresh copy of that parser. However, the parser
233 // is not mutated on many of the iterations, particularly when dealing with
236 // macro_rules! foo {
240 // // ... etc. (maybe hundreds more)
243 // as seen in the `html5ever` benchmark. We use a `Cow` so that the base
244 // parser is only cloned when necessary (upon mutation). Furthermore, we
245 // reinitialize the `Cow` with the base parser at the start of every
246 // iteration, so that any mutated parsers are not reused. This is all quite
247 // hacky, but speeds up the `html5ever` benchmark significantly. (Issue
248 // 68836 suggests a more comprehensive but more complex change to deal with
250 let parser = parser_from_cx(sess, arg.clone());
252 // Try each arm's matchers.
253 let mut tt_parser = TtParser::new(name);
254 for (i, lhs) in lhses.iter().enumerate() {
255 // Take a snapshot of the state of pre-expansion gating at this point.
256 // This is used so that if a matcher is not `Success(..)`ful,
257 // then the spans which became gated when parsing the unsuccessful matcher
258 // are not recorded. On the first `Success(..)`ful matcher, the spans are merged.
259 let mut gated_spans_snapshot = mem::take(&mut *sess.gated_spans.spans.borrow_mut());
261 match tt_parser.parse_tt(&mut Cow::Borrowed(&parser), lhs) {
262 Success(named_matches) => {
263 // The matcher was `Success(..)`ful.
264 // Merge the gated spans from parsing the matcher with the pre-existing ones.
265 sess.gated_spans.merge(gated_spans_snapshot);
267 let (rhs, rhs_span): (&mbe::Delimited, DelimSpan) = match &rhses[i] {
268 mbe::TokenTree::Delimited(span, delimited) => (&delimited, *span),
269 _ => cx.span_bug(sp, "malformed macro rhs"),
271 let arm_span = rhses[i].span();
273 let rhs_spans = rhs.tts.iter().map(|t| t.span()).collect::<Vec<_>>();
274 // rhs has holes ( `$id` and `$(...)` that need filled)
275 let mut tts = match transcribe(cx, &named_matches, &rhs, rhs_span, transparency) {
279 return DummyResult::any(arm_span);
283 // Replace all the tokens for the corresponding positions in the macro, to maintain
284 // proper positions in error reporting, while maintaining the macro_backtrace.
285 if rhs_spans.len() == tts.len() {
286 tts = tts.map_enumerated(|i, tt| {
287 let mut tt = tt.clone();
288 let mut sp = rhs_spans[i];
289 sp = sp.with_ctxt(tt.span().ctxt());
295 if cx.trace_macros() {
296 let msg = format!("to `{}`", pprust::tts_to_string(&tts));
297 trace_macros_note(&mut cx.expansions, sp, msg);
300 let mut p = Parser::new(sess, tts, false, None);
301 p.last_type_ascription = cx.current_expansion.prior_type_ascription;
304 cx.resolver.record_macro_rule_usage(node_id, i);
307 // Let the context choose how to interpret the result.
308 // Weird, but useful for X-macros.
309 return Box::new(ParserAnyMacro {
312 // Pass along the original expansion site and the name of the macro
313 // so we can print a useful error message if the parse of the expanded
314 // macro leaves unparsed tokens.
317 lint_node_id: cx.current_expansion.lint_node_id,
318 is_trailing_mac: cx.current_expansion.is_trailing_mac,
323 Failure(token, msg) => match best_failure {
324 Some((ref best_token, _)) if best_token.span.lo() >= token.span.lo() => {}
325 _ => best_failure = Some((token, msg)),
327 Error(err_sp, ref msg) => {
328 let span = err_sp.substitute_dummy(sp);
329 cx.struct_span_err(span, &msg).emit();
330 return DummyResult::any(span);
332 ErrorReported => return DummyResult::any(sp),
335 // The matcher was not `Success(..)`ful.
336 // Restore to the state before snapshotting and maybe try again.
337 mem::swap(&mut gated_spans_snapshot, &mut sess.gated_spans.spans.borrow_mut());
341 let (token, label) = best_failure.expect("ran no matchers");
342 let span = token.span.substitute_dummy(sp);
343 let mut err = cx.struct_span_err(span, &parse_failure_msg(&token));
344 err.span_label(span, label);
345 if !def_span.is_dummy() && !cx.source_map().is_imported(def_span) {
346 err.span_label(cx.source_map().guess_head_span(def_span), "when calling this macro");
349 // Check whether there's a missing comma in this macro call, like `println!("{}" a);`
350 if let Some((arg, comma_span)) = arg.add_comma() {
352 let parser = parser_from_cx(sess, arg.clone());
353 if let Success(_) = tt_parser.parse_tt(&mut Cow::Borrowed(&parser), lhs) {
354 if comma_span.is_dummy() {
355 err.note("you might be missing a comma");
357 err.span_suggestion_short(
359 "missing comma here",
361 Applicability::MachineApplicable,
368 cx.trace_macros_diag();
372 // Note that macro-by-example's input is also matched against a token tree:
373 // $( $lhs:tt => $rhs:tt );+
375 // Holy self-referential!
377 /// Converts a macro item into a syntax extension.
378 pub fn compile_declarative_macro(
383 ) -> (SyntaxExtension, Vec<Span>) {
384 debug!("compile_declarative_macro: {:?}", def);
385 let mk_syn_ext = |expander| {
386 SyntaxExtension::new(
388 SyntaxExtensionKind::LegacyBang(expander),
396 let dummy_syn_ext = || (mk_syn_ext(Box::new(macro_rules_dummy_expander)), Vec::new());
398 let diag = &sess.parse_sess.span_diagnostic;
399 let lhs_nm = Ident::new(sym::lhs, def.span);
400 let rhs_nm = Ident::new(sym::rhs, def.span);
401 let tt_spec = Some(NonterminalKind::TT);
403 // Parse the macro_rules! invocation
404 let (macro_rules, body) = match &def.kind {
405 ast::ItemKind::MacroDef(def) => (def.macro_rules, def.body.inner_tokens()),
409 // The pattern that macro_rules matches.
410 // The grammar for macro_rules! is:
411 // $( $lhs:tt => $rhs:tt );+
412 // ...quasiquoting this would be nice.
413 // These spans won't matter, anyways
414 let argument_gram = vec![
415 mbe::TokenTree::Sequence(
417 mbe::SequenceRepetition {
419 mbe::TokenTree::MetaVarDecl(def.span, lhs_nm, tt_spec),
420 mbe::TokenTree::token(token::FatArrow, def.span),
421 mbe::TokenTree::MetaVarDecl(def.span, rhs_nm, tt_spec),
423 separator: Some(Token::new(
424 if macro_rules { token::Semi } else { token::Comma },
427 kleene: mbe::KleeneToken::new(mbe::KleeneOp::OneOrMore, def.span),
431 // to phase into semicolon-termination instead of semicolon-separation
432 mbe::TokenTree::Sequence(
434 mbe::SequenceRepetition {
435 tts: vec![mbe::TokenTree::token(
436 if macro_rules { token::Semi } else { token::Comma },
440 kleene: mbe::KleeneToken::new(mbe::KleeneOp::ZeroOrMore, def.span),
445 // Convert it into `MatcherLoc` form.
446 let argument_gram = mbe::macro_parser::compute_locs(&argument_gram);
448 let parser = Parser::new(&sess.parse_sess, body, true, rustc_parse::MACRO_ARGUMENTS);
450 TtParser::new(Ident::with_dummy_span(if macro_rules { kw::MacroRules } else { kw::Macro }));
451 let argument_map = match tt_parser.parse_tt(&mut Cow::Borrowed(&parser), &argument_gram) {
453 Failure(token, msg) => {
454 let s = parse_failure_msg(&token);
455 let sp = token.span.substitute_dummy(def.span);
456 sess.parse_sess.span_diagnostic.struct_span_err(sp, &s).span_label(sp, msg).emit();
457 return dummy_syn_ext();
462 .struct_span_err(sp.substitute_dummy(def.span), &msg)
464 return dummy_syn_ext();
467 return dummy_syn_ext();
471 let mut valid = true;
473 // Extract the arguments:
474 let lhses = match argument_map[&MacroRulesNormalizedIdent::new(lhs_nm)] {
475 MatchedSeq(ref s) => s
478 if let MatchedTokenTree(ref tt) = *m {
479 let tt = mbe::quoted::parse(
489 valid &= check_lhs_nt_follows(&sess.parse_sess, &def, &tt);
492 sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs")
494 .collect::<Vec<mbe::TokenTree>>(),
495 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs"),
498 let rhses = match argument_map[&MacroRulesNormalizedIdent::new(rhs_nm)] {
499 MatchedSeq(ref s) => s
502 if let MatchedTokenTree(ref tt) = *m {
503 return mbe::quoted::parse(
514 sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs")
516 .collect::<Vec<mbe::TokenTree>>(),
517 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured rhs"),
521 valid &= check_rhs(&sess.parse_sess, rhs);
524 // don't abort iteration early, so that errors for multiple lhses can be reported
526 valid &= check_lhs_no_empty_seq(&sess.parse_sess, slice::from_ref(lhs));
529 valid &= macro_check::check_meta_variables(&sess.parse_sess, def.id, def.span, &lhses, &rhses);
531 let (transparency, transparency_error) = attr::find_transparency(&def.attrs, macro_rules);
532 match transparency_error {
533 Some(TransparencyError::UnknownTransparency(value, span)) => {
534 diag.span_err(span, &format!("unknown macro transparency: `{}`", value));
536 Some(TransparencyError::MultipleTransparencyAttrs(old_span, new_span)) => {
537 diag.span_err(vec![old_span, new_span], "multiple macro transparency attributes");
542 // Compute the spans of the macro rules
543 // We only take the span of the lhs here,
544 // so that the spans of created warnings are smaller.
545 let rule_spans = if def.id != DUMMY_NODE_ID {
546 lhses.iter().map(|lhs| lhs.span()).collect::<Vec<_>>()
551 // Convert the lhses into `MatcherLoc` form, which is better for doing the
552 // actual matching. Unless the matcher is invalid.
553 let lhses = if valid {
557 // Ignore the delimiters around the matcher.
559 mbe::TokenTree::Delimited(_, delimited) => {
560 mbe::macro_parser::compute_locs(&delimited.tts)
562 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "malformed macro lhs"),
570 let expander = Box::new(MacroRulesMacroExpander {
579 (mk_syn_ext(expander), rule_spans)
582 fn check_lhs_nt_follows(sess: &ParseSess, def: &ast::Item, lhs: &mbe::TokenTree) -> bool {
583 // lhs is going to be like TokenTree::Delimited(...), where the
584 // entire lhs is those tts. Or, it can be a "bare sequence", not wrapped in parens.
585 if let mbe::TokenTree::Delimited(_, delimited) = lhs {
586 check_matcher(sess, def, &delimited.tts)
588 let msg = "invalid macro matcher; matchers must be contained in balanced delimiters";
589 sess.span_diagnostic.span_err(lhs.span(), msg);
592 // we don't abort on errors on rejection, the driver will do that for us
593 // after parsing/expansion. we can report every error in every macro this way.
596 /// Checks that the lhs contains no repetition which could match an empty token
597 /// tree, because then the matcher would hang indefinitely.
598 fn check_lhs_no_empty_seq(sess: &ParseSess, tts: &[mbe::TokenTree]) -> bool {
603 | TokenTree::MetaVar(..)
604 | TokenTree::MetaVarDecl(..)
605 | TokenTree::MetaVarExpr(..) => (),
606 TokenTree::Delimited(_, ref del) => {
607 if !check_lhs_no_empty_seq(sess, &del.tts) {
611 TokenTree::Sequence(span, ref seq) => {
612 if seq.separator.is_none()
613 && seq.tts.iter().all(|seq_tt| match *seq_tt {
614 TokenTree::MetaVarDecl(_, _, Some(NonterminalKind::Vis)) => true,
615 TokenTree::Sequence(_, ref sub_seq) => {
616 sub_seq.kleene.op == mbe::KleeneOp::ZeroOrMore
617 || sub_seq.kleene.op == mbe::KleeneOp::ZeroOrOne
622 let sp = span.entire();
623 sess.span_diagnostic.span_err(sp, "repetition matches empty token tree");
626 if !check_lhs_no_empty_seq(sess, &seq.tts) {
636 fn check_rhs(sess: &ParseSess, rhs: &mbe::TokenTree) -> bool {
638 mbe::TokenTree::Delimited(..) => return true,
640 sess.span_diagnostic.span_err(rhs.span(), "macro rhs must be delimited");
646 fn check_matcher(sess: &ParseSess, def: &ast::Item, matcher: &[mbe::TokenTree]) -> bool {
647 let first_sets = FirstSets::new(matcher);
648 let empty_suffix = TokenSet::empty();
649 let err = sess.span_diagnostic.err_count();
650 check_matcher_core(sess, def, &first_sets, matcher, &empty_suffix);
651 err == sess.span_diagnostic.err_count()
654 // `The FirstSets` for a matcher is a mapping from subsequences in the
655 // matcher to the FIRST set for that subsequence.
657 // This mapping is partially precomputed via a backwards scan over the
658 // token trees of the matcher, which provides a mapping from each
659 // repetition sequence to its *first* set.
661 // (Hypothetically, sequences should be uniquely identifiable via their
662 // spans, though perhaps that is false, e.g., for macro-generated macros
663 // that do not try to inject artificial span information. My plan is
664 // to try to catch such cases ahead of time and not include them in
665 // the precomputed mapping.)
666 struct FirstSets<'tt> {
667 // this maps each TokenTree::Sequence `$(tt ...) SEP OP` that is uniquely identified by its
668 // span in the original matcher to the First set for the inner sequence `tt ...`.
670 // If two sequences have the same span in a matcher, then map that
671 // span to None (invalidating the mapping here and forcing the code to
673 first: FxHashMap<Span, Option<TokenSet<'tt>>>,
676 impl<'tt> FirstSets<'tt> {
677 fn new(tts: &'tt [mbe::TokenTree]) -> FirstSets<'tt> {
680 let mut sets = FirstSets { first: FxHashMap::default() };
681 build_recur(&mut sets, tts);
684 // walks backward over `tts`, returning the FIRST for `tts`
685 // and updating `sets` at the same time for all sequence
686 // substructure we find within `tts`.
687 fn build_recur<'tt>(sets: &mut FirstSets<'tt>, tts: &'tt [TokenTree]) -> TokenSet<'tt> {
688 let mut first = TokenSet::empty();
689 for tt in tts.iter().rev() {
692 | TokenTree::MetaVar(..)
693 | TokenTree::MetaVarDecl(..)
694 | TokenTree::MetaVarExpr(..) => {
695 first.replace_with(TtHandle::TtRef(tt));
697 TokenTree::Delimited(span, ref delimited) => {
698 build_recur(sets, &delimited.tts);
699 first.replace_with(TtHandle::from_token_kind(
700 token::OpenDelim(delimited.delim),
704 TokenTree::Sequence(sp, ref seq_rep) => {
705 let subfirst = build_recur(sets, &seq_rep.tts);
707 match sets.first.entry(sp.entire()) {
708 Entry::Vacant(vac) => {
709 vac.insert(Some(subfirst.clone()));
711 Entry::Occupied(mut occ) => {
712 // if there is already an entry, then a span must have collided.
713 // This should not happen with typical macro_rules macros,
714 // but syntax extensions need not maintain distinct spans,
715 // so distinct syntax trees can be assigned the same span.
716 // In such a case, the map cannot be trusted; so mark this
717 // entry as unusable.
722 // If the sequence contents can be empty, then the first
723 // token could be the separator token itself.
725 if let (Some(sep), true) = (&seq_rep.separator, subfirst.maybe_empty) {
726 first.add_one_maybe(TtHandle::from_token(sep.clone()));
729 // Reverse scan: Sequence comes before `first`.
730 if subfirst.maybe_empty
731 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrMore
732 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrOne
734 // If sequence is potentially empty, then
735 // union them (preserving first emptiness).
736 first.add_all(&TokenSet { maybe_empty: true, ..subfirst });
738 // Otherwise, sequence guaranteed
739 // non-empty; replace first.
750 // walks forward over `tts` until all potential FIRST tokens are
752 fn first(&self, tts: &'tt [mbe::TokenTree]) -> TokenSet<'tt> {
755 let mut first = TokenSet::empty();
756 for tt in tts.iter() {
757 assert!(first.maybe_empty);
760 | TokenTree::MetaVar(..)
761 | TokenTree::MetaVarDecl(..)
762 | TokenTree::MetaVarExpr(..) => {
763 first.add_one(TtHandle::TtRef(tt));
766 TokenTree::Delimited(span, ref delimited) => {
767 first.add_one(TtHandle::from_token_kind(
768 token::OpenDelim(delimited.delim),
773 TokenTree::Sequence(sp, ref seq_rep) => {
775 let subfirst = match self.first.get(&sp.entire()) {
776 Some(&Some(ref subfirst)) => subfirst,
778 subfirst_owned = self.first(&seq_rep.tts);
782 panic!("We missed a sequence during FirstSets construction");
786 // If the sequence contents can be empty, then the first
787 // token could be the separator token itself.
788 if let (Some(sep), true) = (&seq_rep.separator, subfirst.maybe_empty) {
789 first.add_one_maybe(TtHandle::from_token(sep.clone()));
792 assert!(first.maybe_empty);
793 first.add_all(subfirst);
794 if subfirst.maybe_empty
795 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrMore
796 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrOne
798 // Continue scanning for more first
799 // tokens, but also make sure we
800 // restore empty-tracking state.
801 first.maybe_empty = true;
810 // we only exit the loop if `tts` was empty or if every
811 // element of `tts` matches the empty sequence.
812 assert!(first.maybe_empty);
817 // Most `mbe::TokenTree`s are pre-existing in the matcher, but some are defined
818 // implicitly, such as opening/closing delimiters and sequence repetition ops.
819 // This type encapsulates both kinds. It implements `Clone` while avoiding the
820 // need for `mbe::TokenTree` to implement `Clone`.
823 /// This is used in most cases.
824 TtRef(&'tt mbe::TokenTree),
826 /// This is only used for implicit token trees. The `mbe::TokenTree` *must*
827 /// be `mbe::TokenTree::Token`. No other variants are allowed. We store an
828 /// `mbe::TokenTree` rather than a `Token` so that `get()` can return a
829 /// `&mbe::TokenTree`.
830 Token(mbe::TokenTree),
833 impl<'tt> TtHandle<'tt> {
834 fn from_token(tok: Token) -> Self {
835 TtHandle::Token(mbe::TokenTree::Token(tok))
838 fn from_token_kind(kind: TokenKind, span: Span) -> Self {
839 TtHandle::from_token(Token::new(kind, span))
842 // Get a reference to a token tree.
843 fn get(&'tt self) -> &'tt mbe::TokenTree {
845 TtHandle::TtRef(tt) => tt,
846 TtHandle::Token(token_tt) => &token_tt,
851 impl<'tt> PartialEq for TtHandle<'tt> {
852 fn eq(&self, other: &TtHandle<'tt>) -> bool {
853 self.get() == other.get()
857 impl<'tt> Clone for TtHandle<'tt> {
858 fn clone(&self) -> Self {
860 TtHandle::TtRef(tt) => TtHandle::TtRef(tt),
862 // This variant *must* contain a `mbe::TokenTree::Token`, and not
863 // any other variant of `mbe::TokenTree`.
864 TtHandle::Token(mbe::TokenTree::Token(tok)) => {
865 TtHandle::Token(mbe::TokenTree::Token(tok.clone()))
873 // A set of `mbe::TokenTree`s, which may include `TokenTree::Match`s
874 // (for macro-by-example syntactic variables). It also carries the
875 // `maybe_empty` flag; that is true if and only if the matcher can
876 // match an empty token sequence.
878 // The First set is computed on submatchers like `$($a:expr b),* $(c)* d`,
879 // which has corresponding FIRST = {$a:expr, c, d}.
880 // Likewise, `$($a:expr b),* $(c)+ d` has FIRST = {$a:expr, c}.
882 // (Notably, we must allow for *-op to occur zero times.)
883 #[derive(Clone, Debug)]
884 struct TokenSet<'tt> {
885 tokens: Vec<TtHandle<'tt>>,
889 impl<'tt> TokenSet<'tt> {
890 // Returns a set for the empty sequence.
892 TokenSet { tokens: Vec::new(), maybe_empty: true }
895 // Returns the set `{ tok }` for the single-token (and thus
896 // non-empty) sequence [tok].
897 fn singleton(tt: TtHandle<'tt>) -> Self {
898 TokenSet { tokens: vec![tt], maybe_empty: false }
901 // Changes self to be the set `{ tok }`.
902 // Since `tok` is always present, marks self as non-empty.
903 fn replace_with(&mut self, tt: TtHandle<'tt>) {
905 self.tokens.push(tt);
906 self.maybe_empty = false;
909 // Changes self to be the empty set `{}`; meant for use when
910 // the particular token does not matter, but we want to
911 // record that it occurs.
912 fn replace_with_irrelevant(&mut self) {
914 self.maybe_empty = false;
917 // Adds `tok` to the set for `self`, marking sequence as non-empy.
918 fn add_one(&mut self, tt: TtHandle<'tt>) {
919 if !self.tokens.contains(&tt) {
920 self.tokens.push(tt);
922 self.maybe_empty = false;
925 // Adds `tok` to the set for `self`. (Leaves `maybe_empty` flag alone.)
926 fn add_one_maybe(&mut self, tt: TtHandle<'tt>) {
927 if !self.tokens.contains(&tt) {
928 self.tokens.push(tt);
932 // Adds all elements of `other` to this.
934 // (Since this is a set, we filter out duplicates.)
936 // If `other` is potentially empty, then preserves the previous
937 // setting of the empty flag of `self`. If `other` is guaranteed
938 // non-empty, then `self` is marked non-empty.
939 fn add_all(&mut self, other: &Self) {
940 for tt in &other.tokens {
941 if !self.tokens.contains(tt) {
942 self.tokens.push(tt.clone());
945 if !other.maybe_empty {
946 self.maybe_empty = false;
951 // Checks that `matcher` is internally consistent and that it
952 // can legally be followed by a token `N`, for all `N` in `follow`.
953 // (If `follow` is empty, then it imposes no constraint on
956 // Returns the set of NT tokens that could possibly come last in
957 // `matcher`. (If `matcher` matches the empty sequence, then
958 // `maybe_empty` will be set to true.)
960 // Requires that `first_sets` is pre-computed for `matcher`;
961 // see `FirstSets::new`.
962 fn check_matcher_core<'tt>(
965 first_sets: &FirstSets<'tt>,
966 matcher: &'tt [mbe::TokenTree],
967 follow: &TokenSet<'tt>,
971 let mut last = TokenSet::empty();
973 // 2. For each token and suffix [T, SUFFIX] in M:
974 // ensure that T can be followed by SUFFIX, and if SUFFIX may be empty,
975 // then ensure T can also be followed by any element of FOLLOW.
976 'each_token: for i in 0..matcher.len() {
977 let token = &matcher[i];
978 let suffix = &matcher[i + 1..];
980 let build_suffix_first = || {
981 let mut s = first_sets.first(suffix);
988 // (we build `suffix_first` on demand below; you can tell
989 // which cases are supposed to fall through by looking for the
990 // initialization of this variable.)
993 // First, update `last` so that it corresponds to the set
994 // of NT tokens that might end the sequence `... token`.
997 | TokenTree::MetaVar(..)
998 | TokenTree::MetaVarDecl(..)
999 | TokenTree::MetaVarExpr(..) => {
1000 if token_can_be_followed_by_any(token) {
1001 // don't need to track tokens that work with any,
1002 last.replace_with_irrelevant();
1003 // ... and don't need to check tokens that can be
1004 // followed by anything against SUFFIX.
1005 continue 'each_token;
1007 last.replace_with(TtHandle::TtRef(token));
1008 suffix_first = build_suffix_first();
1011 TokenTree::Delimited(span, ref d) => {
1012 let my_suffix = TokenSet::singleton(TtHandle::from_token_kind(
1013 token::CloseDelim(d.delim),
1016 check_matcher_core(sess, def, first_sets, &d.tts, &my_suffix);
1017 // don't track non NT tokens
1018 last.replace_with_irrelevant();
1020 // also, we don't need to check delimited sequences
1022 continue 'each_token;
1024 TokenTree::Sequence(_, ref seq_rep) => {
1025 suffix_first = build_suffix_first();
1026 // The trick here: when we check the interior, we want
1027 // to include the separator (if any) as a potential
1028 // (but not guaranteed) element of FOLLOW. So in that
1029 // case, we make a temp copy of suffix and stuff
1030 // delimiter in there.
1032 // FIXME: Should I first scan suffix_first to see if
1033 // delimiter is already in it before I go through the
1034 // work of cloning it? But then again, this way I may
1035 // get a "tighter" span?
1037 let my_suffix = if let Some(sep) = &seq_rep.separator {
1038 new = suffix_first.clone();
1039 new.add_one_maybe(TtHandle::from_token(sep.clone()));
1045 // At this point, `suffix_first` is built, and
1046 // `my_suffix` is some TokenSet that we can use
1047 // for checking the interior of `seq_rep`.
1048 let next = check_matcher_core(sess, def, first_sets, &seq_rep.tts, my_suffix);
1049 if next.maybe_empty {
1050 last.add_all(&next);
1055 // the recursive call to check_matcher_core already ran the 'each_last
1056 // check below, so we can just keep going forward here.
1057 continue 'each_token;
1061 // (`suffix_first` guaranteed initialized once reaching here.)
1063 // Now `last` holds the complete set of NT tokens that could
1064 // end the sequence before SUFFIX. Check that every one works with `suffix`.
1065 for tt in &last.tokens {
1066 if let &TokenTree::MetaVarDecl(span, name, Some(kind)) = tt.get() {
1067 for next_token in &suffix_first.tokens {
1068 let next_token = next_token.get();
1070 // Check if the old pat is used and the next token is `|`
1071 // to warn about incompatibility with Rust 2021.
1072 // We only emit this lint if we're parsing the original
1073 // definition of this macro_rules, not while (re)parsing
1074 // the macro when compiling another crate that is using the
1075 // macro. (See #86567.)
1076 // Macros defined in the current crate have a real node id,
1077 // whereas macros from an external crate have a dummy id.
1078 if def.id != DUMMY_NODE_ID
1079 && matches!(kind, NonterminalKind::PatParam { inferred: true })
1080 && matches!(next_token, TokenTree::Token(token) if token.kind == BinOp(token::BinOpToken::Or))
1082 // It is suggestion to use pat_param, for example: $x:pat -> $x:pat_param.
1083 let suggestion = quoted_tt_to_string(&TokenTree::MetaVarDecl(
1086 Some(NonterminalKind::PatParam { inferred: false }),
1088 sess.buffer_lint_with_diagnostic(
1089 &RUST_2021_INCOMPATIBLE_OR_PATTERNS,
1092 "the meaning of the `pat` fragment specifier is changing in Rust 2021, which may affect this macro",
1093 BuiltinLintDiagnostics::OrPatternsBackCompat(span, suggestion),
1096 match is_in_follow(next_token, kind) {
1097 IsInFollow::Yes => {}
1098 IsInFollow::No(possible) => {
1099 let may_be = if last.tokens.len() == 1 && suffix_first.tokens.len() == 1
1106 let sp = next_token.span();
1107 let mut err = sess.span_diagnostic.struct_span_err(
1110 "`${name}:{frag}` {may_be} followed by `{next}`, which \
1111 is not allowed for `{frag}` fragments",
1114 next = quoted_tt_to_string(next_token),
1118 err.span_label(sp, format!("not allowed after `{}` fragments", kind));
1120 if kind == NonterminalKind::PatWithOr
1121 && sess.edition.rust_2021()
1122 && next_token.is_token(&BinOp(token::BinOpToken::Or))
1124 let suggestion = quoted_tt_to_string(&TokenTree::MetaVarDecl(
1127 Some(NonterminalKind::PatParam { inferred: false }),
1129 err.span_suggestion(
1131 "try a `pat_param` fragment specifier instead",
1133 Applicability::MaybeIncorrect,
1137 let msg = "allowed there are: ";
1142 "only {} is allowed after `{}` fragments",
1153 .collect::<Vec<_>>()
1169 fn token_can_be_followed_by_any(tok: &mbe::TokenTree) -> bool {
1170 if let mbe::TokenTree::MetaVarDecl(_, _, Some(kind)) = *tok {
1171 frag_can_be_followed_by_any(kind)
1173 // (Non NT's can always be followed by anything in matchers.)
1178 /// Returns `true` if a fragment of type `frag` can be followed by any sort of
1179 /// token. We use this (among other things) as a useful approximation
1180 /// for when `frag` can be followed by a repetition like `$(...)*` or
1181 /// `$(...)+`. In general, these can be a bit tricky to reason about,
1182 /// so we adopt a conservative position that says that any fragment
1183 /// specifier which consumes at most one token tree can be followed by
1184 /// a fragment specifier (indeed, these fragments can be followed by
1185 /// ANYTHING without fear of future compatibility hazards).
1186 fn frag_can_be_followed_by_any(kind: NonterminalKind) -> bool {
1189 NonterminalKind::Item // always terminated by `}` or `;`
1190 | NonterminalKind::Block // exactly one token tree
1191 | NonterminalKind::Ident // exactly one token tree
1192 | NonterminalKind::Literal // exactly one token tree
1193 | NonterminalKind::Meta // exactly one token tree
1194 | NonterminalKind::Lifetime // exactly one token tree
1195 | NonterminalKind::TT // exactly one token tree
1201 No(&'static [&'static str]),
1204 /// Returns `true` if `frag` can legally be followed by the token `tok`. For
1205 /// fragments that can consume an unbounded number of tokens, `tok`
1206 /// must be within a well-defined follow set. This is intended to
1207 /// guarantee future compatibility: for example, without this rule, if
1208 /// we expanded `expr` to include a new binary operator, we might
1209 /// break macros that were relying on that binary operator as a
1211 // when changing this do not forget to update doc/book/macros.md!
1212 fn is_in_follow(tok: &mbe::TokenTree, kind: NonterminalKind) -> IsInFollow {
1215 if let TokenTree::Token(Token { kind: token::CloseDelim(_), .. }) = *tok {
1216 // closing a token tree can never be matched by any fragment;
1217 // iow, we always require that `(` and `)` match, etc.
1221 NonterminalKind::Item => {
1222 // since items *must* be followed by either a `;` or a `}`, we can
1223 // accept anything after them
1226 NonterminalKind::Block => {
1227 // anything can follow block, the braces provide an easy boundary to
1231 NonterminalKind::Stmt | NonterminalKind::Expr => {
1232 const TOKENS: &[&str] = &["`=>`", "`,`", "`;`"];
1234 TokenTree::Token(token) => match token.kind {
1235 FatArrow | Comma | Semi => IsInFollow::Yes,
1236 _ => IsInFollow::No(TOKENS),
1238 _ => IsInFollow::No(TOKENS),
1241 NonterminalKind::PatParam { .. } => {
1242 const TOKENS: &[&str] = &["`=>`", "`,`", "`=`", "`|`", "`if`", "`in`"];
1244 TokenTree::Token(token) => match token.kind {
1245 FatArrow | Comma | Eq | BinOp(token::Or) => IsInFollow::Yes,
1246 Ident(name, false) if name == kw::If || name == kw::In => IsInFollow::Yes,
1247 _ => IsInFollow::No(TOKENS),
1249 _ => IsInFollow::No(TOKENS),
1252 NonterminalKind::PatWithOr { .. } => {
1253 const TOKENS: &[&str] = &["`=>`", "`,`", "`=`", "`if`", "`in`"];
1255 TokenTree::Token(token) => match token.kind {
1256 FatArrow | Comma | Eq => IsInFollow::Yes,
1257 Ident(name, false) if name == kw::If || name == kw::In => IsInFollow::Yes,
1258 _ => IsInFollow::No(TOKENS),
1260 _ => IsInFollow::No(TOKENS),
1263 NonterminalKind::Path | NonterminalKind::Ty => {
1264 const TOKENS: &[&str] = &[
1265 "`{`", "`[`", "`=>`", "`,`", "`>`", "`=`", "`:`", "`;`", "`|`", "`as`",
1269 TokenTree::Token(token) => match token.kind {
1270 OpenDelim(Delimiter::Brace)
1271 | OpenDelim(Delimiter::Bracket)
1279 | BinOp(token::Or) => IsInFollow::Yes,
1280 Ident(name, false) if name == kw::As || name == kw::Where => {
1283 _ => IsInFollow::No(TOKENS),
1285 TokenTree::MetaVarDecl(_, _, Some(NonterminalKind::Block)) => IsInFollow::Yes,
1286 _ => IsInFollow::No(TOKENS),
1289 NonterminalKind::Ident | NonterminalKind::Lifetime => {
1290 // being a single token, idents and lifetimes are harmless
1293 NonterminalKind::Literal => {
1294 // literals may be of a single token, or two tokens (negative numbers)
1297 NonterminalKind::Meta | NonterminalKind::TT => {
1298 // being either a single token or a delimited sequence, tt is
1302 NonterminalKind::Vis => {
1303 // Explicitly disallow `priv`, on the off chance it comes back.
1304 const TOKENS: &[&str] = &["`,`", "an ident", "a type"];
1306 TokenTree::Token(token) => match token.kind {
1307 Comma => IsInFollow::Yes,
1308 Ident(name, is_raw) if is_raw || name != kw::Priv => IsInFollow::Yes,
1310 if token.can_begin_type() {
1313 IsInFollow::No(TOKENS)
1317 TokenTree::MetaVarDecl(
1320 Some(NonterminalKind::Ident | NonterminalKind::Ty | NonterminalKind::Path),
1321 ) => IsInFollow::Yes,
1322 _ => IsInFollow::No(TOKENS),
1329 fn quoted_tt_to_string(tt: &mbe::TokenTree) -> String {
1331 mbe::TokenTree::Token(ref token) => pprust::token_to_string(&token).into(),
1332 mbe::TokenTree::MetaVar(_, name) => format!("${}", name),
1333 mbe::TokenTree::MetaVarDecl(_, name, Some(kind)) => format!("${}:{}", name, kind),
1334 mbe::TokenTree::MetaVarDecl(_, name, None) => format!("${}:", name),
1337 "unexpected mbe::TokenTree::{Sequence or Delimited} \
1338 in follow set checker"
1343 fn parser_from_cx(sess: &ParseSess, tts: TokenStream) -> Parser<'_> {
1344 Parser::new(sess, tts, true, rustc_parse::MACRO_ARGUMENTS)
1347 /// Generates an appropriate parsing failure message. For EOF, this is "unexpected end...". For
1348 /// other tokens, this is "unexpected token...".
1349 fn parse_failure_msg(tok: &Token) -> String {
1351 token::Eof => "unexpected end of macro invocation".to_string(),
1352 _ => format!("no rules expected the token `{}`", pprust::token_to_string(tok),),