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::parse_tt;
7 use crate::mbe::macro_parser::{Error, ErrorReported, Failure, Success};
8 use crate::mbe::macro_parser::{MatchedNonterminal, MatchedSeq};
9 use crate::mbe::transcribe::transcribe;
12 use rustc_ast::token::{self, NonterminalKind, NtTT, Token, 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;
18 use rustc_data_structures::sync::Lrc;
19 use rustc_errors::{Applicability, DiagnosticBuilder};
20 use rustc_feature::Features;
21 use rustc_lint_defs::builtin::{
22 RUST_2021_INCOMPATIBLE_OR_PATTERNS, SEMICOLON_IN_EXPRESSIONS_FROM_MACROS,
24 use rustc_lint_defs::BuiltinLintDiagnostics;
25 use rustc_parse::parser::Parser;
26 use rustc_session::parse::ParseSess;
27 use rustc_session::Session;
28 use rustc_span::edition::Edition;
29 use rustc_span::hygiene::Transparency;
30 use rustc_span::symbol::{kw, sym, Ident, MacroRulesNormalizedIdent};
34 use std::collections::hash_map::Entry;
35 use std::{mem, slice};
38 crate struct ParserAnyMacro<'a> {
41 /// Span of the expansion site of the macro this parser is for
43 /// The ident of the macro we're parsing
49 crate fn annotate_err_with_kind(
50 err: &mut DiagnosticBuilder<'_>,
51 kind: AstFragmentKind,
55 AstFragmentKind::Ty => {
56 err.span_label(span, "this macro call doesn't expand to a type");
58 AstFragmentKind::Pat => {
59 err.span_label(span, "this macro call doesn't expand to a pattern");
65 fn emit_frag_parse_err(
66 mut e: DiagnosticBuilder<'_>,
68 orig_parser: &mut Parser<'_>,
71 kind: AstFragmentKind,
73 if parser.token == token::Eof && e.message().ends_with(", found `<eof>`") {
74 if !e.span.is_dummy() {
75 // early end of macro arm (#52866)
76 e.replace_span_with(parser.sess.source_map().next_point(parser.token.span));
78 let msg = &e.message[0];
81 "macro expansion ends with an incomplete expression: {}",
82 msg.0.replace(", found `<eof>`", ""),
87 if e.span.is_dummy() {
88 // Get around lack of span in error (#30128)
89 e.replace_span_with(site_span);
90 if !parser.sess.source_map().is_imported(arm_span) {
91 e.span_label(arm_span, "in this macro arm");
93 } else if parser.sess.source_map().is_imported(parser.token.span) {
94 e.span_label(site_span, "in this macro invocation");
97 // Try a statement if an expression is wanted but failed and suggest adding `;` to call.
98 AstFragmentKind::Expr => match parse_ast_fragment(orig_parser, AstFragmentKind::Stmts) {
99 Err(mut err) => err.cancel(),
102 "the macro call doesn't expand to an expression, but it can expand to a statement",
104 e.span_suggestion_verbose(
105 site_span.shrink_to_hi(),
106 "add `;` to interpret the expansion as a statement",
108 Applicability::MaybeIncorrect,
112 _ => annotate_err_with_kind(&mut e, kind, site_span),
117 impl<'a> ParserAnyMacro<'a> {
118 crate fn make(mut self: Box<ParserAnyMacro<'a>>, kind: AstFragmentKind) -> AstFragment {
119 let ParserAnyMacro { site_span, macro_ident, ref mut parser, lint_node_id, arm_span } =
121 let snapshot = &mut parser.clone();
122 let fragment = match parse_ast_fragment(parser, kind) {
125 emit_frag_parse_err(err, parser, snapshot, site_span, arm_span, kind);
126 return kind.dummy(site_span);
130 // We allow semicolons at the end of expressions -- e.g., the semicolon in
131 // `macro_rules! m { () => { panic!(); } }` isn't parsed by `.parse_expr()`,
132 // but `m!()` is allowed in expression positions (cf. issue #34706).
133 if kind == AstFragmentKind::Expr && parser.token == token::Semi {
134 parser.sess.buffer_lint(
135 SEMICOLON_IN_EXPRESSIONS_FROM_MACROS,
138 "trailing semicolon in macro used in expression position",
143 // Make sure we don't have any tokens left to parse so we don't silently drop anything.
144 let path = ast::Path::from_ident(macro_ident.with_span_pos(site_span));
145 ensure_complete_parse(parser, &path, kind.name(), site_span);
150 struct MacroRulesMacroExpander {
153 transparency: Transparency,
154 lhses: Vec<mbe::TokenTree>,
155 rhses: Vec<mbe::TokenTree>,
159 impl TTMacroExpander for MacroRulesMacroExpander {
162 cx: &'cx mut ExtCtxt<'_>,
165 ) -> Box<dyn MacResult + 'cx> {
167 return DummyResult::any(sp);
182 fn macro_rules_dummy_expander<'cx>(
183 _: &'cx mut ExtCtxt<'_>,
186 ) -> Box<dyn MacResult + 'cx> {
187 DummyResult::any(span)
190 fn trace_macros_note(cx_expansions: &mut FxHashMap<Span, Vec<String>>, sp: Span, message: String) {
191 let sp = sp.macro_backtrace().last().map_or(sp, |trace| trace.call_site);
192 cx_expansions.entry(sp).or_default().push(message);
195 /// Given `lhses` and `rhses`, this is the new macro we create
196 fn generic_extension<'cx>(
197 cx: &'cx mut ExtCtxt<'_>,
201 transparency: Transparency,
203 lhses: &[mbe::TokenTree],
204 rhses: &[mbe::TokenTree],
205 ) -> Box<dyn MacResult + 'cx> {
206 let sess = &cx.sess.parse_sess;
208 if cx.trace_macros() {
209 let msg = format!("expanding `{}! {{ {} }}`", name, pprust::tts_to_string(&arg));
210 trace_macros_note(&mut cx.expansions, sp, msg);
213 // Which arm's failure should we report? (the one furthest along)
214 let mut best_failure: Option<(Token, &str)> = None;
216 // We create a base parser that can be used for the "black box" parts.
217 // Every iteration needs a fresh copy of that parser. However, the parser
218 // is not mutated on many of the iterations, particularly when dealing with
221 // macro_rules! foo {
225 // // ... etc. (maybe hundreds more)
228 // as seen in the `html5ever` benchmark. We use a `Cow` so that the base
229 // parser is only cloned when necessary (upon mutation). Furthermore, we
230 // reinitialize the `Cow` with the base parser at the start of every
231 // iteration, so that any mutated parsers are not reused. This is all quite
232 // hacky, but speeds up the `html5ever` benchmark significantly. (Issue
233 // 68836 suggests a more comprehensive but more complex change to deal with
235 let parser = parser_from_cx(sess, arg.clone());
237 for (i, lhs) in lhses.iter().enumerate() {
238 // try each arm's matchers
239 let lhs_tt = match *lhs {
240 mbe::TokenTree::Delimited(_, ref delim) => &delim.tts[..],
241 _ => cx.span_bug(sp, "malformed macro lhs"),
244 // Take a snapshot of the state of pre-expansion gating at this point.
245 // This is used so that if a matcher is not `Success(..)`ful,
246 // then the spans which became gated when parsing the unsuccessful matcher
247 // are not recorded. On the first `Success(..)`ful matcher, the spans are merged.
248 let mut gated_spans_snapshot = mem::take(&mut *sess.gated_spans.spans.borrow_mut());
250 match parse_tt(&mut Cow::Borrowed(&parser), lhs_tt, name) {
251 Success(named_matches) => {
252 // The matcher was `Success(..)`ful.
253 // Merge the gated spans from parsing the matcher with the pre-existing ones.
254 sess.gated_spans.merge(gated_spans_snapshot);
256 let rhs = match rhses[i] {
258 mbe::TokenTree::Delimited(_, ref delimed) => delimed.tts.clone(),
259 _ => cx.span_bug(sp, "malformed macro rhs"),
261 let arm_span = rhses[i].span();
263 let rhs_spans = rhs.iter().map(|t| t.span()).collect::<Vec<_>>();
264 // rhs has holes ( `$id` and `$(...)` that need filled)
265 let mut tts = match transcribe(cx, &named_matches, rhs, transparency) {
269 return DummyResult::any(arm_span);
273 // Replace all the tokens for the corresponding positions in the macro, to maintain
274 // proper positions in error reporting, while maintaining the macro_backtrace.
275 if rhs_spans.len() == tts.len() {
276 tts = tts.map_enumerated(|i, tt| {
277 let mut tt = tt.clone();
278 let mut sp = rhs_spans[i];
279 sp = sp.with_ctxt(tt.span().ctxt());
285 if cx.trace_macros() {
286 let msg = format!("to `{}`", pprust::tts_to_string(&tts));
287 trace_macros_note(&mut cx.expansions, sp, msg);
290 let mut p = Parser::new(sess, tts, false, None);
291 p.last_type_ascription = cx.current_expansion.prior_type_ascription;
293 // Let the context choose how to interpret the result.
294 // Weird, but useful for X-macros.
295 return Box::new(ParserAnyMacro {
298 // Pass along the original expansion site and the name of the macro
299 // so we can print a useful error message if the parse of the expanded
300 // macro leaves unparsed tokens.
303 lint_node_id: cx.current_expansion.lint_node_id,
307 Failure(token, msg) => match best_failure {
308 Some((ref best_token, _)) if best_token.span.lo() >= token.span.lo() => {}
309 _ => best_failure = Some((token, msg)),
311 Error(err_sp, ref msg) => {
312 let span = err_sp.substitute_dummy(sp);
313 cx.struct_span_err(span, &msg).emit();
314 return DummyResult::any(span);
316 ErrorReported => return DummyResult::any(sp),
319 // The matcher was not `Success(..)`ful.
320 // Restore to the state before snapshotting and maybe try again.
321 mem::swap(&mut gated_spans_snapshot, &mut sess.gated_spans.spans.borrow_mut());
325 let (token, label) = best_failure.expect("ran no matchers");
326 let span = token.span.substitute_dummy(sp);
327 let mut err = cx.struct_span_err(span, &parse_failure_msg(&token));
328 err.span_label(span, label);
329 if !def_span.is_dummy() && !cx.source_map().is_imported(def_span) {
330 err.span_label(cx.source_map().guess_head_span(def_span), "when calling this macro");
333 // Check whether there's a missing comma in this macro call, like `println!("{}" a);`
334 if let Some((arg, comma_span)) = arg.add_comma() {
336 // try each arm's matchers
337 let lhs_tt = match *lhs {
338 mbe::TokenTree::Delimited(_, ref delim) => &delim.tts[..],
342 parse_tt(&mut Cow::Borrowed(&parser_from_cx(sess, arg.clone())), lhs_tt, name)
344 if comma_span.is_dummy() {
345 err.note("you might be missing a comma");
347 err.span_suggestion_short(
349 "missing comma here",
351 Applicability::MachineApplicable,
358 cx.trace_macros_diag();
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 {
374 debug!("compile_declarative_macro: {:?}", def);
375 let mk_syn_ext = |expander| {
376 SyntaxExtension::new(
378 SyntaxExtensionKind::LegacyBang(expander),
387 let diag = &sess.parse_sess.span_diagnostic;
388 let lhs_nm = Ident::new(sym::lhs, def.span);
389 let rhs_nm = Ident::new(sym::rhs, def.span);
390 let tt_spec = Some(NonterminalKind::TT);
392 // Parse the macro_rules! invocation
393 let (macro_rules, body) = match &def.kind {
394 ast::ItemKind::MacroDef(def) => (def.macro_rules, def.body.inner_tokens()),
398 // The pattern that macro_rules matches.
399 // The grammar for macro_rules! is:
400 // $( $lhs:tt => $rhs:tt );+
401 // ...quasiquoting this would be nice.
402 // These spans won't matter, anyways
403 let argument_gram = vec![
404 mbe::TokenTree::Sequence(
406 Lrc::new(mbe::SequenceRepetition {
408 mbe::TokenTree::MetaVarDecl(def.span, lhs_nm, tt_spec),
409 mbe::TokenTree::token(token::FatArrow, def.span),
410 mbe::TokenTree::MetaVarDecl(def.span, rhs_nm, tt_spec),
412 separator: Some(Token::new(
413 if macro_rules { token::Semi } else { token::Comma },
416 kleene: mbe::KleeneToken::new(mbe::KleeneOp::OneOrMore, def.span),
420 // to phase into semicolon-termination instead of semicolon-separation
421 mbe::TokenTree::Sequence(
423 Lrc::new(mbe::SequenceRepetition {
424 tts: vec![mbe::TokenTree::token(
425 if macro_rules { token::Semi } else { token::Comma },
429 kleene: mbe::KleeneToken::new(mbe::KleeneOp::ZeroOrMore, def.span),
435 let parser = Parser::new(&sess.parse_sess, body, true, rustc_parse::MACRO_ARGUMENTS);
436 let argument_map = match parse_tt(&mut Cow::Borrowed(&parser), &argument_gram, def.ident) {
438 Failure(token, msg) => {
439 let s = parse_failure_msg(&token);
440 let sp = token.span.substitute_dummy(def.span);
441 sess.parse_sess.span_diagnostic.struct_span_err(sp, &s).span_label(sp, msg).emit();
442 return mk_syn_ext(Box::new(macro_rules_dummy_expander));
447 .struct_span_err(sp.substitute_dummy(def.span), &msg)
449 return mk_syn_ext(Box::new(macro_rules_dummy_expander));
452 return mk_syn_ext(Box::new(macro_rules_dummy_expander));
456 let mut valid = true;
458 // Extract the arguments:
459 let lhses = match argument_map[&MacroRulesNormalizedIdent::new(lhs_nm)] {
460 MatchedSeq(ref s) => s
463 if let MatchedNonterminal(ref nt) = *m {
464 if let NtTT(ref tt) = **nt {
465 let tt = mbe::quoted::parse(
475 valid &= check_lhs_nt_follows(&sess.parse_sess, features, &def, &tt);
479 sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs")
481 .collect::<Vec<mbe::TokenTree>>(),
482 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs"),
485 let rhses = match argument_map[&MacroRulesNormalizedIdent::new(rhs_nm)] {
486 MatchedSeq(ref s) => s
489 if let MatchedNonterminal(ref nt) = *m {
490 if let NtTT(ref tt) = **nt {
491 return mbe::quoted::parse(
503 sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs")
505 .collect::<Vec<mbe::TokenTree>>(),
506 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured rhs"),
510 valid &= check_rhs(&sess.parse_sess, rhs);
513 // don't abort iteration early, so that errors for multiple lhses can be reported
515 valid &= check_lhs_no_empty_seq(&sess.parse_sess, slice::from_ref(lhs));
518 valid &= macro_check::check_meta_variables(&sess.parse_sess, def.id, def.span, &lhses, &rhses);
520 let (transparency, transparency_error) = attr::find_transparency(sess, &def.attrs, macro_rules);
521 match transparency_error {
522 Some(TransparencyError::UnknownTransparency(value, span)) => {
523 diag.span_err(span, &format!("unknown macro transparency: `{}`", value))
525 Some(TransparencyError::MultipleTransparencyAttrs(old_span, new_span)) => {
526 diag.span_err(vec![old_span, new_span], "multiple macro transparency attributes")
531 mk_syn_ext(Box::new(MacroRulesMacroExpander {
541 fn check_lhs_nt_follows(
545 lhs: &mbe::TokenTree,
547 // lhs is going to be like TokenTree::Delimited(...), where the
548 // entire lhs is those tts. Or, it can be a "bare sequence", not wrapped in parens.
549 if let mbe::TokenTree::Delimited(_, ref tts) = *lhs {
550 check_matcher(sess, features, def, &tts.tts)
552 let msg = "invalid macro matcher; matchers must be contained in balanced delimiters";
553 sess.span_diagnostic.span_err(lhs.span(), msg);
556 // we don't abort on errors on rejection, the driver will do that for us
557 // after parsing/expansion. we can report every error in every macro this way.
560 /// Checks that the lhs contains no repetition which could match an empty token
561 /// tree, because then the matcher would hang indefinitely.
562 fn check_lhs_no_empty_seq(sess: &ParseSess, tts: &[mbe::TokenTree]) -> bool {
566 TokenTree::Token(..) | TokenTree::MetaVar(..) | TokenTree::MetaVarDecl(..) => (),
567 TokenTree::Delimited(_, ref del) => {
568 if !check_lhs_no_empty_seq(sess, &del.tts) {
572 TokenTree::Sequence(span, ref seq) => {
573 if seq.separator.is_none()
574 && seq.tts.iter().all(|seq_tt| match *seq_tt {
575 TokenTree::MetaVarDecl(_, _, Some(NonterminalKind::Vis)) => true,
576 TokenTree::Sequence(_, ref sub_seq) => {
577 sub_seq.kleene.op == mbe::KleeneOp::ZeroOrMore
578 || sub_seq.kleene.op == mbe::KleeneOp::ZeroOrOne
583 let sp = span.entire();
584 sess.span_diagnostic.span_err(sp, "repetition matches empty token tree");
587 if !check_lhs_no_empty_seq(sess, &seq.tts) {
597 fn check_rhs(sess: &ParseSess, rhs: &mbe::TokenTree) -> bool {
599 mbe::TokenTree::Delimited(..) => return true,
600 _ => sess.span_diagnostic.span_err(rhs.span(), "macro rhs must be delimited"),
609 matcher: &[mbe::TokenTree],
611 let first_sets = FirstSets::new(matcher);
612 let empty_suffix = TokenSet::empty();
613 let err = sess.span_diagnostic.err_count();
614 check_matcher_core(sess, features, def, &first_sets, matcher, &empty_suffix);
615 err == sess.span_diagnostic.err_count()
618 // `The FirstSets` for a matcher is a mapping from subsequences in the
619 // matcher to the FIRST set for that subsequence.
621 // This mapping is partially precomputed via a backwards scan over the
622 // token trees of the matcher, which provides a mapping from each
623 // repetition sequence to its *first* set.
625 // (Hypothetically, sequences should be uniquely identifiable via their
626 // spans, though perhaps that is false, e.g., for macro-generated macros
627 // that do not try to inject artificial span information. My plan is
628 // to try to catch such cases ahead of time and not include them in
629 // the precomputed mapping.)
631 // this maps each TokenTree::Sequence `$(tt ...) SEP OP` that is uniquely identified by its
632 // span in the original matcher to the First set for the inner sequence `tt ...`.
634 // If two sequences have the same span in a matcher, then map that
635 // span to None (invalidating the mapping here and forcing the code to
637 first: FxHashMap<Span, Option<TokenSet>>,
641 fn new(tts: &[mbe::TokenTree]) -> FirstSets {
644 let mut sets = FirstSets { first: FxHashMap::default() };
645 build_recur(&mut sets, tts);
648 // walks backward over `tts`, returning the FIRST for `tts`
649 // and updating `sets` at the same time for all sequence
650 // substructure we find within `tts`.
651 fn build_recur(sets: &mut FirstSets, tts: &[TokenTree]) -> TokenSet {
652 let mut first = TokenSet::empty();
653 for tt in tts.iter().rev() {
655 TokenTree::Token(..) | TokenTree::MetaVar(..) | TokenTree::MetaVarDecl(..) => {
656 first.replace_with(tt.clone());
658 TokenTree::Delimited(span, ref delimited) => {
659 build_recur(sets, &delimited.tts[..]);
660 first.replace_with(delimited.open_tt(span));
662 TokenTree::Sequence(sp, ref seq_rep) => {
663 let subfirst = build_recur(sets, &seq_rep.tts[..]);
665 match sets.first.entry(sp.entire()) {
666 Entry::Vacant(vac) => {
667 vac.insert(Some(subfirst.clone()));
669 Entry::Occupied(mut occ) => {
670 // if there is already an entry, then a span must have collided.
671 // This should not happen with typical macro_rules macros,
672 // but syntax extensions need not maintain distinct spans,
673 // so distinct syntax trees can be assigned the same span.
674 // In such a case, the map cannot be trusted; so mark this
675 // entry as unusable.
680 // If the sequence contents can be empty, then the first
681 // token could be the separator token itself.
683 if let (Some(sep), true) = (&seq_rep.separator, subfirst.maybe_empty) {
684 first.add_one_maybe(TokenTree::Token(sep.clone()));
687 // Reverse scan: Sequence comes before `first`.
688 if subfirst.maybe_empty
689 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrMore
690 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrOne
692 // If sequence is potentially empty, then
693 // union them (preserving first emptiness).
694 first.add_all(&TokenSet { maybe_empty: true, ..subfirst });
696 // Otherwise, sequence guaranteed
697 // non-empty; replace first.
708 // walks forward over `tts` until all potential FIRST tokens are
710 fn first(&self, tts: &[mbe::TokenTree]) -> TokenSet {
713 let mut first = TokenSet::empty();
714 for tt in tts.iter() {
715 assert!(first.maybe_empty);
717 TokenTree::Token(..) | TokenTree::MetaVar(..) | TokenTree::MetaVarDecl(..) => {
718 first.add_one(tt.clone());
721 TokenTree::Delimited(span, ref delimited) => {
722 first.add_one(delimited.open_tt(span));
725 TokenTree::Sequence(sp, ref seq_rep) => {
727 let subfirst = match self.first.get(&sp.entire()) {
728 Some(&Some(ref subfirst)) => subfirst,
730 subfirst_owned = self.first(&seq_rep.tts[..]);
734 panic!("We missed a sequence during FirstSets construction");
738 // If the sequence contents can be empty, then the first
739 // token could be the separator token itself.
740 if let (Some(sep), true) = (&seq_rep.separator, subfirst.maybe_empty) {
741 first.add_one_maybe(TokenTree::Token(sep.clone()));
744 assert!(first.maybe_empty);
745 first.add_all(subfirst);
746 if subfirst.maybe_empty
747 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrMore
748 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrOne
750 // Continue scanning for more first
751 // tokens, but also make sure we
752 // restore empty-tracking state.
753 first.maybe_empty = true;
762 // we only exit the loop if `tts` was empty or if every
763 // element of `tts` matches the empty sequence.
764 assert!(first.maybe_empty);
769 // A set of `mbe::TokenTree`s, which may include `TokenTree::Match`s
770 // (for macro-by-example syntactic variables). It also carries the
771 // `maybe_empty` flag; that is true if and only if the matcher can
772 // match an empty token sequence.
774 // The First set is computed on submatchers like `$($a:expr b),* $(c)* d`,
775 // which has corresponding FIRST = {$a:expr, c, d}.
776 // Likewise, `$($a:expr b),* $(c)+ d` has FIRST = {$a:expr, c}.
778 // (Notably, we must allow for *-op to occur zero times.)
779 #[derive(Clone, Debug)]
781 tokens: Vec<mbe::TokenTree>,
786 // Returns a set for the empty sequence.
788 TokenSet { tokens: Vec::new(), maybe_empty: true }
791 // Returns the set `{ tok }` for the single-token (and thus
792 // non-empty) sequence [tok].
793 fn singleton(tok: mbe::TokenTree) -> Self {
794 TokenSet { tokens: vec![tok], maybe_empty: false }
797 // Changes self to be the set `{ tok }`.
798 // Since `tok` is always present, marks self as non-empty.
799 fn replace_with(&mut self, tok: mbe::TokenTree) {
801 self.tokens.push(tok);
802 self.maybe_empty = false;
805 // Changes self to be the empty set `{}`; meant for use when
806 // the particular token does not matter, but we want to
807 // record that it occurs.
808 fn replace_with_irrelevant(&mut self) {
810 self.maybe_empty = false;
813 // Adds `tok` to the set for `self`, marking sequence as non-empy.
814 fn add_one(&mut self, tok: mbe::TokenTree) {
815 if !self.tokens.contains(&tok) {
816 self.tokens.push(tok);
818 self.maybe_empty = false;
821 // Adds `tok` to the set for `self`. (Leaves `maybe_empty` flag alone.)
822 fn add_one_maybe(&mut self, tok: mbe::TokenTree) {
823 if !self.tokens.contains(&tok) {
824 self.tokens.push(tok);
828 // Adds all elements of `other` to this.
830 // (Since this is a set, we filter out duplicates.)
832 // If `other` is potentially empty, then preserves the previous
833 // setting of the empty flag of `self`. If `other` is guaranteed
834 // non-empty, then `self` is marked non-empty.
835 fn add_all(&mut self, other: &Self) {
836 for tok in &other.tokens {
837 if !self.tokens.contains(tok) {
838 self.tokens.push(tok.clone());
841 if !other.maybe_empty {
842 self.maybe_empty = false;
847 // Checks that `matcher` is internally consistent and that it
848 // can legally be followed by a token `N`, for all `N` in `follow`.
849 // (If `follow` is empty, then it imposes no constraint on
852 // Returns the set of NT tokens that could possibly come last in
853 // `matcher`. (If `matcher` matches the empty sequence, then
854 // `maybe_empty` will be set to true.)
856 // Requires that `first_sets` is pre-computed for `matcher`;
857 // see `FirstSets::new`.
858 fn check_matcher_core(
862 first_sets: &FirstSets,
863 matcher: &[mbe::TokenTree],
868 let mut last = TokenSet::empty();
870 // 2. For each token and suffix [T, SUFFIX] in M:
871 // ensure that T can be followed by SUFFIX, and if SUFFIX may be empty,
872 // then ensure T can also be followed by any element of FOLLOW.
873 'each_token: for i in 0..matcher.len() {
874 let token = &matcher[i];
875 let suffix = &matcher[i + 1..];
877 let build_suffix_first = || {
878 let mut s = first_sets.first(suffix);
885 // (we build `suffix_first` on demand below; you can tell
886 // which cases are supposed to fall through by looking for the
887 // initialization of this variable.)
890 // First, update `last` so that it corresponds to the set
891 // of NT tokens that might end the sequence `... token`.
893 TokenTree::Token(..) | TokenTree::MetaVar(..) | TokenTree::MetaVarDecl(..) => {
894 if token_can_be_followed_by_any(token) {
895 // don't need to track tokens that work with any,
896 last.replace_with_irrelevant();
897 // ... and don't need to check tokens that can be
898 // followed by anything against SUFFIX.
899 continue 'each_token;
901 last.replace_with(token.clone());
902 suffix_first = build_suffix_first();
905 TokenTree::Delimited(span, ref d) => {
906 let my_suffix = TokenSet::singleton(d.close_tt(span));
907 check_matcher_core(sess, features, def, first_sets, &d.tts, &my_suffix);
908 // don't track non NT tokens
909 last.replace_with_irrelevant();
911 // also, we don't need to check delimited sequences
913 continue 'each_token;
915 TokenTree::Sequence(_, ref seq_rep) => {
916 suffix_first = build_suffix_first();
917 // The trick here: when we check the interior, we want
918 // to include the separator (if any) as a potential
919 // (but not guaranteed) element of FOLLOW. So in that
920 // case, we make a temp copy of suffix and stuff
921 // delimiter in there.
923 // FIXME: Should I first scan suffix_first to see if
924 // delimiter is already in it before I go through the
925 // work of cloning it? But then again, this way I may
926 // get a "tighter" span?
928 let my_suffix = if let Some(sep) = &seq_rep.separator {
929 new = suffix_first.clone();
930 new.add_one_maybe(TokenTree::Token(sep.clone()));
936 // At this point, `suffix_first` is built, and
937 // `my_suffix` is some TokenSet that we can use
938 // for checking the interior of `seq_rep`.
940 check_matcher_core(sess, features, def, first_sets, &seq_rep.tts, my_suffix);
941 if next.maybe_empty {
947 // the recursive call to check_matcher_core already ran the 'each_last
948 // check below, so we can just keep going forward here.
949 continue 'each_token;
953 // (`suffix_first` guaranteed initialized once reaching here.)
955 // Now `last` holds the complete set of NT tokens that could
956 // end the sequence before SUFFIX. Check that every one works with `suffix`.
957 for token in &last.tokens {
958 if let TokenTree::MetaVarDecl(span, name, Some(kind)) = *token {
959 for next_token in &suffix_first.tokens {
960 // Check if the old pat is used and the next token is `|`
961 // to warn about incompatibility with Rust 2021.
962 // We only emit this lint if we're parsing the original
963 // definition of this macro_rules, not while (re)parsing
964 // the macro when compiling another crate that is using the
965 // macro. (See #86567.)
966 // Macros defined in the current crate have a real node id,
967 // whereas macros from an external crate have a dummy id.
968 if def.id != DUMMY_NODE_ID
969 && matches!(kind, NonterminalKind::PatParam { inferred: true })
970 && matches!(next_token, TokenTree::Token(token) if token.kind == BinOp(token::BinOpToken::Or))
972 // It is suggestion to use pat_param, for example: $x:pat -> $x:pat_param.
973 let suggestion = quoted_tt_to_string(&TokenTree::MetaVarDecl(
976 Some(NonterminalKind::PatParam { inferred: false }),
978 sess.buffer_lint_with_diagnostic(
979 &RUST_2021_INCOMPATIBLE_OR_PATTERNS,
982 "the meaning of the `pat` fragment specifier is changing in Rust 2021, which may affect this macro",
983 BuiltinLintDiagnostics::OrPatternsBackCompat(span, suggestion),
986 match is_in_follow(next_token, kind) {
987 IsInFollow::Yes => {}
988 IsInFollow::No(possible) => {
989 let may_be = if last.tokens.len() == 1 && suffix_first.tokens.len() == 1
996 let sp = next_token.span();
997 let mut err = sess.span_diagnostic.struct_span_err(
1000 "`${name}:{frag}` {may_be} followed by `{next}`, which \
1001 is not allowed for `{frag}` fragments",
1004 next = quoted_tt_to_string(next_token),
1008 err.span_label(sp, format!("not allowed after `{}` fragments", kind));
1009 let msg = "allowed there are: ";
1014 "only {} is allowed after `{}` fragments",
1025 .collect::<Vec<_>>()
1041 fn token_can_be_followed_by_any(tok: &mbe::TokenTree) -> bool {
1042 if let mbe::TokenTree::MetaVarDecl(_, _, Some(kind)) = *tok {
1043 frag_can_be_followed_by_any(kind)
1045 // (Non NT's can always be followed by anything in matchers.)
1050 /// Returns `true` if a fragment of type `frag` can be followed by any sort of
1051 /// token. We use this (among other things) as a useful approximation
1052 /// for when `frag` can be followed by a repetition like `$(...)*` or
1053 /// `$(...)+`. In general, these can be a bit tricky to reason about,
1054 /// so we adopt a conservative position that says that any fragment
1055 /// specifier which consumes at most one token tree can be followed by
1056 /// a fragment specifier (indeed, these fragments can be followed by
1057 /// ANYTHING without fear of future compatibility hazards).
1058 fn frag_can_be_followed_by_any(kind: NonterminalKind) -> bool {
1061 NonterminalKind::Item // always terminated by `}` or `;`
1062 | NonterminalKind::Block // exactly one token tree
1063 | NonterminalKind::Ident // exactly one token tree
1064 | NonterminalKind::Literal // exactly one token tree
1065 | NonterminalKind::Meta // exactly one token tree
1066 | NonterminalKind::Lifetime // exactly one token tree
1067 | NonterminalKind::TT // exactly one token tree
1073 No(&'static [&'static str]),
1076 /// Returns `true` if `frag` can legally be followed by the token `tok`. For
1077 /// fragments that can consume an unbounded number of tokens, `tok`
1078 /// must be within a well-defined follow set. This is intended to
1079 /// guarantee future compatibility: for example, without this rule, if
1080 /// we expanded `expr` to include a new binary operator, we might
1081 /// break macros that were relying on that binary operator as a
1083 // when changing this do not forget to update doc/book/macros.md!
1084 fn is_in_follow(tok: &mbe::TokenTree, kind: NonterminalKind) -> IsInFollow {
1087 if let TokenTree::Token(Token { kind: token::CloseDelim(_), .. }) = *tok {
1088 // closing a token tree can never be matched by any fragment;
1089 // iow, we always require that `(` and `)` match, etc.
1093 NonterminalKind::Item => {
1094 // since items *must* be followed by either a `;` or a `}`, we can
1095 // accept anything after them
1098 NonterminalKind::Block => {
1099 // anything can follow block, the braces provide an easy boundary to
1103 NonterminalKind::Stmt | NonterminalKind::Expr => {
1104 const TOKENS: &[&str] = &["`=>`", "`,`", "`;`"];
1106 TokenTree::Token(token) => match token.kind {
1107 FatArrow | Comma | Semi => IsInFollow::Yes,
1108 _ => IsInFollow::No(TOKENS),
1110 _ => IsInFollow::No(TOKENS),
1113 NonterminalKind::PatParam { .. } => {
1114 const TOKENS: &[&str] = &["`=>`", "`,`", "`=`", "`|`", "`if`", "`in`"];
1116 TokenTree::Token(token) => match token.kind {
1117 FatArrow | Comma | Eq | BinOp(token::Or) => IsInFollow::Yes,
1118 Ident(name, false) if name == kw::If || name == kw::In => IsInFollow::Yes,
1119 _ => IsInFollow::No(TOKENS),
1121 _ => IsInFollow::No(TOKENS),
1124 NonterminalKind::PatWithOr { .. } => {
1125 const TOKENS: &[&str] = &["`=>`", "`,`", "`=`", "`if`", "`in`"];
1127 TokenTree::Token(token) => match token.kind {
1128 FatArrow | Comma | Eq => IsInFollow::Yes,
1129 Ident(name, false) if name == kw::If || name == kw::In => IsInFollow::Yes,
1130 _ => IsInFollow::No(TOKENS),
1132 _ => IsInFollow::No(TOKENS),
1135 NonterminalKind::Path | NonterminalKind::Ty => {
1136 const TOKENS: &[&str] = &[
1137 "`{`", "`[`", "`=>`", "`,`", "`>`", "`=`", "`:`", "`;`", "`|`", "`as`",
1141 TokenTree::Token(token) => match token.kind {
1142 OpenDelim(token::DelimToken::Brace)
1143 | OpenDelim(token::DelimToken::Bracket)
1151 | BinOp(token::Or) => IsInFollow::Yes,
1152 Ident(name, false) if name == kw::As || name == kw::Where => {
1155 _ => IsInFollow::No(TOKENS),
1157 TokenTree::MetaVarDecl(_, _, Some(NonterminalKind::Block)) => IsInFollow::Yes,
1158 _ => IsInFollow::No(TOKENS),
1161 NonterminalKind::Ident | NonterminalKind::Lifetime => {
1162 // being a single token, idents and lifetimes are harmless
1165 NonterminalKind::Literal => {
1166 // literals may be of a single token, or two tokens (negative numbers)
1169 NonterminalKind::Meta | NonterminalKind::TT => {
1170 // being either a single token or a delimited sequence, tt is
1174 NonterminalKind::Vis => {
1175 // Explicitly disallow `priv`, on the off chance it comes back.
1176 const TOKENS: &[&str] = &["`,`", "an ident", "a type"];
1178 TokenTree::Token(token) => match token.kind {
1179 Comma => IsInFollow::Yes,
1180 Ident(name, is_raw) if is_raw || name != kw::Priv => IsInFollow::Yes,
1182 if token.can_begin_type() {
1185 IsInFollow::No(TOKENS)
1189 TokenTree::MetaVarDecl(
1192 Some(NonterminalKind::Ident | NonterminalKind::Ty | NonterminalKind::Path),
1193 ) => IsInFollow::Yes,
1194 _ => IsInFollow::No(TOKENS),
1201 fn quoted_tt_to_string(tt: &mbe::TokenTree) -> String {
1203 mbe::TokenTree::Token(ref token) => pprust::token_to_string(&token),
1204 mbe::TokenTree::MetaVar(_, name) => format!("${}", name),
1205 mbe::TokenTree::MetaVarDecl(_, name, Some(kind)) => format!("${}:{}", name, kind),
1206 mbe::TokenTree::MetaVarDecl(_, name, None) => format!("${}:", name),
1209 "unexpected mbe::TokenTree::{Sequence or Delimited} \
1210 in follow set checker"
1215 fn parser_from_cx(sess: &ParseSess, tts: TokenStream) -> Parser<'_> {
1216 Parser::new(sess, tts, true, rustc_parse::MACRO_ARGUMENTS)
1219 /// Generates an appropriate parsing failure message. For EOF, this is "unexpected end...". For
1220 /// other tokens, this is "unexpected token...".
1221 fn parse_failure_msg(tok: &Token) -> String {
1223 token::Eof => "unexpected end of macro invocation".to_string(),
1224 _ => format!("no rules expected the token `{}`", pprust::token_to_string(tok),),