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
46 is_trailing_mac: bool,
50 crate fn annotate_err_with_kind(
51 err: &mut DiagnosticBuilder<'_>,
52 kind: AstFragmentKind,
56 AstFragmentKind::Ty => {
57 err.span_label(span, "this macro call doesn't expand to a type");
59 AstFragmentKind::Pat => {
60 err.span_label(span, "this macro call doesn't expand to a pattern");
66 fn emit_frag_parse_err(
67 mut e: DiagnosticBuilder<'_>,
69 orig_parser: &mut Parser<'_>,
72 kind: AstFragmentKind,
74 if parser.token == token::Eof && e.message().ends_with(", found `<eof>`") {
75 if !e.span.is_dummy() {
76 // early end of macro arm (#52866)
77 e.replace_span_with(parser.sess.source_map().next_point(parser.token.span));
79 let msg = &e.message[0];
82 "macro expansion ends with an incomplete expression: {}",
83 msg.0.replace(", found `<eof>`", ""),
88 if e.span.is_dummy() {
89 // Get around lack of span in error (#30128)
90 e.replace_span_with(site_span);
91 if !parser.sess.source_map().is_imported(arm_span) {
92 e.span_label(arm_span, "in this macro arm");
94 } else if parser.sess.source_map().is_imported(parser.token.span) {
95 e.span_label(site_span, "in this macro invocation");
98 // Try a statement if an expression is wanted but failed and suggest adding `;` to call.
99 AstFragmentKind::Expr => match parse_ast_fragment(orig_parser, AstFragmentKind::Stmts) {
100 Err(mut err) => err.cancel(),
103 "the macro call doesn't expand to an expression, but it can expand to a statement",
105 e.span_suggestion_verbose(
106 site_span.shrink_to_hi(),
107 "add `;` to interpret the expansion as a statement",
109 Applicability::MaybeIncorrect,
113 _ => annotate_err_with_kind(&mut e, kind, site_span),
118 impl<'a> ParserAnyMacro<'a> {
119 crate fn make(mut self: Box<ParserAnyMacro<'a>>, kind: AstFragmentKind) -> AstFragment {
128 let snapshot = &mut parser.clone();
129 let fragment = match parse_ast_fragment(parser, kind) {
132 emit_frag_parse_err(err, parser, snapshot, site_span, arm_span, kind);
133 return kind.dummy(site_span);
137 // We allow semicolons at the end of expressions -- e.g., the semicolon in
138 // `macro_rules! m { () => { panic!(); } }` isn't parsed by `.parse_expr()`,
139 // but `m!()` is allowed in expression positions (cf. issue #34706).
140 if kind == AstFragmentKind::Expr && parser.token == token::Semi {
141 parser.sess.buffer_lint_with_diagnostic(
142 SEMICOLON_IN_EXPRESSIONS_FROM_MACROS,
145 "trailing semicolon in macro used in expression position",
146 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 {
161 transparency: Transparency,
162 lhses: Vec<mbe::TokenTree>,
163 rhses: Vec<mbe::TokenTree>,
167 impl TTMacroExpander for MacroRulesMacroExpander {
170 cx: &'cx mut ExtCtxt<'_>,
173 ) -> Box<dyn MacResult + 'cx> {
175 return DummyResult::any(sp);
190 fn macro_rules_dummy_expander<'cx>(
191 _: &'cx mut ExtCtxt<'_>,
194 ) -> Box<dyn MacResult + 'cx> {
195 DummyResult::any(span)
198 fn trace_macros_note(cx_expansions: &mut FxHashMap<Span, Vec<String>>, sp: Span, message: String) {
199 let sp = sp.macro_backtrace().last().map_or(sp, |trace| trace.call_site);
200 cx_expansions.entry(sp).or_default().push(message);
203 /// Given `lhses` and `rhses`, this is the new macro we create
204 fn generic_extension<'cx>(
205 cx: &'cx mut ExtCtxt<'_>,
209 transparency: Transparency,
211 lhses: &[mbe::TokenTree],
212 rhses: &[mbe::TokenTree],
213 ) -> Box<dyn MacResult + 'cx> {
214 let sess = &cx.sess.parse_sess;
216 if cx.trace_macros() {
217 let msg = format!("expanding `{}! {{ {} }}`", name, pprust::tts_to_string(&arg));
218 trace_macros_note(&mut cx.expansions, sp, msg);
221 // Which arm's failure should we report? (the one furthest along)
222 let mut best_failure: Option<(Token, &str)> = None;
224 // We create a base parser that can be used for the "black box" parts.
225 // Every iteration needs a fresh copy of that parser. However, the parser
226 // is not mutated on many of the iterations, particularly when dealing with
229 // macro_rules! foo {
233 // // ... etc. (maybe hundreds more)
236 // as seen in the `html5ever` benchmark. We use a `Cow` so that the base
237 // parser is only cloned when necessary (upon mutation). Furthermore, we
238 // reinitialize the `Cow` with the base parser at the start of every
239 // iteration, so that any mutated parsers are not reused. This is all quite
240 // hacky, but speeds up the `html5ever` benchmark significantly. (Issue
241 // 68836 suggests a more comprehensive but more complex change to deal with
243 let parser = parser_from_cx(sess, arg.clone());
245 for (i, lhs) in lhses.iter().enumerate() {
246 // try each arm's matchers
247 let lhs_tt = match *lhs {
248 mbe::TokenTree::Delimited(_, ref delim) => &delim.tts[..],
249 _ => cx.span_bug(sp, "malformed macro lhs"),
252 // Take a snapshot of the state of pre-expansion gating at this point.
253 // This is used so that if a matcher is not `Success(..)`ful,
254 // then the spans which became gated when parsing the unsuccessful matcher
255 // are not recorded. On the first `Success(..)`ful matcher, the spans are merged.
256 let mut gated_spans_snapshot = mem::take(&mut *sess.gated_spans.spans.borrow_mut());
258 match parse_tt(&mut Cow::Borrowed(&parser), lhs_tt, name) {
259 Success(named_matches) => {
260 // The matcher was `Success(..)`ful.
261 // Merge the gated spans from parsing the matcher with the pre-existing ones.
262 sess.gated_spans.merge(gated_spans_snapshot);
264 let rhs = match rhses[i] {
266 mbe::TokenTree::Delimited(_, ref delimed) => delimed.tts.clone(),
267 _ => cx.span_bug(sp, "malformed macro rhs"),
269 let arm_span = rhses[i].span();
271 let rhs_spans = rhs.iter().map(|t| t.span()).collect::<Vec<_>>();
272 // rhs has holes ( `$id` and `$(...)` that need filled)
273 let mut tts = match transcribe(cx, &named_matches, rhs, transparency) {
277 return DummyResult::any(arm_span);
281 // Replace all the tokens for the corresponding positions in the macro, to maintain
282 // proper positions in error reporting, while maintaining the macro_backtrace.
283 if rhs_spans.len() == tts.len() {
284 tts = tts.map_enumerated(|i, tt| {
285 let mut tt = tt.clone();
286 let mut sp = rhs_spans[i];
287 sp = sp.with_ctxt(tt.span().ctxt());
293 if cx.trace_macros() {
294 let msg = format!("to `{}`", pprust::tts_to_string(&tts));
295 trace_macros_note(&mut cx.expansions, sp, msg);
298 let mut p = Parser::new(sess, tts, false, None);
299 p.last_type_ascription = cx.current_expansion.prior_type_ascription;
301 // Let the context choose how to interpret the result.
302 // Weird, but useful for X-macros.
303 return Box::new(ParserAnyMacro {
306 // Pass along the original expansion site and the name of the macro
307 // so we can print a useful error message if the parse of the expanded
308 // macro leaves unparsed tokens.
311 lint_node_id: cx.current_expansion.lint_node_id,
312 is_trailing_mac: cx.current_expansion.is_trailing_mac,
316 Failure(token, msg) => match best_failure {
317 Some((ref best_token, _)) if best_token.span.lo() >= token.span.lo() => {}
318 _ => best_failure = Some((token, msg)),
320 Error(err_sp, ref msg) => {
321 let span = err_sp.substitute_dummy(sp);
322 cx.struct_span_err(span, &msg).emit();
323 return DummyResult::any(span);
325 ErrorReported => return DummyResult::any(sp),
328 // The matcher was not `Success(..)`ful.
329 // Restore to the state before snapshotting and maybe try again.
330 mem::swap(&mut gated_spans_snapshot, &mut sess.gated_spans.spans.borrow_mut());
334 let (token, label) = best_failure.expect("ran no matchers");
335 let span = token.span.substitute_dummy(sp);
336 let mut err = cx.struct_span_err(span, &parse_failure_msg(&token));
337 err.span_label(span, label);
338 if !def_span.is_dummy() && !cx.source_map().is_imported(def_span) {
339 err.span_label(cx.source_map().guess_head_span(def_span), "when calling this macro");
342 // Check whether there's a missing comma in this macro call, like `println!("{}" a);`
343 if let Some((arg, comma_span)) = arg.add_comma() {
345 // try each arm's matchers
346 let lhs_tt = match *lhs {
347 mbe::TokenTree::Delimited(_, ref delim) => &delim.tts[..],
351 parse_tt(&mut Cow::Borrowed(&parser_from_cx(sess, arg.clone())), lhs_tt, name)
353 if comma_span.is_dummy() {
354 err.note("you might be missing a comma");
356 err.span_suggestion_short(
358 "missing comma here",
360 Applicability::MachineApplicable,
367 cx.trace_macros_diag();
371 // Note that macro-by-example's input is also matched against a token tree:
372 // $( $lhs:tt => $rhs:tt );+
374 // Holy self-referential!
376 /// Converts a macro item into a syntax extension.
377 pub fn compile_declarative_macro(
382 ) -> SyntaxExtension {
383 debug!("compile_declarative_macro: {:?}", def);
384 let mk_syn_ext = |expander| {
385 SyntaxExtension::new(
387 SyntaxExtensionKind::LegacyBang(expander),
396 let diag = &sess.parse_sess.span_diagnostic;
397 let lhs_nm = Ident::new(sym::lhs, def.span);
398 let rhs_nm = Ident::new(sym::rhs, def.span);
399 let tt_spec = Some(NonterminalKind::TT);
401 // Parse the macro_rules! invocation
402 let (macro_rules, body) = match &def.kind {
403 ast::ItemKind::MacroDef(def) => (def.macro_rules, def.body.inner_tokens()),
407 // The pattern that macro_rules matches.
408 // The grammar for macro_rules! is:
409 // $( $lhs:tt => $rhs:tt );+
410 // ...quasiquoting this would be nice.
411 // These spans won't matter, anyways
412 let argument_gram = vec![
413 mbe::TokenTree::Sequence(
415 Lrc::new(mbe::SequenceRepetition {
417 mbe::TokenTree::MetaVarDecl(def.span, lhs_nm, tt_spec),
418 mbe::TokenTree::token(token::FatArrow, def.span),
419 mbe::TokenTree::MetaVarDecl(def.span, rhs_nm, tt_spec),
421 separator: Some(Token::new(
422 if macro_rules { token::Semi } else { token::Comma },
425 kleene: mbe::KleeneToken::new(mbe::KleeneOp::OneOrMore, def.span),
429 // to phase into semicolon-termination instead of semicolon-separation
430 mbe::TokenTree::Sequence(
432 Lrc::new(mbe::SequenceRepetition {
433 tts: vec![mbe::TokenTree::token(
434 if macro_rules { token::Semi } else { token::Comma },
438 kleene: mbe::KleeneToken::new(mbe::KleeneOp::ZeroOrMore, def.span),
444 let parser = Parser::new(&sess.parse_sess, body, true, rustc_parse::MACRO_ARGUMENTS);
445 let argument_map = match parse_tt(&mut Cow::Borrowed(&parser), &argument_gram, def.ident) {
447 Failure(token, msg) => {
448 let s = parse_failure_msg(&token);
449 let sp = token.span.substitute_dummy(def.span);
450 sess.parse_sess.span_diagnostic.struct_span_err(sp, &s).span_label(sp, msg).emit();
451 return mk_syn_ext(Box::new(macro_rules_dummy_expander));
456 .struct_span_err(sp.substitute_dummy(def.span), &msg)
458 return mk_syn_ext(Box::new(macro_rules_dummy_expander));
461 return mk_syn_ext(Box::new(macro_rules_dummy_expander));
465 let mut valid = true;
467 // Extract the arguments:
468 let lhses = match argument_map[&MacroRulesNormalizedIdent::new(lhs_nm)] {
469 MatchedSeq(ref s) => s
472 if let MatchedNonterminal(ref nt) = *m {
473 if let NtTT(ref tt) = **nt {
474 let tt = mbe::quoted::parse(
484 valid &= check_lhs_nt_follows(&sess.parse_sess, features, &def, &tt);
488 sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs")
490 .collect::<Vec<mbe::TokenTree>>(),
491 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs"),
494 let rhses = match argument_map[&MacroRulesNormalizedIdent::new(rhs_nm)] {
495 MatchedSeq(ref s) => s
498 if let MatchedNonterminal(ref nt) = *m {
499 if let NtTT(ref tt) = **nt {
500 return mbe::quoted::parse(
512 sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs")
514 .collect::<Vec<mbe::TokenTree>>(),
515 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured rhs"),
519 valid &= check_rhs(&sess.parse_sess, rhs);
522 // don't abort iteration early, so that errors for multiple lhses can be reported
524 valid &= check_lhs_no_empty_seq(&sess.parse_sess, slice::from_ref(lhs));
527 valid &= macro_check::check_meta_variables(&sess.parse_sess, def.id, def.span, &lhses, &rhses);
529 let (transparency, transparency_error) = attr::find_transparency(sess, &def.attrs, macro_rules);
530 match transparency_error {
531 Some(TransparencyError::UnknownTransparency(value, span)) => {
532 diag.span_err(span, &format!("unknown macro transparency: `{}`", value))
534 Some(TransparencyError::MultipleTransparencyAttrs(old_span, new_span)) => {
535 diag.span_err(vec![old_span, new_span], "multiple macro transparency attributes")
540 mk_syn_ext(Box::new(MacroRulesMacroExpander {
550 fn check_lhs_nt_follows(
554 lhs: &mbe::TokenTree,
556 // lhs is going to be like TokenTree::Delimited(...), where the
557 // entire lhs is those tts. Or, it can be a "bare sequence", not wrapped in parens.
558 if let mbe::TokenTree::Delimited(_, ref tts) = *lhs {
559 check_matcher(sess, features, def, &tts.tts)
561 let msg = "invalid macro matcher; matchers must be contained in balanced delimiters";
562 sess.span_diagnostic.span_err(lhs.span(), msg);
565 // we don't abort on errors on rejection, the driver will do that for us
566 // after parsing/expansion. we can report every error in every macro this way.
569 /// Checks that the lhs contains no repetition which could match an empty token
570 /// tree, because then the matcher would hang indefinitely.
571 fn check_lhs_no_empty_seq(sess: &ParseSess, tts: &[mbe::TokenTree]) -> bool {
575 TokenTree::Token(..) | TokenTree::MetaVar(..) | TokenTree::MetaVarDecl(..) => (),
576 TokenTree::Delimited(_, ref del) => {
577 if !check_lhs_no_empty_seq(sess, &del.tts) {
581 TokenTree::Sequence(span, ref seq) => {
582 if seq.separator.is_none()
583 && seq.tts.iter().all(|seq_tt| match *seq_tt {
584 TokenTree::MetaVarDecl(_, _, Some(NonterminalKind::Vis)) => true,
585 TokenTree::Sequence(_, ref sub_seq) => {
586 sub_seq.kleene.op == mbe::KleeneOp::ZeroOrMore
587 || sub_seq.kleene.op == mbe::KleeneOp::ZeroOrOne
592 let sp = span.entire();
593 sess.span_diagnostic.span_err(sp, "repetition matches empty token tree");
596 if !check_lhs_no_empty_seq(sess, &seq.tts) {
606 fn check_rhs(sess: &ParseSess, rhs: &mbe::TokenTree) -> bool {
608 mbe::TokenTree::Delimited(..) => return true,
609 _ => sess.span_diagnostic.span_err(rhs.span(), "macro rhs must be delimited"),
618 matcher: &[mbe::TokenTree],
620 let first_sets = FirstSets::new(matcher);
621 let empty_suffix = TokenSet::empty();
622 let err = sess.span_diagnostic.err_count();
623 check_matcher_core(sess, features, def, &first_sets, matcher, &empty_suffix);
624 err == sess.span_diagnostic.err_count()
627 // `The FirstSets` for a matcher is a mapping from subsequences in the
628 // matcher to the FIRST set for that subsequence.
630 // This mapping is partially precomputed via a backwards scan over the
631 // token trees of the matcher, which provides a mapping from each
632 // repetition sequence to its *first* set.
634 // (Hypothetically, sequences should be uniquely identifiable via their
635 // spans, though perhaps that is false, e.g., for macro-generated macros
636 // that do not try to inject artificial span information. My plan is
637 // to try to catch such cases ahead of time and not include them in
638 // the precomputed mapping.)
640 // this maps each TokenTree::Sequence `$(tt ...) SEP OP` that is uniquely identified by its
641 // span in the original matcher to the First set for the inner sequence `tt ...`.
643 // If two sequences have the same span in a matcher, then map that
644 // span to None (invalidating the mapping here and forcing the code to
646 first: FxHashMap<Span, Option<TokenSet>>,
650 fn new(tts: &[mbe::TokenTree]) -> FirstSets {
653 let mut sets = FirstSets { first: FxHashMap::default() };
654 build_recur(&mut sets, tts);
657 // walks backward over `tts`, returning the FIRST for `tts`
658 // and updating `sets` at the same time for all sequence
659 // substructure we find within `tts`.
660 fn build_recur(sets: &mut FirstSets, tts: &[TokenTree]) -> TokenSet {
661 let mut first = TokenSet::empty();
662 for tt in tts.iter().rev() {
664 TokenTree::Token(..) | TokenTree::MetaVar(..) | TokenTree::MetaVarDecl(..) => {
665 first.replace_with(tt.clone());
667 TokenTree::Delimited(span, ref delimited) => {
668 build_recur(sets, &delimited.tts[..]);
669 first.replace_with(delimited.open_tt(span));
671 TokenTree::Sequence(sp, ref seq_rep) => {
672 let subfirst = build_recur(sets, &seq_rep.tts[..]);
674 match sets.first.entry(sp.entire()) {
675 Entry::Vacant(vac) => {
676 vac.insert(Some(subfirst.clone()));
678 Entry::Occupied(mut occ) => {
679 // if there is already an entry, then a span must have collided.
680 // This should not happen with typical macro_rules macros,
681 // but syntax extensions need not maintain distinct spans,
682 // so distinct syntax trees can be assigned the same span.
683 // In such a case, the map cannot be trusted; so mark this
684 // entry as unusable.
689 // If the sequence contents can be empty, then the first
690 // token could be the separator token itself.
692 if let (Some(sep), true) = (&seq_rep.separator, subfirst.maybe_empty) {
693 first.add_one_maybe(TokenTree::Token(sep.clone()));
696 // Reverse scan: Sequence comes before `first`.
697 if subfirst.maybe_empty
698 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrMore
699 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrOne
701 // If sequence is potentially empty, then
702 // union them (preserving first emptiness).
703 first.add_all(&TokenSet { maybe_empty: true, ..subfirst });
705 // Otherwise, sequence guaranteed
706 // non-empty; replace first.
717 // walks forward over `tts` until all potential FIRST tokens are
719 fn first(&self, tts: &[mbe::TokenTree]) -> TokenSet {
722 let mut first = TokenSet::empty();
723 for tt in tts.iter() {
724 assert!(first.maybe_empty);
726 TokenTree::Token(..) | TokenTree::MetaVar(..) | TokenTree::MetaVarDecl(..) => {
727 first.add_one(tt.clone());
730 TokenTree::Delimited(span, ref delimited) => {
731 first.add_one(delimited.open_tt(span));
734 TokenTree::Sequence(sp, ref seq_rep) => {
736 let subfirst = match self.first.get(&sp.entire()) {
737 Some(&Some(ref subfirst)) => subfirst,
739 subfirst_owned = self.first(&seq_rep.tts[..]);
743 panic!("We missed a sequence during FirstSets construction");
747 // If the sequence contents can be empty, then the first
748 // token could be the separator token itself.
749 if let (Some(sep), true) = (&seq_rep.separator, subfirst.maybe_empty) {
750 first.add_one_maybe(TokenTree::Token(sep.clone()));
753 assert!(first.maybe_empty);
754 first.add_all(subfirst);
755 if subfirst.maybe_empty
756 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrMore
757 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrOne
759 // Continue scanning for more first
760 // tokens, but also make sure we
761 // restore empty-tracking state.
762 first.maybe_empty = true;
771 // we only exit the loop if `tts` was empty or if every
772 // element of `tts` matches the empty sequence.
773 assert!(first.maybe_empty);
778 // A set of `mbe::TokenTree`s, which may include `TokenTree::Match`s
779 // (for macro-by-example syntactic variables). It also carries the
780 // `maybe_empty` flag; that is true if and only if the matcher can
781 // match an empty token sequence.
783 // The First set is computed on submatchers like `$($a:expr b),* $(c)* d`,
784 // which has corresponding FIRST = {$a:expr, c, d}.
785 // Likewise, `$($a:expr b),* $(c)+ d` has FIRST = {$a:expr, c}.
787 // (Notably, we must allow for *-op to occur zero times.)
788 #[derive(Clone, Debug)]
790 tokens: Vec<mbe::TokenTree>,
795 // Returns a set for the empty sequence.
797 TokenSet { tokens: Vec::new(), maybe_empty: true }
800 // Returns the set `{ tok }` for the single-token (and thus
801 // non-empty) sequence [tok].
802 fn singleton(tok: mbe::TokenTree) -> Self {
803 TokenSet { tokens: vec![tok], maybe_empty: false }
806 // Changes self to be the set `{ tok }`.
807 // Since `tok` is always present, marks self as non-empty.
808 fn replace_with(&mut self, tok: mbe::TokenTree) {
810 self.tokens.push(tok);
811 self.maybe_empty = false;
814 // Changes self to be the empty set `{}`; meant for use when
815 // the particular token does not matter, but we want to
816 // record that it occurs.
817 fn replace_with_irrelevant(&mut self) {
819 self.maybe_empty = false;
822 // Adds `tok` to the set for `self`, marking sequence as non-empy.
823 fn add_one(&mut self, tok: mbe::TokenTree) {
824 if !self.tokens.contains(&tok) {
825 self.tokens.push(tok);
827 self.maybe_empty = false;
830 // Adds `tok` to the set for `self`. (Leaves `maybe_empty` flag alone.)
831 fn add_one_maybe(&mut self, tok: mbe::TokenTree) {
832 if !self.tokens.contains(&tok) {
833 self.tokens.push(tok);
837 // Adds all elements of `other` to this.
839 // (Since this is a set, we filter out duplicates.)
841 // If `other` is potentially empty, then preserves the previous
842 // setting of the empty flag of `self`. If `other` is guaranteed
843 // non-empty, then `self` is marked non-empty.
844 fn add_all(&mut self, other: &Self) {
845 for tok in &other.tokens {
846 if !self.tokens.contains(tok) {
847 self.tokens.push(tok.clone());
850 if !other.maybe_empty {
851 self.maybe_empty = false;
856 // Checks that `matcher` is internally consistent and that it
857 // can legally be followed by a token `N`, for all `N` in `follow`.
858 // (If `follow` is empty, then it imposes no constraint on
861 // Returns the set of NT tokens that could possibly come last in
862 // `matcher`. (If `matcher` matches the empty sequence, then
863 // `maybe_empty` will be set to true.)
865 // Requires that `first_sets` is pre-computed for `matcher`;
866 // see `FirstSets::new`.
867 fn check_matcher_core(
871 first_sets: &FirstSets,
872 matcher: &[mbe::TokenTree],
877 let mut last = TokenSet::empty();
879 // 2. For each token and suffix [T, SUFFIX] in M:
880 // ensure that T can be followed by SUFFIX, and if SUFFIX may be empty,
881 // then ensure T can also be followed by any element of FOLLOW.
882 'each_token: for i in 0..matcher.len() {
883 let token = &matcher[i];
884 let suffix = &matcher[i + 1..];
886 let build_suffix_first = || {
887 let mut s = first_sets.first(suffix);
894 // (we build `suffix_first` on demand below; you can tell
895 // which cases are supposed to fall through by looking for the
896 // initialization of this variable.)
899 // First, update `last` so that it corresponds to the set
900 // of NT tokens that might end the sequence `... token`.
902 TokenTree::Token(..) | TokenTree::MetaVar(..) | TokenTree::MetaVarDecl(..) => {
903 if token_can_be_followed_by_any(token) {
904 // don't need to track tokens that work with any,
905 last.replace_with_irrelevant();
906 // ... and don't need to check tokens that can be
907 // followed by anything against SUFFIX.
908 continue 'each_token;
910 last.replace_with(token.clone());
911 suffix_first = build_suffix_first();
914 TokenTree::Delimited(span, ref d) => {
915 let my_suffix = TokenSet::singleton(d.close_tt(span));
916 check_matcher_core(sess, features, def, first_sets, &d.tts, &my_suffix);
917 // don't track non NT tokens
918 last.replace_with_irrelevant();
920 // also, we don't need to check delimited sequences
922 continue 'each_token;
924 TokenTree::Sequence(_, ref seq_rep) => {
925 suffix_first = build_suffix_first();
926 // The trick here: when we check the interior, we want
927 // to include the separator (if any) as a potential
928 // (but not guaranteed) element of FOLLOW. So in that
929 // case, we make a temp copy of suffix and stuff
930 // delimiter in there.
932 // FIXME: Should I first scan suffix_first to see if
933 // delimiter is already in it before I go through the
934 // work of cloning it? But then again, this way I may
935 // get a "tighter" span?
937 let my_suffix = if let Some(sep) = &seq_rep.separator {
938 new = suffix_first.clone();
939 new.add_one_maybe(TokenTree::Token(sep.clone()));
945 // At this point, `suffix_first` is built, and
946 // `my_suffix` is some TokenSet that we can use
947 // for checking the interior of `seq_rep`.
949 check_matcher_core(sess, features, def, first_sets, &seq_rep.tts, my_suffix);
950 if next.maybe_empty {
956 // the recursive call to check_matcher_core already ran the 'each_last
957 // check below, so we can just keep going forward here.
958 continue 'each_token;
962 // (`suffix_first` guaranteed initialized once reaching here.)
964 // Now `last` holds the complete set of NT tokens that could
965 // end the sequence before SUFFIX. Check that every one works with `suffix`.
966 for token in &last.tokens {
967 if let TokenTree::MetaVarDecl(span, name, Some(kind)) = *token {
968 for next_token in &suffix_first.tokens {
969 // Check if the old pat is used and the next token is `|`
970 // to warn about incompatibility with Rust 2021.
971 // We only emit this lint if we're parsing the original
972 // definition of this macro_rules, not while (re)parsing
973 // the macro when compiling another crate that is using the
974 // macro. (See #86567.)
975 // Macros defined in the current crate have a real node id,
976 // whereas macros from an external crate have a dummy id.
977 if def.id != DUMMY_NODE_ID
978 && matches!(kind, NonterminalKind::PatParam { inferred: true })
979 && matches!(next_token, TokenTree::Token(token) if token.kind == BinOp(token::BinOpToken::Or))
981 // It is suggestion to use pat_param, for example: $x:pat -> $x:pat_param.
982 let suggestion = quoted_tt_to_string(&TokenTree::MetaVarDecl(
985 Some(NonterminalKind::PatParam { inferred: false }),
987 sess.buffer_lint_with_diagnostic(
988 &RUST_2021_INCOMPATIBLE_OR_PATTERNS,
991 "the meaning of the `pat` fragment specifier is changing in Rust 2021, which may affect this macro",
992 BuiltinLintDiagnostics::OrPatternsBackCompat(span, suggestion),
995 match is_in_follow(next_token, kind) {
996 IsInFollow::Yes => {}
997 IsInFollow::No(possible) => {
998 let may_be = if last.tokens.len() == 1 && suffix_first.tokens.len() == 1
1005 let sp = next_token.span();
1006 let mut err = sess.span_diagnostic.struct_span_err(
1009 "`${name}:{frag}` {may_be} followed by `{next}`, which \
1010 is not allowed for `{frag}` fragments",
1013 next = quoted_tt_to_string(next_token),
1017 err.span_label(sp, format!("not allowed after `{}` fragments", kind));
1018 let msg = "allowed there are: ";
1023 "only {} is allowed after `{}` fragments",
1034 .collect::<Vec<_>>()
1050 fn token_can_be_followed_by_any(tok: &mbe::TokenTree) -> bool {
1051 if let mbe::TokenTree::MetaVarDecl(_, _, Some(kind)) = *tok {
1052 frag_can_be_followed_by_any(kind)
1054 // (Non NT's can always be followed by anything in matchers.)
1059 /// Returns `true` if a fragment of type `frag` can be followed by any sort of
1060 /// token. We use this (among other things) as a useful approximation
1061 /// for when `frag` can be followed by a repetition like `$(...)*` or
1062 /// `$(...)+`. In general, these can be a bit tricky to reason about,
1063 /// so we adopt a conservative position that says that any fragment
1064 /// specifier which consumes at most one token tree can be followed by
1065 /// a fragment specifier (indeed, these fragments can be followed by
1066 /// ANYTHING without fear of future compatibility hazards).
1067 fn frag_can_be_followed_by_any(kind: NonterminalKind) -> bool {
1070 NonterminalKind::Item // always terminated by `}` or `;`
1071 | NonterminalKind::Block // exactly one token tree
1072 | NonterminalKind::Ident // exactly one token tree
1073 | NonterminalKind::Literal // exactly one token tree
1074 | NonterminalKind::Meta // exactly one token tree
1075 | NonterminalKind::Lifetime // exactly one token tree
1076 | NonterminalKind::TT // exactly one token tree
1082 No(&'static [&'static str]),
1085 /// Returns `true` if `frag` can legally be followed by the token `tok`. For
1086 /// fragments that can consume an unbounded number of tokens, `tok`
1087 /// must be within a well-defined follow set. This is intended to
1088 /// guarantee future compatibility: for example, without this rule, if
1089 /// we expanded `expr` to include a new binary operator, we might
1090 /// break macros that were relying on that binary operator as a
1092 // when changing this do not forget to update doc/book/macros.md!
1093 fn is_in_follow(tok: &mbe::TokenTree, kind: NonterminalKind) -> IsInFollow {
1096 if let TokenTree::Token(Token { kind: token::CloseDelim(_), .. }) = *tok {
1097 // closing a token tree can never be matched by any fragment;
1098 // iow, we always require that `(` and `)` match, etc.
1102 NonterminalKind::Item => {
1103 // since items *must* be followed by either a `;` or a `}`, we can
1104 // accept anything after them
1107 NonterminalKind::Block => {
1108 // anything can follow block, the braces provide an easy boundary to
1112 NonterminalKind::Stmt | NonterminalKind::Expr => {
1113 const TOKENS: &[&str] = &["`=>`", "`,`", "`;`"];
1115 TokenTree::Token(token) => match token.kind {
1116 FatArrow | Comma | Semi => IsInFollow::Yes,
1117 _ => IsInFollow::No(TOKENS),
1119 _ => IsInFollow::No(TOKENS),
1122 NonterminalKind::PatParam { .. } => {
1123 const TOKENS: &[&str] = &["`=>`", "`,`", "`=`", "`|`", "`if`", "`in`"];
1125 TokenTree::Token(token) => match token.kind {
1126 FatArrow | Comma | Eq | BinOp(token::Or) => IsInFollow::Yes,
1127 Ident(name, false) if name == kw::If || name == kw::In => IsInFollow::Yes,
1128 _ => IsInFollow::No(TOKENS),
1130 _ => IsInFollow::No(TOKENS),
1133 NonterminalKind::PatWithOr { .. } => {
1134 const TOKENS: &[&str] = &["`=>`", "`,`", "`=`", "`if`", "`in`"];
1136 TokenTree::Token(token) => match token.kind {
1137 FatArrow | Comma | Eq => IsInFollow::Yes,
1138 Ident(name, false) if name == kw::If || name == kw::In => IsInFollow::Yes,
1139 _ => IsInFollow::No(TOKENS),
1141 _ => IsInFollow::No(TOKENS),
1144 NonterminalKind::Path | NonterminalKind::Ty => {
1145 const TOKENS: &[&str] = &[
1146 "`{`", "`[`", "`=>`", "`,`", "`>`", "`=`", "`:`", "`;`", "`|`", "`as`",
1150 TokenTree::Token(token) => match token.kind {
1151 OpenDelim(token::DelimToken::Brace)
1152 | OpenDelim(token::DelimToken::Bracket)
1160 | BinOp(token::Or) => IsInFollow::Yes,
1161 Ident(name, false) if name == kw::As || name == kw::Where => {
1164 _ => IsInFollow::No(TOKENS),
1166 TokenTree::MetaVarDecl(_, _, Some(NonterminalKind::Block)) => IsInFollow::Yes,
1167 _ => IsInFollow::No(TOKENS),
1170 NonterminalKind::Ident | NonterminalKind::Lifetime => {
1171 // being a single token, idents and lifetimes are harmless
1174 NonterminalKind::Literal => {
1175 // literals may be of a single token, or two tokens (negative numbers)
1178 NonterminalKind::Meta | NonterminalKind::TT => {
1179 // being either a single token or a delimited sequence, tt is
1183 NonterminalKind::Vis => {
1184 // Explicitly disallow `priv`, on the off chance it comes back.
1185 const TOKENS: &[&str] = &["`,`", "an ident", "a type"];
1187 TokenTree::Token(token) => match token.kind {
1188 Comma => IsInFollow::Yes,
1189 Ident(name, is_raw) if is_raw || name != kw::Priv => IsInFollow::Yes,
1191 if token.can_begin_type() {
1194 IsInFollow::No(TOKENS)
1198 TokenTree::MetaVarDecl(
1201 Some(NonterminalKind::Ident | NonterminalKind::Ty | NonterminalKind::Path),
1202 ) => IsInFollow::Yes,
1203 _ => IsInFollow::No(TOKENS),
1210 fn quoted_tt_to_string(tt: &mbe::TokenTree) -> String {
1212 mbe::TokenTree::Token(ref token) => pprust::token_to_string(&token),
1213 mbe::TokenTree::MetaVar(_, name) => format!("${}", name),
1214 mbe::TokenTree::MetaVarDecl(_, name, Some(kind)) => format!("${}:{}", name, kind),
1215 mbe::TokenTree::MetaVarDecl(_, name, None) => format!("${}:", name),
1218 "unexpected mbe::TokenTree::{Sequence or Delimited} \
1219 in follow set checker"
1224 fn parser_from_cx(sess: &ParseSess, tts: TokenStream) -> Parser<'_> {
1225 Parser::new(sess, tts, true, rustc_parse::MACRO_ARGUMENTS)
1228 /// Generates an appropriate parsing failure message. For EOF, this is "unexpected end...". For
1229 /// other tokens, this is "unexpected token...".
1230 fn parse_failure_msg(tok: &Token) -> String {
1232 token::Eof => "unexpected end of macro invocation".to_string(),
1233 _ => format!("no rules expected the token `{}`", pprust::token_to_string(tok),),