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, FxIndexMap};
17 use rustc_errors::{Applicability, Diagnostic, DiagnosticBuilder, DiagnosticMessage};
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::source_map::SourceMap;
29 use rustc_span::symbol::{kw, sym, Ident, MacroRulesNormalizedIdent};
33 use std::collections::hash_map::Entry;
34 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
72 && let DiagnosticMessage::Str(message) = &e.message[0].0
73 && message.ends_with(", found `<eof>`")
75 let msg = &e.message[0];
77 DiagnosticMessage::Str(format!(
78 "macro expansion ends with an incomplete expression: {}",
79 message.replace(", found `<eof>`", ""),
83 if !e.span.is_dummy() {
84 // early end of macro arm (#52866)
85 e.replace_span_with(parser.token.span.shrink_to_hi());
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(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 pub(crate) fn make(mut self: Box<ParserAnyMacro<'a>>, kind: AstFragmentKind) -> AstFragment {
129 let snapshot = &mut parser.create_snapshot_for_diagnostic();
130 let fragment = match parse_ast_fragment(parser, kind) {
133 emit_frag_parse_err(err, parser, snapshot, site_span, arm_span, kind);
134 return kind.dummy(site_span);
138 // We allow semicolons at the end of expressions -- e.g., the semicolon in
139 // `macro_rules! m { () => { panic!(); } }` isn't parsed by `.parse_expr()`,
140 // but `m!()` is allowed in expression positions (cf. issue #34706).
141 if kind == AstFragmentKind::Expr && parser.token == token::Semi {
143 parser.sess.buffer_lint_with_diagnostic(
144 SEMICOLON_IN_EXPRESSIONS_FROM_MACROS,
147 "trailing semicolon in macro used in expression position",
148 BuiltinLintDiagnostics::TrailingMacro(is_trailing_mac, macro_ident),
154 // Make sure we don't have any tokens left to parse so we don't silently drop anything.
155 let path = ast::Path::from_ident(macro_ident.with_span_pos(site_span));
156 ensure_complete_parse(parser, &path, kind.name(), site_span);
161 struct MacroRulesMacroExpander {
165 transparency: Transparency,
166 lhses: Vec<Vec<MatcherLoc>>,
167 rhses: Vec<mbe::TokenTree>,
171 impl TTMacroExpander for MacroRulesMacroExpander {
174 cx: &'cx mut ExtCtxt<'_>,
177 ) -> Box<dyn MacResult + 'cx> {
179 return DummyResult::any(sp);
195 fn macro_rules_dummy_expander<'cx>(
196 _: &'cx mut ExtCtxt<'_>,
199 ) -> Box<dyn MacResult + 'cx> {
200 DummyResult::any(span)
203 fn trace_macros_note(cx_expansions: &mut FxIndexMap<Span, Vec<String>>, sp: Span, message: String) {
204 let sp = sp.macro_backtrace().last().map_or(sp, |trace| trace.call_site);
205 cx_expansions.entry(sp).or_default().push(message);
208 /// Expands the rules based macro defined by `lhses` and `rhses` for a given
210 fn expand_macro<'cx>(
211 cx: &'cx mut ExtCtxt<'_>,
216 transparency: Transparency,
218 lhses: &[Vec<MatcherLoc>],
219 rhses: &[mbe::TokenTree],
220 ) -> Box<dyn MacResult + 'cx> {
221 let sess = &cx.sess.parse_sess;
222 // Macros defined in the current crate have a real node id,
223 // whereas macros from an external crate have a dummy id.
224 let is_local = node_id != DUMMY_NODE_ID;
226 if cx.trace_macros() {
227 let msg = format!("expanding `{}! {{ {} }}`", name, pprust::tts_to_string(&arg));
228 trace_macros_note(&mut cx.expansions, sp, msg);
231 // Which arm's failure should we report? (the one furthest along)
232 let mut best_failure: Option<(Token, &str)> = None;
234 // We create a base parser that can be used for the "black box" parts.
235 // Every iteration needs a fresh copy of that parser. However, the parser
236 // is not mutated on many of the iterations, particularly when dealing with
239 // macro_rules! foo {
243 // // ... etc. (maybe hundreds more)
246 // as seen in the `html5ever` benchmark. We use a `Cow` so that the base
247 // parser is only cloned when necessary (upon mutation). Furthermore, we
248 // reinitialize the `Cow` with the base parser at the start of every
249 // iteration, so that any mutated parsers are not reused. This is all quite
250 // hacky, but speeds up the `html5ever` benchmark significantly. (Issue
251 // 68836 suggests a more comprehensive but more complex change to deal with
253 let parser = parser_from_cx(sess, arg.clone());
255 // Try each arm's matchers.
256 let mut tt_parser = TtParser::new(name);
257 for (i, lhs) in lhses.iter().enumerate() {
258 // Take a snapshot of the state of pre-expansion gating at this point.
259 // This is used so that if a matcher is not `Success(..)`ful,
260 // then the spans which became gated when parsing the unsuccessful matcher
261 // are not recorded. On the first `Success(..)`ful matcher, the spans are merged.
262 let mut gated_spans_snapshot = mem::take(&mut *sess.gated_spans.spans.borrow_mut());
264 match tt_parser.parse_tt(&mut Cow::Borrowed(&parser), lhs) {
265 Success(named_matches) => {
266 // The matcher was `Success(..)`ful.
267 // Merge the gated spans from parsing the matcher with the pre-existing ones.
268 sess.gated_spans.merge(gated_spans_snapshot);
270 let (rhs, rhs_span): (&mbe::Delimited, DelimSpan) = match &rhses[i] {
271 mbe::TokenTree::Delimited(span, delimited) => (&delimited, *span),
272 _ => cx.span_bug(sp, "malformed macro rhs"),
274 let arm_span = rhses[i].span();
276 let rhs_spans = rhs.tts.iter().map(|t| t.span()).collect::<Vec<_>>();
277 // rhs has holes ( `$id` and `$(...)` that need filled)
278 let mut tts = match transcribe(cx, &named_matches, &rhs, rhs_span, transparency) {
282 return DummyResult::any(arm_span);
286 // Replace all the tokens for the corresponding positions in the macro, to maintain
287 // proper positions in error reporting, while maintaining the macro_backtrace.
288 if rhs_spans.len() == tts.len() {
289 tts = tts.map_enumerated(|i, tt| {
290 let mut tt = tt.clone();
291 let mut sp = rhs_spans[i];
292 sp = sp.with_ctxt(tt.span().ctxt());
298 if cx.trace_macros() {
299 let msg = format!("to `{}`", pprust::tts_to_string(&tts));
300 trace_macros_note(&mut cx.expansions, sp, msg);
303 let mut p = Parser::new(sess, tts, false, None);
304 p.last_type_ascription = cx.current_expansion.prior_type_ascription;
307 cx.resolver.record_macro_rule_usage(node_id, i);
310 // Let the context choose how to interpret the result.
311 // Weird, but useful for X-macros.
312 return Box::new(ParserAnyMacro {
315 // Pass along the original expansion site and the name of the macro
316 // so we can print a useful error message if the parse of the expanded
317 // macro leaves unparsed tokens.
320 lint_node_id: cx.current_expansion.lint_node_id,
321 is_trailing_mac: cx.current_expansion.is_trailing_mac,
326 Failure(token, msg) => match best_failure {
327 Some((ref best_token, _)) if best_token.span.lo() >= token.span.lo() => {}
328 _ => best_failure = Some((token, msg)),
330 Error(err_sp, ref msg) => {
331 let span = err_sp.substitute_dummy(sp);
332 cx.struct_span_err(span, &msg).emit();
333 return DummyResult::any(span);
335 ErrorReported => return DummyResult::any(sp),
338 // The matcher was not `Success(..)`ful.
339 // Restore to the state before snapshotting and maybe try again.
340 mem::swap(&mut gated_spans_snapshot, &mut sess.gated_spans.spans.borrow_mut());
344 let (token, label) = best_failure.expect("ran no matchers");
345 let span = token.span.substitute_dummy(sp);
346 let mut err = cx.struct_span_err(span, &parse_failure_msg(&token));
347 err.span_label(span, label);
348 if !def_span.is_dummy() && !cx.source_map().is_imported(def_span) {
349 err.span_label(cx.source_map().guess_head_span(def_span), "when calling this macro");
351 annotate_doc_comment(&mut err, sess.source_map(), span);
352 // Check whether there's a missing comma in this macro call, like `println!("{}" a);`
353 if let Some((arg, comma_span)) = arg.add_comma() {
355 let parser = parser_from_cx(sess, arg.clone());
356 if let Success(_) = tt_parser.parse_tt(&mut Cow::Borrowed(&parser), lhs) {
357 if comma_span.is_dummy() {
358 err.note("you might be missing a comma");
360 err.span_suggestion_short(
362 "missing comma here",
364 Applicability::MachineApplicable,
371 cx.trace_macros_diag();
375 // Note that macro-by-example's input is also matched against a token tree:
376 // $( $lhs:tt => $rhs:tt );+
378 // Holy self-referential!
380 /// Converts a macro item into a syntax extension.
381 pub fn compile_declarative_macro(
386 ) -> (SyntaxExtension, Vec<(usize, Span)>) {
387 debug!("compile_declarative_macro: {:?}", def);
388 let mk_syn_ext = |expander| {
389 SyntaxExtension::new(
391 SyntaxExtensionKind::LegacyBang(expander),
399 let dummy_syn_ext = || (mk_syn_ext(Box::new(macro_rules_dummy_expander)), Vec::new());
401 let diag = &sess.parse_sess.span_diagnostic;
402 let lhs_nm = Ident::new(sym::lhs, def.span);
403 let rhs_nm = Ident::new(sym::rhs, def.span);
404 let tt_spec = Some(NonterminalKind::TT);
406 // Parse the macro_rules! invocation
407 let (macro_rules, body) = match &def.kind {
408 ast::ItemKind::MacroDef(def) => (def.macro_rules, def.body.inner_tokens()),
412 // The pattern that macro_rules matches.
413 // The grammar for macro_rules! is:
414 // $( $lhs:tt => $rhs:tt );+
415 // ...quasiquoting this would be nice.
416 // These spans won't matter, anyways
417 let argument_gram = vec![
418 mbe::TokenTree::Sequence(
420 mbe::SequenceRepetition {
422 mbe::TokenTree::MetaVarDecl(def.span, lhs_nm, tt_spec),
423 mbe::TokenTree::token(token::FatArrow, def.span),
424 mbe::TokenTree::MetaVarDecl(def.span, rhs_nm, tt_spec),
426 separator: Some(Token::new(
427 if macro_rules { token::Semi } else { token::Comma },
430 kleene: mbe::KleeneToken::new(mbe::KleeneOp::OneOrMore, def.span),
434 // to phase into semicolon-termination instead of semicolon-separation
435 mbe::TokenTree::Sequence(
437 mbe::SequenceRepetition {
438 tts: vec![mbe::TokenTree::token(
439 if macro_rules { token::Semi } else { token::Comma },
443 kleene: mbe::KleeneToken::new(mbe::KleeneOp::ZeroOrMore, def.span),
448 // Convert it into `MatcherLoc` form.
449 let argument_gram = mbe::macro_parser::compute_locs(&argument_gram);
451 let parser = Parser::new(&sess.parse_sess, body, true, rustc_parse::MACRO_ARGUMENTS);
453 TtParser::new(Ident::with_dummy_span(if macro_rules { kw::MacroRules } else { kw::Macro }));
454 let argument_map = match tt_parser.parse_tt(&mut Cow::Borrowed(&parser), &argument_gram) {
456 Failure(token, msg) => {
457 let s = parse_failure_msg(&token);
458 let sp = token.span.substitute_dummy(def.span);
459 let mut err = sess.parse_sess.span_diagnostic.struct_span_err(sp, &s);
460 err.span_label(sp, msg);
461 annotate_doc_comment(&mut err, sess.source_map(), sp);
463 return dummy_syn_ext();
468 .struct_span_err(sp.substitute_dummy(def.span), &msg)
470 return dummy_syn_ext();
473 return dummy_syn_ext();
477 let mut valid = true;
479 // Extract the arguments:
480 let lhses = match argument_map[&MacroRulesNormalizedIdent::new(lhs_nm)] {
481 MatchedSeq(ref s) => s
484 if let MatchedTokenTree(ref tt) = *m {
485 let tt = mbe::quoted::parse(
486 TokenStream::new(vec![tt.clone()]),
495 valid &= check_lhs_nt_follows(&sess.parse_sess, &def, &tt);
498 sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs")
500 .collect::<Vec<mbe::TokenTree>>(),
501 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs"),
504 let rhses = match argument_map[&MacroRulesNormalizedIdent::new(rhs_nm)] {
505 MatchedSeq(ref s) => s
508 if let MatchedTokenTree(ref tt) = *m {
509 return mbe::quoted::parse(
510 TokenStream::new(vec![tt.clone()]),
520 sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs")
522 .collect::<Vec<mbe::TokenTree>>(),
523 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured rhs"),
527 valid &= check_rhs(&sess.parse_sess, rhs);
530 // don't abort iteration early, so that errors for multiple lhses can be reported
532 valid &= check_lhs_no_empty_seq(&sess.parse_sess, slice::from_ref(lhs));
535 valid &= macro_check::check_meta_variables(&sess.parse_sess, def.id, def.span, &lhses, &rhses);
537 let (transparency, transparency_error) = attr::find_transparency(&def.attrs, macro_rules);
538 match transparency_error {
539 Some(TransparencyError::UnknownTransparency(value, span)) => {
540 diag.span_err(span, &format!("unknown macro transparency: `{}`", value));
542 Some(TransparencyError::MultipleTransparencyAttrs(old_span, new_span)) => {
543 diag.span_err(vec![old_span, new_span], "multiple macro transparency attributes");
548 // Compute the spans of the macro rules for unused rule linting.
549 // To avoid warning noise, only consider the rules of this
550 // macro for the lint, if all rules are valid.
551 // Also, we are only interested in non-foreign macros.
552 let rule_spans = if valid && def.id != DUMMY_NODE_ID {
557 // If the rhs contains an invocation like compile_error!,
558 // don't consider the rule for the unused rule lint.
559 .filter(|(_idx, (_lhs, rhs))| !has_compile_error_macro(rhs))
560 // We only take the span of the lhs here,
561 // so that the spans of created warnings are smaller.
562 .map(|(idx, (lhs, _rhs))| (idx, lhs.span()))
568 // Convert the lhses into `MatcherLoc` form, which is better for doing the
569 // actual matching. Unless the matcher is invalid.
570 let lhses = if valid {
574 // Ignore the delimiters around the matcher.
576 mbe::TokenTree::Delimited(_, delimited) => {
577 mbe::macro_parser::compute_locs(&delimited.tts)
579 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "malformed macro lhs"),
587 let expander = Box::new(MacroRulesMacroExpander {
596 (mk_syn_ext(expander), rule_spans)
599 #[derive(Subdiagnostic)]
600 enum ExplainDocComment {
601 #[label(expand_explain_doc_comment_inner)]
606 #[label(expand_explain_doc_comment_outer)]
613 fn annotate_doc_comment(err: &mut Diagnostic, sm: &SourceMap, span: Span) {
614 if let Ok(src) = sm.span_to_snippet(span) {
615 if src.starts_with("///") || src.starts_with("/**") {
616 err.subdiagnostic(ExplainDocComment::Outer { span });
617 } else if src.starts_with("//!") || src.starts_with("/*!") {
618 err.subdiagnostic(ExplainDocComment::Inner { span });
623 fn check_lhs_nt_follows(sess: &ParseSess, def: &ast::Item, lhs: &mbe::TokenTree) -> bool {
624 // lhs is going to be like TokenTree::Delimited(...), where the
625 // entire lhs is those tts. Or, it can be a "bare sequence", not wrapped in parens.
626 if let mbe::TokenTree::Delimited(_, delimited) = lhs {
627 check_matcher(sess, def, &delimited.tts)
629 let msg = "invalid macro matcher; matchers must be contained in balanced delimiters";
630 sess.span_diagnostic.span_err(lhs.span(), msg);
633 // we don't abort on errors on rejection, the driver will do that for us
634 // after parsing/expansion. we can report every error in every macro this way.
637 /// Checks that the lhs contains no repetition which could match an empty token
638 /// tree, because then the matcher would hang indefinitely.
639 fn check_lhs_no_empty_seq(sess: &ParseSess, tts: &[mbe::TokenTree]) -> bool {
644 | TokenTree::MetaVar(..)
645 | TokenTree::MetaVarDecl(..)
646 | TokenTree::MetaVarExpr(..) => (),
647 TokenTree::Delimited(_, ref del) => {
648 if !check_lhs_no_empty_seq(sess, &del.tts) {
652 TokenTree::Sequence(span, ref seq) => {
653 if seq.separator.is_none()
654 && seq.tts.iter().all(|seq_tt| match *seq_tt {
655 TokenTree::MetaVarDecl(_, _, Some(NonterminalKind::Vis)) => true,
656 TokenTree::Sequence(_, ref sub_seq) => {
657 sub_seq.kleene.op == mbe::KleeneOp::ZeroOrMore
658 || sub_seq.kleene.op == mbe::KleeneOp::ZeroOrOne
663 let sp = span.entire();
664 sess.span_diagnostic.span_err(sp, "repetition matches empty token tree");
667 if !check_lhs_no_empty_seq(sess, &seq.tts) {
677 fn check_rhs(sess: &ParseSess, rhs: &mbe::TokenTree) -> bool {
679 mbe::TokenTree::Delimited(..) => return true,
681 sess.span_diagnostic.span_err(rhs.span(), "macro rhs must be delimited");
687 fn check_matcher(sess: &ParseSess, def: &ast::Item, matcher: &[mbe::TokenTree]) -> bool {
688 let first_sets = FirstSets::new(matcher);
689 let empty_suffix = TokenSet::empty();
690 let err = sess.span_diagnostic.err_count();
691 check_matcher_core(sess, def, &first_sets, matcher, &empty_suffix);
692 err == sess.span_diagnostic.err_count()
695 fn has_compile_error_macro(rhs: &mbe::TokenTree) -> bool {
697 mbe::TokenTree::Delimited(_sp, d) => {
698 let has_compile_error = d.tts.array_windows::<3>().any(|[ident, bang, args]| {
699 if let mbe::TokenTree::Token(ident) = ident &&
700 let TokenKind::Ident(ident, _) = ident.kind &&
701 ident == sym::compile_error &&
702 let mbe::TokenTree::Token(bang) = bang &&
703 let TokenKind::Not = bang.kind &&
704 let mbe::TokenTree::Delimited(_, del) = args &&
705 del.delim != Delimiter::Invisible
712 if has_compile_error { true } else { d.tts.iter().any(has_compile_error_macro) }
718 // `The FirstSets` for a matcher is a mapping from subsequences in the
719 // matcher to the FIRST set for that subsequence.
721 // This mapping is partially precomputed via a backwards scan over the
722 // token trees of the matcher, which provides a mapping from each
723 // repetition sequence to its *first* set.
725 // (Hypothetically, sequences should be uniquely identifiable via their
726 // spans, though perhaps that is false, e.g., for macro-generated macros
727 // that do not try to inject artificial span information. My plan is
728 // to try to catch such cases ahead of time and not include them in
729 // the precomputed mapping.)
730 struct FirstSets<'tt> {
731 // this maps each TokenTree::Sequence `$(tt ...) SEP OP` that is uniquely identified by its
732 // span in the original matcher to the First set for the inner sequence `tt ...`.
734 // If two sequences have the same span in a matcher, then map that
735 // span to None (invalidating the mapping here and forcing the code to
737 first: FxHashMap<Span, Option<TokenSet<'tt>>>,
740 impl<'tt> FirstSets<'tt> {
741 fn new(tts: &'tt [mbe::TokenTree]) -> FirstSets<'tt> {
744 let mut sets = FirstSets { first: FxHashMap::default() };
745 build_recur(&mut sets, tts);
748 // walks backward over `tts`, returning the FIRST for `tts`
749 // and updating `sets` at the same time for all sequence
750 // substructure we find within `tts`.
751 fn build_recur<'tt>(sets: &mut FirstSets<'tt>, tts: &'tt [TokenTree]) -> TokenSet<'tt> {
752 let mut first = TokenSet::empty();
753 for tt in tts.iter().rev() {
756 | TokenTree::MetaVar(..)
757 | TokenTree::MetaVarDecl(..)
758 | TokenTree::MetaVarExpr(..) => {
759 first.replace_with(TtHandle::TtRef(tt));
761 TokenTree::Delimited(span, ref delimited) => {
762 build_recur(sets, &delimited.tts);
763 first.replace_with(TtHandle::from_token_kind(
764 token::OpenDelim(delimited.delim),
768 TokenTree::Sequence(sp, ref seq_rep) => {
769 let subfirst = build_recur(sets, &seq_rep.tts);
771 match sets.first.entry(sp.entire()) {
772 Entry::Vacant(vac) => {
773 vac.insert(Some(subfirst.clone()));
775 Entry::Occupied(mut occ) => {
776 // if there is already an entry, then a span must have collided.
777 // This should not happen with typical macro_rules macros,
778 // but syntax extensions need not maintain distinct spans,
779 // so distinct syntax trees can be assigned the same span.
780 // In such a case, the map cannot be trusted; so mark this
781 // entry as unusable.
786 // If the sequence contents can be empty, then the first
787 // token could be the separator token itself.
789 if let (Some(sep), true) = (&seq_rep.separator, subfirst.maybe_empty) {
790 first.add_one_maybe(TtHandle::from_token(sep.clone()));
793 // Reverse scan: Sequence comes before `first`.
794 if subfirst.maybe_empty
795 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrMore
796 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrOne
798 // If sequence is potentially empty, then
799 // union them (preserving first emptiness).
800 first.add_all(&TokenSet { maybe_empty: true, ..subfirst });
802 // Otherwise, sequence guaranteed
803 // non-empty; replace first.
814 // walks forward over `tts` until all potential FIRST tokens are
816 fn first(&self, tts: &'tt [mbe::TokenTree]) -> TokenSet<'tt> {
819 let mut first = TokenSet::empty();
820 for tt in tts.iter() {
821 assert!(first.maybe_empty);
824 | TokenTree::MetaVar(..)
825 | TokenTree::MetaVarDecl(..)
826 | TokenTree::MetaVarExpr(..) => {
827 first.add_one(TtHandle::TtRef(tt));
830 TokenTree::Delimited(span, ref delimited) => {
831 first.add_one(TtHandle::from_token_kind(
832 token::OpenDelim(delimited.delim),
837 TokenTree::Sequence(sp, ref seq_rep) => {
839 let subfirst = match self.first.get(&sp.entire()) {
840 Some(&Some(ref subfirst)) => subfirst,
842 subfirst_owned = self.first(&seq_rep.tts);
846 panic!("We missed a sequence during FirstSets construction");
850 // If the sequence contents can be empty, then the first
851 // token could be the separator token itself.
852 if let (Some(sep), true) = (&seq_rep.separator, subfirst.maybe_empty) {
853 first.add_one_maybe(TtHandle::from_token(sep.clone()));
856 assert!(first.maybe_empty);
857 first.add_all(subfirst);
858 if subfirst.maybe_empty
859 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrMore
860 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrOne
862 // Continue scanning for more first
863 // tokens, but also make sure we
864 // restore empty-tracking state.
865 first.maybe_empty = true;
874 // we only exit the loop if `tts` was empty or if every
875 // element of `tts` matches the empty sequence.
876 assert!(first.maybe_empty);
881 // Most `mbe::TokenTree`s are pre-existing in the matcher, but some are defined
882 // implicitly, such as opening/closing delimiters and sequence repetition ops.
883 // This type encapsulates both kinds. It implements `Clone` while avoiding the
884 // need for `mbe::TokenTree` to implement `Clone`.
887 /// This is used in most cases.
888 TtRef(&'tt mbe::TokenTree),
890 /// This is only used for implicit token trees. The `mbe::TokenTree` *must*
891 /// be `mbe::TokenTree::Token`. No other variants are allowed. We store an
892 /// `mbe::TokenTree` rather than a `Token` so that `get()` can return a
893 /// `&mbe::TokenTree`.
894 Token(mbe::TokenTree),
897 impl<'tt> TtHandle<'tt> {
898 fn from_token(tok: Token) -> Self {
899 TtHandle::Token(mbe::TokenTree::Token(tok))
902 fn from_token_kind(kind: TokenKind, span: Span) -> Self {
903 TtHandle::from_token(Token::new(kind, span))
906 // Get a reference to a token tree.
907 fn get(&'tt self) -> &'tt mbe::TokenTree {
909 TtHandle::TtRef(tt) => tt,
910 TtHandle::Token(token_tt) => &token_tt,
915 impl<'tt> PartialEq for TtHandle<'tt> {
916 fn eq(&self, other: &TtHandle<'tt>) -> bool {
917 self.get() == other.get()
921 impl<'tt> Clone for TtHandle<'tt> {
922 fn clone(&self) -> Self {
924 TtHandle::TtRef(tt) => TtHandle::TtRef(tt),
926 // This variant *must* contain a `mbe::TokenTree::Token`, and not
927 // any other variant of `mbe::TokenTree`.
928 TtHandle::Token(mbe::TokenTree::Token(tok)) => {
929 TtHandle::Token(mbe::TokenTree::Token(tok.clone()))
937 // A set of `mbe::TokenTree`s, which may include `TokenTree::Match`s
938 // (for macro-by-example syntactic variables). It also carries the
939 // `maybe_empty` flag; that is true if and only if the matcher can
940 // match an empty token sequence.
942 // The First set is computed on submatchers like `$($a:expr b),* $(c)* d`,
943 // which has corresponding FIRST = {$a:expr, c, d}.
944 // Likewise, `$($a:expr b),* $(c)+ d` has FIRST = {$a:expr, c}.
946 // (Notably, we must allow for *-op to occur zero times.)
947 #[derive(Clone, Debug)]
948 struct TokenSet<'tt> {
949 tokens: Vec<TtHandle<'tt>>,
953 impl<'tt> TokenSet<'tt> {
954 // Returns a set for the empty sequence.
956 TokenSet { tokens: Vec::new(), maybe_empty: true }
959 // Returns the set `{ tok }` for the single-token (and thus
960 // non-empty) sequence [tok].
961 fn singleton(tt: TtHandle<'tt>) -> Self {
962 TokenSet { tokens: vec![tt], maybe_empty: false }
965 // Changes self to be the set `{ tok }`.
966 // Since `tok` is always present, marks self as non-empty.
967 fn replace_with(&mut self, tt: TtHandle<'tt>) {
969 self.tokens.push(tt);
970 self.maybe_empty = false;
973 // Changes self to be the empty set `{}`; meant for use when
974 // the particular token does not matter, but we want to
975 // record that it occurs.
976 fn replace_with_irrelevant(&mut self) {
978 self.maybe_empty = false;
981 // Adds `tok` to the set for `self`, marking sequence as non-empty.
982 fn add_one(&mut self, tt: TtHandle<'tt>) {
983 if !self.tokens.contains(&tt) {
984 self.tokens.push(tt);
986 self.maybe_empty = false;
989 // Adds `tok` to the set for `self`. (Leaves `maybe_empty` flag alone.)
990 fn add_one_maybe(&mut self, tt: TtHandle<'tt>) {
991 if !self.tokens.contains(&tt) {
992 self.tokens.push(tt);
996 // Adds all elements of `other` to this.
998 // (Since this is a set, we filter out duplicates.)
1000 // If `other` is potentially empty, then preserves the previous
1001 // setting of the empty flag of `self`. If `other` is guaranteed
1002 // non-empty, then `self` is marked non-empty.
1003 fn add_all(&mut self, other: &Self) {
1004 for tt in &other.tokens {
1005 if !self.tokens.contains(tt) {
1006 self.tokens.push(tt.clone());
1009 if !other.maybe_empty {
1010 self.maybe_empty = false;
1015 // Checks that `matcher` is internally consistent and that it
1016 // can legally be followed by a token `N`, for all `N` in `follow`.
1017 // (If `follow` is empty, then it imposes no constraint on
1020 // Returns the set of NT tokens that could possibly come last in
1021 // `matcher`. (If `matcher` matches the empty sequence, then
1022 // `maybe_empty` will be set to true.)
1024 // Requires that `first_sets` is pre-computed for `matcher`;
1025 // see `FirstSets::new`.
1026 fn check_matcher_core<'tt>(
1029 first_sets: &FirstSets<'tt>,
1030 matcher: &'tt [mbe::TokenTree],
1031 follow: &TokenSet<'tt>,
1032 ) -> TokenSet<'tt> {
1035 let mut last = TokenSet::empty();
1037 // 2. For each token and suffix [T, SUFFIX] in M:
1038 // ensure that T can be followed by SUFFIX, and if SUFFIX may be empty,
1039 // then ensure T can also be followed by any element of FOLLOW.
1040 'each_token: for i in 0..matcher.len() {
1041 let token = &matcher[i];
1042 let suffix = &matcher[i + 1..];
1044 let build_suffix_first = || {
1045 let mut s = first_sets.first(suffix);
1052 // (we build `suffix_first` on demand below; you can tell
1053 // which cases are supposed to fall through by looking for the
1054 // initialization of this variable.)
1057 // First, update `last` so that it corresponds to the set
1058 // of NT tokens that might end the sequence `... token`.
1060 TokenTree::Token(..)
1061 | TokenTree::MetaVar(..)
1062 | TokenTree::MetaVarDecl(..)
1063 | TokenTree::MetaVarExpr(..) => {
1064 if token_can_be_followed_by_any(token) {
1065 // don't need to track tokens that work with any,
1066 last.replace_with_irrelevant();
1067 // ... and don't need to check tokens that can be
1068 // followed by anything against SUFFIX.
1069 continue 'each_token;
1071 last.replace_with(TtHandle::TtRef(token));
1072 suffix_first = build_suffix_first();
1075 TokenTree::Delimited(span, ref d) => {
1076 let my_suffix = TokenSet::singleton(TtHandle::from_token_kind(
1077 token::CloseDelim(d.delim),
1080 check_matcher_core(sess, def, first_sets, &d.tts, &my_suffix);
1081 // don't track non NT tokens
1082 last.replace_with_irrelevant();
1084 // also, we don't need to check delimited sequences
1086 continue 'each_token;
1088 TokenTree::Sequence(_, ref seq_rep) => {
1089 suffix_first = build_suffix_first();
1090 // The trick here: when we check the interior, we want
1091 // to include the separator (if any) as a potential
1092 // (but not guaranteed) element of FOLLOW. So in that
1093 // case, we make a temp copy of suffix and stuff
1094 // delimiter in there.
1096 // FIXME: Should I first scan suffix_first to see if
1097 // delimiter is already in it before I go through the
1098 // work of cloning it? But then again, this way I may
1099 // get a "tighter" span?
1101 let my_suffix = if let Some(sep) = &seq_rep.separator {
1102 new = suffix_first.clone();
1103 new.add_one_maybe(TtHandle::from_token(sep.clone()));
1109 // At this point, `suffix_first` is built, and
1110 // `my_suffix` is some TokenSet that we can use
1111 // for checking the interior of `seq_rep`.
1112 let next = check_matcher_core(sess, def, first_sets, &seq_rep.tts, my_suffix);
1113 if next.maybe_empty {
1114 last.add_all(&next);
1119 // the recursive call to check_matcher_core already ran the 'each_last
1120 // check below, so we can just keep going forward here.
1121 continue 'each_token;
1125 // (`suffix_first` guaranteed initialized once reaching here.)
1127 // Now `last` holds the complete set of NT tokens that could
1128 // end the sequence before SUFFIX. Check that every one works with `suffix`.
1129 for tt in &last.tokens {
1130 if let &TokenTree::MetaVarDecl(span, name, Some(kind)) = tt.get() {
1131 for next_token in &suffix_first.tokens {
1132 let next_token = next_token.get();
1134 // Check if the old pat is used and the next token is `|`
1135 // to warn about incompatibility with Rust 2021.
1136 // We only emit this lint if we're parsing the original
1137 // definition of this macro_rules, not while (re)parsing
1138 // the macro when compiling another crate that is using the
1139 // macro. (See #86567.)
1140 // Macros defined in the current crate have a real node id,
1141 // whereas macros from an external crate have a dummy id.
1142 if def.id != DUMMY_NODE_ID
1143 && matches!(kind, NonterminalKind::PatParam { inferred: true })
1144 && matches!(next_token, TokenTree::Token(token) if token.kind == BinOp(token::BinOpToken::Or))
1146 // It is suggestion to use pat_param, for example: $x:pat -> $x:pat_param.
1147 let suggestion = quoted_tt_to_string(&TokenTree::MetaVarDecl(
1150 Some(NonterminalKind::PatParam { inferred: false }),
1152 sess.buffer_lint_with_diagnostic(
1153 &RUST_2021_INCOMPATIBLE_OR_PATTERNS,
1156 "the meaning of the `pat` fragment specifier is changing in Rust 2021, which may affect this macro",
1157 BuiltinLintDiagnostics::OrPatternsBackCompat(span, suggestion),
1160 match is_in_follow(next_token, kind) {
1161 IsInFollow::Yes => {}
1162 IsInFollow::No(possible) => {
1163 let may_be = if last.tokens.len() == 1 && suffix_first.tokens.len() == 1
1170 let sp = next_token.span();
1171 let mut err = sess.span_diagnostic.struct_span_err(
1174 "`${name}:{frag}` {may_be} followed by `{next}`, which \
1175 is not allowed for `{frag}` fragments",
1178 next = quoted_tt_to_string(next_token),
1182 err.span_label(sp, format!("not allowed after `{}` fragments", kind));
1184 if kind == NonterminalKind::PatWithOr
1185 && sess.edition.rust_2021()
1186 && next_token.is_token(&BinOp(token::BinOpToken::Or))
1188 let suggestion = quoted_tt_to_string(&TokenTree::MetaVarDecl(
1191 Some(NonterminalKind::PatParam { inferred: false }),
1193 err.span_suggestion(
1195 "try a `pat_param` fragment specifier instead",
1197 Applicability::MaybeIncorrect,
1201 let msg = "allowed there are: ";
1206 "only {} is allowed after `{}` fragments",
1217 .collect::<Vec<_>>()
1233 fn token_can_be_followed_by_any(tok: &mbe::TokenTree) -> bool {
1234 if let mbe::TokenTree::MetaVarDecl(_, _, Some(kind)) = *tok {
1235 frag_can_be_followed_by_any(kind)
1237 // (Non NT's can always be followed by anything in matchers.)
1242 /// Returns `true` if a fragment of type `frag` can be followed by any sort of
1243 /// token. We use this (among other things) as a useful approximation
1244 /// for when `frag` can be followed by a repetition like `$(...)*` or
1245 /// `$(...)+`. In general, these can be a bit tricky to reason about,
1246 /// so we adopt a conservative position that says that any fragment
1247 /// specifier which consumes at most one token tree can be followed by
1248 /// a fragment specifier (indeed, these fragments can be followed by
1249 /// ANYTHING without fear of future compatibility hazards).
1250 fn frag_can_be_followed_by_any(kind: NonterminalKind) -> bool {
1253 NonterminalKind::Item // always terminated by `}` or `;`
1254 | NonterminalKind::Block // exactly one token tree
1255 | NonterminalKind::Ident // exactly one token tree
1256 | NonterminalKind::Literal // exactly one token tree
1257 | NonterminalKind::Meta // exactly one token tree
1258 | NonterminalKind::Lifetime // exactly one token tree
1259 | NonterminalKind::TT // exactly one token tree
1265 No(&'static [&'static str]),
1268 /// Returns `true` if `frag` can legally be followed by the token `tok`. For
1269 /// fragments that can consume an unbounded number of tokens, `tok`
1270 /// must be within a well-defined follow set. This is intended to
1271 /// guarantee future compatibility: for example, without this rule, if
1272 /// we expanded `expr` to include a new binary operator, we might
1273 /// break macros that were relying on that binary operator as a
1275 // when changing this do not forget to update doc/book/macros.md!
1276 fn is_in_follow(tok: &mbe::TokenTree, kind: NonterminalKind) -> IsInFollow {
1279 if let TokenTree::Token(Token { kind: token::CloseDelim(_), .. }) = *tok {
1280 // closing a token tree can never be matched by any fragment;
1281 // iow, we always require that `(` and `)` match, etc.
1285 NonterminalKind::Item => {
1286 // since items *must* be followed by either a `;` or a `}`, we can
1287 // accept anything after them
1290 NonterminalKind::Block => {
1291 // anything can follow block, the braces provide an easy boundary to
1295 NonterminalKind::Stmt | NonterminalKind::Expr => {
1296 const TOKENS: &[&str] = &["`=>`", "`,`", "`;`"];
1298 TokenTree::Token(token) => match token.kind {
1299 FatArrow | Comma | Semi => IsInFollow::Yes,
1300 _ => IsInFollow::No(TOKENS),
1302 _ => IsInFollow::No(TOKENS),
1305 NonterminalKind::PatParam { .. } => {
1306 const TOKENS: &[&str] = &["`=>`", "`,`", "`=`", "`|`", "`if`", "`in`"];
1308 TokenTree::Token(token) => match token.kind {
1309 FatArrow | Comma | Eq | BinOp(token::Or) => IsInFollow::Yes,
1310 Ident(name, false) if name == kw::If || name == kw::In => IsInFollow::Yes,
1311 _ => IsInFollow::No(TOKENS),
1313 _ => IsInFollow::No(TOKENS),
1316 NonterminalKind::PatWithOr { .. } => {
1317 const TOKENS: &[&str] = &["`=>`", "`,`", "`=`", "`if`", "`in`"];
1319 TokenTree::Token(token) => match token.kind {
1320 FatArrow | Comma | Eq => IsInFollow::Yes,
1321 Ident(name, false) if name == kw::If || name == kw::In => IsInFollow::Yes,
1322 _ => IsInFollow::No(TOKENS),
1324 _ => IsInFollow::No(TOKENS),
1327 NonterminalKind::Path | NonterminalKind::Ty => {
1328 const TOKENS: &[&str] = &[
1329 "`{`", "`[`", "`=>`", "`,`", "`>`", "`=`", "`:`", "`;`", "`|`", "`as`",
1333 TokenTree::Token(token) => match token.kind {
1334 OpenDelim(Delimiter::Brace)
1335 | OpenDelim(Delimiter::Bracket)
1343 | BinOp(token::Or) => IsInFollow::Yes,
1344 Ident(name, false) if name == kw::As || name == kw::Where => {
1347 _ => IsInFollow::No(TOKENS),
1349 TokenTree::MetaVarDecl(_, _, Some(NonterminalKind::Block)) => IsInFollow::Yes,
1350 _ => IsInFollow::No(TOKENS),
1353 NonterminalKind::Ident | NonterminalKind::Lifetime => {
1354 // being a single token, idents and lifetimes are harmless
1357 NonterminalKind::Literal => {
1358 // literals may be of a single token, or two tokens (negative numbers)
1361 NonterminalKind::Meta | NonterminalKind::TT => {
1362 // being either a single token or a delimited sequence, tt is
1366 NonterminalKind::Vis => {
1367 // Explicitly disallow `priv`, on the off chance it comes back.
1368 const TOKENS: &[&str] = &["`,`", "an ident", "a type"];
1370 TokenTree::Token(token) => match token.kind {
1371 Comma => IsInFollow::Yes,
1372 Ident(name, is_raw) if is_raw || name != kw::Priv => IsInFollow::Yes,
1374 if token.can_begin_type() {
1377 IsInFollow::No(TOKENS)
1381 TokenTree::MetaVarDecl(
1384 Some(NonterminalKind::Ident | NonterminalKind::Ty | NonterminalKind::Path),
1385 ) => IsInFollow::Yes,
1386 _ => IsInFollow::No(TOKENS),
1393 fn quoted_tt_to_string(tt: &mbe::TokenTree) -> String {
1395 mbe::TokenTree::Token(ref token) => pprust::token_to_string(&token).into(),
1396 mbe::TokenTree::MetaVar(_, name) => format!("${}", name),
1397 mbe::TokenTree::MetaVarDecl(_, name, Some(kind)) => format!("${}:{}", name, kind),
1398 mbe::TokenTree::MetaVarDecl(_, name, None) => format!("${}:", name),
1401 "unexpected mbe::TokenTree::{Sequence or Delimited} \
1402 in follow set checker"
1407 fn parser_from_cx(sess: &ParseSess, tts: TokenStream) -> Parser<'_> {
1408 Parser::new(sess, tts, true, rustc_parse::MACRO_ARGUMENTS)
1411 /// Generates an appropriate parsing failure message. For EOF, this is "unexpected end...". For
1412 /// other tokens, this is "unexpected token...".
1413 fn parse_failure_msg(tok: &Token) -> String {
1415 token::Eof => "unexpected end of macro invocation".to_string(),
1416 _ => format!("no rules expected the token `{}`", pprust::token_to_string(tok),),