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};
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};
37 pub(crate) struct ParserAnyMacro<'a> {
40 /// Span of the expansion site of the macro this parser is for
42 /// The ident of the macro we're parsing
45 is_trailing_mac: bool,
47 /// Whether or not this macro is defined in the current crate
51 pub(crate) fn annotate_err_with_kind(err: &mut Diagnostic, kind: AstFragmentKind, span: Span) {
53 AstFragmentKind::Ty => {
54 err.span_label(span, "this macro call doesn't expand to a type");
56 AstFragmentKind::Pat => {
57 err.span_label(span, "this macro call doesn't expand to a pattern");
63 fn emit_frag_parse_err(
64 mut e: DiagnosticBuilder<'_, rustc_errors::ErrorGuaranteed>,
66 orig_parser: &mut Parser<'_>,
69 kind: AstFragmentKind,
71 // FIXME(davidtwco): avoid depending on the error message text
72 if parser.token == token::Eof && e.message[0].0.expect_str().ends_with(", found `<eof>`") {
73 if !e.span.is_dummy() {
74 // early end of macro arm (#52866)
75 e.replace_span_with(parser.sess.source_map().next_point(parser.token.span));
77 let msg = &e.message[0];
79 rustc_errors::DiagnosticMessage::Str(format!(
80 "macro expansion ends with an incomplete expression: {}",
81 msg.0.expect_str().replace(", found `<eof>`", ""),
86 if e.span.is_dummy() {
87 // Get around lack of span in error (#30128)
88 e.replace_span_with(site_span);
89 if !parser.sess.source_map().is_imported(arm_span) {
90 e.span_label(arm_span, "in this macro arm");
92 } else if parser.sess.source_map().is_imported(parser.token.span) {
93 e.span_label(site_span, "in this macro invocation");
96 // Try a statement if an expression is wanted but failed and suggest adding `;` to call.
97 AstFragmentKind::Expr => match parse_ast_fragment(orig_parser, AstFragmentKind::Stmts) {
98 Err(err) => err.cancel(),
101 "the macro call doesn't expand to an expression, but it can expand to a statement",
103 e.span_suggestion_verbose(
104 site_span.shrink_to_hi(),
105 "add `;` to interpret the expansion as a statement",
107 Applicability::MaybeIncorrect,
111 _ => annotate_err_with_kind(&mut e, kind, site_span),
116 impl<'a> ParserAnyMacro<'a> {
117 pub(crate) fn make(mut self: Box<ParserAnyMacro<'a>>, kind: AstFragmentKind) -> AstFragment {
127 let snapshot = &mut parser.create_snapshot_for_diagnostic();
128 let fragment = match parse_ast_fragment(parser, kind) {
131 emit_frag_parse_err(err, parser, snapshot, site_span, arm_span, kind);
132 return kind.dummy(site_span);
136 // We allow semicolons at the end of expressions -- e.g., the semicolon in
137 // `macro_rules! m { () => { panic!(); } }` isn't parsed by `.parse_expr()`,
138 // but `m!()` is allowed in expression positions (cf. issue #34706).
139 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),
152 // Make sure we don't have any tokens left to parse so we don't silently drop anything.
153 let path = ast::Path::from_ident(macro_ident.with_span_pos(site_span));
154 ensure_complete_parse(parser, &path, kind.name(), site_span);
159 struct MacroRulesMacroExpander {
163 transparency: Transparency,
164 lhses: Vec<Vec<MatcherLoc>>,
165 rhses: Vec<mbe::TokenTree>,
169 impl TTMacroExpander for MacroRulesMacroExpander {
172 cx: &'cx mut ExtCtxt<'_>,
175 ) -> Box<dyn MacResult + 'cx> {
177 return DummyResult::any(sp);
193 fn macro_rules_dummy_expander<'cx>(
194 _: &'cx mut ExtCtxt<'_>,
197 ) -> Box<dyn MacResult + 'cx> {
198 DummyResult::any(span)
201 fn trace_macros_note(cx_expansions: &mut FxIndexMap<Span, Vec<String>>, sp: Span, message: String) {
202 let sp = sp.macro_backtrace().last().map_or(sp, |trace| trace.call_site);
203 cx_expansions.entry(sp).or_default().push(message);
206 /// Expands the rules based macro defined by `lhses` and `rhses` for a given
208 fn expand_macro<'cx>(
209 cx: &'cx mut ExtCtxt<'_>,
214 transparency: Transparency,
216 lhses: &[Vec<MatcherLoc>],
217 rhses: &[mbe::TokenTree],
218 ) -> Box<dyn MacResult + 'cx> {
219 let sess = &cx.sess.parse_sess;
220 // Macros defined in the current crate have a real node id,
221 // whereas macros from an external crate have a dummy id.
222 let is_local = node_id != DUMMY_NODE_ID;
224 if cx.trace_macros() {
225 let msg = format!("expanding `{}! {{ {} }}`", name, pprust::tts_to_string(&arg));
226 trace_macros_note(&mut cx.expansions, sp, msg);
229 // Which arm's failure should we report? (the one furthest along)
230 let mut best_failure: Option<(Token, &str)> = None;
232 // We create a base parser that can be used for the "black box" parts.
233 // Every iteration needs a fresh copy of that parser. However, the parser
234 // is not mutated on many of the iterations, particularly when dealing with
237 // macro_rules! foo {
241 // // ... etc. (maybe hundreds more)
244 // as seen in the `html5ever` benchmark. We use a `Cow` so that the base
245 // parser is only cloned when necessary (upon mutation). Furthermore, we
246 // reinitialize the `Cow` with the base parser at the start of every
247 // iteration, so that any mutated parsers are not reused. This is all quite
248 // hacky, but speeds up the `html5ever` benchmark significantly. (Issue
249 // 68836 suggests a more comprehensive but more complex change to deal with
251 let parser = parser_from_cx(sess, arg.clone());
253 // Try each arm's matchers.
254 let mut tt_parser = TtParser::new(name);
255 for (i, lhs) in lhses.iter().enumerate() {
256 // Take a snapshot of the state of pre-expansion gating at this point.
257 // This is used so that if a matcher is not `Success(..)`ful,
258 // then the spans which became gated when parsing the unsuccessful matcher
259 // are not recorded. On the first `Success(..)`ful matcher, the spans are merged.
260 let mut gated_spans_snapshot = mem::take(&mut *sess.gated_spans.spans.borrow_mut());
262 match tt_parser.parse_tt(&mut Cow::Borrowed(&parser), lhs) {
263 Success(named_matches) => {
264 // The matcher was `Success(..)`ful.
265 // Merge the gated spans from parsing the matcher with the pre-existing ones.
266 sess.gated_spans.merge(gated_spans_snapshot);
268 let (rhs, rhs_span): (&mbe::Delimited, DelimSpan) = match &rhses[i] {
269 mbe::TokenTree::Delimited(span, delimited) => (&delimited, *span),
270 _ => cx.span_bug(sp, "malformed macro rhs"),
272 let arm_span = rhses[i].span();
274 let rhs_spans = rhs.tts.iter().map(|t| t.span()).collect::<Vec<_>>();
275 // rhs has holes ( `$id` and `$(...)` that need filled)
276 let mut tts = match transcribe(cx, &named_matches, &rhs, rhs_span, transparency) {
280 return DummyResult::any(arm_span);
284 // Replace all the tokens for the corresponding positions in the macro, to maintain
285 // proper positions in error reporting, while maintaining the macro_backtrace.
286 if rhs_spans.len() == tts.len() {
287 tts = tts.map_enumerated(|i, tt| {
288 let mut tt = tt.clone();
289 let mut sp = rhs_spans[i];
290 sp = sp.with_ctxt(tt.span().ctxt());
296 if cx.trace_macros() {
297 let msg = format!("to `{}`", pprust::tts_to_string(&tts));
298 trace_macros_note(&mut cx.expansions, sp, msg);
301 let mut p = Parser::new(sess, tts, false, None);
302 p.last_type_ascription = cx.current_expansion.prior_type_ascription;
305 cx.resolver.record_macro_rule_usage(node_id, i);
308 // Let the context choose how to interpret the result.
309 // Weird, but useful for X-macros.
310 return Box::new(ParserAnyMacro {
313 // Pass along the original expansion site and the name of the macro
314 // so we can print a useful error message if the parse of the expanded
315 // macro leaves unparsed tokens.
318 lint_node_id: cx.current_expansion.lint_node_id,
319 is_trailing_mac: cx.current_expansion.is_trailing_mac,
324 Failure(token, msg) => match best_failure {
325 Some((ref best_token, _)) if best_token.span.lo() >= token.span.lo() => {}
326 _ => best_failure = Some((token, msg)),
328 Error(err_sp, ref msg) => {
329 let span = err_sp.substitute_dummy(sp);
330 cx.struct_span_err(span, &msg).emit();
331 return DummyResult::any(span);
333 ErrorReported => return DummyResult::any(sp),
336 // The matcher was not `Success(..)`ful.
337 // Restore to the state before snapshotting and maybe try again.
338 mem::swap(&mut gated_spans_snapshot, &mut sess.gated_spans.spans.borrow_mut());
342 let (token, label) = best_failure.expect("ran no matchers");
343 let span = token.span.substitute_dummy(sp);
344 let mut err = cx.struct_span_err(span, &parse_failure_msg(&token));
345 err.span_label(span, label);
346 if !def_span.is_dummy() && !cx.source_map().is_imported(def_span) {
347 err.span_label(cx.source_map().guess_head_span(def_span), "when calling this macro");
349 annotate_doc_comment(&mut err, sess.source_map(), span);
350 // Check whether there's a missing comma in this macro call, like `println!("{}" a);`
351 if let Some((arg, comma_span)) = arg.add_comma() {
353 let parser = parser_from_cx(sess, arg.clone());
354 if let Success(_) = tt_parser.parse_tt(&mut Cow::Borrowed(&parser), lhs) {
355 if comma_span.is_dummy() {
356 err.note("you might be missing a comma");
358 err.span_suggestion_short(
360 "missing comma here",
362 Applicability::MachineApplicable,
369 cx.trace_macros_diag();
373 // Note that macro-by-example's input is also matched against a token tree:
374 // $( $lhs:tt => $rhs:tt );+
376 // Holy self-referential!
378 /// Converts a macro item into a syntax extension.
379 pub fn compile_declarative_macro(
384 ) -> (SyntaxExtension, Vec<(usize, Span)>) {
385 debug!("compile_declarative_macro: {:?}", def);
386 let mk_syn_ext = |expander| {
387 SyntaxExtension::new(
389 SyntaxExtensionKind::LegacyBang(expander),
397 let dummy_syn_ext = || (mk_syn_ext(Box::new(macro_rules_dummy_expander)), Vec::new());
399 let diag = &sess.parse_sess.span_diagnostic;
400 let lhs_nm = Ident::new(sym::lhs, def.span);
401 let rhs_nm = Ident::new(sym::rhs, def.span);
402 let tt_spec = Some(NonterminalKind::TT);
404 // Parse the macro_rules! invocation
405 let (macro_rules, body) = match &def.kind {
406 ast::ItemKind::MacroDef(def) => (def.macro_rules, def.body.inner_tokens()),
410 // The pattern that macro_rules matches.
411 // The grammar for macro_rules! is:
412 // $( $lhs:tt => $rhs:tt );+
413 // ...quasiquoting this would be nice.
414 // These spans won't matter, anyways
415 let argument_gram = vec![
416 mbe::TokenTree::Sequence(
418 mbe::SequenceRepetition {
420 mbe::TokenTree::MetaVarDecl(def.span, lhs_nm, tt_spec),
421 mbe::TokenTree::token(token::FatArrow, def.span),
422 mbe::TokenTree::MetaVarDecl(def.span, rhs_nm, tt_spec),
424 separator: Some(Token::new(
425 if macro_rules { token::Semi } else { token::Comma },
428 kleene: mbe::KleeneToken::new(mbe::KleeneOp::OneOrMore, def.span),
432 // to phase into semicolon-termination instead of semicolon-separation
433 mbe::TokenTree::Sequence(
435 mbe::SequenceRepetition {
436 tts: vec![mbe::TokenTree::token(
437 if macro_rules { token::Semi } else { token::Comma },
441 kleene: mbe::KleeneToken::new(mbe::KleeneOp::ZeroOrMore, def.span),
446 // Convert it into `MatcherLoc` form.
447 let argument_gram = mbe::macro_parser::compute_locs(&argument_gram);
449 let parser = Parser::new(&sess.parse_sess, body, true, rustc_parse::MACRO_ARGUMENTS);
451 TtParser::new(Ident::with_dummy_span(if macro_rules { kw::MacroRules } else { kw::Macro }));
452 let argument_map = match tt_parser.parse_tt(&mut Cow::Borrowed(&parser), &argument_gram) {
454 Failure(token, msg) => {
455 let s = parse_failure_msg(&token);
456 let sp = token.span.substitute_dummy(def.span);
457 let mut err = sess.parse_sess.span_diagnostic.struct_span_err(sp, &s);
458 err.span_label(sp, msg);
459 annotate_doc_comment(&mut err, sess.source_map(), sp);
461 return dummy_syn_ext();
466 .struct_span_err(sp.substitute_dummy(def.span), &msg)
468 return dummy_syn_ext();
471 return dummy_syn_ext();
475 let mut valid = true;
477 // Extract the arguments:
478 let lhses = match argument_map[&MacroRulesNormalizedIdent::new(lhs_nm)] {
479 MatchedSeq(ref s) => s
482 if let MatchedTokenTree(ref tt) = *m {
483 let tt = mbe::quoted::parse(
484 TokenStream::new(vec![tt.clone()]),
493 valid &= check_lhs_nt_follows(&sess.parse_sess, &def, &tt);
496 sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs")
498 .collect::<Vec<mbe::TokenTree>>(),
499 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs"),
502 let rhses = match argument_map[&MacroRulesNormalizedIdent::new(rhs_nm)] {
503 MatchedSeq(ref s) => s
506 if let MatchedTokenTree(ref tt) = *m {
507 return mbe::quoted::parse(
508 TokenStream::new(vec![tt.clone()]),
518 sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs")
520 .collect::<Vec<mbe::TokenTree>>(),
521 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "wrong-structured rhs"),
525 valid &= check_rhs(&sess.parse_sess, rhs);
528 // don't abort iteration early, so that errors for multiple lhses can be reported
530 valid &= check_lhs_no_empty_seq(&sess.parse_sess, slice::from_ref(lhs));
533 valid &= macro_check::check_meta_variables(&sess.parse_sess, def.id, def.span, &lhses, &rhses);
535 let (transparency, transparency_error) = attr::find_transparency(&def.attrs, macro_rules);
536 match transparency_error {
537 Some(TransparencyError::UnknownTransparency(value, span)) => {
538 diag.span_err(span, &format!("unknown macro transparency: `{}`", value));
540 Some(TransparencyError::MultipleTransparencyAttrs(old_span, new_span)) => {
541 diag.span_err(vec![old_span, new_span], "multiple macro transparency attributes");
546 // Compute the spans of the macro rules for unused rule linting.
547 // To avoid warning noise, only consider the rules of this
548 // macro for the lint, if all rules are valid.
549 // Also, we are only interested in non-foreign macros.
550 let rule_spans = if valid && def.id != DUMMY_NODE_ID {
555 // If the rhs contains an invocation like compile_error!,
556 // don't consider the rule for the unused rule lint.
557 .filter(|(_idx, (_lhs, rhs))| !has_compile_error_macro(rhs))
558 // We only take the span of the lhs here,
559 // so that the spans of created warnings are smaller.
560 .map(|(idx, (lhs, _rhs))| (idx, lhs.span()))
566 // Convert the lhses into `MatcherLoc` form, which is better for doing the
567 // actual matching. Unless the matcher is invalid.
568 let lhses = if valid {
572 // Ignore the delimiters around the matcher.
574 mbe::TokenTree::Delimited(_, delimited) => {
575 mbe::macro_parser::compute_locs(&delimited.tts)
577 _ => sess.parse_sess.span_diagnostic.span_bug(def.span, "malformed macro lhs"),
585 let expander = Box::new(MacroRulesMacroExpander {
594 (mk_syn_ext(expander), rule_spans)
597 #[derive(SessionSubdiagnostic)]
598 enum ExplainDocComment {
599 #[label(expand::explain_doc_comment_inner)]
604 #[label(expand::explain_doc_comment_outer)]
611 fn annotate_doc_comment(err: &mut Diagnostic, sm: &SourceMap, span: Span) {
612 if let Ok(src) = sm.span_to_snippet(span) {
613 if src.starts_with("///") || src.starts_with("/**") {
614 err.subdiagnostic(ExplainDocComment::Outer { span });
615 } else if src.starts_with("//!") || src.starts_with("/*!") {
616 err.subdiagnostic(ExplainDocComment::Inner { span });
621 fn check_lhs_nt_follows(sess: &ParseSess, def: &ast::Item, lhs: &mbe::TokenTree) -> bool {
622 // lhs is going to be like TokenTree::Delimited(...), where the
623 // entire lhs is those tts. Or, it can be a "bare sequence", not wrapped in parens.
624 if let mbe::TokenTree::Delimited(_, delimited) = lhs {
625 check_matcher(sess, def, &delimited.tts)
627 let msg = "invalid macro matcher; matchers must be contained in balanced delimiters";
628 sess.span_diagnostic.span_err(lhs.span(), msg);
631 // we don't abort on errors on rejection, the driver will do that for us
632 // after parsing/expansion. we can report every error in every macro this way.
635 /// Checks that the lhs contains no repetition which could match an empty token
636 /// tree, because then the matcher would hang indefinitely.
637 fn check_lhs_no_empty_seq(sess: &ParseSess, tts: &[mbe::TokenTree]) -> bool {
642 | TokenTree::MetaVar(..)
643 | TokenTree::MetaVarDecl(..)
644 | TokenTree::MetaVarExpr(..) => (),
645 TokenTree::Delimited(_, ref del) => {
646 if !check_lhs_no_empty_seq(sess, &del.tts) {
650 TokenTree::Sequence(span, ref seq) => {
651 if seq.separator.is_none()
652 && seq.tts.iter().all(|seq_tt| match *seq_tt {
653 TokenTree::MetaVarDecl(_, _, Some(NonterminalKind::Vis)) => true,
654 TokenTree::Sequence(_, ref sub_seq) => {
655 sub_seq.kleene.op == mbe::KleeneOp::ZeroOrMore
656 || sub_seq.kleene.op == mbe::KleeneOp::ZeroOrOne
661 let sp = span.entire();
662 sess.span_diagnostic.span_err(sp, "repetition matches empty token tree");
665 if !check_lhs_no_empty_seq(sess, &seq.tts) {
675 fn check_rhs(sess: &ParseSess, rhs: &mbe::TokenTree) -> bool {
677 mbe::TokenTree::Delimited(..) => return true,
679 sess.span_diagnostic.span_err(rhs.span(), "macro rhs must be delimited");
685 fn check_matcher(sess: &ParseSess, def: &ast::Item, matcher: &[mbe::TokenTree]) -> bool {
686 let first_sets = FirstSets::new(matcher);
687 let empty_suffix = TokenSet::empty();
688 let err = sess.span_diagnostic.err_count();
689 check_matcher_core(sess, def, &first_sets, matcher, &empty_suffix);
690 err == sess.span_diagnostic.err_count()
693 fn has_compile_error_macro(rhs: &mbe::TokenTree) -> bool {
695 mbe::TokenTree::Delimited(_sp, d) => {
696 let has_compile_error = d.tts.array_windows::<3>().any(|[ident, bang, args]| {
697 if let mbe::TokenTree::Token(ident) = ident &&
698 let TokenKind::Ident(ident, _) = ident.kind &&
699 ident == sym::compile_error &&
700 let mbe::TokenTree::Token(bang) = bang &&
701 let TokenKind::Not = bang.kind &&
702 let mbe::TokenTree::Delimited(_, del) = args &&
703 del.delim != Delimiter::Invisible
710 if has_compile_error { true } else { d.tts.iter().any(has_compile_error_macro) }
716 // `The FirstSets` for a matcher is a mapping from subsequences in the
717 // matcher to the FIRST set for that subsequence.
719 // This mapping is partially precomputed via a backwards scan over the
720 // token trees of the matcher, which provides a mapping from each
721 // repetition sequence to its *first* set.
723 // (Hypothetically, sequences should be uniquely identifiable via their
724 // spans, though perhaps that is false, e.g., for macro-generated macros
725 // that do not try to inject artificial span information. My plan is
726 // to try to catch such cases ahead of time and not include them in
727 // the precomputed mapping.)
728 struct FirstSets<'tt> {
729 // this maps each TokenTree::Sequence `$(tt ...) SEP OP` that is uniquely identified by its
730 // span in the original matcher to the First set for the inner sequence `tt ...`.
732 // If two sequences have the same span in a matcher, then map that
733 // span to None (invalidating the mapping here and forcing the code to
735 first: FxHashMap<Span, Option<TokenSet<'tt>>>,
738 impl<'tt> FirstSets<'tt> {
739 fn new(tts: &'tt [mbe::TokenTree]) -> FirstSets<'tt> {
742 let mut sets = FirstSets { first: FxHashMap::default() };
743 build_recur(&mut sets, tts);
746 // walks backward over `tts`, returning the FIRST for `tts`
747 // and updating `sets` at the same time for all sequence
748 // substructure we find within `tts`.
749 fn build_recur<'tt>(sets: &mut FirstSets<'tt>, tts: &'tt [TokenTree]) -> TokenSet<'tt> {
750 let mut first = TokenSet::empty();
751 for tt in tts.iter().rev() {
754 | TokenTree::MetaVar(..)
755 | TokenTree::MetaVarDecl(..)
756 | TokenTree::MetaVarExpr(..) => {
757 first.replace_with(TtHandle::TtRef(tt));
759 TokenTree::Delimited(span, ref delimited) => {
760 build_recur(sets, &delimited.tts);
761 first.replace_with(TtHandle::from_token_kind(
762 token::OpenDelim(delimited.delim),
766 TokenTree::Sequence(sp, ref seq_rep) => {
767 let subfirst = build_recur(sets, &seq_rep.tts);
769 match sets.first.entry(sp.entire()) {
770 Entry::Vacant(vac) => {
771 vac.insert(Some(subfirst.clone()));
773 Entry::Occupied(mut occ) => {
774 // if there is already an entry, then a span must have collided.
775 // This should not happen with typical macro_rules macros,
776 // but syntax extensions need not maintain distinct spans,
777 // so distinct syntax trees can be assigned the same span.
778 // In such a case, the map cannot be trusted; so mark this
779 // entry as unusable.
784 // If the sequence contents can be empty, then the first
785 // token could be the separator token itself.
787 if let (Some(sep), true) = (&seq_rep.separator, subfirst.maybe_empty) {
788 first.add_one_maybe(TtHandle::from_token(sep.clone()));
791 // Reverse scan: Sequence comes before `first`.
792 if subfirst.maybe_empty
793 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrMore
794 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrOne
796 // If sequence is potentially empty, then
797 // union them (preserving first emptiness).
798 first.add_all(&TokenSet { maybe_empty: true, ..subfirst });
800 // Otherwise, sequence guaranteed
801 // non-empty; replace first.
812 // walks forward over `tts` until all potential FIRST tokens are
814 fn first(&self, tts: &'tt [mbe::TokenTree]) -> TokenSet<'tt> {
817 let mut first = TokenSet::empty();
818 for tt in tts.iter() {
819 assert!(first.maybe_empty);
822 | TokenTree::MetaVar(..)
823 | TokenTree::MetaVarDecl(..)
824 | TokenTree::MetaVarExpr(..) => {
825 first.add_one(TtHandle::TtRef(tt));
828 TokenTree::Delimited(span, ref delimited) => {
829 first.add_one(TtHandle::from_token_kind(
830 token::OpenDelim(delimited.delim),
835 TokenTree::Sequence(sp, ref seq_rep) => {
837 let subfirst = match self.first.get(&sp.entire()) {
838 Some(&Some(ref subfirst)) => subfirst,
840 subfirst_owned = self.first(&seq_rep.tts);
844 panic!("We missed a sequence during FirstSets construction");
848 // If the sequence contents can be empty, then the first
849 // token could be the separator token itself.
850 if let (Some(sep), true) = (&seq_rep.separator, subfirst.maybe_empty) {
851 first.add_one_maybe(TtHandle::from_token(sep.clone()));
854 assert!(first.maybe_empty);
855 first.add_all(subfirst);
856 if subfirst.maybe_empty
857 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrMore
858 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrOne
860 // Continue scanning for more first
861 // tokens, but also make sure we
862 // restore empty-tracking state.
863 first.maybe_empty = true;
872 // we only exit the loop if `tts` was empty or if every
873 // element of `tts` matches the empty sequence.
874 assert!(first.maybe_empty);
879 // Most `mbe::TokenTree`s are pre-existing in the matcher, but some are defined
880 // implicitly, such as opening/closing delimiters and sequence repetition ops.
881 // This type encapsulates both kinds. It implements `Clone` while avoiding the
882 // need for `mbe::TokenTree` to implement `Clone`.
885 /// This is used in most cases.
886 TtRef(&'tt mbe::TokenTree),
888 /// This is only used for implicit token trees. The `mbe::TokenTree` *must*
889 /// be `mbe::TokenTree::Token`. No other variants are allowed. We store an
890 /// `mbe::TokenTree` rather than a `Token` so that `get()` can return a
891 /// `&mbe::TokenTree`.
892 Token(mbe::TokenTree),
895 impl<'tt> TtHandle<'tt> {
896 fn from_token(tok: Token) -> Self {
897 TtHandle::Token(mbe::TokenTree::Token(tok))
900 fn from_token_kind(kind: TokenKind, span: Span) -> Self {
901 TtHandle::from_token(Token::new(kind, span))
904 // Get a reference to a token tree.
905 fn get(&'tt self) -> &'tt mbe::TokenTree {
907 TtHandle::TtRef(tt) => tt,
908 TtHandle::Token(token_tt) => &token_tt,
913 impl<'tt> PartialEq for TtHandle<'tt> {
914 fn eq(&self, other: &TtHandle<'tt>) -> bool {
915 self.get() == other.get()
919 impl<'tt> Clone for TtHandle<'tt> {
920 fn clone(&self) -> Self {
922 TtHandle::TtRef(tt) => TtHandle::TtRef(tt),
924 // This variant *must* contain a `mbe::TokenTree::Token`, and not
925 // any other variant of `mbe::TokenTree`.
926 TtHandle::Token(mbe::TokenTree::Token(tok)) => {
927 TtHandle::Token(mbe::TokenTree::Token(tok.clone()))
935 // A set of `mbe::TokenTree`s, which may include `TokenTree::Match`s
936 // (for macro-by-example syntactic variables). It also carries the
937 // `maybe_empty` flag; that is true if and only if the matcher can
938 // match an empty token sequence.
940 // The First set is computed on submatchers like `$($a:expr b),* $(c)* d`,
941 // which has corresponding FIRST = {$a:expr, c, d}.
942 // Likewise, `$($a:expr b),* $(c)+ d` has FIRST = {$a:expr, c}.
944 // (Notably, we must allow for *-op to occur zero times.)
945 #[derive(Clone, Debug)]
946 struct TokenSet<'tt> {
947 tokens: Vec<TtHandle<'tt>>,
951 impl<'tt> TokenSet<'tt> {
952 // Returns a set for the empty sequence.
954 TokenSet { tokens: Vec::new(), maybe_empty: true }
957 // Returns the set `{ tok }` for the single-token (and thus
958 // non-empty) sequence [tok].
959 fn singleton(tt: TtHandle<'tt>) -> Self {
960 TokenSet { tokens: vec![tt], maybe_empty: false }
963 // Changes self to be the set `{ tok }`.
964 // Since `tok` is always present, marks self as non-empty.
965 fn replace_with(&mut self, tt: TtHandle<'tt>) {
967 self.tokens.push(tt);
968 self.maybe_empty = false;
971 // Changes self to be the empty set `{}`; meant for use when
972 // the particular token does not matter, but we want to
973 // record that it occurs.
974 fn replace_with_irrelevant(&mut self) {
976 self.maybe_empty = false;
979 // Adds `tok` to the set for `self`, marking sequence as non-empy.
980 fn add_one(&mut self, tt: TtHandle<'tt>) {
981 if !self.tokens.contains(&tt) {
982 self.tokens.push(tt);
984 self.maybe_empty = false;
987 // Adds `tok` to the set for `self`. (Leaves `maybe_empty` flag alone.)
988 fn add_one_maybe(&mut self, tt: TtHandle<'tt>) {
989 if !self.tokens.contains(&tt) {
990 self.tokens.push(tt);
994 // Adds all elements of `other` to this.
996 // (Since this is a set, we filter out duplicates.)
998 // If `other` is potentially empty, then preserves the previous
999 // setting of the empty flag of `self`. If `other` is guaranteed
1000 // non-empty, then `self` is marked non-empty.
1001 fn add_all(&mut self, other: &Self) {
1002 for tt in &other.tokens {
1003 if !self.tokens.contains(tt) {
1004 self.tokens.push(tt.clone());
1007 if !other.maybe_empty {
1008 self.maybe_empty = false;
1013 // Checks that `matcher` is internally consistent and that it
1014 // can legally be followed by a token `N`, for all `N` in `follow`.
1015 // (If `follow` is empty, then it imposes no constraint on
1018 // Returns the set of NT tokens that could possibly come last in
1019 // `matcher`. (If `matcher` matches the empty sequence, then
1020 // `maybe_empty` will be set to true.)
1022 // Requires that `first_sets` is pre-computed for `matcher`;
1023 // see `FirstSets::new`.
1024 fn check_matcher_core<'tt>(
1027 first_sets: &FirstSets<'tt>,
1028 matcher: &'tt [mbe::TokenTree],
1029 follow: &TokenSet<'tt>,
1030 ) -> TokenSet<'tt> {
1033 let mut last = TokenSet::empty();
1035 // 2. For each token and suffix [T, SUFFIX] in M:
1036 // ensure that T can be followed by SUFFIX, and if SUFFIX may be empty,
1037 // then ensure T can also be followed by any element of FOLLOW.
1038 'each_token: for i in 0..matcher.len() {
1039 let token = &matcher[i];
1040 let suffix = &matcher[i + 1..];
1042 let build_suffix_first = || {
1043 let mut s = first_sets.first(suffix);
1050 // (we build `suffix_first` on demand below; you can tell
1051 // which cases are supposed to fall through by looking for the
1052 // initialization of this variable.)
1055 // First, update `last` so that it corresponds to the set
1056 // of NT tokens that might end the sequence `... token`.
1058 TokenTree::Token(..)
1059 | TokenTree::MetaVar(..)
1060 | TokenTree::MetaVarDecl(..)
1061 | TokenTree::MetaVarExpr(..) => {
1062 if token_can_be_followed_by_any(token) {
1063 // don't need to track tokens that work with any,
1064 last.replace_with_irrelevant();
1065 // ... and don't need to check tokens that can be
1066 // followed by anything against SUFFIX.
1067 continue 'each_token;
1069 last.replace_with(TtHandle::TtRef(token));
1070 suffix_first = build_suffix_first();
1073 TokenTree::Delimited(span, ref d) => {
1074 let my_suffix = TokenSet::singleton(TtHandle::from_token_kind(
1075 token::CloseDelim(d.delim),
1078 check_matcher_core(sess, def, first_sets, &d.tts, &my_suffix);
1079 // don't track non NT tokens
1080 last.replace_with_irrelevant();
1082 // also, we don't need to check delimited sequences
1084 continue 'each_token;
1086 TokenTree::Sequence(_, ref seq_rep) => {
1087 suffix_first = build_suffix_first();
1088 // The trick here: when we check the interior, we want
1089 // to include the separator (if any) as a potential
1090 // (but not guaranteed) element of FOLLOW. So in that
1091 // case, we make a temp copy of suffix and stuff
1092 // delimiter in there.
1094 // FIXME: Should I first scan suffix_first to see if
1095 // delimiter is already in it before I go through the
1096 // work of cloning it? But then again, this way I may
1097 // get a "tighter" span?
1099 let my_suffix = if let Some(sep) = &seq_rep.separator {
1100 new = suffix_first.clone();
1101 new.add_one_maybe(TtHandle::from_token(sep.clone()));
1107 // At this point, `suffix_first` is built, and
1108 // `my_suffix` is some TokenSet that we can use
1109 // for checking the interior of `seq_rep`.
1110 let next = check_matcher_core(sess, def, first_sets, &seq_rep.tts, my_suffix);
1111 if next.maybe_empty {
1112 last.add_all(&next);
1117 // the recursive call to check_matcher_core already ran the 'each_last
1118 // check below, so we can just keep going forward here.
1119 continue 'each_token;
1123 // (`suffix_first` guaranteed initialized once reaching here.)
1125 // Now `last` holds the complete set of NT tokens that could
1126 // end the sequence before SUFFIX. Check that every one works with `suffix`.
1127 for tt in &last.tokens {
1128 if let &TokenTree::MetaVarDecl(span, name, Some(kind)) = tt.get() {
1129 for next_token in &suffix_first.tokens {
1130 let next_token = next_token.get();
1132 // Check if the old pat is used and the next token is `|`
1133 // to warn about incompatibility with Rust 2021.
1134 // We only emit this lint if we're parsing the original
1135 // definition of this macro_rules, not while (re)parsing
1136 // the macro when compiling another crate that is using the
1137 // macro. (See #86567.)
1138 // Macros defined in the current crate have a real node id,
1139 // whereas macros from an external crate have a dummy id.
1140 if def.id != DUMMY_NODE_ID
1141 && matches!(kind, NonterminalKind::PatParam { inferred: true })
1142 && matches!(next_token, TokenTree::Token(token) if token.kind == BinOp(token::BinOpToken::Or))
1144 // It is suggestion to use pat_param, for example: $x:pat -> $x:pat_param.
1145 let suggestion = quoted_tt_to_string(&TokenTree::MetaVarDecl(
1148 Some(NonterminalKind::PatParam { inferred: false }),
1150 sess.buffer_lint_with_diagnostic(
1151 &RUST_2021_INCOMPATIBLE_OR_PATTERNS,
1154 "the meaning of the `pat` fragment specifier is changing in Rust 2021, which may affect this macro",
1155 BuiltinLintDiagnostics::OrPatternsBackCompat(span, suggestion),
1158 match is_in_follow(next_token, kind) {
1159 IsInFollow::Yes => {}
1160 IsInFollow::No(possible) => {
1161 let may_be = if last.tokens.len() == 1 && suffix_first.tokens.len() == 1
1168 let sp = next_token.span();
1169 let mut err = sess.span_diagnostic.struct_span_err(
1172 "`${name}:{frag}` {may_be} followed by `{next}`, which \
1173 is not allowed for `{frag}` fragments",
1176 next = quoted_tt_to_string(next_token),
1180 err.span_label(sp, format!("not allowed after `{}` fragments", kind));
1182 if kind == NonterminalKind::PatWithOr
1183 && sess.edition.rust_2021()
1184 && next_token.is_token(&BinOp(token::BinOpToken::Or))
1186 let suggestion = quoted_tt_to_string(&TokenTree::MetaVarDecl(
1189 Some(NonterminalKind::PatParam { inferred: false }),
1191 err.span_suggestion(
1193 "try a `pat_param` fragment specifier instead",
1195 Applicability::MaybeIncorrect,
1199 let msg = "allowed there are: ";
1204 "only {} is allowed after `{}` fragments",
1215 .collect::<Vec<_>>()
1231 fn token_can_be_followed_by_any(tok: &mbe::TokenTree) -> bool {
1232 if let mbe::TokenTree::MetaVarDecl(_, _, Some(kind)) = *tok {
1233 frag_can_be_followed_by_any(kind)
1235 // (Non NT's can always be followed by anything in matchers.)
1240 /// Returns `true` if a fragment of type `frag` can be followed by any sort of
1241 /// token. We use this (among other things) as a useful approximation
1242 /// for when `frag` can be followed by a repetition like `$(...)*` or
1243 /// `$(...)+`. In general, these can be a bit tricky to reason about,
1244 /// so we adopt a conservative position that says that any fragment
1245 /// specifier which consumes at most one token tree can be followed by
1246 /// a fragment specifier (indeed, these fragments can be followed by
1247 /// ANYTHING without fear of future compatibility hazards).
1248 fn frag_can_be_followed_by_any(kind: NonterminalKind) -> bool {
1251 NonterminalKind::Item // always terminated by `}` or `;`
1252 | NonterminalKind::Block // exactly one token tree
1253 | NonterminalKind::Ident // exactly one token tree
1254 | NonterminalKind::Literal // exactly one token tree
1255 | NonterminalKind::Meta // exactly one token tree
1256 | NonterminalKind::Lifetime // exactly one token tree
1257 | NonterminalKind::TT // exactly one token tree
1263 No(&'static [&'static str]),
1266 /// Returns `true` if `frag` can legally be followed by the token `tok`. For
1267 /// fragments that can consume an unbounded number of tokens, `tok`
1268 /// must be within a well-defined follow set. This is intended to
1269 /// guarantee future compatibility: for example, without this rule, if
1270 /// we expanded `expr` to include a new binary operator, we might
1271 /// break macros that were relying on that binary operator as a
1273 // when changing this do not forget to update doc/book/macros.md!
1274 fn is_in_follow(tok: &mbe::TokenTree, kind: NonterminalKind) -> IsInFollow {
1277 if let TokenTree::Token(Token { kind: token::CloseDelim(_), .. }) = *tok {
1278 // closing a token tree can never be matched by any fragment;
1279 // iow, we always require that `(` and `)` match, etc.
1283 NonterminalKind::Item => {
1284 // since items *must* be followed by either a `;` or a `}`, we can
1285 // accept anything after them
1288 NonterminalKind::Block => {
1289 // anything can follow block, the braces provide an easy boundary to
1293 NonterminalKind::Stmt | NonterminalKind::Expr => {
1294 const TOKENS: &[&str] = &["`=>`", "`,`", "`;`"];
1296 TokenTree::Token(token) => match token.kind {
1297 FatArrow | Comma | Semi => IsInFollow::Yes,
1298 _ => IsInFollow::No(TOKENS),
1300 _ => IsInFollow::No(TOKENS),
1303 NonterminalKind::PatParam { .. } => {
1304 const TOKENS: &[&str] = &["`=>`", "`,`", "`=`", "`|`", "`if`", "`in`"];
1306 TokenTree::Token(token) => match token.kind {
1307 FatArrow | Comma | Eq | BinOp(token::Or) => IsInFollow::Yes,
1308 Ident(name, false) if name == kw::If || name == kw::In => IsInFollow::Yes,
1309 _ => IsInFollow::No(TOKENS),
1311 _ => IsInFollow::No(TOKENS),
1314 NonterminalKind::PatWithOr { .. } => {
1315 const TOKENS: &[&str] = &["`=>`", "`,`", "`=`", "`if`", "`in`"];
1317 TokenTree::Token(token) => match token.kind {
1318 FatArrow | Comma | Eq => IsInFollow::Yes,
1319 Ident(name, false) if name == kw::If || name == kw::In => IsInFollow::Yes,
1320 _ => IsInFollow::No(TOKENS),
1322 _ => IsInFollow::No(TOKENS),
1325 NonterminalKind::Path | NonterminalKind::Ty => {
1326 const TOKENS: &[&str] = &[
1327 "`{`", "`[`", "`=>`", "`,`", "`>`", "`=`", "`:`", "`;`", "`|`", "`as`",
1331 TokenTree::Token(token) => match token.kind {
1332 OpenDelim(Delimiter::Brace)
1333 | OpenDelim(Delimiter::Bracket)
1341 | BinOp(token::Or) => IsInFollow::Yes,
1342 Ident(name, false) if name == kw::As || name == kw::Where => {
1345 _ => IsInFollow::No(TOKENS),
1347 TokenTree::MetaVarDecl(_, _, Some(NonterminalKind::Block)) => IsInFollow::Yes,
1348 _ => IsInFollow::No(TOKENS),
1351 NonterminalKind::Ident | NonterminalKind::Lifetime => {
1352 // being a single token, idents and lifetimes are harmless
1355 NonterminalKind::Literal => {
1356 // literals may be of a single token, or two tokens (negative numbers)
1359 NonterminalKind::Meta | NonterminalKind::TT => {
1360 // being either a single token or a delimited sequence, tt is
1364 NonterminalKind::Vis => {
1365 // Explicitly disallow `priv`, on the off chance it comes back.
1366 const TOKENS: &[&str] = &["`,`", "an ident", "a type"];
1368 TokenTree::Token(token) => match token.kind {
1369 Comma => IsInFollow::Yes,
1370 Ident(name, is_raw) if is_raw || name != kw::Priv => IsInFollow::Yes,
1372 if token.can_begin_type() {
1375 IsInFollow::No(TOKENS)
1379 TokenTree::MetaVarDecl(
1382 Some(NonterminalKind::Ident | NonterminalKind::Ty | NonterminalKind::Path),
1383 ) => IsInFollow::Yes,
1384 _ => IsInFollow::No(TOKENS),
1391 fn quoted_tt_to_string(tt: &mbe::TokenTree) -> String {
1393 mbe::TokenTree::Token(ref token) => pprust::token_to_string(&token).into(),
1394 mbe::TokenTree::MetaVar(_, name) => format!("${}", name),
1395 mbe::TokenTree::MetaVarDecl(_, name, Some(kind)) => format!("${}:{}", name, kind),
1396 mbe::TokenTree::MetaVarDecl(_, name, None) => format!("${}:", name),
1399 "unexpected mbe::TokenTree::{Sequence or Delimited} \
1400 in follow set checker"
1405 fn parser_from_cx(sess: &ParseSess, tts: TokenStream) -> Parser<'_> {
1406 Parser::new(sess, tts, true, rustc_parse::MACRO_ARGUMENTS)
1409 /// Generates an appropriate parsing failure message. For EOF, this is "unexpected end...". For
1410 /// other tokens, this is "unexpected token...".
1411 fn parse_failure_msg(tok: &Token) -> String {
1413 token::Eof => "unexpected end of macro invocation".to_string(),
1414 _ => format!("no rules expected the token `{}`", pprust::token_to_string(tok),),