1 use crate::base::{DummyResult, ExpansionData, 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, Failure, Success};
8 use crate::mbe::macro_parser::{MatchedNonterminal, MatchedSeq};
9 use crate::mbe::transcribe::transcribe;
11 use rustc_ast_pretty::pprust;
12 use rustc_attr::{self as attr, TransparencyError};
13 use rustc_data_structures::fx::FxHashMap;
14 use rustc_data_structures::sync::Lrc;
15 use rustc_errors::{Applicability, DiagnosticBuilder, FatalError};
16 use rustc_feature::Features;
17 use rustc_parse::parser::Parser;
18 use rustc_parse::Directory;
19 use rustc_session::parse::ParseSess;
20 use rustc_span::edition::Edition;
21 use rustc_span::hygiene::Transparency;
22 use rustc_span::symbol::{kw, sym, Symbol};
25 use syntax::token::{self, NtTT, Token, TokenKind::*};
26 use syntax::tokenstream::{DelimSpan, TokenStream};
30 use std::collections::hash_map::Entry;
31 use std::{mem, slice};
33 const VALID_FRAGMENT_NAMES_MSG: &str = "valid fragment specifiers are \
34 `ident`, `block`, `stmt`, `expr`, `pat`, `ty`, `lifetime`, \
35 `literal`, `path`, `meta`, `tt`, `item` and `vis`";
37 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
43 macro_ident: ast::Ident,
47 crate fn annotate_err_with_kind(
48 err: &mut DiagnosticBuilder<'_>,
49 kind: AstFragmentKind,
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 /// Instead of e.g. `vec![a, b, c]` in a pattern context, suggest `[a, b, c]`.
64 fn suggest_slice_pat(e: &mut DiagnosticBuilder<'_>, site_span: Span, parser: &Parser<'_>) {
65 let mut suggestion = None;
66 if let Ok(code) = parser.sess.source_map().span_to_snippet(site_span) {
67 if let Some(bang) = code.find('!') {
68 suggestion = Some(code[bang + 1..].to_string());
71 if let Some(suggestion) = suggestion {
74 "use a slice pattern here instead",
76 Applicability::MachineApplicable,
79 e.span_label(site_span, "use a slice pattern here instead");
82 "for more information, see https://doc.rust-lang.org/edition-guide/\
83 rust-2018/slice-patterns.html",
87 impl<'a> ParserAnyMacro<'a> {
88 crate fn make(mut self: Box<ParserAnyMacro<'a>>, kind: AstFragmentKind) -> AstFragment {
89 let ParserAnyMacro { site_span, macro_ident, ref mut parser, arm_span } = *self;
90 let fragment = panictry!(parse_ast_fragment(parser, kind).map_err(|mut e| {
91 if parser.token == token::Eof && e.message().ends_with(", found `<eof>`") {
92 if !e.span.is_dummy() {
93 // early end of macro arm (#52866)
94 e.replace_span_with(parser.sess.source_map().next_point(parser.token.span));
96 let msg = &e.message[0];
99 "macro expansion ends with an incomplete expression: {}",
100 msg.0.replace(", found `<eof>`", ""),
105 if e.span.is_dummy() {
106 // Get around lack of span in error (#30128)
107 e.replace_span_with(site_span);
108 if parser.sess.source_map().span_to_filename(arm_span).is_real() {
109 e.span_label(arm_span, "in this macro arm");
111 } else if !parser.sess.source_map().span_to_filename(parser.token.span).is_real() {
112 e.span_label(site_span, "in this macro invocation");
115 AstFragmentKind::Pat if macro_ident.name == sym::vec => {
116 suggest_slice_pat(&mut e, site_span, parser);
118 _ => annotate_err_with_kind(&mut e, kind, site_span),
123 // We allow semicolons at the end of expressions -- e.g., the semicolon in
124 // `macro_rules! m { () => { panic!(); } }` isn't parsed by `.parse_expr()`,
125 // but `m!()` is allowed in expression positions (cf. issue #34706).
126 if kind == AstFragmentKind::Expr && parser.token == token::Semi {
130 // Make sure we don't have any tokens left to parse so we don't silently drop anything.
131 let path = ast::Path::from_ident(macro_ident.with_span_pos(site_span));
132 ensure_complete_parse(parser, &path, kind.name(), site_span);
137 struct MacroRulesMacroExpander {
140 transparency: Transparency,
141 lhses: Vec<mbe::TokenTree>,
142 rhses: Vec<mbe::TokenTree>,
146 impl TTMacroExpander for MacroRulesMacroExpander {
149 cx: &'cx mut ExtCtxt<'_>,
152 ) -> Box<dyn MacResult + 'cx> {
154 return DummyResult::any(sp);
169 fn trace_macros_note(cx_expansions: &mut FxHashMap<Span, Vec<String>>, sp: Span, message: String) {
170 let sp = sp.macro_backtrace().last().map(|trace| trace.call_site).unwrap_or(sp);
171 cx_expansions.entry(sp).or_default().push(message);
174 /// Given `lhses` and `rhses`, this is the new macro we create
175 fn generic_extension<'cx>(
176 cx: &'cx mut ExtCtxt<'_>,
180 transparency: Transparency,
182 lhses: &[mbe::TokenTree],
183 rhses: &[mbe::TokenTree],
184 ) -> Box<dyn MacResult + 'cx> {
185 if cx.trace_macros() {
186 let msg = format!("expanding `{}! {{ {} }}`", name, pprust::tts_to_string(arg.clone()));
187 trace_macros_note(&mut cx.expansions, sp, msg);
190 // Which arm's failure should we report? (the one furthest along)
191 let mut best_failure: Option<(Token, &str)> = None;
193 // We create a base parser that can be used for the "black box" parts.
194 // Every iteration needs a fresh copy of that base parser. However, the
195 // parser is not mutated on many of the iterations, particularly when
196 // dealing with macros like this:
198 // macro_rules! foo {
202 // // ... etc. (maybe hundreds more)
205 // as seen in the `html5ever` benchmark. We use a `Cow` so that the base
206 // parser is only cloned when necessary (upon mutation). Furthermore, we
207 // reinitialize the `Cow` with the base parser at the start of every
208 // iteration, so that any mutated parsers are not reused. This is all quite
209 // hacky, but speeds up the `html5ever` benchmark significantly. (Issue
210 // 68836 suggests a more comprehensive but more complex change to deal with
212 let base_parser = base_parser_from_cx(&cx.current_expansion, &cx.parse_sess, arg.clone());
214 for (i, lhs) in lhses.iter().enumerate() {
215 let mut parser = Cow::Borrowed(&base_parser);
217 // try each arm's matchers
218 let lhs_tt = match *lhs {
219 mbe::TokenTree::Delimited(_, ref delim) => &delim.tts[..],
220 _ => cx.span_bug(sp, "malformed macro lhs"),
223 // Take a snapshot of the state of pre-expansion gating at this point.
224 // This is used so that if a matcher is not `Success(..)`ful,
225 // then the spans which became gated when parsing the unsuccessful matcher
226 // are not recorded. On the first `Success(..)`ful matcher, the spans are merged.
227 let mut gated_spans_snaphot = mem::take(&mut *cx.parse_sess.gated_spans.spans.borrow_mut());
229 match parse_tt(&mut parser, lhs_tt) {
230 Success(named_matches) => {
231 // The matcher was `Success(..)`ful.
232 // Merge the gated spans from parsing the matcher with the pre-existing ones.
233 cx.parse_sess.gated_spans.merge(gated_spans_snaphot);
235 let rhs = match rhses[i] {
237 mbe::TokenTree::Delimited(_, ref delimed) => delimed.tts.clone(),
238 _ => cx.span_bug(sp, "malformed macro rhs"),
240 let arm_span = rhses[i].span();
242 let rhs_spans = rhs.iter().map(|t| t.span()).collect::<Vec<_>>();
243 // rhs has holes ( `$id` and `$(...)` that need filled)
244 let mut tts = transcribe(cx, &named_matches, rhs, transparency);
246 // Replace all the tokens for the corresponding positions in the macro, to maintain
247 // proper positions in error reporting, while maintaining the macro_backtrace.
248 if rhs_spans.len() == tts.len() {
249 tts = tts.map_enumerated(|i, mut tt| {
250 let mut sp = rhs_spans[i];
251 sp = sp.with_ctxt(tt.span().ctxt());
257 if cx.trace_macros() {
258 let msg = format!("to `{}`", pprust::tts_to_string(tts.clone()));
259 trace_macros_note(&mut cx.expansions, sp, msg);
262 let directory = Directory {
263 path: cx.current_expansion.module.directory.clone(),
264 ownership: cx.current_expansion.directory_ownership,
266 let mut p = Parser::new(cx.parse_sess(), tts, Some(directory), true, false, None);
268 cx.current_expansion.module.mod_path.last().map(|id| id.to_string());
269 p.last_type_ascription = cx.current_expansion.prior_type_ascription;
271 p.process_potential_macro_variable();
272 // Let the context choose how to interpret the result.
273 // Weird, but useful for X-macros.
274 return Box::new(ParserAnyMacro {
277 // Pass along the original expansion site and the name of the macro
278 // so we can print a useful error message if the parse of the expanded
279 // macro leaves unparsed tokens.
285 Failure(token, msg) => match best_failure {
286 Some((ref best_token, _)) if best_token.span.lo() >= token.span.lo() => {}
287 _ => best_failure = Some((token, msg)),
289 Error(err_sp, ref msg) => cx.span_fatal(err_sp.substitute_dummy(sp), &msg[..]),
292 // The matcher was not `Success(..)`ful.
293 // Restore to the state before snapshotting and maybe try again.
294 mem::swap(&mut gated_spans_snaphot, &mut cx.parse_sess.gated_spans.spans.borrow_mut());
298 let (token, label) = best_failure.expect("ran no matchers");
299 let span = token.span.substitute_dummy(sp);
300 let mut err = cx.struct_span_err(span, &parse_failure_msg(&token));
301 err.span_label(span, label);
302 if !def_span.is_dummy() && cx.source_map().span_to_filename(def_span).is_real() {
303 err.span_label(cx.source_map().def_span(def_span), "when calling this macro");
306 // Check whether there's a missing comma in this macro call, like `println!("{}" a);`
307 if let Some((arg, comma_span)) = arg.add_comma() {
309 // try each arm's matchers
310 let lhs_tt = match *lhs {
311 mbe::TokenTree::Delimited(_, ref delim) => &delim.tts[..],
315 base_parser_from_cx(&cx.current_expansion, &cx.parse_sess, arg.clone());
316 match parse_tt(&mut Cow::Borrowed(&base_parser), lhs_tt) {
318 if comma_span.is_dummy() {
319 err.note("you might be missing a comma");
321 err.span_suggestion_short(
323 "missing comma here",
325 Applicability::MachineApplicable,
334 cx.trace_macros_diag();
338 // Note that macro-by-example's input is also matched against a token tree:
339 // $( $lhs:tt => $rhs:tt );+
341 // Holy self-referential!
343 /// Converts a macro item into a syntax extension.
344 pub fn compile_declarative_macro(
349 ) -> SyntaxExtension {
350 let diag = &sess.span_diagnostic;
351 let lhs_nm = ast::Ident::new(sym::lhs, def.span);
352 let rhs_nm = ast::Ident::new(sym::rhs, def.span);
353 let tt_spec = ast::Ident::new(sym::tt, def.span);
355 // Parse the macro_rules! invocation
356 let (is_legacy, body) = match &def.kind {
357 ast::ItemKind::MacroDef(macro_def) => (macro_def.legacy, macro_def.body.inner_tokens()),
361 // The pattern that macro_rules matches.
362 // The grammar for macro_rules! is:
363 // $( $lhs:tt => $rhs:tt );+
364 // ...quasiquoting this would be nice.
365 // These spans won't matter, anyways
366 let argument_gram = vec![
367 mbe::TokenTree::Sequence(
369 Lrc::new(mbe::SequenceRepetition {
371 mbe::TokenTree::MetaVarDecl(def.span, lhs_nm, tt_spec),
372 mbe::TokenTree::token(token::FatArrow, def.span),
373 mbe::TokenTree::MetaVarDecl(def.span, rhs_nm, tt_spec),
375 separator: Some(Token::new(
376 if is_legacy { token::Semi } else { token::Comma },
379 kleene: mbe::KleeneToken::new(mbe::KleeneOp::OneOrMore, def.span),
383 // to phase into semicolon-termination instead of semicolon-separation
384 mbe::TokenTree::Sequence(
386 Lrc::new(mbe::SequenceRepetition {
387 tts: vec![mbe::TokenTree::token(
388 if is_legacy { token::Semi } else { token::Comma },
392 kleene: mbe::KleeneToken::new(mbe::KleeneOp::ZeroOrMore, def.span),
398 let base_parser = Parser::new(sess, body, None, true, true, rustc_parse::MACRO_ARGUMENTS);
399 let argument_map = match parse_tt(&mut Cow::Borrowed(&base_parser), &argument_gram) {
401 Failure(token, msg) => {
402 let s = parse_failure_msg(&token);
403 let sp = token.span.substitute_dummy(def.span);
404 let mut err = sess.span_diagnostic.struct_span_fatal(sp, &s);
405 err.span_label(sp, msg);
410 sess.span_diagnostic.span_fatal(sp.substitute_dummy(def.span), &s).raise();
414 let mut valid = true;
416 // Extract the arguments:
417 let lhses = match argument_map[&lhs_nm] {
418 MatchedSeq(ref s) => s
421 if let MatchedNonterminal(ref nt) = *m {
422 if let NtTT(ref tt) = **nt {
423 let tt = mbe::quoted::parse(tt.clone().into(), true, sess).pop().unwrap();
424 valid &= check_lhs_nt_follows(sess, features, &def.attrs, &tt);
428 sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs")
430 .collect::<Vec<mbe::TokenTree>>(),
431 _ => sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs"),
434 let rhses = match argument_map[&rhs_nm] {
435 MatchedSeq(ref s) => s
438 if let MatchedNonterminal(ref nt) = *m {
439 if let NtTT(ref tt) = **nt {
440 return mbe::quoted::parse(tt.clone().into(), false, sess).pop().unwrap();
443 sess.span_diagnostic.span_bug(def.span, "wrong-structured lhs")
445 .collect::<Vec<mbe::TokenTree>>(),
446 _ => sess.span_diagnostic.span_bug(def.span, "wrong-structured rhs"),
450 valid &= check_rhs(sess, rhs);
453 // don't abort iteration early, so that errors for multiple lhses can be reported
455 valid &= check_lhs_no_empty_seq(sess, slice::from_ref(lhs));
458 // We use CRATE_NODE_ID instead of `def.id` otherwise we may emit buffered lints for a node id
459 // that is not lint-checked and trigger the "failed to process buffered lint here" bug.
460 valid &= macro_check::check_meta_variables(sess, ast::CRATE_NODE_ID, def.span, &lhses, &rhses);
462 let (transparency, transparency_error) = attr::find_transparency(&def.attrs, is_legacy);
463 match transparency_error {
464 Some(TransparencyError::UnknownTransparency(value, span)) => {
465 diag.span_err(span, &format!("unknown macro transparency: `{}`", value))
467 Some(TransparencyError::MultipleTransparencyAttrs(old_span, new_span)) => {
468 diag.span_err(vec![old_span, new_span], "multiple macro transparency attributes")
473 let expander: Box<_> = Box::new(MacroRulesMacroExpander {
482 SyntaxExtension::new(
484 SyntaxExtensionKind::LegacyBang(expander),
493 fn check_lhs_nt_follows(
496 attrs: &[ast::Attribute],
497 lhs: &mbe::TokenTree,
499 // lhs is going to be like TokenTree::Delimited(...), where the
500 // entire lhs is those tts. Or, it can be a "bare sequence", not wrapped in parens.
501 if let mbe::TokenTree::Delimited(_, ref tts) = *lhs {
502 check_matcher(sess, features, attrs, &tts.tts)
504 let msg = "invalid macro matcher; matchers must be contained in balanced delimiters";
505 sess.span_diagnostic.span_err(lhs.span(), msg);
508 // we don't abort on errors on rejection, the driver will do that for us
509 // after parsing/expansion. we can report every error in every macro this way.
512 /// Checks that the lhs contains no repetition which could match an empty token
513 /// tree, because then the matcher would hang indefinitely.
514 fn check_lhs_no_empty_seq(sess: &ParseSess, tts: &[mbe::TokenTree]) -> bool {
518 TokenTree::Token(..) | TokenTree::MetaVar(..) | TokenTree::MetaVarDecl(..) => (),
519 TokenTree::Delimited(_, ref del) => {
520 if !check_lhs_no_empty_seq(sess, &del.tts) {
524 TokenTree::Sequence(span, ref seq) => {
525 if seq.separator.is_none()
526 && seq.tts.iter().all(|seq_tt| match *seq_tt {
527 TokenTree::MetaVarDecl(_, _, id) => id.name == sym::vis,
528 TokenTree::Sequence(_, ref sub_seq) => {
529 sub_seq.kleene.op == mbe::KleeneOp::ZeroOrMore
530 || sub_seq.kleene.op == mbe::KleeneOp::ZeroOrOne
535 let sp = span.entire();
536 sess.span_diagnostic.span_err(sp, "repetition matches empty token tree");
539 if !check_lhs_no_empty_seq(sess, &seq.tts) {
549 fn check_rhs(sess: &ParseSess, rhs: &mbe::TokenTree) -> bool {
551 mbe::TokenTree::Delimited(..) => return true,
552 _ => sess.span_diagnostic.span_err(rhs.span(), "macro rhs must be delimited"),
560 attrs: &[ast::Attribute],
561 matcher: &[mbe::TokenTree],
563 let first_sets = FirstSets::new(matcher);
564 let empty_suffix = TokenSet::empty();
565 let err = sess.span_diagnostic.err_count();
566 check_matcher_core(sess, features, attrs, &first_sets, matcher, &empty_suffix);
567 err == sess.span_diagnostic.err_count()
570 // `The FirstSets` for a matcher is a mapping from subsequences in the
571 // matcher to the FIRST set for that subsequence.
573 // This mapping is partially precomputed via a backwards scan over the
574 // token trees of the matcher, which provides a mapping from each
575 // repetition sequence to its *first* set.
577 // (Hypothetically, sequences should be uniquely identifiable via their
578 // spans, though perhaps that is false, e.g., for macro-generated macros
579 // that do not try to inject artificial span information. My plan is
580 // to try to catch such cases ahead of time and not include them in
581 // the precomputed mapping.)
583 // this maps each TokenTree::Sequence `$(tt ...) SEP OP` that is uniquely identified by its
584 // span in the original matcher to the First set for the inner sequence `tt ...`.
586 // If two sequences have the same span in a matcher, then map that
587 // span to None (invalidating the mapping here and forcing the code to
589 first: FxHashMap<Span, Option<TokenSet>>,
593 fn new(tts: &[mbe::TokenTree]) -> FirstSets {
596 let mut sets = FirstSets { first: FxHashMap::default() };
597 build_recur(&mut sets, tts);
600 // walks backward over `tts`, returning the FIRST for `tts`
601 // and updating `sets` at the same time for all sequence
602 // substructure we find within `tts`.
603 fn build_recur(sets: &mut FirstSets, tts: &[TokenTree]) -> TokenSet {
604 let mut first = TokenSet::empty();
605 for tt in tts.iter().rev() {
607 TokenTree::Token(..) | TokenTree::MetaVar(..) | TokenTree::MetaVarDecl(..) => {
608 first.replace_with(tt.clone());
610 TokenTree::Delimited(span, ref delimited) => {
611 build_recur(sets, &delimited.tts[..]);
612 first.replace_with(delimited.open_tt(span));
614 TokenTree::Sequence(sp, ref seq_rep) => {
615 let subfirst = build_recur(sets, &seq_rep.tts[..]);
617 match sets.first.entry(sp.entire()) {
618 Entry::Vacant(vac) => {
619 vac.insert(Some(subfirst.clone()));
621 Entry::Occupied(mut occ) => {
622 // if there is already an entry, then a span must have collided.
623 // This should not happen with typical macro_rules macros,
624 // but syntax extensions need not maintain distinct spans,
625 // so distinct syntax trees can be assigned the same span.
626 // In such a case, the map cannot be trusted; so mark this
627 // entry as unusable.
632 // If the sequence contents can be empty, then the first
633 // token could be the separator token itself.
635 if let (Some(sep), true) = (&seq_rep.separator, subfirst.maybe_empty) {
636 first.add_one_maybe(TokenTree::Token(sep.clone()));
639 // Reverse scan: Sequence comes before `first`.
640 if subfirst.maybe_empty
641 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrMore
642 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrOne
644 // If sequence is potentially empty, then
645 // union them (preserving first emptiness).
646 first.add_all(&TokenSet { maybe_empty: true, ..subfirst });
648 // Otherwise, sequence guaranteed
649 // non-empty; replace first.
660 // walks forward over `tts` until all potential FIRST tokens are
662 fn first(&self, tts: &[mbe::TokenTree]) -> TokenSet {
665 let mut first = TokenSet::empty();
666 for tt in tts.iter() {
667 assert!(first.maybe_empty);
669 TokenTree::Token(..) | TokenTree::MetaVar(..) | TokenTree::MetaVarDecl(..) => {
670 first.add_one(tt.clone());
673 TokenTree::Delimited(span, ref delimited) => {
674 first.add_one(delimited.open_tt(span));
677 TokenTree::Sequence(sp, ref seq_rep) => {
679 let subfirst = match self.first.get(&sp.entire()) {
680 Some(&Some(ref subfirst)) => subfirst,
682 subfirst_owned = self.first(&seq_rep.tts[..]);
686 panic!("We missed a sequence during FirstSets construction");
690 // If the sequence contents can be empty, then the first
691 // 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 assert!(first.maybe_empty);
697 first.add_all(subfirst);
698 if subfirst.maybe_empty
699 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrMore
700 || seq_rep.kleene.op == mbe::KleeneOp::ZeroOrOne
702 // Continue scanning for more first
703 // tokens, but also make sure we
704 // restore empty-tracking state.
705 first.maybe_empty = true;
714 // we only exit the loop if `tts` was empty or if every
715 // element of `tts` matches the empty sequence.
716 assert!(first.maybe_empty);
721 // A set of `mbe::TokenTree`s, which may include `TokenTree::Match`s
722 // (for macro-by-example syntactic variables). It also carries the
723 // `maybe_empty` flag; that is true if and only if the matcher can
724 // match an empty token sequence.
726 // The First set is computed on submatchers like `$($a:expr b),* $(c)* d`,
727 // which has corresponding FIRST = {$a:expr, c, d}.
728 // Likewise, `$($a:expr b),* $(c)+ d` has FIRST = {$a:expr, c}.
730 // (Notably, we must allow for *-op to occur zero times.)
731 #[derive(Clone, Debug)]
733 tokens: Vec<mbe::TokenTree>,
738 // Returns a set for the empty sequence.
740 TokenSet { tokens: Vec::new(), maybe_empty: true }
743 // Returns the set `{ tok }` for the single-token (and thus
744 // non-empty) sequence [tok].
745 fn singleton(tok: mbe::TokenTree) -> Self {
746 TokenSet { tokens: vec![tok], maybe_empty: false }
749 // Changes self to be the set `{ tok }`.
750 // Since `tok` is always present, marks self as non-empty.
751 fn replace_with(&mut self, tok: mbe::TokenTree) {
753 self.tokens.push(tok);
754 self.maybe_empty = false;
757 // Changes self to be the empty set `{}`; meant for use when
758 // the particular token does not matter, but we want to
759 // record that it occurs.
760 fn replace_with_irrelevant(&mut self) {
762 self.maybe_empty = false;
765 // Adds `tok` to the set for `self`, marking sequence as non-empy.
766 fn add_one(&mut self, tok: mbe::TokenTree) {
767 if !self.tokens.contains(&tok) {
768 self.tokens.push(tok);
770 self.maybe_empty = false;
773 // Adds `tok` to the set for `self`. (Leaves `maybe_empty` flag alone.)
774 fn add_one_maybe(&mut self, tok: mbe::TokenTree) {
775 if !self.tokens.contains(&tok) {
776 self.tokens.push(tok);
780 // Adds all elements of `other` to this.
782 // (Since this is a set, we filter out duplicates.)
784 // If `other` is potentially empty, then preserves the previous
785 // setting of the empty flag of `self`. If `other` is guaranteed
786 // non-empty, then `self` is marked non-empty.
787 fn add_all(&mut self, other: &Self) {
788 for tok in &other.tokens {
789 if !self.tokens.contains(tok) {
790 self.tokens.push(tok.clone());
793 if !other.maybe_empty {
794 self.maybe_empty = false;
799 // Checks that `matcher` is internally consistent and that it
800 // can legally be followed by a token `N`, for all `N` in `follow`.
801 // (If `follow` is empty, then it imposes no constraint on
804 // Returns the set of NT tokens that could possibly come last in
805 // `matcher`. (If `matcher` matches the empty sequence, then
806 // `maybe_empty` will be set to true.)
808 // Requires that `first_sets` is pre-computed for `matcher`;
809 // see `FirstSets::new`.
810 fn check_matcher_core(
813 attrs: &[ast::Attribute],
814 first_sets: &FirstSets,
815 matcher: &[mbe::TokenTree],
820 let mut last = TokenSet::empty();
822 // 2. For each token and suffix [T, SUFFIX] in M:
823 // ensure that T can be followed by SUFFIX, and if SUFFIX may be empty,
824 // then ensure T can also be followed by any element of FOLLOW.
825 'each_token: for i in 0..matcher.len() {
826 let token = &matcher[i];
827 let suffix = &matcher[i + 1..];
829 let build_suffix_first = || {
830 let mut s = first_sets.first(suffix);
837 // (we build `suffix_first` on demand below; you can tell
838 // which cases are supposed to fall through by looking for the
839 // initialization of this variable.)
842 // First, update `last` so that it corresponds to the set
843 // of NT tokens that might end the sequence `... token`.
845 TokenTree::Token(..) | TokenTree::MetaVar(..) | TokenTree::MetaVarDecl(..) => {
846 let can_be_followed_by_any;
847 if let Err(bad_frag) = has_legal_fragment_specifier(sess, features, attrs, token) {
848 let msg = format!("invalid fragment specifier `{}`", bad_frag);
850 .struct_span_err(token.span(), &msg)
851 .help(VALID_FRAGMENT_NAMES_MSG)
853 // (This eliminates false positives and duplicates
854 // from error messages.)
855 can_be_followed_by_any = true;
857 can_be_followed_by_any = token_can_be_followed_by_any(token);
860 if can_be_followed_by_any {
861 // don't need to track tokens that work with any,
862 last.replace_with_irrelevant();
863 // ... and don't need to check tokens that can be
864 // followed by anything against SUFFIX.
865 continue 'each_token;
867 last.replace_with(token.clone());
868 suffix_first = build_suffix_first();
871 TokenTree::Delimited(span, ref d) => {
872 let my_suffix = TokenSet::singleton(d.close_tt(span));
873 check_matcher_core(sess, features, attrs, first_sets, &d.tts, &my_suffix);
874 // don't track non NT tokens
875 last.replace_with_irrelevant();
877 // also, we don't need to check delimited sequences
879 continue 'each_token;
881 TokenTree::Sequence(_, ref seq_rep) => {
882 suffix_first = build_suffix_first();
883 // The trick here: when we check the interior, we want
884 // to include the separator (if any) as a potential
885 // (but not guaranteed) element of FOLLOW. So in that
886 // case, we make a temp copy of suffix and stuff
887 // delimiter in there.
889 // FIXME: Should I first scan suffix_first to see if
890 // delimiter is already in it before I go through the
891 // work of cloning it? But then again, this way I may
892 // get a "tighter" span?
894 let my_suffix = if let Some(sep) = &seq_rep.separator {
895 new = suffix_first.clone();
896 new.add_one_maybe(TokenTree::Token(sep.clone()));
902 // At this point, `suffix_first` is built, and
903 // `my_suffix` is some TokenSet that we can use
904 // for checking the interior of `seq_rep`.
906 check_matcher_core(sess, features, attrs, first_sets, &seq_rep.tts, my_suffix);
907 if next.maybe_empty {
913 // the recursive call to check_matcher_core already ran the 'each_last
914 // check below, so we can just keep going forward here.
915 continue 'each_token;
919 // (`suffix_first` guaranteed initialized once reaching here.)
921 // Now `last` holds the complete set of NT tokens that could
922 // end the sequence before SUFFIX. Check that every one works with `suffix`.
923 'each_last: for token in &last.tokens {
924 if let TokenTree::MetaVarDecl(_, name, frag_spec) = *token {
925 for next_token in &suffix_first.tokens {
926 match is_in_follow(next_token, frag_spec.name) {
927 IsInFollow::Invalid(msg, help) => {
929 .struct_span_err(next_token.span(), &msg)
932 // don't bother reporting every source of
933 // conflict for a particular element of `last`.
936 IsInFollow::Yes => {}
937 IsInFollow::No(possible) => {
938 let may_be = if last.tokens.len() == 1 && suffix_first.tokens.len() == 1
945 let sp = next_token.span();
946 let mut err = sess.span_diagnostic.struct_span_err(
949 "`${name}:{frag}` {may_be} followed by `{next}`, which \
950 is not allowed for `{frag}` fragments",
953 next = quoted_tt_to_string(next_token),
959 format!("not allowed after `{}` fragments", frag_spec),
961 let msg = "allowed there are: ";
966 "only {} is allowed after `{}` fragments",
993 fn token_can_be_followed_by_any(tok: &mbe::TokenTree) -> bool {
994 if let mbe::TokenTree::MetaVarDecl(_, _, frag_spec) = *tok {
995 frag_can_be_followed_by_any(frag_spec.name)
997 // (Non NT's can always be followed by anthing in matchers.)
1002 /// Returns `true` if a fragment of type `frag` can be followed by any sort of
1003 /// token. We use this (among other things) as a useful approximation
1004 /// for when `frag` can be followed by a repetition like `$(...)*` or
1005 /// `$(...)+`. In general, these can be a bit tricky to reason about,
1006 /// so we adopt a conservative position that says that any fragment
1007 /// specifier which consumes at most one token tree can be followed by
1008 /// a fragment specifier (indeed, these fragments can be followed by
1009 /// ANYTHING without fear of future compatibility hazards).
1010 fn frag_can_be_followed_by_any(frag: Symbol) -> bool {
1012 sym::item | // always terminated by `}` or `;`
1013 sym::block | // exactly one token tree
1014 sym::ident | // exactly one token tree
1015 sym::literal | // exactly one token tree
1016 sym::meta | // exactly one token tree
1017 sym::lifetime | // exactly one token tree
1018 sym::tt => // exactly one token tree
1028 No(&'static [&'static str]),
1029 Invalid(String, &'static str),
1032 /// Returns `true` if `frag` can legally be followed by the token `tok`. For
1033 /// fragments that can consume an unbounded number of tokens, `tok`
1034 /// must be within a well-defined follow set. This is intended to
1035 /// guarantee future compatibility: for example, without this rule, if
1036 /// we expanded `expr` to include a new binary operator, we might
1037 /// break macros that were relying on that binary operator as a
1039 // when changing this do not forget to update doc/book/macros.md!
1040 fn is_in_follow(tok: &mbe::TokenTree, frag: Symbol) -> IsInFollow {
1043 if let TokenTree::Token(Token { kind: token::CloseDelim(_), .. }) = *tok {
1044 // closing a token tree can never be matched by any fragment;
1045 // iow, we always require that `(` and `)` match, etc.
1050 // since items *must* be followed by either a `;` or a `}`, we can
1051 // accept anything after them
1055 // anything can follow block, the braces provide an easy boundary to
1059 sym::stmt | sym::expr => {
1060 const TOKENS: &[&str] = &["`=>`", "`,`", "`;`"];
1062 TokenTree::Token(token) => match token.kind {
1063 FatArrow | Comma | Semi => IsInFollow::Yes,
1064 _ => IsInFollow::No(TOKENS),
1066 _ => IsInFollow::No(TOKENS),
1070 const TOKENS: &[&str] = &["`=>`", "`,`", "`=`", "`|`", "`if`", "`in`"];
1072 TokenTree::Token(token) => match token.kind {
1073 FatArrow | Comma | Eq | BinOp(token::Or) => IsInFollow::Yes,
1074 Ident(name, false) if name == kw::If || name == kw::In => IsInFollow::Yes,
1075 _ => IsInFollow::No(TOKENS),
1077 _ => IsInFollow::No(TOKENS),
1080 sym::path | sym::ty => {
1081 const TOKENS: &[&str] = &[
1082 "`{`", "`[`", "`=>`", "`,`", "`>`", "`=`", "`:`", "`;`", "`|`", "`as`",
1086 TokenTree::Token(token) => match token.kind {
1087 OpenDelim(token::DelimToken::Brace)
1088 | OpenDelim(token::DelimToken::Bracket)
1096 | BinOp(token::Or) => IsInFollow::Yes,
1097 Ident(name, false) if name == kw::As || name == kw::Where => {
1100 _ => IsInFollow::No(TOKENS),
1102 TokenTree::MetaVarDecl(_, _, frag) if frag.name == sym::block => {
1105 _ => IsInFollow::No(TOKENS),
1108 sym::ident | sym::lifetime => {
1109 // being a single token, idents and lifetimes are harmless
1113 // literals may be of a single token, or two tokens (negative numbers)
1116 sym::meta | sym::tt => {
1117 // being either a single token or a delimited sequence, tt is
1122 // Explicitly disallow `priv`, on the off chance it comes back.
1123 const TOKENS: &[&str] = &["`,`", "an ident", "a type"];
1125 TokenTree::Token(token) => match token.kind {
1126 Comma => IsInFollow::Yes,
1127 Ident(name, is_raw) if is_raw || name != kw::Priv => IsInFollow::Yes,
1129 if token.can_begin_type() {
1132 IsInFollow::No(TOKENS)
1136 TokenTree::MetaVarDecl(_, _, frag)
1137 if frag.name == sym::ident
1138 || frag.name == sym::ty
1139 || frag.name == sym::path =>
1143 _ => IsInFollow::No(TOKENS),
1146 kw::Invalid => IsInFollow::Yes,
1147 _ => IsInFollow::Invalid(
1148 format!("invalid fragment specifier `{}`", frag),
1149 VALID_FRAGMENT_NAMES_MSG,
1155 fn has_legal_fragment_specifier(
1157 features: &Features,
1158 attrs: &[ast::Attribute],
1159 tok: &mbe::TokenTree,
1160 ) -> Result<(), String> {
1161 debug!("has_legal_fragment_specifier({:?})", tok);
1162 if let mbe::TokenTree::MetaVarDecl(_, _, ref frag_spec) = *tok {
1163 let frag_span = tok.span();
1164 if !is_legal_fragment_specifier(sess, features, attrs, frag_spec.name, frag_span) {
1165 return Err(frag_spec.to_string());
1171 fn is_legal_fragment_specifier(
1173 _features: &Features,
1174 _attrs: &[ast::Attribute],
1179 * If new fragment specifiers are invented in nightly, `_sess`,
1180 * `_features`, `_attrs`, and `_frag_span` will be useful here
1181 * for checking against feature gates. See past versions of
1198 | kw::Invalid => true,
1203 fn quoted_tt_to_string(tt: &mbe::TokenTree) -> String {
1205 mbe::TokenTree::Token(ref token) => pprust::token_to_string(&token),
1206 mbe::TokenTree::MetaVar(_, name) => format!("${}", name),
1207 mbe::TokenTree::MetaVarDecl(_, name, kind) => format!("${}:{}", name, kind),
1209 "unexpected mbe::TokenTree::{{Sequence or Delimited}} \
1210 in follow set checker"
1215 fn base_parser_from_cx<'cx>(
1216 current_expansion: &'cx ExpansionData,
1217 sess: &'cx ParseSess,
1220 let directory = Directory {
1221 path: current_expansion.module.directory.clone(),
1222 ownership: current_expansion.directory_ownership,
1224 Parser::new(sess, tts, Some(directory), true, true, rustc_parse::MACRO_ARGUMENTS)
1227 /// Generates an appropriate parsing failure message. For EOF, this is "unexpected end...". For
1228 /// other tokens, this is "unexpected token...".
1229 fn parse_failure_msg(tok: &Token) -> String {
1231 token::Eof => "unexpected end of macro invocation".to_string(),
1232 _ => format!("no rules expected the token `{}`", pprust::token_to_string(tok),),