1 //! The main parser interface.
3 #![feature(bool_to_option)]
4 #![feature(crate_visibility_modifier)]
5 #![feature(bindings_after_at)]
6 #![feature(iter_order_by)]
7 #![feature(or_patterns)]
10 use rustc_ast::token::{self, DelimToken, Nonterminal, Token, TokenKind};
11 use rustc_ast::tokenstream::{self, TokenStream, TokenTree};
12 use rustc_ast_pretty::pprust;
13 use rustc_data_structures::sync::Lrc;
14 use rustc_errors::{Diagnostic, FatalError, Level, PResult};
15 use rustc_session::parse::ParseSess;
16 use rustc_span::{symbol::kw, FileName, SourceFile, Span, DUMMY_SP};
18 use smallvec::SmallVec;
23 use tracing::{debug, info};
25 pub const MACRO_ARGUMENTS: Option<&'static str> = Some("macro arguments");
29 use parser::{emit_unclosed_delims, make_unclosed_delims_error, Parser};
31 pub mod validate_attr;
33 // A bunch of utility functions of the form `parse_<thing>_from_<source>`
34 // where <thing> includes crate, expr, item, stmt, tts, and one that
35 // uses a HOF to parse anything, and <source> includes file and
38 /// A variant of 'panictry!' that works on a Vec<Diagnostic> instead of a single DiagnosticBuilder.
39 macro_rules! panictry_buffer {
40 ($handler:expr, $e:expr) => {{
41 use rustc_errors::FatalError;
42 use std::result::Result::{Err, Ok};
47 $handler.emit_diagnostic(&e);
55 pub fn parse_crate_from_file<'a>(input: &Path, sess: &'a ParseSess) -> PResult<'a, ast::Crate> {
56 let mut parser = new_parser_from_file(sess, input, None);
57 parser.parse_crate_mod()
60 pub fn parse_crate_attrs_from_file<'a>(
63 ) -> PResult<'a, Vec<ast::Attribute>> {
64 let mut parser = new_parser_from_file(sess, input, None);
65 parser.parse_inner_attributes()
68 pub fn parse_crate_from_source_str(
72 ) -> PResult<'_, ast::Crate> {
73 new_parser_from_source_str(sess, name, source).parse_crate_mod()
76 pub fn parse_crate_attrs_from_source_str(
80 ) -> PResult<'_, Vec<ast::Attribute>> {
81 new_parser_from_source_str(sess, name, source).parse_inner_attributes()
84 pub fn parse_stream_from_source_str(
88 override_span: Option<Span>,
90 let (stream, mut errors) =
91 source_file_to_stream(sess, sess.source_map().new_source_file(name, source), override_span);
92 emit_unclosed_delims(&mut errors, &sess);
96 /// Creates a new parser from a source string.
97 pub fn new_parser_from_source_str(sess: &ParseSess, name: FileName, source: String) -> Parser<'_> {
98 panictry_buffer!(&sess.span_diagnostic, maybe_new_parser_from_source_str(sess, name, source))
101 /// Creates a new parser from a source string. Returns any buffered errors from lexing the initial
103 pub fn maybe_new_parser_from_source_str(
107 ) -> Result<Parser<'_>, Vec<Diagnostic>> {
108 maybe_source_file_to_parser(sess, sess.source_map().new_source_file(name, source))
111 /// Creates a new parser, handling errors as appropriate if the file doesn't exist.
112 /// If a span is given, that is used on an error as the source of the problem.
113 pub fn new_parser_from_file<'a>(sess: &'a ParseSess, path: &Path, sp: Option<Span>) -> Parser<'a> {
114 source_file_to_parser(sess, file_to_source_file(sess, path, sp))
117 /// Given a `source_file` and config, returns a parser.
118 fn source_file_to_parser(sess: &ParseSess, source_file: Lrc<SourceFile>) -> Parser<'_> {
119 panictry_buffer!(&sess.span_diagnostic, maybe_source_file_to_parser(sess, source_file))
122 /// Given a `source_file` and config, return a parser. Returns any buffered errors from lexing the
123 /// initial token stream.
124 fn maybe_source_file_to_parser(
126 source_file: Lrc<SourceFile>,
127 ) -> Result<Parser<'_>, Vec<Diagnostic>> {
128 let end_pos = source_file.end_pos;
129 let (stream, unclosed_delims) = maybe_file_to_stream(sess, source_file, None)?;
130 let mut parser = stream_to_parser(sess, stream, None);
131 parser.unclosed_delims = unclosed_delims;
132 if parser.token == token::Eof {
133 parser.token.span = Span::new(end_pos, end_pos, parser.token.span.ctxt());
141 /// Given a session and a path and an optional span (for error reporting),
142 /// add the path to the session's source_map and return the new source_file or
143 /// error when a file can't be read.
144 fn try_file_to_source_file(
147 spanopt: Option<Span>,
148 ) -> Result<Lrc<SourceFile>, Diagnostic> {
149 sess.source_map().load_file(path).map_err(|e| {
150 let msg = format!("couldn't read {}: {}", path.display(), e);
151 let mut diag = Diagnostic::new(Level::Fatal, &msg);
152 if let Some(sp) = spanopt {
159 /// Given a session and a path and an optional span (for error reporting),
160 /// adds the path to the session's `source_map` and returns the new `source_file`.
161 fn file_to_source_file(sess: &ParseSess, path: &Path, spanopt: Option<Span>) -> Lrc<SourceFile> {
162 match try_file_to_source_file(sess, path, spanopt) {
163 Ok(source_file) => source_file,
165 sess.span_diagnostic.emit_diagnostic(&d);
171 /// Given a `source_file`, produces a sequence of token trees.
172 pub fn source_file_to_stream(
174 source_file: Lrc<SourceFile>,
175 override_span: Option<Span>,
176 ) -> (TokenStream, Vec<lexer::UnmatchedBrace>) {
177 panictry_buffer!(&sess.span_diagnostic, maybe_file_to_stream(sess, source_file, override_span))
180 /// Given a source file, produces a sequence of token trees. Returns any buffered errors from
181 /// parsing the token stream.
182 pub fn maybe_file_to_stream(
184 source_file: Lrc<SourceFile>,
185 override_span: Option<Span>,
186 ) -> Result<(TokenStream, Vec<lexer::UnmatchedBrace>), Vec<Diagnostic>> {
187 let src = source_file.src.as_ref().unwrap_or_else(|| {
189 .bug(&format!("cannot lex `source_file` without source: {}", source_file.name));
192 let (token_trees, unmatched_braces) =
193 lexer::parse_token_trees(sess, src.as_str(), source_file.start_pos, override_span);
196 Ok(stream) => Ok((stream, unmatched_braces)),
198 let mut buffer = Vec::with_capacity(1);
199 err.buffer(&mut buffer);
200 // Not using `emit_unclosed_delims` to use `db.buffer`
201 for unmatched in unmatched_braces {
202 if let Some(err) = make_unclosed_delims_error(unmatched, &sess) {
203 err.buffer(&mut buffer);
211 /// Given a stream and the `ParseSess`, produces a parser.
212 pub fn stream_to_parser<'a>(
215 subparser_name: Option<&'static str>,
217 Parser::new(sess, stream, false, subparser_name)
220 /// Runs the given subparser `f` on the tokens of the given `attr`'s item.
221 pub fn parse_in<'a, T>(
225 mut f: impl FnMut(&mut Parser<'a>) -> PResult<'a, T>,
226 ) -> PResult<'a, T> {
227 let mut parser = Parser::new(sess, tts, false, Some(name));
228 let result = f(&mut parser)?;
229 if parser.token != token::Eof {
230 parser.unexpected()?;
235 // NOTE(Centril): The following probably shouldn't be here but it acknowledges the
236 // fact that architecturally, we are using parsing (read on below to understand why).
238 pub fn nt_to_tokenstream(nt: &Nonterminal, sess: &ParseSess, span: Span) -> TokenStream {
239 // A `Nonterminal` is often a parsed AST item. At this point we now
240 // need to convert the parsed AST to an actual token stream, e.g.
241 // un-parse it basically.
243 // Unfortunately there's not really a great way to do that in a
244 // guaranteed lossless fashion right now. The fallback here is to just
245 // stringify the AST node and reparse it, but this loses all span
248 // As a result, some AST nodes are annotated with the token stream they
249 // came from. Here we attempt to extract these lossless token streams
250 // before we fall back to the stringification.
251 let tokens = match *nt {
252 Nonterminal::NtItem(ref item) => {
253 prepend_attrs(sess, &item.attrs, item.tokens.as_ref(), span)
255 Nonterminal::NtBlock(ref block) => block.tokens.clone(),
256 Nonterminal::NtStmt(ref stmt) => {
257 // FIXME: We currently only collect tokens for `:stmt`
258 // matchers in `macro_rules!` macros. When we start collecting
259 // tokens for attributes on statements, we will need to prepend
263 Nonterminal::NtPat(ref pat) => pat.tokens.clone(),
264 Nonterminal::NtTy(ref ty) => ty.tokens.clone(),
265 Nonterminal::NtIdent(ident, is_raw) => {
266 Some(tokenstream::TokenTree::token(token::Ident(ident.name, is_raw), ident.span).into())
268 Nonterminal::NtLifetime(ident) => {
269 Some(tokenstream::TokenTree::token(token::Lifetime(ident.name), ident.span).into())
271 Nonterminal::NtMeta(ref attr) => attr.tokens.clone(),
272 Nonterminal::NtPath(ref path) => path.tokens.clone(),
273 Nonterminal::NtVis(ref vis) => vis.tokens.clone(),
274 Nonterminal::NtTT(ref tt) => Some(tt.clone().into()),
275 Nonterminal::NtExpr(ref expr) | Nonterminal::NtLiteral(ref expr) => {
276 if expr.tokens.is_none() {
277 debug!("missing tokens for expr {:?}", expr);
279 prepend_attrs(sess, &expr.attrs, expr.tokens.as_ref(), span)
283 // FIXME(#43081): Avoid this pretty-print + reparse hack
284 // Pretty-print the AST struct without inserting any parenthesis
285 // beyond those explicitly written by the user (e.g. `ExpnKind::Paren`).
286 // The resulting stream may have incorrect precedence, but it's only
287 // ever used for a comparison against the capture tokenstream.
288 let source = pprust::nonterminal_to_string_no_extra_parens(nt);
289 let filename = FileName::macro_expansion_source_code(&source);
290 let reparsed_tokens = parse_stream_from_source_str(filename, source, sess, Some(span));
292 // During early phases of the compiler the AST could get modified
293 // directly (e.g., attributes added or removed) and the internal cache
294 // of tokens my not be invalidated or updated. Consequently if the
295 // "lossless" token stream disagrees with our actual stringification
296 // (which has historically been much more battle-tested) then we go
297 // with the lossy stream anyway (losing span information).
299 // Note that the comparison isn't `==` here to avoid comparing spans,
300 // but it *also* is a "probable" equality which is a pretty weird
301 // definition. We mostly want to catch actual changes to the AST
302 // like a `#[cfg]` being processed or some weird `macro_rules!`
305 // What we *don't* want to catch is the fact that a user-defined
306 // literal like `0xf` is stringified as `15`, causing the cached token
307 // stream to not be literal `==` token-wise (ignoring spans) to the
308 // token stream we got from stringification.
310 // Instead the "probably equal" check here is "does each token
311 // recursively have the same discriminant?" We basically don't look at
312 // the token values here and assume that such fine grained token stream
313 // modifications, including adding/removing typically non-semantic
314 // tokens such as extra braces and commas, don't happen.
315 if let Some(tokens) = tokens {
316 // Compare with a non-relaxed delim match to start.
317 if tokenstream_probably_equal_for_proc_macro(&tokens, &reparsed_tokens, sess, false) {
321 // The check failed. This time, we pretty-print the AST struct with parenthesis
322 // inserted to preserve precedence. This may cause `None`-delimiters in the captured
323 // token stream to match up with inserted parenthesis in the reparsed stream.
324 let source_with_parens = pprust::nonterminal_to_string(nt);
325 let filename_with_parens = FileName::macro_expansion_source_code(&source_with_parens);
326 let reparsed_tokens_with_parens = parse_stream_from_source_str(
327 filename_with_parens,
333 // Compare with a relaxed delim match - we want inserted parenthesis in the
334 // reparsed stream to match `None`-delimiters in the original stream.
335 if tokenstream_probably_equal_for_proc_macro(
337 &reparsed_tokens_with_parens,
345 "cached tokens found, but they're not \"probably equal\", \
346 going with stringified version"
348 info!("cached tokens: {}", pprust::tts_to_string(&tokens));
349 info!("reparsed tokens: {}", pprust::tts_to_string(&reparsed_tokens_with_parens));
351 info!("cached tokens debug: {:?}", tokens);
352 info!("reparsed tokens debug: {:?}", reparsed_tokens_with_parens);
357 // See comments in `Nonterminal::to_tokenstream` for why we care about
358 // *probably* equal here rather than actual equality
360 // This is otherwise the same as `eq_unspanned`, only recursing with a
362 pub fn tokenstream_probably_equal_for_proc_macro(
363 tokens: &TokenStream,
364 reparsed_tokens: &TokenStream,
366 relaxed_delim_match: bool,
368 // When checking for `probably_eq`, we ignore certain tokens that aren't
369 // preserved in the AST. Because they are not preserved, the pretty
370 // printer arbitrarily adds or removes them when printing as token
371 // streams, making a comparison between a token stream generated from an
372 // AST and a token stream which was parsed into an AST more reliable.
373 fn semantic_tree(tree: &TokenTree) -> bool {
374 if let TokenTree::Token(token) = tree {
376 // The pretty printer tends to add trailing commas to
377 // everything, and in particular, after struct fields.
379 // The pretty printer collapses many semicolons into one.
381 // We don't preserve leading `|` tokens in patterns, so
382 // we ignore them entirely
383 | token::BinOp(token::BinOpToken::Or)
384 // We don't preserve trailing '+' tokens in trait bounds,
385 // so we ignore them entirely
386 | token::BinOp(token::BinOpToken::Plus)
387 // The pretty printer can turn `$crate` into `::crate_name`
388 | token::ModSep = token.kind {
395 // When comparing two `TokenStream`s, we ignore the `IsJoint` information.
397 // However, `rustc_parse::lexer::tokentrees::TokenStreamBuilder` will
398 // use `Token.glue` on adjacent tokens with the proper `IsJoint`.
399 // Since we are ignoreing `IsJoint`, a 'glued' token (e.g. `BinOp(Shr)`)
400 // and its 'split'/'unglued' compoenents (e.g. `Gt, Gt`) are equivalent
401 // when determining if two `TokenStream`s are 'probably equal'.
403 // Therefore, we use `break_two_token_op` to convert all tokens
404 // to the 'unglued' form (if it exists). This ensures that two
405 // `TokenStream`s which differ only in how their tokens are glued
406 // will be considered 'probably equal', which allows us to keep spans.
408 // This is important when the original `TokenStream` contained
409 // extra spaces (e.g. `f :: < Vec < _ > > ( ) ;'). These extra spaces
410 // will be omitted when we pretty-print, which can cause the original
411 // and reparsed `TokenStream`s to differ in the assignment of `IsJoint`,
412 // leading to some tokens being 'glued' together in one stream but not
413 // the other. See #68489 for more details.
414 fn break_tokens(tree: TokenTree) -> impl Iterator<Item = TokenTree> {
415 // In almost all cases, we should have either zero or one levels
416 // of 'unglueing'. However, in some unusual cases, we may need
417 // to iterate breaking tokens mutliple times. For example:
418 // '[BinOpEq(Shr)] => [Gt, Ge] -> [Gt, Gt, Eq]'
419 let mut token_trees: SmallVec<[_; 2]>;
420 if let TokenTree::Token(token) = &tree {
421 let mut out = SmallVec::<[_; 2]>::new();
422 out.push(token.clone());
423 // Iterate to fixpoint:
424 // * We start off with 'out' containing our initial token, and `temp` empty
425 // * If we are able to break any tokens in `out`, then `out` will have
426 // at least one more element than 'temp', so we will try to break tokens
428 // * If we cannot break any tokens in 'out', we are done
430 let mut temp = SmallVec::<[_; 2]>::new();
431 let mut changed = false;
433 for token in out.into_iter() {
434 if let Some((first, second)) = token.kind.break_two_token_op() {
435 temp.push(Token::new(first, DUMMY_SP));
436 temp.push(Token::new(second, DUMMY_SP));
447 token_trees = out.into_iter().map(TokenTree::Token).collect();
449 token_trees = SmallVec::new();
450 token_trees.push(tree);
452 token_trees.into_iter()
455 fn expand_token(tree: TokenTree, sess: &ParseSess) -> impl Iterator<Item = TokenTree> {
456 // When checking tokenstreams for 'probable equality', we are comparing
457 // a captured (from parsing) `TokenStream` to a reparsed tokenstream.
458 // The reparsed Tokenstream will never have `None`-delimited groups,
459 // since they are only ever inserted as a result of macro expansion.
460 // Therefore, inserting a `None`-delimtied group here (when we
461 // convert a nested `Nonterminal` to a tokenstream) would cause
462 // a mismatch with the reparsed tokenstream.
464 // Note that we currently do not handle the case where the
465 // reparsed stream has a `Parenthesis`-delimited group
466 // inserted. This will cause a spurious mismatch:
467 // issue #75734 tracks resolving this.
469 let expanded: SmallVec<[_; 1]> =
470 if let TokenTree::Token(Token { kind: TokenKind::Interpolated(nt), span }) = &tree {
471 nt_to_tokenstream(nt, sess, *span)
473 .flat_map(|t| expand_token(t, sess))
476 // Filter before and after breaking tokens,
477 // since we may want to ignore both glued and unglued tokens.
478 std::iter::once(tree)
479 .filter(semantic_tree)
480 .flat_map(break_tokens)
481 .filter(semantic_tree)
487 // Break tokens after we expand any nonterminals, so that we break tokens
488 // that are produced as a result of nonterminal expansion.
489 let tokens = tokens.trees().flat_map(|t| expand_token(t, sess));
490 let reparsed_tokens = reparsed_tokens.trees().flat_map(|t| expand_token(t, sess));
492 tokens.eq_by(reparsed_tokens, |t, rt| {
493 tokentree_probably_equal_for_proc_macro(&t, &rt, sess, relaxed_delim_match)
497 // See comments in `Nonterminal::to_tokenstream` for why we care about
498 // *probably* equal here rather than actual equality
500 // This is otherwise the same as `eq_unspanned`, only recursing with a
502 pub fn tokentree_probably_equal_for_proc_macro(
504 reparsed_token: &TokenTree,
506 relaxed_delim_match: bool,
508 match (token, reparsed_token) {
509 (TokenTree::Token(token), TokenTree::Token(reparsed_token)) => {
510 token_probably_equal_for_proc_macro(token, reparsed_token)
513 TokenTree::Delimited(_, delim, tokens),
514 TokenTree::Delimited(_, reparsed_delim, reparsed_tokens),
515 ) if delim == reparsed_delim => tokenstream_probably_equal_for_proc_macro(
521 (TokenTree::Delimited(_, DelimToken::NoDelim, tokens), reparsed_token) => {
522 if relaxed_delim_match {
523 if let TokenTree::Delimited(_, DelimToken::Paren, reparsed_tokens) = reparsed_token
525 if tokenstream_probably_equal_for_proc_macro(
536 && tokentree_probably_equal_for_proc_macro(
537 &tokens.trees().next().unwrap(),
547 // See comments in `Nonterminal::to_tokenstream` for why we care about
548 // *probably* equal here rather than actual equality
549 fn token_probably_equal_for_proc_macro(first: &Token, other: &Token) -> bool {
550 if mem::discriminant(&first.kind) != mem::discriminant(&other.kind) {
553 use rustc_ast::token::TokenKind::*;
554 match (&first.kind, &other.kind) {
569 | (&DotDotDot, &DotDotDot)
570 | (&DotDotEq, &DotDotEq)
577 | (&FatArrow, &FatArrow)
580 | (&Question, &Question)
581 | (&Eof, &Eof) => true,
583 (&BinOp(a), &BinOp(b)) | (&BinOpEq(a), &BinOpEq(b)) => a == b,
585 (&OpenDelim(a), &OpenDelim(b)) | (&CloseDelim(a), &CloseDelim(b)) => a == b,
587 (&DocComment(a1, a2, a3), &DocComment(b1, b2, b3)) => a1 == b1 && a2 == b2 && a3 == b3,
589 (&Literal(a), &Literal(b)) => a == b,
591 (&Lifetime(a), &Lifetime(b)) => a == b,
592 (&Ident(a, b), &Ident(c, d)) => {
593 b == d && (a == c || a == kw::DollarCrate || c == kw::DollarCrate)
596 (&Interpolated(..), &Interpolated(..)) => panic!("Unexpanded Interpolated!"),
598 _ => panic!("forgot to add a token?"),
604 attrs: &[ast::Attribute],
605 tokens: Option<&tokenstream::TokenStream>,
606 span: rustc_span::Span,
607 ) -> Option<tokenstream::TokenStream> {
608 let tokens = tokens?;
609 if attrs.is_empty() {
610 return Some(tokens.clone());
612 let mut builder = tokenstream::TokenStreamBuilder::new();
616 ast::AttrStyle::Outer,
617 "inner attributes should prevent cached tokens from existing"
620 let source = pprust::attribute_to_string(attr);
621 let macro_filename = FileName::macro_expansion_source_code(&source);
623 let item = match attr.kind {
624 ast::AttrKind::Normal(ref item) => item,
625 ast::AttrKind::DocComment(..) => {
626 let stream = parse_stream_from_source_str(macro_filename, source, sess, Some(span));
627 builder.push(stream);
632 // synthesize # [ $path $tokens ] manually here
633 let mut brackets = tokenstream::TokenStreamBuilder::new();
635 // For simple paths, push the identifier directly
636 if item.path.segments.len() == 1 && item.path.segments[0].args.is_none() {
637 let ident = item.path.segments[0].ident;
638 let token = token::Ident(ident.name, ident.as_str().starts_with("r#"));
639 brackets.push(tokenstream::TokenTree::token(token, ident.span));
641 // ... and for more complicated paths, fall back to a reparse hack that
642 // should eventually be removed.
644 let stream = parse_stream_from_source_str(macro_filename, source, sess, Some(span));
645 brackets.push(stream);
648 brackets.push(item.args.outer_tokens());
650 // The span we list here for `#` and for `[ ... ]` are both wrong in
651 // that it encompasses more than each token, but it hopefully is "good
652 // enough" for now at least.
653 builder.push(tokenstream::TokenTree::token(token::Pound, attr.span));
654 let delim_span = tokenstream::DelimSpan::from_single(attr.span);
655 builder.push(tokenstream::TokenTree::Delimited(
657 token::DelimToken::Bracket,
661 builder.push(tokens.clone());
662 Some(builder.build())