1 use super::{Capturing, FlatToken, ForceCollect, Parser, ReplaceRange, TokenCursor, TrailingToken};
2 use rustc_ast::token::{self, DelimToken, Token, TokenKind};
3 use rustc_ast::tokenstream::{AttrAnnotatedTokenStream, AttributesData, CreateTokenStream};
4 use rustc_ast::tokenstream::{AttrAnnotatedTokenTree, DelimSpan, LazyTokenStream, Spacing};
5 use rustc_ast::{self as ast};
6 use rustc_ast::{AstLike, AttrVec, Attribute};
7 use rustc_errors::PResult;
8 use rustc_span::{sym, Span, DUMMY_SP};
10 use std::convert::TryInto;
13 /// A wrapper type to ensure that the parser handles outer attributes correctly.
14 /// When we parse outer attributes, we need to ensure that we capture tokens
15 /// for the attribute target. This allows us to perform cfg-expansion on
16 /// a token stream before we invoke a derive proc-macro.
18 /// This wrapper prevents direct access to the underlying `Vec<ast::Attribute>`.
19 /// Parsing code can only get access to the underlying attributes
20 /// by passing an `AttrWrapper` to `collect_tokens_trailing_tokens`.
21 /// This makes it difficult to accidentally construct an AST node
22 /// (which stores a `Vec<ast::Attribute>`) without first collecting tokens.
24 /// This struct has its own module, to ensure that the parser code
25 /// cannot directly access the `attrs` field
26 #[derive(Debug, Clone)]
27 pub struct AttrWrapper {
29 // The start of the outer attributes in the token cursor.
30 // This allows us to create a `ReplaceRange` for the entire attribute
31 // target, including outer attributes.
35 // This struct is passed around very frequently,
36 // so make sure it doesn't accidentally get larger
37 #[cfg(all(target_arch = "x86_64", target_pointer_width = "64"))]
38 rustc_data_structures::static_assert_size!(AttrWrapper, 16);
41 pub(super) fn new(attrs: AttrVec, start_pos: usize) -> AttrWrapper {
42 AttrWrapper { attrs, start_pos }
44 pub fn empty() -> AttrWrapper {
45 AttrWrapper { attrs: AttrVec::new(), start_pos: usize::MAX }
47 // FIXME: Delay span bug here?
48 pub(crate) fn take_for_recovery(self) -> AttrVec {
52 // FIXME: require passing an NT to prevent misuse of this method
53 pub(crate) fn prepend_to_nt_inner(self, attrs: &mut Vec<Attribute>) {
54 let mut self_attrs: Vec<_> = self.attrs.into();
55 std::mem::swap(attrs, &mut self_attrs);
56 attrs.extend(self_attrs);
59 pub fn is_empty(&self) -> bool {
63 pub fn maybe_needs_tokens(&self) -> bool {
64 crate::parser::attr::maybe_needs_tokens(&self.attrs)
68 /// Returns `true` if `attrs` contains a `cfg` or `cfg_attr` attribute
69 fn has_cfg_or_cfg_attr(attrs: &[Attribute]) -> bool {
70 // NOTE: Builtin attributes like `cfg` and `cfg_attr` cannot be renamed via imports.
71 // Therefore, the absence of a literal `cfg` or `cfg_attr` guarantees that
72 // we don't need to do any eager expansion.
73 attrs.iter().any(|attr| {
74 attr.ident().map_or(false, |ident| ident.name == sym::cfg || ident.name == sym::cfg_attr)
78 // Produces a `TokenStream` on-demand. Using `cursor_snapshot`
79 // and `num_calls`, we can reconstruct the `TokenStream` seen
80 // by the callback. This allows us to avoid producing a `TokenStream`
81 // if it is never needed - for example, a captured `macro_rules!`
82 // argument that is never passed to a proc macro.
83 // In practice token stream creation happens rarely compared to
84 // calls to `collect_tokens` (see some statistics in #78736),
85 // so we are doing as little up-front work as possible.
87 // This also makes `Parser` very cheap to clone, since
88 // there is no intermediate collection buffer to clone.
90 struct LazyTokenStreamImpl {
91 start_token: (Token, Spacing),
92 cursor_snapshot: TokenCursor,
94 break_last_token: bool,
95 replace_ranges: Box<[ReplaceRange]>,
98 #[cfg(all(target_arch = "x86_64", target_pointer_width = "64"))]
99 rustc_data_structures::static_assert_size!(LazyTokenStreamImpl, 144);
101 impl CreateTokenStream for LazyTokenStreamImpl {
102 fn create_token_stream(&self) -> AttrAnnotatedTokenStream {
103 // The token produced by the final call to `{,inlined_}next` or
104 // `{,inlined_}next_desugared` was not actually consumed by the
105 // callback. The combination of chaining the initial token and using
106 // `take` produces the desired result - we produce an empty
107 // `TokenStream` if no calls were made, and omit the final token
109 let mut cursor_snapshot = self.cursor_snapshot.clone();
111 std::iter::once((FlatToken::Token(self.start_token.0.clone()), self.start_token.1))
112 .chain((0..self.num_calls).map(|_| {
113 let token = if cursor_snapshot.desugar_doc_comments {
114 cursor_snapshot.next_desugared()
116 cursor_snapshot.next()
118 (FlatToken::Token(token.0), token.1)
120 .take(self.num_calls);
122 if !self.replace_ranges.is_empty() {
123 let mut tokens: Vec<_> = tokens.collect();
124 let mut replace_ranges = self.replace_ranges.clone();
125 replace_ranges.sort_by_key(|(range, _)| range.start);
127 #[cfg(debug_assertions)]
129 for [(range, tokens), (next_range, next_tokens)] in replace_ranges.array_windows() {
131 range.end <= next_range.start || range.end >= next_range.end,
132 "Replace ranges should either be disjoint or nested: ({:?}, {:?}) ({:?}, {:?})",
141 // Process the replace ranges, starting from the highest start
142 // position and working our way back. If have tokens like:
144 // `#[cfg(FALSE)]` struct Foo { #[cfg(FALSE)] field: bool }`
146 // Then we will generate replace ranges for both
147 // the `#[cfg(FALSE)] field: bool` and the entire
148 // `#[cfg(FALSE)]` struct Foo { #[cfg(FALSE)] field: bool }`
150 // By starting processing from the replace range with the greatest
151 // start position, we ensure that any replace range which encloses
152 // another replace range will capture the *replaced* tokens for the inner
153 // range, not the original tokens.
154 for (range, new_tokens) in replace_ranges.iter().rev() {
155 assert!(!range.is_empty(), "Cannot replace an empty range: {:?}", range);
156 // Replace ranges are only allowed to decrease the number of tokens.
158 range.len() >= new_tokens.len(),
159 "Range {:?} has greater len than {:?}",
164 // Replace any removed tokens with `FlatToken::Empty`.
165 // This keeps the total length of `tokens` constant throughout the
166 // replacement process, allowing us to use all of the `ReplaceRanges` entries
167 // without adjusting indices.
168 let filler = std::iter::repeat((FlatToken::Empty, Spacing::Alone))
169 .take(range.len() - new_tokens.len());
172 (range.start as usize)..(range.end as usize),
173 new_tokens.clone().into_iter().chain(filler),
176 make_token_stream(tokens.into_iter(), self.break_last_token)
178 make_token_stream(tokens, self.break_last_token)
183 impl<'a> Parser<'a> {
184 /// Records all tokens consumed by the provided callback,
185 /// including the current token. These tokens are collected
186 /// into a `LazyTokenStream`, and returned along with the result
189 /// Note: If your callback consumes an opening delimiter
190 /// (including the case where you call `collect_tokens`
191 /// when the current token is an opening delimiter),
192 /// you must also consume the corresponding closing delimiter.
194 /// That is, you can consume
195 /// `something ([{ }])` or `([{}])`, but not `([{}]`
197 /// This restriction shouldn't be an issue in practice,
198 /// since this function is used to record the tokens for
199 /// a parsed AST item, which always has matching delimiters.
200 pub fn collect_tokens_trailing_token<R: AstLike>(
203 force_collect: ForceCollect,
204 f: impl FnOnce(&mut Self, Vec<ast::Attribute>) -> PResult<'a, (R, TrailingToken)>,
205 ) -> PResult<'a, R> {
206 // We only bail out when nothing could possibly observe the collected tokens:
207 // 1. We cannot be force collecting tokens (since force-collecting requires tokens
209 if matches!(force_collect, ForceCollect::No)
210 // None of our outer attributes can require tokens (e.g. a proc-macro)
211 && !attrs.maybe_needs_tokens()
212 // If our target supports custom inner attributes, then we cannot bail
213 // out early, since we may need to capture tokens for a custom inner attribute
215 && !R::SUPPORTS_CUSTOM_INNER_ATTRS
216 // Never bail out early in `capture_cfg` mode, since there might be `#[cfg]`
217 // or `#[cfg_attr]` attributes.
220 return Ok(f(self, attrs.attrs.into())?.0);
223 let start_token = (self.token.clone(), self.token_spacing);
224 let cursor_snapshot = self.token_cursor.clone();
226 let has_outer_attrs = !attrs.attrs.is_empty();
227 let prev_capturing = std::mem::replace(&mut self.capture_state.capturing, Capturing::Yes);
228 let replace_ranges_start = self.capture_state.replace_ranges.len();
230 let ret = f(self, attrs.attrs.into());
232 self.capture_state.capturing = prev_capturing;
234 let (mut ret, trailing) = ret?;
236 // When we're not in `capture-cfg` mode, then bail out early if:
237 // 1. Our target doesn't support tokens at all (e.g we're parsing an `NtIdent`)
238 // so there's nothing for us to do.
239 // 2. Our target already has tokens set (e.g. we've parsed something
240 // like `#[my_attr] $item`. The actual parsing code takes care of prepending
241 // any attributes to the nonterminal, so we don't need to modify the
242 // already captured tokens.
243 // Note that this check is independent of `force_collect`- if we already
244 // have tokens, or can't even store them, then there's never a need to
245 // force collection of new tokens.
246 if !self.capture_cfg && matches!(ret.tokens_mut(), None | Some(Some(_))) {
250 // This is very similar to the bail out check at the start of this function.
251 // Now that we've parsed an AST node, we have more information available.
252 if matches!(force_collect, ForceCollect::No)
253 // We now have inner attributes available, so this check is more precise
254 // than `attrs.maybe_needs_tokens()` at the start of the function.
255 // As a result, we don't need to check `R::SUPPORTS_CUSTOM_INNER_ATTRS`
256 && !crate::parser::attr::maybe_needs_tokens(ret.attrs())
257 // Subtle: We call `has_cfg_or_cfg_attr` with the attrs from `ret`.
258 // This ensures that we consider inner attributes (e.g. `#![cfg]`),
259 // which require us to have tokens available
260 // We also call `has_cfg_or_cfg_attr` at the beginning of this function,
261 // but we only bail out if there's no possibility of inner attributes
262 // (!R::SUPPORTS_CUSTOM_INNER_ATTRS)
263 // We only capture about `#[cfg]` or `#[cfg_attr]` in `capture_cfg`
264 // mode - during normal parsing, we don't need any special capturing
265 // for those attributes, since they're builtin.
266 && !(self.capture_cfg && has_cfg_or_cfg_attr(ret.attrs()))
271 let mut inner_attr_replace_ranges = Vec::new();
272 // Take the captured ranges for any inner attributes that we parsed.
273 for inner_attr in ret.attrs().iter().filter(|a| a.style == ast::AttrStyle::Inner) {
274 if let Some(attr_range) = self.capture_state.inner_attr_ranges.remove(&inner_attr.id) {
275 inner_attr_replace_ranges.push(attr_range);
279 .delay_span_bug(inner_attr.span, "Missing token range for attribute");
283 let replace_ranges_end = self.capture_state.replace_ranges.len();
285 let cursor_snapshot_next_calls = cursor_snapshot.num_next_calls;
286 let mut end_pos = self.token_cursor.num_next_calls;
288 // Capture a trailing token if requested by the callback 'f'
290 TrailingToken::None => {}
291 TrailingToken::Semi => {
292 assert_eq!(self.token.kind, token::Semi);
295 TrailingToken::MaybeComma => {
296 if self.token.kind == token::Comma {
302 // If we 'broke' the last token (e.g. breaking a '>>' token to two '>' tokens),
303 // then extend the range of captured tokens to include it, since the parser
304 // was not actually bumped past it. When the `LazyTokenStream` gets converted
305 // into an `AttrAnnotatedTokenStream`, we will create the proper token.
306 if self.token_cursor.break_last_token {
310 "Cannot set `break_last_token` and have trailing token"
315 let num_calls = end_pos - cursor_snapshot_next_calls;
317 // If we have no attributes, then we will never need to
318 // use any replace ranges.
319 let replace_ranges: Box<[ReplaceRange]> = if ret.attrs().is_empty() && !self.capture_cfg {
322 // Grab any replace ranges that occur *inside* the current AST node.
323 // We will perform the actual replacement when we convert the `LazyTokenStream`
324 // to an `AttrAnnotatedTokenStream`
325 let start_calls: u32 = cursor_snapshot_next_calls.try_into().unwrap();
326 self.capture_state.replace_ranges[replace_ranges_start..replace_ranges_end]
329 .chain(inner_attr_replace_ranges.clone().into_iter())
330 .map(|(range, tokens)| {
331 ((range.start - start_calls)..(range.end - start_calls), tokens)
336 let tokens = LazyTokenStream::new(LazyTokenStreamImpl {
340 break_last_token: self.token_cursor.break_last_token,
344 // If we support tokens at all
345 if let Some(target_tokens) = ret.tokens_mut() {
346 if target_tokens.is_none() {
347 // Store se our newly captured tokens into the AST node
348 *target_tokens = Some(tokens.clone());
352 let final_attrs = ret.attrs();
354 // If `capture_cfg` is set and we're inside a recursive call to
355 // `collect_tokens_trailing_token`, then we need to register a replace range
356 // if we have `#[cfg]` or `#[cfg_attr]`. This allows us to run eager cfg-expansion
357 // on the captured token stream.
359 && matches!(self.capture_state.capturing, Capturing::Yes)
360 && has_cfg_or_cfg_attr(&final_attrs)
362 let attr_data = AttributesData { attrs: final_attrs.to_vec().into(), tokens };
364 // Replace the entire AST node that we just parsed, including attributes,
365 // with a `FlatToken::AttrTarget`. If this AST node is inside an item
366 // that has `#[derive]`, then this will allow us to cfg-expand this
369 if has_outer_attrs { attrs.start_pos } else { cursor_snapshot_next_calls };
370 let new_tokens = vec![(FlatToken::AttrTarget(attr_data), Spacing::Alone)];
373 !self.token_cursor.break_last_token,
374 "Should not have unglued last token with cfg attr"
376 let range: Range<u32> = (start_pos.try_into().unwrap())..(end_pos.try_into().unwrap());
377 self.capture_state.replace_ranges.push((range, new_tokens));
378 self.capture_state.replace_ranges.extend(inner_attr_replace_ranges);
381 // Only clear our `replace_ranges` when we're finished capturing entirely.
382 if matches!(self.capture_state.capturing, Capturing::No) {
383 self.capture_state.replace_ranges.clear();
384 // We don't clear `inner_attr_ranges`, as doing so repeatedly
385 // had a measurable performance impact. Most inner attributes that
386 // we insert will get removed - when we drop the parser, we'll free
387 // up the memory used by any attributes that we didn't remove from the map.
393 /// Converts a flattened iterator of tokens (including open and close delimiter tokens)
394 /// into a `TokenStream`, creating a `TokenTree::Delimited` for each matching pair
395 /// of open and close delims.
396 // FIXME(#67062): Currently, we don't parse `None`-delimited groups correctly,
397 // which can cause us to end up with mismatched `None` delimiters in our
398 // captured tokens. This function contains several hacks to work around this -
399 // essentially, we throw away mismatched `None` delimiters when we encounter them.
400 // Once we properly parse `None` delimiters, they can be captured just like any
401 // other tokens, and these hacks can be removed.
402 fn make_token_stream(
403 mut iter: impl Iterator<Item = (FlatToken, Spacing)>,
404 break_last_token: bool,
405 ) -> AttrAnnotatedTokenStream {
409 open_delim: DelimToken,
410 inner: Vec<(AttrAnnotatedTokenTree, Spacing)>,
413 vec![FrameData { open: DUMMY_SP, open_delim: DelimToken::NoDelim, inner: vec![] }];
414 let mut token_and_spacing = iter.next();
415 while let Some((token, spacing)) = token_and_spacing {
417 FlatToken::Token(Token { kind: TokenKind::OpenDelim(delim), span }) => {
418 stack.push(FrameData { open: span, open_delim: delim, inner: vec![] });
420 FlatToken::Token(Token { kind: TokenKind::CloseDelim(delim), span }) => {
421 // HACK: If we encounter a mismatched `None` delimiter at the top
422 // level, just ignore it.
423 if matches!(delim, DelimToken::NoDelim)
425 || !matches!(stack.last_mut().unwrap().open_delim, DelimToken::NoDelim))
427 token_and_spacing = iter.next();
430 let frame_data = stack
432 .unwrap_or_else(|| panic!("Token stack was empty for token: {:?}", token));
434 // HACK: If our current frame has a mismatched opening `None` delimiter,
435 // merge our current frame with the one above it. That is, transform
436 // `[ { < first second } third ]` into `[ { first second } third ]`
437 if !matches!(delim, DelimToken::NoDelim)
438 && matches!(frame_data.open_delim, DelimToken::NoDelim)
440 stack.last_mut().unwrap().inner.extend(frame_data.inner);
441 // Process our closing delimiter again, this time at the previous
442 // frame in the stack
443 token_and_spacing = Some((token, spacing));
448 frame_data.open_delim, delim,
449 "Mismatched open/close delims: open={:?} close={:?}",
450 frame_data.open, span
452 let dspan = DelimSpan::from_pair(frame_data.open, span);
453 let stream = AttrAnnotatedTokenStream::new(frame_data.inner);
454 let delimited = AttrAnnotatedTokenTree::Delimited(dspan, delim, stream);
458 panic!("Bottom token frame is missing for token: {:?}", token)
461 .push((delimited, Spacing::Alone));
463 FlatToken::Token(token) => stack
465 .expect("Bottom token frame is missing!")
467 .push((AttrAnnotatedTokenTree::Token(token), spacing)),
468 FlatToken::AttrTarget(data) => stack
470 .expect("Bottom token frame is missing!")
472 .push((AttrAnnotatedTokenTree::Attributes(data), spacing)),
473 FlatToken::Empty => {}
475 token_and_spacing = iter.next();
477 // HACK: If we don't have a closing `None` delimiter for our last
478 // frame, merge the frame with the top-level frame. That is,
479 // turn `< first second` into `first second`
480 if stack.len() == 2 && stack[1].open_delim == DelimToken::NoDelim {
481 let temp_buf = stack.pop().unwrap();
482 stack.last_mut().unwrap().inner.extend(temp_buf.inner);
484 let mut final_buf = stack.pop().expect("Missing final buf!");
485 if break_last_token {
486 let (last_token, spacing) = final_buf.inner.pop().unwrap();
487 if let AttrAnnotatedTokenTree::Token(last_token) = last_token {
488 let unglued_first = last_token.kind.break_two_token_op().unwrap().0;
490 // An 'unglued' token is always two ASCII characters
491 let mut first_span = last_token.span.shrink_to_lo();
492 first_span = first_span.with_hi(first_span.lo() + rustc_span::BytePos(1));
494 final_buf.inner.push((
495 AttrAnnotatedTokenTree::Token(Token::new(unglued_first, first_span)),
499 panic!("Unexpected last token {:?}", last_token)
502 assert!(stack.is_empty(), "Stack should be empty: final_buf={:?} stack={:?}", final_buf, stack);
503 AttrAnnotatedTokenStream::new(final_buf.inner)