1 use super::{Capturing, FlatToken, ForceCollect, Parser, ReplaceRange, TokenCursor, TrailingToken};
2 use rustc_ast::token::{self, Delimiter, Token, TokenKind};
3 use rustc_ast::tokenstream::{AttrTokenStream, AttributesData, ToAttrTokenStream};
4 use rustc_ast::tokenstream::{AttrTokenTree, DelimSpan, LazyAttrTokenStream, Spacing};
5 use rustc_ast::{self as ast};
6 use rustc_ast::{AttrVec, Attribute, HasAttrs, HasTokens};
7 use rustc_errors::PResult;
8 use rustc_span::{sym, Span};
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 `ast::AttrVec>`.
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 an `ast::AttrVec`) 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.
36 pub(super) fn new(attrs: AttrVec, start_pos: usize) -> AttrWrapper {
37 AttrWrapper { attrs, start_pos }
39 pub fn empty() -> AttrWrapper {
40 AttrWrapper { attrs: AttrVec::new(), start_pos: usize::MAX }
42 // FIXME: Delay span bug here?
43 pub(crate) fn take_for_recovery(self) -> AttrVec {
47 // Prepend `self.attrs` to `attrs`.
48 // FIXME: require passing an NT to prevent misuse of this method
49 pub(crate) fn prepend_to_nt_inner(self, attrs: &mut AttrVec) {
50 let mut self_attrs = self.attrs;
51 std::mem::swap(attrs, &mut self_attrs);
52 attrs.extend(self_attrs);
55 pub fn is_empty(&self) -> bool {
59 pub fn maybe_needs_tokens(&self) -> bool {
60 crate::parser::attr::maybe_needs_tokens(&self.attrs)
64 /// Returns `true` if `attrs` contains a `cfg` or `cfg_attr` attribute
65 fn has_cfg_or_cfg_attr(attrs: &[Attribute]) -> bool {
66 // NOTE: Builtin attributes like `cfg` and `cfg_attr` cannot be renamed via imports.
67 // Therefore, the absence of a literal `cfg` or `cfg_attr` guarantees that
68 // we don't need to do any eager expansion.
69 attrs.iter().any(|attr| {
70 attr.ident().map_or(false, |ident| ident.name == sym::cfg || ident.name == sym::cfg_attr)
74 // Produces a `TokenStream` on-demand. Using `cursor_snapshot`
75 // and `num_calls`, we can reconstruct the `TokenStream` seen
76 // by the callback. This allows us to avoid producing a `TokenStream`
77 // if it is never needed - for example, a captured `macro_rules!`
78 // argument that is never passed to a proc macro.
79 // In practice token stream creation happens rarely compared to
80 // calls to `collect_tokens` (see some statistics in #78736),
81 // so we are doing as little up-front work as possible.
83 // This also makes `Parser` very cheap to clone, since
84 // there is no intermediate collection buffer to clone.
86 struct LazyAttrTokenStreamImpl {
87 start_token: (Token, Spacing),
88 cursor_snapshot: TokenCursor,
90 break_last_token: bool,
91 replace_ranges: Box<[ReplaceRange]>,
94 impl ToAttrTokenStream for LazyAttrTokenStreamImpl {
95 fn to_attr_token_stream(&self) -> AttrTokenStream {
96 // The token produced by the final call to `{,inlined_}next` was not
97 // actually consumed by the callback. The combination of chaining the
98 // initial token and using `take` produces the desired result - we
99 // produce an empty `TokenStream` if no calls were made, and omit the
100 // final token otherwise.
101 let mut cursor_snapshot = self.cursor_snapshot.clone();
103 std::iter::once((FlatToken::Token(self.start_token.0.clone()), self.start_token.1))
104 .chain((0..self.num_calls).map(|_| {
105 let token = cursor_snapshot.next(cursor_snapshot.desugar_doc_comments);
106 (FlatToken::Token(token.0), token.1)
108 .take(self.num_calls);
110 if !self.replace_ranges.is_empty() {
111 let mut tokens: Vec<_> = tokens.collect();
112 let mut replace_ranges = self.replace_ranges.to_vec();
113 replace_ranges.sort_by_key(|(range, _)| range.start);
115 #[cfg(debug_assertions)]
117 for [(range, tokens), (next_range, next_tokens)] in replace_ranges.array_windows() {
119 range.end <= next_range.start || range.end >= next_range.end,
120 "Replace ranges should either be disjoint or nested: ({:?}, {:?}) ({:?}, {:?})",
129 // Process the replace ranges, starting from the highest start
130 // position and working our way back. If have tokens like:
132 // `#[cfg(FALSE)]` struct Foo { #[cfg(FALSE)] field: bool }`
134 // Then we will generate replace ranges for both
135 // the `#[cfg(FALSE)] field: bool` and the entire
136 // `#[cfg(FALSE)]` struct Foo { #[cfg(FALSE)] field: bool }`
138 // By starting processing from the replace range with the greatest
139 // start position, we ensure that any replace range which encloses
140 // another replace range will capture the *replaced* tokens for the inner
141 // range, not the original tokens.
142 for (range, new_tokens) in replace_ranges.into_iter().rev() {
143 assert!(!range.is_empty(), "Cannot replace an empty range: {:?}", range);
144 // Replace ranges are only allowed to decrease the number of tokens.
146 range.len() >= new_tokens.len(),
147 "Range {:?} has greater len than {:?}",
152 // Replace any removed tokens with `FlatToken::Empty`.
153 // This keeps the total length of `tokens` constant throughout the
154 // replacement process, allowing us to use all of the `ReplaceRanges` entries
155 // without adjusting indices.
156 let filler = std::iter::repeat((FlatToken::Empty, Spacing::Alone))
157 .take(range.len() - new_tokens.len());
160 (range.start as usize)..(range.end as usize),
161 new_tokens.into_iter().chain(filler),
164 make_token_stream(tokens.into_iter(), self.break_last_token)
166 make_token_stream(tokens, self.break_last_token)
171 impl<'a> Parser<'a> {
172 /// Records all tokens consumed by the provided callback,
173 /// including the current token. These tokens are collected
174 /// into a `LazyAttrTokenStream`, and returned along with the result
177 /// Note: If your callback consumes an opening delimiter
178 /// (including the case where you call `collect_tokens`
179 /// when the current token is an opening delimiter),
180 /// you must also consume the corresponding closing delimiter.
182 /// That is, you can consume
183 /// `something ([{ }])` or `([{}])`, but not `([{}]`
185 /// This restriction shouldn't be an issue in practice,
186 /// since this function is used to record the tokens for
187 /// a parsed AST item, which always has matching delimiters.
188 pub fn collect_tokens_trailing_token<R: HasAttrs + HasTokens>(
191 force_collect: ForceCollect,
192 f: impl FnOnce(&mut Self, ast::AttrVec) -> PResult<'a, (R, TrailingToken)>,
193 ) -> PResult<'a, R> {
194 // We only bail out when nothing could possibly observe the collected tokens:
195 // 1. We cannot be force collecting tokens (since force-collecting requires tokens
197 if matches!(force_collect, ForceCollect::No)
198 // None of our outer attributes can require tokens (e.g. a proc-macro)
199 && !attrs.maybe_needs_tokens()
200 // If our target supports custom inner attributes, then we cannot bail
201 // out early, since we may need to capture tokens for a custom inner attribute
203 && !R::SUPPORTS_CUSTOM_INNER_ATTRS
204 // Never bail out early in `capture_cfg` mode, since there might be `#[cfg]`
205 // or `#[cfg_attr]` attributes.
208 return Ok(f(self, attrs.attrs)?.0);
211 let start_token = (self.token.clone(), self.token_spacing);
212 let cursor_snapshot = self.token_cursor.clone();
214 let has_outer_attrs = !attrs.attrs.is_empty();
215 let prev_capturing = std::mem::replace(&mut self.capture_state.capturing, Capturing::Yes);
216 let replace_ranges_start = self.capture_state.replace_ranges.len();
218 let ret = f(self, attrs.attrs);
220 self.capture_state.capturing = prev_capturing;
222 let (mut ret, trailing) = ret?;
224 // When we're not in `capture-cfg` mode, then bail out early if:
225 // 1. Our target doesn't support tokens at all (e.g we're parsing an `NtIdent`)
226 // so there's nothing for us to do.
227 // 2. Our target already has tokens set (e.g. we've parsed something
228 // like `#[my_attr] $item`. The actual parsing code takes care of prepending
229 // any attributes to the nonterminal, so we don't need to modify the
230 // already captured tokens.
231 // Note that this check is independent of `force_collect`- if we already
232 // have tokens, or can't even store them, then there's never a need to
233 // force collection of new tokens.
234 if !self.capture_cfg && matches!(ret.tokens_mut(), None | Some(Some(_))) {
238 // This is very similar to the bail out check at the start of this function.
239 // Now that we've parsed an AST node, we have more information available.
240 if matches!(force_collect, ForceCollect::No)
241 // We now have inner attributes available, so this check is more precise
242 // than `attrs.maybe_needs_tokens()` at the start of the function.
243 // As a result, we don't need to check `R::SUPPORTS_CUSTOM_INNER_ATTRS`
244 && !crate::parser::attr::maybe_needs_tokens(ret.attrs())
245 // Subtle: We call `has_cfg_or_cfg_attr` with the attrs from `ret`.
246 // This ensures that we consider inner attributes (e.g. `#![cfg]`),
247 // which require us to have tokens available
248 // We also call `has_cfg_or_cfg_attr` at the beginning of this function,
249 // but we only bail out if there's no possibility of inner attributes
250 // (!R::SUPPORTS_CUSTOM_INNER_ATTRS)
251 // We only capture about `#[cfg]` or `#[cfg_attr]` in `capture_cfg`
252 // mode - during normal parsing, we don't need any special capturing
253 // for those attributes, since they're builtin.
254 && !(self.capture_cfg && has_cfg_or_cfg_attr(ret.attrs()))
259 let mut inner_attr_replace_ranges = Vec::new();
260 // Take the captured ranges for any inner attributes that we parsed.
261 for inner_attr in ret.attrs().iter().filter(|a| a.style == ast::AttrStyle::Inner) {
262 if let Some(attr_range) = self.capture_state.inner_attr_ranges.remove(&inner_attr.id) {
263 inner_attr_replace_ranges.push(attr_range);
267 .delay_span_bug(inner_attr.span, "Missing token range for attribute");
271 let replace_ranges_end = self.capture_state.replace_ranges.len();
273 let cursor_snapshot_next_calls = cursor_snapshot.num_next_calls;
274 let mut end_pos = self.token_cursor.num_next_calls;
276 let mut captured_trailing = false;
278 // Capture a trailing token if requested by the callback 'f'
280 TrailingToken::None => {}
281 TrailingToken::Gt => {
282 assert_eq!(self.token.kind, token::Gt);
284 TrailingToken::Semi => {
285 assert_eq!(self.token.kind, token::Semi);
287 captured_trailing = true;
289 TrailingToken::MaybeComma => {
290 if self.token.kind == token::Comma {
292 captured_trailing = true;
297 // If we 'broke' the last token (e.g. breaking a '>>' token to two '>' tokens),
298 // then extend the range of captured tokens to include it, since the parser
299 // was not actually bumped past it. When the `LazyAttrTokenStream` gets converted
300 // into an `AttrTokenStream`, we will create the proper token.
301 if self.token_cursor.break_last_token {
302 assert!(!captured_trailing, "Cannot set break_last_token and have trailing token");
306 let num_calls = end_pos - cursor_snapshot_next_calls;
308 // If we have no attributes, then we will never need to
309 // use any replace ranges.
310 let replace_ranges: Box<[ReplaceRange]> = if ret.attrs().is_empty() && !self.capture_cfg {
313 // Grab any replace ranges that occur *inside* the current AST node.
314 // We will perform the actual replacement when we convert the `LazyAttrTokenStream`
315 // to an `AttrTokenStream`.
316 let start_calls: u32 = cursor_snapshot_next_calls.try_into().unwrap();
317 self.capture_state.replace_ranges[replace_ranges_start..replace_ranges_end]
320 .chain(inner_attr_replace_ranges.iter().cloned())
321 .map(|(range, tokens)| {
322 ((range.start - start_calls)..(range.end - start_calls), tokens)
327 let tokens = LazyAttrTokenStream::new(LazyAttrTokenStreamImpl {
331 break_last_token: self.token_cursor.break_last_token,
335 // If we support tokens at all
336 if let Some(target_tokens) = ret.tokens_mut() {
337 if target_tokens.is_none() {
338 // Store se our newly captured tokens into the AST node
339 *target_tokens = Some(tokens.clone());
343 let final_attrs = ret.attrs();
345 // If `capture_cfg` is set and we're inside a recursive call to
346 // `collect_tokens_trailing_token`, then we need to register a replace range
347 // if we have `#[cfg]` or `#[cfg_attr]`. This allows us to run eager cfg-expansion
348 // on the captured token stream.
350 && matches!(self.capture_state.capturing, Capturing::Yes)
351 && has_cfg_or_cfg_attr(final_attrs)
353 let attr_data = AttributesData { attrs: final_attrs.iter().cloned().collect(), tokens };
355 // Replace the entire AST node that we just parsed, including attributes,
356 // with a `FlatToken::AttrTarget`. If this AST node is inside an item
357 // that has `#[derive]`, then this will allow us to cfg-expand this
360 if has_outer_attrs { attrs.start_pos } else { cursor_snapshot_next_calls };
361 let new_tokens = vec![(FlatToken::AttrTarget(attr_data), Spacing::Alone)];
364 !self.token_cursor.break_last_token,
365 "Should not have unglued last token with cfg attr"
367 let range: Range<u32> = (start_pos.try_into().unwrap())..(end_pos.try_into().unwrap());
368 self.capture_state.replace_ranges.push((range, new_tokens));
369 self.capture_state.replace_ranges.extend(inner_attr_replace_ranges);
372 // Only clear our `replace_ranges` when we're finished capturing entirely.
373 if matches!(self.capture_state.capturing, Capturing::No) {
374 self.capture_state.replace_ranges.clear();
375 // We don't clear `inner_attr_ranges`, as doing so repeatedly
376 // had a measurable performance impact. Most inner attributes that
377 // we insert will get removed - when we drop the parser, we'll free
378 // up the memory used by any attributes that we didn't remove from the map.
384 /// Converts a flattened iterator of tokens (including open and close delimiter tokens)
385 /// into a `TokenStream`, creating a `TokenTree::Delimited` for each matching pair
386 /// of open and close delims.
387 fn make_token_stream(
388 mut iter: impl Iterator<Item = (FlatToken, Spacing)>,
389 break_last_token: bool,
390 ) -> AttrTokenStream {
393 // This is `None` for the first frame, `Some` for all others.
394 open_delim_sp: Option<(Delimiter, Span)>,
395 inner: Vec<AttrTokenTree>,
397 let mut stack = vec![FrameData { open_delim_sp: None, inner: vec![] }];
398 let mut token_and_spacing = iter.next();
399 while let Some((token, spacing)) = token_and_spacing {
401 FlatToken::Token(Token { kind: TokenKind::OpenDelim(delim), span }) => {
402 stack.push(FrameData { open_delim_sp: Some((delim, span)), inner: vec![] });
404 FlatToken::Token(Token { kind: TokenKind::CloseDelim(delim), span }) => {
405 let frame_data = stack
407 .unwrap_or_else(|| panic!("Token stack was empty for token: {:?}", token));
409 let (open_delim, open_sp) = frame_data.open_delim_sp.unwrap();
412 "Mismatched open/close delims: open={:?} close={:?}",
415 let dspan = DelimSpan::from_pair(open_sp, span);
416 let stream = AttrTokenStream::new(frame_data.inner);
417 let delimited = AttrTokenTree::Delimited(dspan, delim, stream);
421 panic!("Bottom token frame is missing for token: {:?}", token)
426 FlatToken::Token(token) => stack
428 .expect("Bottom token frame is missing!")
430 .push(AttrTokenTree::Token(token, spacing)),
431 FlatToken::AttrTarget(data) => stack
433 .expect("Bottom token frame is missing!")
435 .push(AttrTokenTree::Attributes(data)),
436 FlatToken::Empty => {}
438 token_and_spacing = iter.next();
440 let mut final_buf = stack.pop().expect("Missing final buf!");
441 if break_last_token {
442 let last_token = final_buf.inner.pop().unwrap();
443 if let AttrTokenTree::Token(last_token, spacing) = last_token {
444 let unglued_first = last_token.kind.break_two_token_op().unwrap().0;
446 // An 'unglued' token is always two ASCII characters
447 let mut first_span = last_token.span.shrink_to_lo();
448 first_span = first_span.with_hi(first_span.lo() + rustc_span::BytePos(1));
452 .push(AttrTokenTree::Token(Token::new(unglued_first, first_span), spacing));
454 panic!("Unexpected last token {:?}", last_token)
457 AttrTokenStream::new(final_buf.inner)
460 // Some types are used a lot. Make sure they don't unintentionally get bigger.
461 #[cfg(all(target_arch = "x86_64", target_pointer_width = "64"))]
464 use rustc_data_structures::static_assert_size;
465 // tidy-alphabetical-start
466 static_assert_size!(AttrWrapper, 16);
467 static_assert_size!(LazyAttrTokenStreamImpl, 144);
468 // tidy-alphabetical-end