compiler/rustc_parse/src/parser/attr_wrapper.rs

   1 use super::{Capturing, FlatToken, ForceCollect, Parser, ReplaceRange, TokenCursor, TrailingToken};
   2 use rustc_ast::token::{self, Delimiter, 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::{AttrVec, Attribute, HasAttrs, HasTokens};
   7 use rustc_errors::PResult;
   8 use rustc_span::{sym, Span};
   9
  10 use std::convert::TryInto;
  11 use std::ops::Range;
  12
  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.
  17 ///
  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.
  23 ///
  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 {
  28     attrs: AttrVec,
  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.
  32     start_pos: usize,
  33 }
  34
  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);
  39
  40 impl AttrWrapper {
  41     pub(super) fn new(attrs: AttrVec, start_pos: usize) -> AttrWrapper {
  42         AttrWrapper { attrs, start_pos }
  43     }
  44     pub fn empty() -> AttrWrapper {
  45         AttrWrapper { attrs: AttrVec::new(), start_pos: usize::MAX }
  46     }
  47     // FIXME: Delay span bug here?
  48     pub(crate) fn take_for_recovery(self) -> AttrVec {
  49         self.attrs
  50     }
  51
  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);
  57     }
  58
  59     pub fn is_empty(&self) -> bool {
  60         self.attrs.is_empty()
  61     }
  62
  63     pub fn maybe_needs_tokens(&self) -> bool {
  64         crate::parser::attr::maybe_needs_tokens(&self.attrs)
  65     }
  66 }
  67
  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)
  75     })
  76 }
  77
  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.
  86 //
  87 // This also makes `Parser` very cheap to clone, since
  88 // there is no intermediate collection buffer to clone.
  89 #[derive(Clone)]
  90 struct LazyTokenStreamImpl {
  91     start_token: (Token, Spacing),
  92     cursor_snapshot: TokenCursor,
  93     num_calls: usize,
  94     break_last_token: bool,
  95     replace_ranges: Box<[ReplaceRange]>,
  96 }
  97
  98 #[cfg(all(target_arch = "x86_64", target_pointer_width = "64"))]
  99 rustc_data_structures::static_assert_size!(LazyTokenStreamImpl, 144);
 100
 101 impl CreateTokenStream for LazyTokenStreamImpl {
 102     fn create_token_stream(&self) -> AttrAnnotatedTokenStream {
 103         // The token produced by the final call to `{,inlined_}next` was not
 104         // actually consumed by the callback. The combination of chaining the
 105         // initial token and using `take` produces the desired result - we
 106         // produce an empty `TokenStream` if no calls were made, and omit the
 107         // final token otherwise.
 108         let mut cursor_snapshot = self.cursor_snapshot.clone();
 109         let tokens =
 110             std::iter::once((FlatToken::Token(self.start_token.0.clone()), self.start_token.1))
 111                 .chain((0..self.num_calls).map(|_| {
 112                     let token = cursor_snapshot.next(cursor_snapshot.desugar_doc_comments);
 113                     (FlatToken::Token(token.0), token.1)
 114                 }))
 115                 .take(self.num_calls);
 116
 117         if !self.replace_ranges.is_empty() {
 118             let mut tokens: Vec<_> = tokens.collect();
 119             let mut replace_ranges = self.replace_ranges.clone();
 120             replace_ranges.sort_by_key(|(range, _)| range.start);
 121
 122             #[cfg(debug_assertions)]
 123             {
 124                 for [(range, tokens), (next_range, next_tokens)] in replace_ranges.array_windows() {
 125                     assert!(
 126                         range.end <= next_range.start || range.end >= next_range.end,
 127                         "Replace ranges should either be disjoint or nested: ({:?}, {:?}) ({:?}, {:?})",
 128                         range,
 129                         tokens,
 130                         next_range,
 131                         next_tokens,
 132                     );
 133                 }
 134             }
 135
 136             // Process the replace ranges, starting from the highest start
 137             // position and working our way back. If have tokens like:
 138             //
 139             // `#[cfg(FALSE)]` struct Foo { #[cfg(FALSE)] field: bool }`
 140             //
 141             // Then we will generate replace ranges for both
 142             // the `#[cfg(FALSE)] field: bool` and the entire
 143             // `#[cfg(FALSE)]` struct Foo { #[cfg(FALSE)] field: bool }`
 144             //
 145             // By starting processing from the replace range with the greatest
 146             // start position, we ensure that any replace range which encloses
 147             // another replace range will capture the *replaced* tokens for the inner
 148             // range, not the original tokens.
 149             for (range, new_tokens) in replace_ranges.iter().rev() {
 150                 assert!(!range.is_empty(), "Cannot replace an empty range: {:?}", range);
 151                 // Replace ranges are only allowed to decrease the number of tokens.
 152                 assert!(
 153                     range.len() >= new_tokens.len(),
 154                     "Range {:?} has greater len than {:?}",
 155                     range,
 156                     new_tokens
 157                 );
 158
 159                 // Replace any removed tokens with `FlatToken::Empty`.
 160                 // This keeps the total length of `tokens` constant throughout the
 161                 // replacement process, allowing us to use all of the `ReplaceRanges` entries
 162                 // without adjusting indices.
 163                 let filler = std::iter::repeat((FlatToken::Empty, Spacing::Alone))
 164                     .take(range.len() - new_tokens.len());
 165
 166                 tokens.splice(
 167                     (range.start as usize)..(range.end as usize),
 168                     new_tokens.clone().into_iter().chain(filler),
 169                 );
 170             }
 171             make_token_stream(tokens.into_iter(), self.break_last_token)
 172         } else {
 173             make_token_stream(tokens, self.break_last_token)
 174         }
 175     }
 176 }
 177
 178 impl<'a> Parser<'a> {
 179     /// Records all tokens consumed by the provided callback,
 180     /// including the current token. These tokens are collected
 181     /// into a `LazyTokenStream`, and returned along with the result
 182     /// of the callback.
 183     ///
 184     /// Note: If your callback consumes an opening delimiter
 185     /// (including the case where you call `collect_tokens`
 186     /// when the current token is an opening delimiter),
 187     /// you must also consume the corresponding closing delimiter.
 188     ///
 189     /// That is, you can consume
 190     /// `something ([{ }])` or `([{}])`, but not `([{}]`
 191     ///
 192     /// This restriction shouldn't be an issue in practice,
 193     /// since this function is used to record the tokens for
 194     /// a parsed AST item, which always has matching delimiters.
 195     pub fn collect_tokens_trailing_token<R: HasAttrs + HasTokens>(
 196         &mut self,
 197         attrs: AttrWrapper,
 198         force_collect: ForceCollect,
 199         f: impl FnOnce(&mut Self, Vec<ast::Attribute>) -> PResult<'a, (R, TrailingToken)>,
 200     ) -> PResult<'a, R> {
 201         // We only bail out when nothing could possibly observe the collected tokens:
 202         // 1. We cannot be force collecting tokens (since force-collecting requires tokens
 203         //    by definition
 204         if matches!(force_collect, ForceCollect::No)
 205             // None of our outer attributes can require tokens (e.g. a proc-macro)
 206             && !attrs.maybe_needs_tokens()
 207             // If our target supports custom inner attributes, then we cannot bail
 208             // out early, since we may need to capture tokens for a custom inner attribute
 209             // invocation.
 210             && !R::SUPPORTS_CUSTOM_INNER_ATTRS
 211             // Never bail out early in `capture_cfg` mode, since there might be `#[cfg]`
 212             // or `#[cfg_attr]` attributes.
 213             && !self.capture_cfg
 214         {
 215             return Ok(f(self, attrs.attrs.into())?.0);
 216         }
 217
 218         let start_token = (self.token.clone(), self.token_spacing);
 219         let cursor_snapshot = self.token_cursor.clone();
 220
 221         let has_outer_attrs = !attrs.attrs.is_empty();
 222         let prev_capturing = std::mem::replace(&mut self.capture_state.capturing, Capturing::Yes);
 223         let replace_ranges_start = self.capture_state.replace_ranges.len();
 224
 225         let ret = f(self, attrs.attrs.into());
 226
 227         self.capture_state.capturing = prev_capturing;
 228
 229         let (mut ret, trailing) = ret?;
 230
 231         // When we're not in `capture-cfg` mode, then bail out early if:
 232         // 1. Our target doesn't support tokens at all (e.g we're parsing an `NtIdent`)
 233         //    so there's nothing for us to do.
 234         // 2. Our target already has tokens set (e.g. we've parsed something
 235         // like `#[my_attr] $item`. The actual parsing code takes care of prepending
 236         // any attributes to the nonterminal, so we don't need to modify the
 237         // already captured tokens.
 238         // Note that this check is independent of `force_collect`- if we already
 239         // have tokens, or can't even store them, then there's never a need to
 240         // force collection of new tokens.
 241         if !self.capture_cfg && matches!(ret.tokens_mut(), None | Some(Some(_))) {
 242             return Ok(ret);
 243         }
 244
 245         // This is very similar to the bail out check at the start of this function.
 246         // Now that we've parsed an AST node, we have more information available.
 247         if matches!(force_collect, ForceCollect::No)
 248             // We now have inner attributes available, so this check is more precise
 249             // than `attrs.maybe_needs_tokens()` at the start of the function.
 250             // As a result, we don't need to check `R::SUPPORTS_CUSTOM_INNER_ATTRS`
 251             && !crate::parser::attr::maybe_needs_tokens(ret.attrs())
 252             // Subtle: We call `has_cfg_or_cfg_attr` with the attrs from `ret`.
 253             // This ensures that we consider inner attributes (e.g. `#![cfg]`),
 254             // which require us to have tokens available
 255             // We also call `has_cfg_or_cfg_attr` at the beginning of this function,
 256             // but we only bail out if there's no possibility of inner attributes
 257             // (!R::SUPPORTS_CUSTOM_INNER_ATTRS)
 258             // We only capture about `#[cfg]` or `#[cfg_attr]` in `capture_cfg`
 259             // mode - during normal parsing, we don't need any special capturing
 260             // for those attributes, since they're builtin.
 261             && !(self.capture_cfg && has_cfg_or_cfg_attr(ret.attrs()))
 262         {
 263             return Ok(ret);
 264         }
 265
 266         let mut inner_attr_replace_ranges = Vec::new();
 267         // Take the captured ranges for any inner attributes that we parsed.
 268         for inner_attr in ret.attrs().iter().filter(|a| a.style == ast::AttrStyle::Inner) {
 269             if let Some(attr_range) = self.capture_state.inner_attr_ranges.remove(&inner_attr.id) {
 270                 inner_attr_replace_ranges.push(attr_range);
 271             } else {
 272                 self.sess
 273                     .span_diagnostic
 274                     .delay_span_bug(inner_attr.span, "Missing token range for attribute");
 275             }
 276         }
 277
 278         let replace_ranges_end = self.capture_state.replace_ranges.len();
 279
 280         let cursor_snapshot_next_calls = cursor_snapshot.num_next_calls;
 281         let mut end_pos = self.token_cursor.num_next_calls;
 282
 283         // Capture a trailing token if requested by the callback 'f'
 284         match trailing {
 285             TrailingToken::None => {}
 286             TrailingToken::Semi => {
 287                 assert_eq!(self.token.kind, token::Semi);
 288                 end_pos += 1;
 289             }
 290             TrailingToken::MaybeComma => {
 291                 if self.token.kind == token::Comma {
 292                     end_pos += 1;
 293                 }
 294             }
 295         }
 296
 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 `LazyTokenStream` gets converted
 300         // into an `AttrAnnotatedTokenStream`, we will create the proper token.
 301         if self.token_cursor.break_last_token {
 302             assert_eq!(
 303                 trailing,
 304                 TrailingToken::None,
 305                 "Cannot set `break_last_token` and have trailing token"
 306             );
 307             end_pos += 1;
 308         }
 309
 310         let num_calls = end_pos - cursor_snapshot_next_calls;
 311
 312         // If we have no attributes, then we will never need to
 313         // use any replace ranges.
 314         let replace_ranges: Box<[ReplaceRange]> = if ret.attrs().is_empty() && !self.capture_cfg {
 315             Box::new([])
 316         } else {
 317             // Grab any replace ranges that occur *inside* the current AST node.
 318             // We will perform the actual replacement when we convert the `LazyTokenStream`
 319             // to an `AttrAnnotatedTokenStream`
 320             let start_calls: u32 = cursor_snapshot_next_calls.try_into().unwrap();
 321             self.capture_state.replace_ranges[replace_ranges_start..replace_ranges_end]
 322                 .iter()
 323                 .cloned()
 324                 .chain(inner_attr_replace_ranges.clone().into_iter())
 325                 .map(|(range, tokens)| {
 326                     ((range.start - start_calls)..(range.end - start_calls), tokens)
 327                 })
 328                 .collect()
 329         };
 330
 331         let tokens = LazyTokenStream::new(LazyTokenStreamImpl {
 332             start_token,
 333             num_calls,
 334             cursor_snapshot,
 335             break_last_token: self.token_cursor.break_last_token,
 336             replace_ranges,
 337         });
 338
 339         // If we support tokens at all
 340         if let Some(target_tokens) = ret.tokens_mut() {
 341             if target_tokens.is_none() {
 342                 // Store se our newly captured tokens into the AST node
 343                 *target_tokens = Some(tokens.clone());
 344             }
 345         }
 346
 347         let final_attrs = ret.attrs();
 348
 349         // If `capture_cfg` is set and we're inside a recursive call to
 350         // `collect_tokens_trailing_token`, then we need to register a replace range
 351         // if we have `#[cfg]` or `#[cfg_attr]`. This allows us to run eager cfg-expansion
 352         // on the captured token stream.
 353         if self.capture_cfg
 354             && matches!(self.capture_state.capturing, Capturing::Yes)
 355             && has_cfg_or_cfg_attr(&final_attrs)
 356         {
 357             let attr_data = AttributesData { attrs: final_attrs.to_vec().into(), tokens };
 358
 359             // Replace the entire AST node that we just parsed, including attributes,
 360             // with a `FlatToken::AttrTarget`. If this AST node is inside an item
 361             // that has `#[derive]`, then this will allow us to cfg-expand this
 362             // AST node.
 363             let start_pos =
 364                 if has_outer_attrs { attrs.start_pos } else { cursor_snapshot_next_calls };
 365             let new_tokens = vec![(FlatToken::AttrTarget(attr_data), Spacing::Alone)];
 366
 367             assert!(
 368                 !self.token_cursor.break_last_token,
 369                 "Should not have unglued last token with cfg attr"
 370             );
 371             let range: Range<u32> = (start_pos.try_into().unwrap())..(end_pos.try_into().unwrap());
 372             self.capture_state.replace_ranges.push((range, new_tokens));
 373             self.capture_state.replace_ranges.extend(inner_attr_replace_ranges);
 374         }
 375
 376         // Only clear our `replace_ranges` when we're finished capturing entirely.
 377         if matches!(self.capture_state.capturing, Capturing::No) {
 378             self.capture_state.replace_ranges.clear();
 379             // We don't clear `inner_attr_ranges`, as doing so repeatedly
 380             // had a measurable performance impact. Most inner attributes that
 381             // we insert will get removed - when we drop the parser, we'll free
 382             // up the memory used by any attributes that we didn't remove from the map.
 383         }
 384         Ok(ret)
 385     }
 386 }
 387
 388 /// Converts a flattened iterator of tokens (including open and close delimiter tokens)
 389 /// into a `TokenStream`, creating a `TokenTree::Delimited` for each matching pair
 390 /// of open and close delims.
 391 fn make_token_stream(
 392     mut iter: impl Iterator<Item = (FlatToken, Spacing)>,
 393     break_last_token: bool,
 394 ) -> AttrAnnotatedTokenStream {
 395     #[derive(Debug)]
 396     struct FrameData {
 397         // This is `None` for the first frame, `Some` for all others.
 398         open_delim_sp: Option<(Delimiter, Span)>,
 399         inner: Vec<(AttrAnnotatedTokenTree, Spacing)>,
 400     }
 401     let mut stack = vec![FrameData { open_delim_sp: None, inner: vec![] }];
 402     let mut token_and_spacing = iter.next();
 403     while let Some((token, spacing)) = token_and_spacing {
 404         match token {
 405             FlatToken::Token(Token { kind: TokenKind::OpenDelim(delim), span }) => {
 406                 stack.push(FrameData { open_delim_sp: Some((delim, span)), inner: vec![] });
 407             }
 408             FlatToken::Token(Token { kind: TokenKind::CloseDelim(delim), span }) => {
 409                 let frame_data = stack
 410                     .pop()
 411                     .unwrap_or_else(|| panic!("Token stack was empty for token: {:?}", token));
 412
 413                 let (open_delim, open_sp) = frame_data.open_delim_sp.unwrap();
 414                 assert_eq!(
 415                     open_delim, delim,
 416                     "Mismatched open/close delims: open={:?} close={:?}",
 417                     open_delim, span
 418                 );
 419                 let dspan = DelimSpan::from_pair(open_sp, span);
 420                 let stream = AttrAnnotatedTokenStream::new(frame_data.inner);
 421                 let delimited = AttrAnnotatedTokenTree::Delimited(dspan, delim, stream);
 422                 stack
 423                     .last_mut()
 424                     .unwrap_or_else(|| {
 425                         panic!("Bottom token frame is missing for token: {:?}", token)
 426                     })
 427                     .inner
 428                     .push((delimited, Spacing::Alone));
 429             }
 430             FlatToken::Token(token) => stack
 431                 .last_mut()
 432                 .expect("Bottom token frame is missing!")
 433                 .inner
 434                 .push((AttrAnnotatedTokenTree::Token(token), spacing)),
 435             FlatToken::AttrTarget(data) => stack
 436                 .last_mut()
 437                 .expect("Bottom token frame is missing!")
 438                 .inner
 439                 .push((AttrAnnotatedTokenTree::Attributes(data), spacing)),
 440             FlatToken::Empty => {}
 441         }
 442         token_and_spacing = iter.next();
 443     }
 444     let mut final_buf = stack.pop().expect("Missing final buf!");
 445     if break_last_token {
 446         let (last_token, spacing) = final_buf.inner.pop().unwrap();
 447         if let AttrAnnotatedTokenTree::Token(last_token) = last_token {
 448             let unglued_first = last_token.kind.break_two_token_op().unwrap().0;
 449
 450             // An 'unglued' token is always two ASCII characters
 451             let mut first_span = last_token.span.shrink_to_lo();
 452             first_span = first_span.with_hi(first_span.lo() + rustc_span::BytePos(1));
 453
 454             final_buf.inner.push((
 455                 AttrAnnotatedTokenTree::Token(Token::new(unglued_first, first_span)),
 456                 spacing,
 457             ));
 458         } else {
 459             panic!("Unexpected last token {:?}", last_token)
 460         }
 461     }
 462     assert!(stack.is_empty(), "Stack should be empty: final_buf={:?} stack={:?}", final_buf, stack);
 463     AttrAnnotatedTokenStream::new(final_buf.inner)
 464 }