In a case like this:
```
mod a {
mod b {
#[cfg_attr(unix, inline)]
fn f() {
#[cfg_attr(linux, inline)]
fn g1() {}
#[cfg_attr(linux, inline)]
fn g2() {}
}
}
}
```
We currently end up with the following replacement ranges.
- The lazy tokens for `f` has replacement ranges for `g1` and `g2`.
- The lazy tokens for `a` has replacement ranges for `f`, `g1`, and
`g2`.
I.e. the replacement ranges for `g1` and `g2` are duplicated. In
general, replacement ranges for inner AST nodes are duplicated up the
chain for each nested `collect_tokens` call. And the code that processes
the replacements is careful about the ordering in which the replacements
are applied, to ensure that inner replacements are applied before outer
replacements.
But all of this is unnecessary. If you apply an inner replacement and
then an outer replacement, the outer replacement completely overwrites
the inner replacement.
This commit avoids the duplication by removing replacements from
`self.capture_state.parser_replacements` when they are used. (The effect
on the example above is that the lazy tokesn for `a` no longer include
replacement ranges for `g1` and `g2`.) This eliminates the possibility
of nested replacements on individual AST nodes, which avoids the need
for careful ordering of replacements.
This example triggers an assertion failure:
```
fn f() -> u32 {
#[cfg_eval] #[cfg(not(FALSE))] 0
}
```
The sequence of events:
- `configure_annotatable` calls `parse_expr_force_collect`, which calls
`collect_tokens`.
- Within that, we end up in `parse_expr_dot_or_call`, which again calls
`collect_tokens`.
- The return value of the `f` call is the expression `0`.
- This inner call collects tokens for `0` (parser range 10..11) and
creates a replacement covering `#[cfg(not(FALSE))] 0` (parser range
0..11).
- We return to the outer `collect_tokens` call. The return value of the
`f` call is *again* the expression `0`, again with the range 10..11,
but the replacement from earlier covers the range 0..11. The code
mistakenly assumes that any attributes from an inner `collect_tokens`
call fit entirely within the body of the result of an outer
`collect_tokens` call. So it adjusts the replacement parser range
0..11 to a node range by subtracting 10, resulting in -10..1. This is
an invalid range and triggers an assertion failure.
It's tricky to follow, but basically things get complicated when an AST
node is returned from an inner `collect_tokens` call and then returned
again from an outer `collect_token` node without being wrapped in any
kind of additional layer.
This commit changes `collect_tokens` to return early in some extra cases,
avoiding the construction of lazy tokens. In the example above, the
outer `collect_tokens` returns earlier because the `0` token already has
tokens and `self.capture_state.capturing` is `Capturing::No`. This early
return avoids the creation of the invalid range and the assertion
failure.
Fixes#129166. Note: these invalid ranges have been happening for a long
time. #128725 looks like it's at fault only because it introduced the
assertion that catches the invalid ranges.
This commit does the following.
- Renames `collect_tokens_trailing_token` as `collect_tokens`, because
(a) it's annoying long, and (b) the `_trailing_token` bit is less
accurate now that its types have changed.
- In `collect_tokens`, adds a `Option<CollectPos>` argument and a
`UsePreAttrPos` in the return type of `f`. These are used in
`parse_expr_force_collect` (for vanilla expressions) and in
`parse_stmt_without_recovery` (for two different cases of expression
statements). Together these ensure are enough to fix all the problems
with token collection and assoc expressions. The changes to the
`stringify.rs` test demonstrate some of these.
- Adds a new test. The code in this test was causing an assertion
failure prior to this commit, due to an invalid `NodeRange`.
The extra complexity is annoying, but necessary to fix the existing
problems.
This pre-existing type is suitable for use with the return value of the
`f` parameter in `collect_tokens_trailing_token`. The more descriptive
name will be useful because the next commit will add another boolean
value to the return value of `f`.
When collecting tokens there are two kinds of range:
- a range relative to the parser's full token stream (which we get when
we are parsing);
- a range relative to a single AST node's token stream (which we use
within `LazyAttrTokenStreamImpl` when replacing tokens).
These are currently both represented with `Range<u32>` and it's easy to
mix them up -- until now I hadn't properly understood the difference.
This commit introduces `ParserRange` and `NodeRange` to distinguish
them. This also requires splitting `ReplaceRange` in two, giving the new
types `ParserReplacement` and `NodeReplacement`. (These latter two names
reduce the overloading of the word "range".)
The commit also rewrites some comments to be clearer.
The end result is a little more verbose, but much clearer.
Remove unnecessary range replacements
This PR removes an unnecessary range replacement in `collect_tokens_trailing_token`, and does a couple of other small cleanups.
r? ````@petrochenkov````
A fully imperative style is easier to read than a half-iterator,
half-imperative style. Also, rename `inner_attr` as `attr` because it
might be an outer attribute.
Imagine you have replace ranges (2..20,X) and (5..15,Y), and these tokens:
```
a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,q,r,s,t,u,v,w,x
```
If we replace (5..15,Y) first, then (2..20,X) we get this sequence
```
a,b,c,d,e,Y,_,_,_,_,_,_,_,_,_,p,q,r,s,t,u,v,w,x
a,b,X,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,u,v,w,x
```
which is what we want.
If we do it in the other order, we get this:
```
a,b,X,_,_,_,_,_,_,_,_,_,_,_,_,p,q,r,s,t,u,v,w,x
a,b,X,_,_,Y,_,_,_,_,_,_,_,_,_,_,_,_,_,_,u,v,w,x
```
which is wrong. So it's true that we need the `.rev()` but the comment
is wrong about why.
The current code is this:
```
self.capture_state.replace_ranges.push((start_pos..end_pos, Some(target)));
self.capture_state.replace_ranges.extend(inner_attr_replace_ranges);
```
What's not obvious is that every range in `inner_attr_replace_ranges`
must be a strict sub-range of `start_pos..end_pos`. Which means, in
`LazyAttrTokenStreamImpl::to_attr_token_stream`, they will be done
first, and then the `start_pos..end_pos` replacement will just overwrite
them. So they aren't needed.
This has been bugging me for a while. I find complex "if any of these
are true" conditions easier to think about than complex "if all of these
are true" conditions, because you can stop as soon as one is true.
Adding details, clarifying lots of little things, etc. In particular,
the commit adds details of an example. I find this very helpful, because
it's taken me a long time to understand how this code works.
The `Option`s within the `ReplaceRange`s within the hashmap are always
`None`. This PR omits them and inserts them when they are extracted from
the hashmap.
It's used in `Parser::collect_tokens_trailing_token` to decide whether
to capture a trailing token. But the callers actually know whether to
capture a trailing token, so it's simpler for them to just pass in a
bool.
Also, the `TrailingToken::Gt` case was weird, because it didn't result
in a trailing token being captured. It could have been subsumed by the
`TrailingToken::MaybeComma` case, and it effectively is in the new code.
The new condition is equivalent in practice, but it's much more obvious
that it would result in an empty range, because the condition lines up
with the contents of the iterator.
There's a comment saying we don't do it for performance reasons, but it
doesn't actually affect performance.
The commit also tweaks the control flow, to make clearer that two code
paths are mutually exclusive.
Currently the second element is a `Vec<(FlatToken, Spacing)>`. But the
vector always has zero or one elements, and the `FlatToken` is always
`FlatToken::AttrTarget` (which contains an `AttributesData`), and the
spacing is always `Alone`. So we can simplify it to
`Option<AttributesData>`.
An assertion in `to_attr_token_stream` can can also be removed, because
`new_tokens.len()` was always 0 or 1, which means than `range.len()`
is always greater than or equal to it, because `range.is_empty()` is
always false (as per the earlier assertion).
And update the comment. Clearly the return type of this function was
changed at some point in the past, but its name and comment weren't
updated to match.
The number of source code bytes can't exceed a `u32`'s range, so a token
position also can't. This reduces the size of `Parser` and
`LazyAttrTokenStreamImpl` by eight bytes each.
Fix duplicated attributes on nonterminal expressions
This PR fixes a long-standing bug (#86055) whereby expression attributes can be duplicated when expanded through declarative macros.
First, consider how items are parsed in declarative macros:
```
Items:
- parse_nonterminal
- parse_item(ForceCollect::Yes)
- parse_item_
- attrs = parse_outer_attributes
- parse_item_common(attrs)
- maybe_whole!
- collect_tokens_trailing_token
```
The important thing is that the parsing of outer attributes is outside token collection, so the item's tokens don't include the attributes. This is how it's supposed to be.
Now consider how expression are parsed in declarative macros:
```
Exprs:
- parse_nonterminal
- parse_expr_force_collect
- collect_tokens_no_attrs
- collect_tokens_trailing_token
- parse_expr
- parse_expr_res(None)
- parse_expr_assoc_with
- parse_expr_prefix
- parse_or_use_outer_attributes
- parse_expr_dot_or_call
```
The important thing is that the parsing of outer attributes is inside token collection, so the the expr's tokens do include the attributes, i.e. in `AttributesData::tokens`.
This PR fixes the bug by rearranging expression parsing to that outer attribute parsing happens outside of token collection. This requires a number of small refactorings because expression parsing is somewhat complicated. While doing so the PR makes the code a bit cleaner and simpler, by eliminating `parse_or_use_outer_attributes` and `Option<AttrWrapper>` arguments (in favour of the simpler `parse_outer_attributes` and `AttrWrapper` arguments), and simplifying `LhsExpr`.
r? `@petrochenkov`
Existing names for values of this type are `sess`, `parse_sess`,
`parse_session`, and `ps`. `sess` is particularly annoying because
that's also used for `Session` values, which are often co-located, and
it can be difficult to know which type a value named `sess` refers to.
(That annoyance is the main motivation for this change.) `psess` is nice
and short, which is good for a name used this much.
The commit also renames some `parse_sess_created` values as
`psess_created`.
