Revise dataflow to use middle::cfg flowgraph to drive analysis. #14873
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
:) |
|
Will this effectively close #13767, given that most (all?) things are driven off dataflow now? |
|
@pcwalton yes it would, IMO |
|
Great! |
|
(updated description with @pcwalton's insight.) |
|
Okay I have now added the tests that I think are reasonable to try to put in with this PR. The main remaining thing to test is the control-flow graph and borrowck behavior associated with |
|
This looks good. r+. Here are some suggested changes: |
|
Ah, one more comment -- the change to introduce LpCopiedUpvar seems to apply equally to all upvars. |
1. After recursively processing an ExprWhile, need to pop loop_scopes the same way we do for ExprLoop. 2. Proposed fix for flowgraph handling of ExprInlineAsm: we need to represent the flow into the subexpressions of an `asm!` block.
Fix rust-lang#6298. This is instead of the prior approach of emulating cfg traversal privately by traversing AST in same way). Of special note, this removes a special case handling of `ExprParen` that was actually injecting a bug (since it was acting like an expression like `(*func)()` was consuming `*func` *twice*: once from `(*func)` and again from `*func`). nikomatsakis was the first one to point out that it might suffice to simply have the outer `ExprParen` do the consumption of the contents (alone). (This version has been updated to incorporate feedback from Niko's review of PR 14873.)
Namely:
1. Now that cfg mod is used for dataflow, we do not need to turn on
the `allow(deadcode)` to placate the linter.
2. remove dead struct defn.
…ext.
Details: in a program like:
```
type T = proc(int) -> int; /* 4 */
pub fn outer(captured /* pat 16 */: T) -> T {
(proc(x /* pat 23 */) {
((captured /* 29 */).foo((x /* 30 */)) /* 28 */)
} /* block 27 */ /* 20 */)
} /* block 19 */ /* 12 */
```
the `captured` arg is moved from the outer fn into the inner proc (id=20).
The old dataflow analysis for flowed_move_data_moves, when looking at
the inner proc, would attempt to add a kill bit for `captured` at the
end of its scope; the problem is that it thought the end of the
`captured` arg's scope was the outer fn (id=12), even though at that
point in the analysis, the `captured` arg's scope should now be
restricted to the proc itself (id=20).
This patch fixes handling of upvars so that dataflow of a fn/proc
should never attempts to add a gen or kill bit to any NodeId outside
of the current fn/proc. It accomplishes this by adding an `LpUpvar`
variant to `borrowck::LoanPath`, so for cases like `captured` above
will carry both their original `var_id`, as before, as well as the
`NodeId` for the closure that is capturing them.
As a drive-by fix to another occurrence of a similar bug that
nikomatsakis pointed out to me earlier, this also fixes
`gather_loans::compute_kill_scope` so that it computes the kill scope
of the `captured` arg to be block 27; that is, the block for the proc
itself (id=20).
(This is an updated version that generalizes the new loan path variant
to cover all upvars, and thus renamed the variant from `LpCopiedUpvar`
to just `LpUpvar`.)
thereof.) PR 14739 injected the new message that this removes from one test case: borrowck-vec-pattern-loan-from-mut.rs When reviewing the test case, I was not able to convince myself that the error message was a legitimate thing to start emitting. Niko did not see an obvious reason for it either, so I am going to remove it and wait for someone (maybe Cameron Zwarich) to explain to me why we should be emitting it.
bors
added a commit
that referenced
this pull request
Jun 18, 2014
…tsakis Fix #6298. Fix #13767. This also includes some drive by fixes for some other issues, noted in the commits. I still need to integrate regression tests for some cases that I noticed were missing from our unit test suite (i.e. things that compiling rustc exposes that should have been exposed when doing `make check-stage1`). So do not land this yet, until I get the chance to add those tests. I just wanted to get the review process started soon, since this has been long in the coming.
|
Filed followup issue #15019 that I decided not to attempt to address before getting this landed. |
|
Filed followup issue #15020 that I did not want to attempt to address before getting this landed. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Fix #6298. Fix #13767.
This also includes some drive by fixes for some other issues, noted in the commits.
I still need to integrate regression tests for some cases that I noticed were missing from our unit test suite (i.e. things that compiling rustc exposes that should have been exposed when doing
make check-stage1). So do not land this yet, until I get the chance to add those tests.I just wanted to get the review process started soon, since this has been long in the coming.