Skip to content

5.9: [MoveOnlyAddressChecker] Fix representation for used fields. #66738

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Conversation

nate-chandler
Copy link
Contributor

Description: Fix a class of miscompiles in the move-only address checker.

The move-only address checker records which instructions use ranges of fields of a move-only value. Previously, it recorded instructions with uses and the used fields via a map from instructions to ranges of fields of the value. But a single instruction can use non-gontiguous fields of the value being checked.

The fix is to change the the value stored corresponding to an instruction to have enough information to store all the non-contiguous fields of the value that it uses. Here, the representation of a bit vector is implemented.
Risk: Low. The change is involved, but the previous representation was wrong.
Scope: Narrow. This only affects move-only types.
Original PR: #66728
Reviewed By: Andrew Trick ( @atrick )
Testing: Added test that the move-only address checker previously miscompiled.
Resolves: rdar://110676577

According to language rules, such lifetimes are fixed and the relative
order of their deinits is guaranteed.

rdar://110913116
The members were declared but undefined.
Its storage vector is intended to be of some type like
`std::vector<std::pair<Key, Optional<Value>>>`, i.e., some collection of
pairs whose `second` is an `Optional<Value>`.  So when constructing a
default instance of that pair, just construct an Optional in the None
case.
FieldSensitivePrunedLiveness is used as a vectorization of
PrunedLiveness.  An instance of FSPL with N elements needs to be able to
represent the same states as N instances of PL.

Previously, it failed to do that in two significant ways:

(1) It attempted to save space for which elements were live by using
    a range.  This failed to account for instructions which are users of
    non-contiguous fields of an aggregate.

    apply(
      @owned (struct_element_addr %s, #S.f1),
      @owned (struct_element_addr %s, #S.f3)
    )

(2) It used a single bit to represent whether the instruction was
    consuming.  This failed to account for instructions which consumed
    some fields and borrowed others.

    apply(
      @owned (struct_element_addr %s, #S.f1),
      @guaranteed (struct_element_addr %s, #S.f2)
    )

The fix for (1) is to use a bit vector to represent which elements
are used by the instruction.  The fix for (2) is to use a second bit
vector to represent which elements are _consumed_ by the instruction.

Adapted the move-checker to use the new representation.

rdar://110909290
@nate-chandler nate-chandler requested a review from a team as a code owner June 17, 2023 21:55
Dumped more info and called llvm_unreachable on bad state.
Previously, the checker inserted destroys after each last use.  Here,
extend the lifetimes of fields as far as possible within their original
(unchecked) limits.

rdar://99681073
It's always the first line of the function, so try to do better.
Passing

```
-Xllvm -move-only-address-checker-disable-lifetime-extension=true
```

will skip the maximization of unconsumed field lifetimes.
@nate-chandler nate-chandler force-pushed the cherrypick/release/5.9/rdar110676577_2 branch from 90ab786 to 882ae97 Compare June 17, 2023 22:01
@nate-chandler nate-chandler requested review from atrick and tbkka June 17, 2023 22:02
The address checker records uses in its livenessUses map.  Previously,
that map mapped from an instruction to a range of fields of the type.
But an instruction can use multiple discontiguous fields of a single
value.  Here, such instructions are properly recorded by fixing the map
to store a bit vector for each instruction.

rdar://110676577
@nate-chandler nate-chandler force-pushed the cherrypick/release/5.9/rdar110676577_2 branch from 882ae97 to bff2838 Compare June 17, 2023 23:05
@nate-chandler
Copy link
Contributor Author

@swift-ci please test

@nate-chandler nate-chandler merged commit 40abe85 into swiftlang:release/5.9 Jun 19, 2023
@nate-chandler nate-chandler deleted the cherrypick/release/5.9/rdar110676577_2 branch June 19, 2023 17:44
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants