[Util] Fix intermittent crash in OptimizeIntArithmeticPass #20637
Conversation
This file contains hidden or 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
Signed-off-by: Jorn Tuyls <jorn.tuyls@gmail.com>
krzysz00
approved these changes
Apr 25, 2025
Yeah, I was concerned about that as well and visually compared the compilation time of the previously failing test case (toy_llama) and it didn't feel any different. And the sharktank CI test wall clock times seem to be similar as well. |
|
Can you merge this or should I? |
KyleHerndon
pushed a commit
to KyleHerndon/iree
that referenced
this pull request
May 7, 2025
…20637) We're seeing intermittent failures in 'sharktank_model_tests :: rocm_hip_w7900' tests on main: - https://github.com/iree-org/iree/actions/runs/14600283542/job/40957129060?pr=20601 - https://github.com/iree-org/iree/actions/runs/14636457142/job/41069048559 - https://github.com/iree-org/iree/actions/runs/14630167718/job/41051068642 The following issue contains more information on the crash and a reproducer: iree-org#20636 I am not entirely sure whether this PR contains a fix or a workaround. On one hand, it seems like correctly updating and invalidating all states manually in the transformations should lead to the same result without needing to recompute all states. On the other hand, the `DataFlowSolver` mentions that all analysis states should be erased when IR changes before a re-run (https://github.com/llvm/llvm-project/blob/12a31658ea36cda74157c6b4e6b6c031e39a19c0/mlir/include/mlir/Analysis/DataFlowFramework.h#L320): > /// Steps to re-run a data-flow analysis when IR changes: /// 1. Erase all analysis states as they are no longer valid. /// 2. Re-run the analysis using `initializeAndRun`. Note: I didn't add a test because I don't have a deterministic reproducer. As the crash is intermittent, the reproducer needs to be run a large number of times and will fail once in 10-200 runs maybe. Signed-off-by: Jorn Tuyls <jorn.tuyls@gmail.com>
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.
We're seeing intermittent failures in 'sharktank_model_tests :: rocm_hip_w7900' tests on main:
The following issue contains more information on the crash and a reproducer: #20636
I am not entirely sure whether this PR contains a fix or a workaround. On one hand, it seems like correctly updating and invalidating all states manually in the transformations should lead to the same result without needing to recompute all states. On the other hand, the
DataFlowSolvermentions that all analysis states should be erased when IR changes before a re-run (https://github.com/llvm/llvm-project/blob/12a31658ea36cda74157c6b4e6b6c031e39a19c0/mlir/include/mlir/Analysis/DataFlowFramework.h#L320):Note: I didn't add a test because I don't have a deterministic reproducer. As the crash is intermittent, the reproducer needs to be run a large number of times and will fail once in 10-200 runs maybe.