[ARM] Fix infinite recursion in unwind #134387
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/134387
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 1de64b9 with merge base 2553278 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
nWEIdia
left a comment
nWEIdia
left a comment
There was a problem hiding this comment.
Geniously speed of light forward fix!
Aidyn-A
added
ciflow/trunk
|
@pytorchbot merge |
Merge failedReason: This PR needs a If not, please add the To add a label, you can comment to pytorchbot, for example For more information, see Details for Dev Infra teamRaised by workflow job |
|
Perhaps as a follow up, we can remove IS_MACOS and the inductor check line? |
|
All tests are green. @pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Fixes pytorch#119905 The `TORCH_SHOW_CPP_STACKTRACES=1` setting on ARM causes infinite recursive unwind because on failure a `StackTraceFetcher` attempts to unwind the <ins>failed instruction</ins>: https://github.com/pytorch/pytorch/blob/5ad759ca33ba8299cf7e1a6bb1dff7c9a5555e29/torch/csrc/profiler/combined_traceback.cpp#L25 then the unwind itself fails: https://github.com/pytorch/pytorch/blob/5ad759ca33ba8299cf7e1a6bb1dff7c9a5555e29/torch/csrc/profiler/unwind/unwind.cpp#L10-L12 and it causes another attempt to unwind the failure in `unwind()`... In summary, the executed instruction is equivalent to: ```C++ std::vector<void*> unwind() { // some instructions ... return unwind(); } ``` This PR replaces `TORCH_CHECK` by `TORCH_WARN_ONCE` as it will not cause an uncontrolled recursion. The only side effect would be an empty back-trace. Huge thanks to @nWEIdia who found the root cause! Pull Request resolved: pytorch#134387 Approved by: https://github.com/eqy, https://github.com/nWEIdia, https://github.com/malfet
6 participants
Fixes #119905
The
TORCH_SHOW_CPP_STACKTRACES=1setting on ARM causes infinite recursive unwind because on failure aStackTraceFetcherattempts to unwind the failed instruction:pytorch/torch/csrc/profiler/combined_traceback.cpp
Line 25 in 5ad759c
then the unwind itself fails:
pytorch/torch/csrc/profiler/unwind/unwind.cpp
Lines 10 to 12 in 5ad759c
and it causes another attempt to unwind the failure in
unwind()...In summary, the executed instruction is equivalent to:
This PR replaces
TORCH_CHECKbyTORCH_WARN_ONCEas it will not cause an uncontrolled recursion. The only side effect would be an empty back-trace.Huge thanks to @nWEIdia who found the root cause!
cc @malfet @snadampal @milpuz01