[AMDGPU] HIP miscompilation bug (aggregate assignment of nested unions)
Open
backend:AMDGPU
miscompilation
- Dominant language
- LLVM
- Stars
- 40.5k
- Forks
- 18.7k
- PR merge metrics
- PR metrics pending
Description
The following HIP program triggers a miscompilation bug. The output should be a deterministic value as it's UB free. The bug manifests by enabling optimisations which cause a divergence in the printed value.
Test File:
[HIPProg.hip](https://gist.github.com/ElliotBLabs/e29fa8bb31c09268321e482de9cbdd2e)
How to reproduce:
1. `clang++ -O0 -x hip HIPProg.hip -Wno-c++11-narrowing --offload-arch=native -o O0.out`
2. `clang++ -O1 -x hip HIPProg.hip -Wno-c++11-narrowing --offload-arch=native -o O1.out`
3. `./O0.out `
4. `./O1.out`
Compare output values:
Bug (-O1): 589824
Ref (-O0): 0
Contributor guide
Assessment
This issue has not been assessed yet.