[Offload][amdgpu] Regression in OpenMP unit tests

[jplehr]

#83131 caused four test failures that are XFAIL via #98899

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

Author: Jan Patrick Lehr (jplehr)

https://github.com//pull/83131 caused four test failures that are `XFAIL` via https://github.com//pull/98899

[llvmbot]

@llvm/issue-subscribers-offload

Author: Jan Patrick Lehr (jplehr)

https://github.com//pull/83131 caused four test failures that are `XFAIL` via https://github.com//pull/98899

[arsenm]

Can you attach bitcode from the before/after failing cases with the invocation

[jhuber6]

Looking at the tests, this might just be yet another case of #64863, It fails with the same message. I'll try to get some IR.

[arsenm]

Looking at the tests, this might just be yet another case of #64863, It fails with the same message. I'll try to get some IR.

It is, but that shouldn't have really changed

[jhuber6]

Can you attach bitcode from the before/after failing cases with the invocation

https://godbolt.org/z/fGMfvdYo3 is the IR for one of the failing tests. Doing llvm-reduce on it just nets the same IR as the existing issue, so it's definitely the same thing. Normally the LTO pass pipeline would run at O2 by default which would at least do DCE even if all the functions had noopt on them from -O0. I wonder what your patch changed that resulted in these not being dropped.