[clang] clang/AMDGPU: Fix test to work with -lgomp (PR #220331)

[Jordan Rupprecht via cfe-commits]

rupprecht wrote:

> > e.g., '-fopenmp=libomp' or '-fopenmp=libiomp5'" if it is built with -lgomp. This adds a default value (libomp) for the -fopenmp flag to prevent this test failure.
> 
> I'm not sure what this means

Not OP but I've seen this sort of failure a few times before so I can mostly explain it.

The issue is that `-fopenmp` as a bare flag does not have a stable default; it defaults to what's been configured via CMake. The default choice in CMake, when unconfigured, is `libomp`. So in a vanilla configuration, `-fopenmp` is equivalent to `-fopenmp=libomp`. However, CMake can be configured to default to `libgomp`, i.e. `-fopenmp` is equivalent to `-fopenmp=libgomp`

Some flags are incomptible w/ `-fopenmp=libgomp`. I guess `-fopenmp-targets` is one of them. I don't see `-fopenmp-targets` mentioned in the test case itself, but I assume this could be implied by another flag (maybe `--offload-arch=`). In which case this test only works in vanilla build configs where the default is `-fopenmp=libomp`, and this PR is just setting that explicitly so the test passes when `-fopenmp=libgomp` is the default instead.

https://github.com/llvm/llvm-project/pull/220331