[llvm] [AMDGPU] IGLP: Fix static variables (PR #137549)

[Frederik Harwath via llvm-commits]

https://github.com/frederik-h commented:

I am not sure that the performance improvements achievable by the caching are worth the added code complexity. In fact, I did a quick experiment and I changed the baseline `MFMASmallGemmSingleWaveOpt::applyIGLPStrategy` implementation to always recompute the analysis data and compared the time spent in the scheduler of that build with the time using a build of this PR. For this I used different IR files from a build of project that uses "IGLP_OPT 1" (composable_kernel). Recomputing was consistently faster for me. Perhaps that's different for the other strategy or perhaps I made a mistake in my ad hoc benchmarking, but in any case I would suggest that you also try to benchmark if you haven't already. I think @jrbyrnes has originally implemented those strategies. @jrbyrnes Do you think the caching is crucial? @ro-i If you do not see a significant performance advantage obtained by the caching, perhaps you could fix the correctness issue first (recomputing the analysis data) and then try to improve the performance?

https://github.com/llvm/llvm-project/pull/137549