[flang][hlfir] No regressions on BLAS/LAPACK

[vzakhari]

Part of https://discourse.llvm.org/t/rfc-enabling-the-hlfir-lowering-by-default/72778:

• @kiranchandramohan: BLAS/LAPACK should continue to work. We are in the process of checking where we are with this.

[kiranchandramohan]

HLFIR lowering looks OK for the BLAS/LAPACK version that we run downstream.
@h-vetinari @mmuetzel Would you like to see whether your changes for OpenBLAS (OpenMathLib/OpenBLAS#4016) work with HLFIR lowering (-flang-experimental-hlfir)?

[h-vetinari]

I'm still stuck trying to get flang 17 to work on windows (spread over a couple of issues; most are referenced from conda-forge/flang-feedstock#28). I could build off of main with some effort, but I first want to be able to get flang 17 to compile a hello world on windows at all (help would be much appreciated...). Once that's done, I can easily throw it at various of our code bases, with or without -flang-experimental-hlfir.

The OpenBLAS folks were further along with this already, perhaps @martin-frbg (or @mmuetzel) could retry building OpenBLAS using a flang build off of main with -flang-experimental-hlfir.

[martin-frbg]

No difference in the LAPACK testsuite results from 17RC3 with and without -flang-experimental-hlfir on Skylake.

(There does however appear to be a regression in LLVM17's clang compilation of the AVX512 kernels for BLAS1 CASUM/ZASUM that looks similar to what Nvidia engineers found from their nvc OpenMathLib/OpenBLAS#4162 (comment) )

[kiranchandramohan]

Thanks @martin-frbg @h-vetinari. I am closing this ticket as completed.

[h-vetinari]

Thanks @martin-frbg!

There does however appear to be a regression in LLVM17's clang compilation of the AVX512 kernels for BLAS1 CASUM/ZASUM that looks similar to what Nvidia engineers found from their nvc xianyi/OpenBLAS#4162 (comment)

Worth another issue perhaps?

[martin-frbg]

Thanks @martin-frbg!

There does however appear to be a regression in LLVM17's clang compilation of the AVX512 kernels for BLAS1 CASUM/ZASUM that looks similar to what Nvidia engineers found from their nvc xianyi/OpenBLAS#4162 (comment)

Worth another issue perhaps?

Just for the record - this was now found to be caused by unsafe coding in OpenBLAS that just happened to work with gcc