Alexey Bader

> Hi [@bader](https://github.com/bader) -- does this comment suggest that moving to clang-linker-wrapper (via `--offload-new-driver`) better positions us to fix this issue while still using llc for the wrapper compilation step?...

> The upstream approach (that is, invoking CodeGen from clang-linker-wrapper) should avoid this kind of issue. Hopefully we can move toward that approach. Please, sync with @YuriPlyakhin on plans regarding...

I think the discussion in https://github.com/intel/llvm/issues/4381 might be related.

As far I can tell, the discussion in #4381 goes beyond CUDA BE/adaptor. I didn't read carefully the whole discussion, so feel free to ignore my comment if it's not...

> We have declared __imf_fsigm* in https://github.com/intel/llvm/pull/20873/files#diff-8da594ced8852cd2e62e9d1fcd013ca0b2256b88b3745b30a239bb01aa7b2011R251 > So, the declarations at the begining is redundant, just remove them. I suggest we keep the first declaration and remove the second...

Post-commit impacted as well - https://github.com/intel/llvm/actions/runs/19653410200/job/56287074948.