[llvm] llvm: honor VFABI mappings without builtin suppression (PR #193884)

[via llvm-commits]

wpcwzy wrote:

Hi all,

Apologies for the delay. I have refined the tests and seperated the "warn missed transformations' changes into a seperate PR. Most importantly, I have updated the OpenMP CodeGen phase to mark math `declare-simd` lib calls as `readnone` when `-ffast-math` is enabled, instead of relaxing the conditions in `LoopAccessAnalysis.cpp`.

For the related OpenMP CodeGen PR, please see llvm-project#193881.

Thanks for your time and review!

https://github.com/llvm/llvm-project/pull/193884