[PATCH] D41599: [X86] Lowering X86 avx512 sqrt intrinsics to IR - LLVM

[Simon Pilgrim via Phabricator via llvm-commits]

RKSimon added a comment.

In https://reviews.llvm.org/D41599#976062, @uriel.k wrote:

> Simon, is there anything else you think that is needed to be changed before accepting the revision?
> Thanks


I'm still a little worried about this - it can create a lot more bit differences in results than previous other intrinsics where we've replaced with generic implementations - I guess _mm_div_ps already does this to an extent (and other fadd/fsub/fmul cases via re-association etc.).

Maybe I'm just being a little over cautious, but at very least I'd like to see https://reviews.llvm.org/D41168 update the intrinsic documentation to explain that -ffast-math may result in rsqrt+nr codgen under some circumstances - it still says that (v)sqrtps will be generated.


https://reviews.llvm.org/D41599