[PATCH] D130817: [X86][FP16] Fix vector_shuffle and lowering without f16c feature problems

Sat Jul 30 06:51:10 PDT 2022

pengfei added inline comments.

================
Comment at: llvm/test/CodeGen/X86/vector-half-conversions.ll:29
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    retq
----------------
RKSimon wrote:
> pengfei wrote:
> > RKSimon wrote:
> > > I'm confused as to under what cases we can't use vcvtph2ps for half -> float given all half values are exactly representable by float?
> > Not sure if I understand you question correctly. `vcvtph2ps` requires `F16C`, which is independent of `AVX` and `AVX2`.
> Sorry - I missed the corresponding change in the attributes for the check prefixes. 
> 
> It might make sense to pre-commit the changes to the RUN/check-prefixes so you can rebase the patch and show the codegen differences more clearly.
Yeah, I tried, but the test failed without the change.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D130817/new/

https://reviews.llvm.org/D130817