[PATCH] D158950: [X86][BF16] Lower FP_EXTEND for vector types under AVX512BF16

[Phoebe Wang via Phabricator via llvm-commits]

pengfei added inline comments.


================
Comment at: llvm/test/CodeGen/X86/bfloat.ll:1518
+; SSE2-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm4[0]
+; SSE2-NEXT:    retq
+;
----------------
RKSimon wrote:
> how come this is scalarized given how easy it is to bit-twiddle bfloat to float?
We cannot do vector fptrunc without AVX512BF16, so I deferred to set vXbf16 legal until the feature ready.
I think we may make it asymmetry between fptrunc and fpext. We can do it in a follow up. This is to address #64460 and backport to 17.0 with a low risk.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D158950/new/

https://reviews.llvm.org/D158950