[llvm-branch-commits] [llvm] [SelectionDAG] avoid rounding exact f32->bf16 (STRICT_)FP_ROUND (PR #230557)

[via llvm-branch-commits]

================
@@ -26,7 +26,9 @@ define bfloat @return_arg_bf16(bfloat %x) #0 {
 ; FAST_ISEL_AVX512BF16-LABEL: return_arg_bf16:
 ; FAST_ISEL_AVX512BF16:       # %bb.0:
 ; FAST_ISEL_AVX512BF16-NEXT:    vpslld $16, %xmm0, %xmm0
-; FAST_ISEL_AVX512BF16-NEXT:    vcvtneps2bf16 %xmm0, %xmm0
----------------
Andarwinux wrote:

I believe the disagreement is about performance. Similar to #221052, this PR will introduce performance regression due to FPU<->GPU moves, but the only improvement is about denormals which is considered not important for bf16 use cases.

Anyway, with #230698, this is no longer a problem, LGTM.

https://github.com/llvm/llvm-project/pull/230557