[llvm] [X86][BF16] Do not combine FP_TRUNC + FP_EXTEND if they come from user (PR #91420)

[Rasmus Munk Larsen via llvm-commits]

================
@@ -2420,3 +2420,47 @@ define <16 x bfloat> @concat_dup_v8bf16(<8 x bfloat> %x, <8 x bfloat> %y) {
   %a = shufflevector <8 x bfloat> %x, <8 x bfloat> %y, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
   ret <16 x bfloat> %a
 }
+
+define float @trunc_ext(float %a) nounwind {
+; X86-LABEL: trunc_ext:
+; X86:       # %bb.0:
+; X86-NEXT:    pushl %eax
+; X86-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-NEXT:    vcvtneps2bf16 %xmm0, %xmm0
+; X86-NEXT:    vmovw %xmm0, %eax
+; X86-NEXT:    shll $16, %eax
+; X86-NEXT:    vmovd %eax, %xmm0
+; X86-NEXT:    vmovd %xmm0, (%esp)
+; X86-NEXT:    flds (%esp)
+; X86-NEXT:    popl %eax
+; X86-NEXT:    retl
+;
+; SSE2-LABEL: trunc_ext:
+; SSE2:       # %bb.0:
+; SSE2-NEXT:    pushq %rax
+; SSE2-NEXT:    callq __truncsfbf2 at PLT
----------------
rmlarsen wrote:

Is there a way to use a rounding conversion here, even if it's slower? Are there tests for this that check that rounding is used?

https://github.com/llvm/llvm-project/pull/91420