[llvm] [InstCombine] Fold shuffles through all trivially vectorizable intrinsics (PR #141979)

[Luke Lau via llvm-commits]

================
@@ -264,23 +264,17 @@ define <3 x i16> @uadd_sat_v3i16(<3 x i16> %arg0, <3 x i16> %arg1) {
 ; GFX8-NEXT:  bb:
 ; GFX8-NEXT:    [[ARG0_2:%.*]] = extractelement <3 x i16> [[ARG0:%.*]], i64 2
 ; GFX8-NEXT:    [[ARG1_2:%.*]] = extractelement <3 x i16> [[ARG1:%.*]], i64 2
-; GFX8-NEXT:    [[TMP0:%.*]] = shufflevector <3 x i16> [[ARG0]], <3 x i16> poison, <2 x i32> <i32 0, i32 1>
-; GFX8-NEXT:    [[TMP1:%.*]] = shufflevector <3 x i16> [[ARG1]], <3 x i16> poison, <2 x i32> <i32 0, i32 1>
-; GFX8-NEXT:    [[TMP2:%.*]] = call <2 x i16> @llvm.uadd.sat.v2i16(<2 x i16> [[TMP0]], <2 x i16> [[TMP1]])
+; GFX8-NEXT:    [[TMP3:%.*]] = call <3 x i16> @llvm.uadd.sat.v3i16(<3 x i16> [[ARG0]], <3 x i16> [[ARG1]])
----------------
lukel97 wrote:

Should still allow the case where we combine into a smaller type, e.g.

```llvm

define <3 x float> @fma_unary_shuffle_ops_widening(<2 x float> %x, <2 x float> %y, <2 x float> %z) {
; CHECK-LABEL: @fma_unary_shuffle_ops_widening(
; CHECK-NEXT:    [[A:%.*]] = shufflevector <2 x float> [[X:%.*]], <2 x float> poison, <3 x i32> <i32 1, i32 0, i32 1>
; CHECK-NEXT:    call void @use_vec3(<3 x float> [[A]])
; CHECK-NEXT:    [[TMP1:%.*]] = call fast <2 x float> @llvm.fma.v2f32(<2 x float> [[X]], <2 x float> [[Y:%.*]], <2 x float> [[Z:%.*]])
; CHECK-NEXT:    [[R:%.*]] = shufflevector <2 x float> [[TMP1]], <2 x float> poison, <3 x i32> <i32 1, i32 0, i32 1>
; CHECK-NEXT:    ret <3 x float> [[R]]
;
  %a = shufflevector <2 x float> %x, <2 x float> poison, <3 x i32> <i32 1, i32 0, i32 1>
  call void @use_vec3(<3 x float> %a)
  %b = shufflevector <2 x float> %y, <2 x float> poison, <3 x i32> <i32 1, i32 0, i32 1>
  %c = shufflevector <2 x float> %z, <2 x float> poison, <3 x i32> <i32 1, i32 0, i32 1>
  %r = call fast <3 x float> @llvm.fma.v3f32(<3 x float> %a, <3 x float> %b, <3 x float> %c)
  ret <3 x float> %r
}
```

I agree either way, this would make it more consistent with the binop combine in InstCombinerImpl::foldVectorBinop which currently only allows the same type for two non-constant operands. For a constant operand, it looks like it's allowed to narrow it.

https://github.com/llvm/llvm-project/pull/141979