[llvm] [VPlan] Simplify BLEND %a, %b, NOT(%m) -> BLEND %b, %a, %m. (PR #128375)

[Florian Hahn via llvm-commits]

================
@@ -36,16 +36,16 @@ define float @test(ptr nocapture readonly %pA, ptr nocapture readonly %pB, i32 %
 ; CHECK-NEXT:    [[NEXT_GEP6:%.*]] = getelementptr i8, ptr [[PB]], i32 [[OFFSET_IDX5]]
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[NEXT_GEP]], align 4
 ; CHECK-NEXT:    [[WIDE_LOAD7:%.*]] = load <4 x float>, ptr [[NEXT_GEP6]], align 4
-; CHECK-NEXT:    [[TMP2:%.*]] = fcmp fast oeq <4 x float> [[WIDE_LOAD]], zeroinitializer
-; CHECK-NEXT:    [[TMP3:%.*]] = fcmp fast oeq <4 x float> [[WIDE_LOAD7]], zeroinitializer
-; CHECK-NEXT:    [[DOTNOT9:%.*]] = select <4 x i1> [[TMP2]], <4 x i1> [[TMP3]], <4 x i1> zeroinitializer
+; CHECK-NEXT:    [[TMP18:%.*]] = fcmp fast une <4 x float> [[WIDE_LOAD]], zeroinitializer
+; CHECK-NEXT:    [[TMP19:%.*]] = fcmp fast une <4 x float> [[WIDE_LOAD7]], zeroinitializer
+; CHECK-NEXT:    [[TMP20:%.*]] = select <4 x i1> [[TMP18]], <4 x i1> splat (i1 true), <4 x i1> [[TMP19]]
----------------
fhahn wrote:

It may not be profitable, depending on the cost to materialize `(splat (i1 true))`, but is is instcombine that rewrites the phi in that way. Such details should probably be dealt with in the backend, where there's more HW specific info available?

After LV, we get

```
<   %19 = xor <4 x i1> %11, splat (i1 true)
<   %predphi = select <4 x i1> %19, <4 x float> %vec.phi, <4 x float> %18
---
>   %predphi = select <4 x i1> %11, <4 x float> %18, <4 x float> %vec.phi
```

https://github.com/llvm/llvm-project/pull/128375