[llvm] [VPlan] Compute blend masks from minimum set of edge masks (PR #201783)

[Florian Hahn via llvm-commits]

================
@@ -1145,14 +1146,16 @@ define float @reduction_conditional(ptr %A, ptr %B, ptr %C, float %S) {
 ; CHECK-INTERLEAVED-NEXT:    [[TMP16:%.*]] = select <4 x i1> [[TMP6]], <4 x i1> [[TMP8]], <4 x i1> zeroinitializer
 ; CHECK-INTERLEAVED-NEXT:    [[TMP17:%.*]] = fadd fast <4 x float> [[VEC_PHI]], [[WIDE_LOAD]]
 ; CHECK-INTERLEAVED-NEXT:    [[TMP18:%.*]] = fadd fast <4 x float> [[VEC_PHI1]], [[WIDE_LOAD2]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP27:%.*]] = xor <4 x i1> [[TMP5]], splat (i1 true)
+; CHECK-INTERLEAVED-NEXT:    [[TMP28:%.*]] = xor <4 x i1> [[TMP6]], splat (i1 true)
 ; CHECK-INTERLEAVED-NEXT:    [[TMP20:%.*]] = select <4 x i1> [[TMP19]], <4 x i1> [[TMP11]], <4 x i1> zeroinitializer
 ; CHECK-INTERLEAVED-NEXT:    [[TMP22:%.*]] = select <4 x i1> [[TMP21]], <4 x i1> [[TMP12]], <4 x i1> zeroinitializer
-; CHECK-INTERLEAVED-NEXT:    [[PREDPHI:%.*]] = select <4 x i1> [[TMP20]], <4 x float> [[VEC_PHI]], <4 x float> [[TMP13]]
-; CHECK-INTERLEAVED-NEXT:    [[PREDPHI5:%.*]] = select <4 x i1> [[TMP15]], <4 x float> [[TMP17]], <4 x float> [[PREDPHI]]
-; CHECK-INTERLEAVED-NEXT:    [[PREDPHI6]] = select <4 x i1> [[TMP5]], <4 x float> [[PREDPHI5]], <4 x float> [[VEC_PHI]]
-; CHECK-INTERLEAVED-NEXT:    [[PREDPHI7:%.*]] = select <4 x i1> [[TMP22]], <4 x float> [[VEC_PHI1]], <4 x float> [[TMP14]]
-; CHECK-INTERLEAVED-NEXT:    [[PREDPHI8:%.*]] = select <4 x i1> [[TMP16]], <4 x float> [[TMP18]], <4 x float> [[PREDPHI7]]
-; CHECK-INTERLEAVED-NEXT:    [[PREDPHI9]] = select <4 x i1> [[TMP6]], <4 x float> [[PREDPHI8]], <4 x float> [[VEC_PHI1]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP25:%.*]] = or <4 x i1> [[TMP20]], [[TMP27]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP26:%.*]] = or <4 x i1> [[TMP22]], [[TMP28]]
+; CHECK-INTERLEAVED-NEXT:    [[PREDPHI:%.*]] = select <4 x i1> [[TMP25]], <4 x float> [[VEC_PHI]], <4 x float> [[TMP13]]
----------------
fhahn wrote:

Ah yes thanks for checking, that should be a bit cheaper overall in the generated assembly, thanks.

I did a run on my corpus on AArch64 with SVE + tail-folding and it looks like there are only a few changes, most of them improvements or neutral.

The only case where things might actually go wrong I could spot is something like  https://llvm.godbolt.org/z/r4189jac4. The assembly for the generated vector IR with this patch has 2 more vector compares, which seem to be slightly more expensive (higher latency, more uops) than value selects, and `llvm-mca` reports ~80% more cycles for the version with the patch.

Not sure if `llvm-mca` is accruate here, but may be worth to quickly double check? It is probably just a lowering issues

https://github.com/llvm/llvm-project/pull/201783