[llvm] [VPlan] Peek through nested phi incoming values in computeBlendMasks (PR #203164)

[Luke Lau via llvm-commits]

================
@@ -97,15 +97,25 @@ define i32 @anyof_three_blend_chain(i1 %c0, i1 %c1, i1 %c2, i32 %n) {
 ; CHECK:       [[VECTOR_PH]]:
 ; CHECK-NEXT:    [[N_MOD_VF:%.*]] = and i32 [[SMAX]], 3
 ; CHECK-NEXT:    [[N_VEC:%.*]] = sub i32 [[SMAX]], [[N_MOD_VF]]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i1> poison, i1 [[C2]], i64 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i1> [[BROADCAST_SPLATINSERT]], <4 x i1> poison, <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <4 x i1> poison, i1 [[C1]], i64 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT2:%.*]] = shufflevector <4 x i1> [[BROADCAST_SPLATINSERT1]], <4 x i1> poison, <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <4 x i1> poison, i1 [[C0]], i64 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT4:%.*]] = shufflevector <4 x i1> [[BROADCAST_SPLATINSERT3]], <4 x i1> poison, <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP7:%.*]] = xor <4 x i1> [[BROADCAST_SPLAT4]], splat (i1 true)
+; CHECK-NEXT:    [[TMP3:%.*]] = xor <4 x i1> [[BROADCAST_SPLAT2]], splat (i1 true)
+; CHECK-NEXT:    [[TMP9:%.*]] = select <4 x i1> [[TMP7]], <4 x i1> [[TMP3]], <4 x i1> zeroinitializer
+; CHECK-NEXT:    [[TMP10:%.*]] = xor <4 x i1> [[BROADCAST_SPLAT]], splat (i1 true)
----------------
lukel97 wrote:

I took a look at this and I think the resulting codegen from these xors + selects isn't bad, they get lowered as mask NOTs and ANDs. I tried scalarizing them but that pretty tricky for not that much obvious benefit, since mask arithmetic is generally cheap. 

https://github.com/llvm/llvm-project/pull/203164