[X86][SSE] Swap X86ISD::BLENDV inputs with an inverted selection mask (PR42825)

As discussed on PR42825, if we are inverting the selection mask we can just swap the inputs and avoid the inversion.

Differential Revision: https://reviews.llvm.org/D65522

llvm-svn: 368438

3 files changed, 19 insertions, 21 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index fa5f788ff9f..6dd5ec87e7b 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -36615,6 +36615,12 @@ static SDValue combineSelect(SDNode *N, SelectionDAG &DAG,
   if (SDValue V = narrowVectorSelect(N, DAG, Subtarget))
     return V;
 
+  // select(~Cond, X, Y) -> select(Cond, Y, X)
+  if (CondVT.getScalarType() != MVT::i1)
+    if (SDValue CondNot = IsNOT(Cond, DAG))
+      return DAG.getNode(N->getOpcode(), DL, VT,
+                         DAG.getBitcast(CondVT, CondNot), RHS, LHS);
+
   // Custom action for SELECT MMX
   if (VT == MVT::x86mmx) {
     LHS = DAG.getBitcast(MVT::i64, LHS);
diff --git a/llvm/test/CodeGen/X86/combine-sse41-intrinsics.ll b/llvm/test/CodeGen/X86/combine-sse41-intrinsics.ll
index 0774f7fe20d..cc4dee33c61 100644
--- a/llvm/test/CodeGen/X86/combine-sse41-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/combine-sse41-intrinsics.ll
@@ -157,10 +157,9 @@ define <16 x i8> @xor_pblendvb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> %a2) {
 ; CHECK-LABEL: xor_pblendvb:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    movdqa %xmm0, %xmm3
-; CHECK-NEXT:    pcmpeqd %xmm0, %xmm0
-; CHECK-NEXT:    pxor %xmm2, %xmm0
-; CHECK-NEXT:    pblendvb %xmm0, %xmm1, %xmm3
-; CHECK-NEXT:    movdqa %xmm3, %xmm0
+; CHECK-NEXT:    movaps %xmm2, %xmm0
+; CHECK-NEXT:    pblendvb %xmm0, %xmm3, %xmm1
+; CHECK-NEXT:    movdqa %xmm1, %xmm0
 ; CHECK-NEXT:    retq
   %1 = xor <16 x i8> %a2, <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>
   %2 = tail call <16 x i8> @llvm.x86.sse41.pblendvb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> %1)
@@ -170,11 +169,10 @@ define <16 x i8> @xor_pblendvb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> %a2) {
 define <4 x float> @xor_blendvps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) {
 ; CHECK-LABEL: xor_blendvps:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    movdqa %xmm0, %xmm3
-; CHECK-NEXT:    pcmpeqd %xmm0, %xmm0
-; CHECK-NEXT:    pxor %xmm2, %xmm0
-; CHECK-NEXT:    blendvps %xmm0, %xmm1, %xmm3
-; CHECK-NEXT:    movaps %xmm3, %xmm0
+; CHECK-NEXT:    movaps %xmm0, %xmm3
+; CHECK-NEXT:    movaps %xmm2, %xmm0
+; CHECK-NEXT:    blendvps %xmm0, %xmm3, %xmm1
+; CHECK-NEXT:    movaps %xmm1, %xmm0
 ; CHECK-NEXT:    retq
   %1 = bitcast <4 x float> %a2 to <4 x i32>
   %2 = xor <4 x i32> %1, <i32 -1, i32 -1, i32 -1, i32 -1>
@@ -186,11 +184,10 @@ define <4 x float> @xor_blendvps(<4 x float> %a0, <4 x float> %a1, <4 x float> %
 define <2 x double> @xor_blendvpd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) {
 ; CHECK-LABEL: xor_blendvpd:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    movdqa %xmm0, %xmm3
-; CHECK-NEXT:    pcmpeqd %xmm0, %xmm0
-; CHECK-NEXT:    pxor %xmm2, %xmm0
-; CHECK-NEXT:    blendvpd %xmm0, %xmm1, %xmm3
-; CHECK-NEXT:    movapd %xmm3, %xmm0
+; CHECK-NEXT:    movapd %xmm0, %xmm3
+; CHECK-NEXT:    movaps %xmm2, %xmm0
+; CHECK-NEXT:    blendvpd %xmm0, %xmm3, %xmm1
+; CHECK-NEXT:    movapd %xmm1, %xmm0
 ; CHECK-NEXT:    retq
   %1 = bitcast <2 x double> %a2 to <4 x i32>
   %2 = xor <4 x i32> %1, <i32 -1, i32 -1, i32 -1, i32 -1>
diff --git a/llvm/test/CodeGen/X86/nontemporal-loads.ll b/llvm/test/CodeGen/X86/nontemporal-loads.ll
index 8f0118d39bd..8af4a680c77 100644
--- a/llvm/test/CodeGen/X86/nontemporal-loads.ll
+++ b/llvm/test/CodeGen/X86/nontemporal-loads.ll
@@ -1852,25 +1852,20 @@ define <16 x i32> @test_masked_v16i32(i8 * %addr, <16 x i32> %old, <16 x i32> %m
 ; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm4
 ; AVX1-NEXT:    vpxor %xmm5, %xmm5, %xmm5
 ; AVX1-NEXT:    vpcmpeqd %xmm5, %xmm4, %xmm4
-; AVX1-NEXT:    vpcmpeqd %xmm6, %xmm6, %xmm6
-; AVX1-NEXT:    vpxor %xmm6, %xmm4, %xmm4
 ; AVX1-NEXT:    vpcmpeqd %xmm5, %xmm3, %xmm3
-; AVX1-NEXT:    vpxor %xmm6, %xmm3, %xmm3
 ; AVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm3, %ymm3
 ; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm4
 ; AVX1-NEXT:    vpcmpeqd %xmm5, %xmm4, %xmm4
-; AVX1-NEXT:    vpxor %xmm6, %xmm4, %xmm4
 ; AVX1-NEXT:    vpcmpeqd %xmm5, %xmm2, %xmm2
-; AVX1-NEXT:    vpxor %xmm6, %xmm2, %xmm2
 ; AVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm2, %ymm2
 ; AVX1-NEXT:    vmovntdqa 32(%rdi), %xmm4
 ; AVX1-NEXT:    vmovntdqa 48(%rdi), %xmm5
 ; AVX1-NEXT:    vinsertf128 $1, %xmm5, %ymm4, %ymm4
-; AVX1-NEXT:    vblendvps %ymm3, %ymm4, %ymm1, %ymm1
+; AVX1-NEXT:    vblendvps %ymm3, %ymm1, %ymm4, %ymm1
 ; AVX1-NEXT:    vmovntdqa (%rdi), %xmm3
 ; AVX1-NEXT:    vmovntdqa 16(%rdi), %xmm4
 ; AVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm3, %ymm3
-; AVX1-NEXT:    vblendvps %ymm2, %ymm3, %ymm0, %ymm0
+; AVX1-NEXT:    vblendvps %ymm2, %ymm0, %ymm3, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: test_masked_v16i32: