[X86] combineSelect - attempt to combine with shuffles (#143753)

Before legalization we will convert to a vector_shuffle node - but afterward we can try to combine the select into an existing target shuffle chain

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -47785,13 +47785,19 @@ static SDValue combineSelect(SDNode *N, SelectionDAG &DAG,
                                                            DL, DAG, Subtarget))
       return V;
 
-  // Convert vselects with constant condition into shuffles.
-  if (CondConstantVector && DCI.isBeforeLegalizeOps() &&
-      (N->getOpcode() == ISD::VSELECT || N->getOpcode() == X86ISD::BLENDV)) {
+  if (N->getOpcode() == ISD::VSELECT || N->getOpcode() == X86ISD::BLENDV) {
     SmallVector<int, 64> Mask;
     if (createShuffleMaskFromVSELECT(Mask, Cond,
-                                     N->getOpcode() == X86ISD::BLENDV))
-      return DAG.getVectorShuffle(VT, DL, LHS, RHS, Mask);
+                                     N->getOpcode() == X86ISD::BLENDV)) {
+      // Convert vselects with constant condition into shuffles.
+      if (DCI.isBeforeLegalizeOps())
+        return DAG.getVectorShuffle(VT, DL, LHS, RHS, Mask);
+
+      // Attempt to combine as shuffle.
+      SDValue Op(N, 0);
+      if (SDValue Res = combineX86ShufflesRecursively(Op, DAG, Subtarget))
+        return Res;
+    }
   }
 
   // fold vselect(cond, pshufb(x), pshufb(y)) -> or (pshufb(x), pshufb(y))

llvm/test/CodeGen/X86/combine-mask-with-shuffle.ll

@@ -67,11 +67,9 @@ define <16 x i32> @combine_mask_with_abs(<16 x i32> %v0) {
 define <16 x i32> @combine_mask_with_umin(<16 x i32> %v0) {
 ; CHECK-LABEL: combine_mask_with_umin:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vpshufd {{.*#+}} zmm1 = zmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
-; CHECK-NEXT:    movw $-21846, %ax # imm = 0xAAAA
-; CHECK-NEXT:    kmovw %eax, %k1
-; CHECK-NEXT:    vpblendmd %zmm1, %zmm0, %zmm2 {%k1}
-; CHECK-NEXT:    vpminud %zmm2, %zmm1, %zmm1
+; CHECK-NEXT:    vpshufd {{.*#+}} zmm1 = zmm0[0,0,2,2,4,4,6,6,8,8,10,10,12,12,14,14]
+; CHECK-NEXT:    vpshufd {{.*#+}} zmm2 = zmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
+; CHECK-NEXT:    vpminud %zmm1, %zmm2, %zmm1
 ; CHECK-NEXT:    movw $-3856, %ax # imm = 0xF0F0
 ; CHECK-NEXT:    kmovw %eax, %k1
 ; CHECK-NEXT:    vpopcntd %zmm0, %zmm1 {%k1}
@@ -88,11 +86,9 @@ define <16 x i32> @combine_mask_with_umin(<16 x i32> %v0) {
 define <16 x i32> @combine_mask_with_umax(<16 x i32> %v0) {
 ; CHECK-LABEL: combine_mask_with_umax:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vpshufd {{.*#+}} zmm1 = zmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
-; CHECK-NEXT:    movw $-21846, %ax # imm = 0xAAAA
-; CHECK-NEXT:    kmovw %eax, %k1
-; CHECK-NEXT:    vpblendmd %zmm1, %zmm0, %zmm2 {%k1}
-; CHECK-NEXT:    vpmaxud %zmm2, %zmm1, %zmm1
+; CHECK-NEXT:    vpshufd {{.*#+}} zmm1 = zmm0[0,0,2,2,4,4,6,6,8,8,10,10,12,12,14,14]
+; CHECK-NEXT:    vpshufd {{.*#+}} zmm2 = zmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
+; CHECK-NEXT:    vpmaxud %zmm1, %zmm2, %zmm1
 ; CHECK-NEXT:    movw $-3856, %ax # imm = 0xF0F0
 ; CHECK-NEXT:    kmovw %eax, %k1
 ; CHECK-NEXT:    vpopcntd %zmm0, %zmm1 {%k1}
@@ -109,11 +105,9 @@ define <16 x i32> @combine_mask_with_umax(<16 x i32> %v0) {
 define <16 x i32> @combine_mask_with_smin(<16 x i32> %v0) {
 ; CHECK-LABEL: combine_mask_with_smin:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vpshufd {{.*#+}} zmm1 = zmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
-; CHECK-NEXT:    movw $-21846, %ax # imm = 0xAAAA
-; CHECK-NEXT:    kmovw %eax, %k1
-; CHECK-NEXT:    vpblendmd %zmm1, %zmm0, %zmm2 {%k1}
-; CHECK-NEXT:    vpminsd %zmm2, %zmm1, %zmm1
+; CHECK-NEXT:    vpshufd {{.*#+}} zmm1 = zmm0[0,0,2,2,4,4,6,6,8,8,10,10,12,12,14,14]
+; CHECK-NEXT:    vpshufd {{.*#+}} zmm2 = zmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
+; CHECK-NEXT:    vpminsd %zmm1, %zmm2, %zmm1
 ; CHECK-NEXT:    movw $-3856, %ax # imm = 0xF0F0
 ; CHECK-NEXT:    kmovw %eax, %k1
 ; CHECK-NEXT:    vpopcntd %zmm0, %zmm1 {%k1}
@@ -130,11 +124,9 @@ define <16 x i32> @combine_mask_with_smin(<16 x i32> %v0) {
 define <16 x i32> @combine_mask_with_smax(<16 x i32> %v0) {
 ; CHECK-LABEL: combine_mask_with_smax:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vpshufd {{.*#+}} zmm1 = zmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
-; CHECK-NEXT:    movw $-21846, %ax # imm = 0xAAAA
-; CHECK-NEXT:    kmovw %eax, %k1
-; CHECK-NEXT:    vpblendmd %zmm1, %zmm0, %zmm2 {%k1}
-; CHECK-NEXT:    vpmaxsd %zmm2, %zmm1, %zmm1
+; CHECK-NEXT:    vpshufd {{.*#+}} zmm1 = zmm0[0,0,2,2,4,4,6,6,8,8,10,10,12,12,14,14]
+; CHECK-NEXT:    vpshufd {{.*#+}} zmm2 = zmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
+; CHECK-NEXT:    vpmaxsd %zmm1, %zmm2, %zmm1
 ; CHECK-NEXT:    movw $-3856, %ax # imm = 0xF0F0
 ; CHECK-NEXT:    kmovw %eax, %k1
 ; CHECK-NEXT:    vpopcntd %zmm0, %zmm1 {%k1}

llvm/test/CodeGen/X86/pr132844.ll

@@ -4,12 +4,11 @@
 define  { ptr, i8 } @PR132844(<4 x ptr> %0, <4 x ptr> %1) {
 ; CHECK-LABEL: PR132844:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm2
-; CHECK-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
-; CHECK-NEXT:    movb $10, %al
-; CHECK-NEXT:    kmovd %eax, %k1
-; CHECK-NEXT:    vinserti64x2 $1, 16, %ymm2, %ymm0 {%k1}
-; CHECK-NEXT:    vmovdqu %ymm0, {{[0-9]+}}(%rsp)
+; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm2
+; CHECK-NEXT:    vinsertf128 $1, 16, %ymm2, %ymm2
+; CHECK-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; CHECK-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5],ymm2[6,7]
+; CHECK-NEXT:    vmovups %ymm0, {{[0-9]+}}(%rsp)
 ; CHECK-NEXT:    xorl %eax, %eax
 ; CHECK-NEXT:    xorl %edx, %edx
 ; CHECK-NEXT:    vzeroupper