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

[RKSimon]

Summary

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

[RKSimon]

ping

[lebedev.ri]

Looks good to me.

[RKSimon]

any more comments?

[spatel]

LGTM

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

if (SDValue V = combineVSelectWithAllOnesOrZeros(N, DAG, DCI, Subtarget))
     return V;
 
   if (SDValue V = combineVSelectToBLENDV(N, DAG, DCI, Subtarget))
     return V;
 
   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);
     RHS = DAG.getBitcast(MVT::i64, RHS);
     SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::i64, Cond, LHS, RHS);
     return DAG.getBitcast(VT, newSelect);
   }
 

llvm/trunk/test/CodeGen/X86/combine-sse41-intrinsics.ll

; CHECK-NEXT:    retq
   %9  = lshr <2 x i64> %8, <i64 11, i64 11>
   ret <2 x i64> %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)
   ret <16 x i8> %2
 }
 
 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>
   %3 = bitcast <4 x i32> %2 to <4 x float>
   %4 = tail call <4 x float> @llvm.x86.sse41.blendvps(<4 x float> %a0, <4 x float> %a1, <4 x float> %3)
   ret <4 x float> %4
 }
 
 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>
   %3 = bitcast <4 x i32> %2 to <2 x double>
   %4 = tail call <2 x double> @llvm.x86.sse41.blendvpd(<2 x double> %a0, <2 x double> %a1, <2 x double> %3)
   ret <2 x double> %4
 }
 
 declare <2 x double> @llvm.x86.sse41.blendpd(<2 x double>, <2 x double>, i32)
 declare <4 x float> @llvm.x86.sse41.blendps(<4 x float>, <4 x float>, i32)
 declare <8 x i16> @llvm.x86.sse41.pblendw(<8 x i16>, <8 x i16>, i32)
 
 declare <2 x double> @llvm.x86.sse41.blendvpd(<2 x double>, <2 x double>, <2 x double>)
 declare <4 x float> @llvm.x86.sse41.blendvps(<4 x float>, <4 x float>, <4 x float>)
 declare <16 x i8> @llvm.x86.sse41.pblendvb(<16 x i8>, <16 x i8>, <16 x i8>)

llvm/trunk/test/CodeGen/X86/nontemporal-loads.ll

 ; SSE41-NEXT:    movaps %xmm4, %xmm3
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: test_masked_v16i32:
 ; AVX1:       # %bb.0:
 ; 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:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpxor %xmm4, %xmm4, %xmm4
 ; AVX2-NEXT:    vpcmpeqd %ymm4, %ymm3, %ymm3
 ; AVX2-NEXT:    vpcmpeqd %ymm4, %ymm2, %ymm2
 ; AVX2-NEXT:    vmovntdqa 32(%rdi), %ymm4