[X86][SSE] Improve support for vselect(Cond, 0, X) -> ANDN(Cond, X)

[RKSimon]

Summary

As discussed on PR28925 and D37849.

[spatel]

LGTM.

lib/Target/X86/X86ISelLowering.cpp

30157–30159

We might be able to remove the legality checks below here now that we have this, but I agree that the bigger cleanup is to make this semi-generic in DAGCombiner, so x86 won't need to do anything else for these patterns.

lib/Target/X86/X86ISelLowering.cpp

 static SDValue combineExtractVectorElt_SSE(SDNode *N, SelectionDAG &DAG,
                                            TargetLowering::DAGCombinerInfo &DCI,
                                            const X86Subtarget &Subtarget) {
   return combineExtractWithShuffle(N, DAG, DCI, Subtarget);
 }
 
 /// If a vector select has an operand that is -1 or 0, try to simplify the
 /// select to a bitwise logic operation.
+/// TODO: Move to DAGCombiner, possibly using TargetLowering::hasAndNot()?
 static SDValue
 combineVSelectWithAllOnesOrZeros(SDNode *N, SelectionDAG &DAG,
                                  TargetLowering::DAGCombinerInfo &DCI,
                                  const X86Subtarget &Subtarget) {
   SDValue Cond = N->getOperand(0);
   SDValue LHS = N->getOperand(1);
   SDValue RHS = N->getOperand(2);
   EVT VT = LHS.getValueType();
if (TValIsAllZeros || FValIsAllOnes) {
       Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1),
                           NewCC);
       std::swap(LHS, RHS);
       TValIsAllOnes = FValIsAllOnes;
       FValIsAllZeros = TValIsAllZeros;
     }
   }
 
+  // Cond value must be 'sign splat' to be converted to a logical op.
+  if (DAG.ComputeNumSignBits(Cond) != CondVT.getScalarSizeInBits())
+    return SDValue();
+
   // vselect Cond, 111..., 000... -> Cond
   if (TValIsAllOnes && FValIsAllZeros)
     return DAG.getBitcast(VT, Cond);
 
   if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(CondVT))
     return SDValue();
 
   // vselect Cond, 111..., X -> or Cond, X
   if (TValIsAllOnes) {
     SDValue CastRHS = DAG.getBitcast(CondVT, RHS);
     SDValue Or = DAG.getNode(ISD::OR, DL, CondVT, Cond, CastRHS);
     return DAG.getBitcast(VT, Or);
   }
 
   // vselect Cond, X, 000... -> and Cond, X
   if (FValIsAllZeros) {
     SDValue CastLHS = DAG.getBitcast(CondVT, LHS);
     SDValue And = DAG.getNode(ISD::AND, DL, CondVT, Cond, CastLHS);
     return DAG.getBitcast(VT, And);
   }
 
+  // vselect Cond, 000..., X -> andn Cond, X
+  if (TValIsAllZeros) {
+    MVT AndNVT = MVT::getVectorVT(MVT::i64, CondVT.getSizeInBits() / 64);
+    SDValue CastCond = DAG.getBitcast(AndNVT, Cond);
+    SDValue CastRHS = DAG.getBitcast(AndNVT, RHS);
+    SDValue AndN = DAG.getNode(X86ISD::ANDNP, DL, AndNVT, CastCond, CastRHS);
+    return DAG.getBitcast(VT, AndN);
+  }
+
   return SDValue();
 }
 
 static SDValue combineSelectOfTwoConstants(SDNode *N, SelectionDAG &DAG) {
   SDValue Cond = N->getOperand(0);
   SDValue LHS = N->getOperand(1);
   SDValue RHS = N->getOperand(2);
   SDLoc DL(N);

test/CodeGen/X86/psubus.ll

 ; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0]
 ; SSE41-NEXT:    pand %xmm5, %xmm4
 ; SSE41-NEXT:    pand %xmm5, %xmm3
 ; SSE41-NEXT:    packuswb %xmm4, %xmm3
 ; SSE41-NEXT:    pand %xmm5, %xmm1
 ; SSE41-NEXT:    pand %xmm5, %xmm2
 ; SSE41-NEXT:    packuswb %xmm2, %xmm1
 ; SSE41-NEXT:    packuswb %xmm3, %xmm1
-; SSE41-NEXT:    pxor %xmm2, %xmm2
-; SSE41-NEXT:    pblendvb %xmm0, %xmm2, %xmm1
-; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pandn %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: test14:
 ; AVX1:       # BB#0: # %vector.ph
 ; AVX1-NEXT:    vpshufd {{.*#+}} xmm3 = xmm0[1,1,2,3]
 ; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm8 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero,xmm3[2],zero,zero,zero,xmm3[3],zero,zero,zero
 ; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm9 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero
 ; AVX1-NEXT:    vpshufd {{.*#+}} xmm5 = xmm0[2,3,0,1]

 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm5 = [255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0]
 ; AVX1-NEXT:    vpand %xmm5, %xmm0, %xmm0
 ; AVX1-NEXT:    vpand %xmm5, %xmm2, %xmm2
 ; AVX1-NEXT:    vpackuswb %xmm0, %xmm2, %xmm0
 ; AVX1-NEXT:    vpand %xmm5, %xmm1, %xmm1
 ; AVX1-NEXT:    vpand %xmm5, %xmm4, %xmm2
 ; AVX1-NEXT:    vpackuswb %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vpackuswb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpblendvb %xmm3, %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpandn %xmm0, %xmm3, %xmm0
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: test14:
 ; AVX2:       # BB#0: # %vector.ph
 ; AVX2-NEXT:    vpshufd {{.*#+}} xmm3 = xmm0[2,3,0,1]
 ; AVX2-NEXT:    vpmovzxbd {{.*#+}} ymm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero,xmm3[2],zero,zero,zero,xmm3[3],zero,zero,zero,xmm3[4],zero,zero,zero,xmm3[5],zero,zero,zero,xmm3[6],zero,zero,zero,xmm3[7],zero,zero,zero
 ; AVX2-NEXT:    vpmovzxbd {{.*#+}} ymm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero

 ; AVX2-NEXT:    vpshufb %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = <0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u>
 ; AVX2-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
 ; AVX2-NEXT:    vpshufb %ymm1, %ymm2, %ymm1
 ; AVX2-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,2,2,3]
 ; AVX2-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
 ; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpblendvb %xmm4, %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpandn %xmm0, %xmm4, %xmm0
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
 vector.ph:
   %rhs = zext <16 x i8> %x to <16 x i32>
   %cond = icmp ult <16 x i32> %y, %rhs
   %sub = sub <16 x i32> %y, %rhs
   %truncsub = trunc <16 x i32> %sub to <16 x i8>
   %res = select <16 x i1> %cond, <16 x i8> zeroinitializer, <16 x i8> %truncsub

test/CodeGen/X86/vselect-zero.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2   | FileCheck %s --check-prefix=SSE --check-prefix=SSE2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.2 | FileCheck %s --check-prefix=SSE --check-prefix=SSE42
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx    | FileCheck %s --check-prefix=AVX --check-prefix=AVX1
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2   | FileCheck %s --check-prefix=AVX --check-prefix=AVX2
 
 ; PR28925
 
 define <4 x i32> @test1(<4 x i1> %cond, <4 x i32> %x) {
-; SSE2-LABEL: test1:
-; SSE2:       # BB#0:
-; SSE2-NEXT:    pslld $31, %xmm0
-; SSE2-NEXT:    psrad $31, %xmm0
-; SSE2-NEXT:    pandn %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE42-LABEL: test1:
-; SSE42:       # BB#0:
-; SSE42-NEXT:    pslld $31, %xmm0
-; SSE42-NEXT:    xorps %xmm2, %xmm2
-; SSE42-NEXT:    blendvps %xmm0, %xmm2, %xmm1
-; SSE42-NEXT:    movaps %xmm1, %xmm0
-; SSE42-NEXT:    retq
+; SSE-LABEL: test1:
+; SSE:       # BB#0:
+; SSE-NEXT:    pslld $31, %xmm0
+; SSE-NEXT:    psrad $31, %xmm0
+; SSE-NEXT:    pandn %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: test1:
 ; AVX:       # BB#0:
 ; AVX-NEXT:    vpslld $31, %xmm0, %xmm0
-; AVX-NEXT:    vxorps %xmm2, %xmm2, %xmm2
-; AVX-NEXT:    vblendvps %xmm0, %xmm2, %xmm1, %xmm0
+; AVX-NEXT:    vpsrad $31, %xmm0, %xmm0
+; AVX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %r = select <4 x i1> %cond, <4 x i32> zeroinitializer, <4 x i32> %x
   ret <4 x i32> %r
 }
 
 define <4 x i32> @test2(<4 x float> %a, <4 x float> %b, <4 x i32> %x) {
 ; SSE-LABEL: test2:
 ; SSE:       # BB#0: