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

RKSimon · RKSimon · commit 00161c996106 · 2017-09-18T14:23:23.000Z
As discussed on PR28925 and D37849. Differential Revision: https://reviews.llvm.org/D37975 llvm-svn: 313532
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -30090,6 +30090,7 @@ static SDValue combineExtractVectorElt_SSE(SDNode *N, SelectionDAG &DAG,
 
 /// 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,
@@ -30153,6 +30154,10 @@ combineVSelectWithAllOnesOrZeros(SDNode *N, SelectionDAG &DAG,
     }
   }
 
+  // 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);
@@ -30174,6 +30179,15 @@ combineVSelectWithAllOnesOrZeros(SDNode *N, SelectionDAG &DAG,
     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();
 }
 
diff --git a/llvm/test/CodeGen/X86/psubus.ll b/llvm/test/CodeGen/X86/psubus.ll
@@ -683,9 +683,7 @@ define <16 x i8> @test14(<16 x i8> %x, <16 x i32> %y) nounwind {
 ; 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:
@@ -727,8 +725,7 @@ define <16 x i8> @test14(<16 x i8> %x, <16 x i32> %y) nounwind {
 ; 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
 ;
@@ -760,8 +757,7 @@ define <16 x i8> @test14(<16 x i8> %x, <16 x i32> %y) nounwind {
 ; 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:
diff --git a/llvm/test/CodeGen/X86/vselect-zero.ll b/llvm/test/CodeGen/X86/vselect-zero.ll
@@ -7,26 +7,18 @@
 ; 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
