feat: move combineVSelectWithAllOnesOrZeros to DAGCombiner and x86 test

Author: woruyu

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -12945,6 +12945,85 @@ SDValue DAGCombiner::visitVP_SELECT(SDNode *N) {
   return SDValue();
 }
 
+static SDValue combineVSelectWithAllOnesOrZeros(SDValue Cond, SDValue TVal,
+                                                SDValue FVal,
+                                                const TargetLowering &TLI,
+                                                SelectionDAG &DAG,
+                                                const SDLoc &DL) {
+  if (!TLI.isTypeLegal(TVal.getValueType()))
+    return SDValue();
+
+  EVT VT = TVal.getValueType();
+  EVT CondVT = Cond.getValueType();
+
+  assert(CondVT.isVector() && "Vector select expects a vector selector!");
+
+  // Classify TVal/FVal content
+  bool IsTAllZero = ISD::isBuildVectorAllZeros(TVal.getNode());
+  bool IsTAllOne = ISD::isBuildVectorAllOnes(TVal.getNode());
+  bool IsFAllZero = ISD::isBuildVectorAllZeros(FVal.getNode());
+  bool IsFAllOne = ISD::isBuildVectorAllOnes(FVal.getNode());
+
+  // no vselect(cond, 0/-1, X) or vselect(cond, X, 0/-1), return
+  if (!(IsTAllZero || IsTAllOne || IsFAllZero || IsFAllOne))
+    return SDValue();
+
+  // select Cond, 0, 0 → 0
+  if (IsTAllZero && IsFAllZero) {
+    return VT.isFloatingPoint() ? DAG.getConstantFP(0.0, DL, VT)
+                                : DAG.getConstant(0, DL, VT);
+  }
+
+  // To use the condition operand as a bitwise mask, it must have elements that
+  // are the same size as the select elements. Ie, the condition operand must
+  // have already been promoted from the IR select condition type <N x i1>.
+  // Don't check if the types themselves are equal because that excludes
+  // vector floating-point selects.
+  if (CondVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
+    return SDValue();
+
+  // Try inverting Cond and swapping T/F if it gives all-ones/all-zeros form
+  if (!IsTAllOne && !IsFAllZero && Cond.hasOneUse() &&
+      Cond.getOpcode() == ISD::SETCC &&
+      TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
+          CondVT) {
+    if (IsTAllZero || IsFAllOne) {
+      SDValue CC = Cond.getOperand(2);
+      ISD::CondCode InverseCC = ISD::getSetCCInverse(
+          cast<CondCodeSDNode>(CC)->get(), Cond.getOperand(0).getValueType());
+      Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1),
+                          InverseCC);
+      std::swap(TVal, FVal);
+      std::swap(IsTAllOne, IsFAllOne);
+      std::swap(IsTAllZero, IsFAllZero);
+    }
+  }
+
+  // Cond value must be 'sign splat' to be converted to a logical op.
+  if (DAG.ComputeNumSignBits(Cond) != CondVT.getScalarSizeInBits())
+    return SDValue();
+
+  // select Cond, -1, 0 → bitcast Cond
+  if (IsTAllOne && IsFAllZero)
+    return DAG.getBitcast(VT, Cond);
+
+  // select Cond, -1, x → or Cond, x
+  if (IsTAllOne) {
+    SDValue X = DAG.getBitcast(CondVT, FVal);
+    SDValue Or = DAG.getNode(ISD::OR, DL, CondVT, Cond, X);
+    return DAG.getBitcast(VT, Or);
+  }
+
+  // select Cond, x, 0 → and Cond, x
+  if (IsFAllZero) {
+    SDValue X = DAG.getBitcast(CondVT, TVal);
+    SDValue And = DAG.getNode(ISD::AND, DL, CondVT, Cond, X);
+    return DAG.getBitcast(VT, And);
+  }
+
+  return SDValue();
+}
+
 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
   SDValue N0 = N->getOperand(0);
   SDValue N1 = N->getOperand(1);
@@ -13213,6 +13292,9 @@ SDValue DAGCombiner::visitVSELECT(SDNode *N) {
   if (SimplifyDemandedVectorElts(SDValue(N, 0)))
     return SDValue(N, 0);
 
+  if (SDValue V = combineVSelectWithAllOnesOrZeros(N0, N1, N2, TLI, DAG, DL))
+    return V;
+
   return SDValue();
 }
 

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -47264,13 +47264,14 @@ static SDValue combineToExtendBoolVectorInReg(
                      DAG.getConstant(EltSizeInBits - 1, DL, VT));
 }
 
-/// If a vector select has an operand that is -1 or 0, try to simplify the
+/// If a vector select has an left operand that is 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, const SDLoc &DL,
-                                 TargetLowering::DAGCombinerInfo &DCI,
-                                 const X86Subtarget &Subtarget) {
+/// TODO: Move to DAGCombiner.combineVSelectWithAllOnesOrZeros, possibly using
+/// TargetLowering::hasAndNot()?
+static SDValue combineVSelectWithLastZeros(SDNode *N, SelectionDAG &DAG,
+                                           const SDLoc &DL,
+                                           TargetLowering::DAGCombinerInfo &DCI,
+                                           const X86Subtarget &Subtarget) {
   SDValue Cond = N->getOperand(0);
   SDValue LHS = N->getOperand(1);
   SDValue RHS = N->getOperand(2);
@@ -47283,20 +47284,6 @@ combineVSelectWithAllOnesOrZeros(SDNode *N, SelectionDAG &DAG, const SDLoc &DL,
 
   assert(CondVT.isVector() && "Vector select expects a vector selector!");
 
-  // TODO: Use isNullOrNullSplat() to distinguish constants with undefs?
-  // TODO: Can we assert that both operands are not zeros (because that should
-  //       get simplified at node creation time)?
-  bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
-  bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
-
-  // If both inputs are 0/undef, create a complete zero vector.
-  // FIXME: As noted above this should be handled by DAGCombiner/getNode.
-  if (TValIsAllZeros && FValIsAllZeros) {
-    if (VT.isFloatingPoint())
-      return DAG.getConstantFP(0.0, DL, VT);
-    return DAG.getConstant(0, DL, VT);
-  }
-
   // To use the condition operand as a bitwise mask, it must have elements that
   // are the same size as the select elements. Ie, the condition operand must
   // have already been promoted from the IR select condition type <N x i1>.
@@ -47305,56 +47292,15 @@ combineVSelectWithAllOnesOrZeros(SDNode *N, SelectionDAG &DAG, const SDLoc &DL,
   if (CondVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
     return SDValue();
 
-  // Try to invert the condition if true value is not all 1s and false value is
-  // not all 0s. Only do this if the condition has one use.
-  bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
-  if (!TValIsAllOnes && !FValIsAllZeros && Cond.hasOneUse() &&
-      // Check if the selector will be produced by CMPP*/PCMP*.
-      Cond.getOpcode() == ISD::SETCC &&
-      // Check if SETCC has already been promoted.
-      TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
-          CondVT) {
-    bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
-
-    if (TValIsAllZeros || FValIsAllOnes) {
-      SDValue CC = Cond.getOperand(2);
-      ISD::CondCode NewCC = ISD::getSetCCInverse(
-          cast<CondCodeSDNode>(CC)->get(), Cond.getOperand(0).getValueType());
-      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 (!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) {
+  if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
     SDValue CastRHS = DAG.getBitcast(CondVT, RHS);
     SDValue AndN;
     // The canonical form differs for i1 vectors - x86andnp is not used
@@ -48117,7 +48063,7 @@ static SDValue combineSelect(SDNode *N, SelectionDAG &DAG,
   if (!TLI.isTypeLegal(VT) || isSoftF16(VT, Subtarget))
     return SDValue();
 
-  if (SDValue V = combineVSelectWithAllOnesOrZeros(N, DAG, DL, DCI, Subtarget))
+  if (SDValue V = combineVSelectWithLastZeros(N, DAG, DL, DCI, Subtarget))
     return V;
 
   if (SDValue V = combineVSelectToBLENDV(N, DAG, DL, DCI, Subtarget))

llvm/test/CodeGen/X86/urem-seteq-vec-tautological.ll

@@ -77,11 +77,9 @@ define <4 x i1> @t1_all_odd_ne(<4 x i32> %X) nounwind {
 ; CHECK-SSE2-LABEL: t1_all_odd_ne:
 ; CHECK-SSE2:       # %bb.0:
 ; CHECK-SSE2-NEXT:    pmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; CHECK-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; CHECK-SSE2-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; CHECK-SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; CHECK-SSE2-NEXT:    pcmpeqd %xmm1, %xmm1
-; CHECK-SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; CHECK-SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; CHECK-SSE2-NEXT:    retq
 ;
 ; CHECK-SSE41-LABEL: t1_all_odd_ne:
@@ -92,7 +90,7 @@ define <4 x i1> @t1_all_odd_ne(<4 x i32> %X) nounwind {
 ; CHECK-SSE41-NEXT:    pcmpeqd %xmm1, %xmm0
 ; CHECK-SSE41-NEXT:    pcmpeqd %xmm1, %xmm1
 ; CHECK-SSE41-NEXT:    pxor %xmm1, %xmm0
-; CHECK-SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
+; CHECK-SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; CHECK-SSE41-NEXT:    retq
 ;
 ; CHECK-AVX1-LABEL: t1_all_odd_ne:
@@ -102,7 +100,7 @@ define <4 x i1> @t1_all_odd_ne(<4 x i32> %X) nounwind {
 ; CHECK-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; CHECK-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
 ; CHECK-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
-; CHECK-AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
+; CHECK-AVX1-NEXT:    vpor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
 ; CHECK-AVX1-NEXT:    retq
 ;
 ; CHECK-AVX2-LABEL: t1_all_odd_ne:
@@ -113,17 +111,16 @@ define <4 x i1> @t1_all_odd_ne(<4 x i32> %X) nounwind {
 ; CHECK-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; CHECK-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
 ; CHECK-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
-; CHECK-AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3]
+; CHECK-AVX2-NEXT:    vpor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
 ; CHECK-AVX2-NEXT:    retq
 ;
 ; CHECK-AVX512VL-LABEL: t1_all_odd_ne:
 ; CHECK-AVX512VL:       # %bb.0:
 ; CHECK-AVX512VL-NEXT:    vpmulld {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0
 ; CHECK-AVX512VL-NEXT:    vpminud {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1
-; CHECK-AVX512VL-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; CHECK-AVX512VL-NEXT:    vpternlogq {{.*#+}} xmm0 = ~xmm0
-; CHECK-AVX512VL-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; CHECK-AVX512VL-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3]
+; CHECK-AVX512VL-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm1
+; CHECK-AVX512VL-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
+; CHECK-AVX512VL-NEXT:    vpternlogq {{.*#+}} xmm0 = m64bcst | (xmm0 ^ xmm1)
 ; CHECK-AVX512VL-NEXT:    retq
   %urem = urem <4 x i32> %X, <i32 3, i32 1, i32 1, i32 9>
   %cmp = icmp ne <4 x i32> %urem, <i32 0, i32 42, i32 0, i32 42>