[X86] isHorizontalBinOp - always create HADD/SUB if it will be merged with another existing HADD/SUB

Fixes some more cases from #34072 where undemanded vector elements prevent HADD/SUB being matched on slow targets

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -51571,7 +51571,8 @@ static SDValue combineVEXTRACT_STORE(SDNode *N, SelectionDAG &DAG,
 static bool isHorizontalBinOp(unsigned HOpcode, SDValue &LHS, SDValue &RHS,
                               SelectionDAG &DAG, const X86Subtarget &Subtarget,
                               bool IsCommutative,
-                              SmallVectorImpl<int> &PostShuffleMask) {
+                              SmallVectorImpl<int> &PostShuffleMask,
+                              bool ForceHorizOp) {
   // If either operand is undef, bail out. The binop should be simplified.
   if (LHS.isUndef() || RHS.isUndef())
     return false;
@@ -51734,13 +51735,12 @@ static bool isHorizontalBinOp(unsigned HOpcode, SDValue &LHS, SDValue &RHS,
 
   // If the source nodes are already used in HorizOps then always accept this.
   // Shuffle folding should merge these back together.
-  bool FoundHorizLHS = llvm::any_of(NewLHS->uses(), [&](SDNode *User) {
+  auto FoundHorizUser = [&](SDNode *User) {
     return User->getOpcode() == HOpcode && User->getValueType(0) == VT;
-  });
-  bool FoundHorizRHS = llvm::any_of(NewRHS->uses(), [&](SDNode *User) {
-    return User->getOpcode() == HOpcode && User->getValueType(0) == VT;
-  });
-  bool ForceHorizOp = FoundHorizLHS && FoundHorizRHS;
+  };
+  ForceHorizOp =
+      ForceHorizOp || (llvm::any_of(NewLHS->uses(), FoundHorizUser) &&
+                       llvm::any_of(NewRHS->uses(), FoundHorizUser));
 
   // Assume a SingleSource HOP if we only shuffle one input and don't need to
   // shuffle the result.
@@ -51763,6 +51763,13 @@ static SDValue combineToHorizontalAddSub(SDNode *N, SelectionDAG &DAG,
   bool IsAdd = (Opcode == ISD::FADD) || (Opcode == ISD::ADD);
   SmallVector<int, 8> PostShuffleMask;
 
+  auto MergableHorizOp = [N](unsigned HorizOpcode) {
+    return N->hasOneUse() &&
+           N->use_begin()->getOpcode() == ISD::VECTOR_SHUFFLE &&
+           (N->use_begin()->getOperand(0).getOpcode() == HorizOpcode ||
+            N->use_begin()->getOperand(1).getOpcode() == HorizOpcode);
+  };
+
   switch (Opcode) {
   case ISD::FADD:
   case ISD::FSUB:
@@ -51772,7 +51779,7 @@ static SDValue combineToHorizontalAddSub(SDNode *N, SelectionDAG &DAG,
       SDValue RHS = N->getOperand(1);
       auto HorizOpcode = IsAdd ? X86ISD::FHADD : X86ISD::FHSUB;
       if (isHorizontalBinOp(HorizOpcode, LHS, RHS, DAG, Subtarget, IsAdd,
-                            PostShuffleMask)) {
+                            PostShuffleMask, MergableHorizOp(HorizOpcode))) {
         SDValue HorizBinOp = DAG.getNode(HorizOpcode, SDLoc(N), VT, LHS, RHS);
         if (!PostShuffleMask.empty())
           HorizBinOp = DAG.getVectorShuffle(VT, SDLoc(HorizBinOp), HorizBinOp,
@@ -51789,7 +51796,7 @@ static SDValue combineToHorizontalAddSub(SDNode *N, SelectionDAG &DAG,
       SDValue RHS = N->getOperand(1);
       auto HorizOpcode = IsAdd ? X86ISD::HADD : X86ISD::HSUB;
       if (isHorizontalBinOp(HorizOpcode, LHS, RHS, DAG, Subtarget, IsAdd,
-                            PostShuffleMask)) {
+                            PostShuffleMask, MergableHorizOp(HorizOpcode))) {
         auto HOpBuilder = [HorizOpcode](SelectionDAG &DAG, const SDLoc &DL,
                                         ArrayRef<SDValue> Ops) {
           return DAG.getNode(HorizOpcode, DL, Ops[0].getValueType(), Ops);

llvm/test/CodeGen/X86/haddsub-undef.ll

@@ -1052,31 +1052,15 @@ define <4 x float> @PR34724_add_v4f32_0u23(<4 x float> %0, <4 x float> %1) {
 }
 
 define <4 x float> @PR34724_add_v4f32_01u3(<4 x float> %0, <4 x float> %1) {
-; SSE-SLOW-LABEL: PR34724_add_v4f32_01u3:
-; SSE-SLOW:       # %bb.0:
-; SSE-SLOW-NEXT:    haddps %xmm0, %xmm0
-; SSE-SLOW-NEXT:    movsldup {{.*#+}} xmm2 = xmm1[0,0,2,2]
-; SSE-SLOW-NEXT:    addps %xmm1, %xmm2
-; SSE-SLOW-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3]
-; SSE-SLOW-NEXT:    retq
-;
-; SSE-FAST-LABEL: PR34724_add_v4f32_01u3:
-; SSE-FAST:       # %bb.0:
-; SSE-FAST-NEXT:    haddps %xmm1, %xmm0
-; SSE-FAST-NEXT:    retq
-;
-; AVX-SLOW-LABEL: PR34724_add_v4f32_01u3:
-; AVX-SLOW:       # %bb.0:
-; AVX-SLOW-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
-; AVX-SLOW-NEXT:    vmovsldup {{.*#+}} xmm2 = xmm1[0,0,2,2]
-; AVX-SLOW-NEXT:    vaddps %xmm1, %xmm2, %xmm1
-; AVX-SLOW-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
-; AVX-SLOW-NEXT:    retq
+; SSE-LABEL: PR34724_add_v4f32_01u3:
+; SSE:       # %bb.0:
+; SSE-NEXT:    haddps %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
-; AVX-FAST-LABEL: PR34724_add_v4f32_01u3:
-; AVX-FAST:       # %bb.0:
-; AVX-FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
-; AVX-FAST-NEXT:    retq
+; AVX-LABEL: PR34724_add_v4f32_01u3:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    retq
   %3 = shufflevector <4 x float> %0, <4 x float> undef, <2 x i32> <i32 0, i32 2>
   %4 = shufflevector <4 x float> %0, <4 x float> undef, <2 x i32> <i32 1, i32 3>
   %5 = fadd <2 x float> %3, %4
@@ -1088,31 +1072,15 @@ define <4 x float> @PR34724_add_v4f32_01u3(<4 x float> %0, <4 x float> %1) {
 }
 
 define <4 x float> @PR34724_add_v4f32_012u(<4 x float> %0, <4 x float> %1) {
-; SSE-SLOW-LABEL: PR34724_add_v4f32_012u:
-; SSE-SLOW:       # %bb.0:
-; SSE-SLOW-NEXT:    haddps %xmm0, %xmm0
-; SSE-SLOW-NEXT:    movshdup {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE-SLOW-NEXT:    addps %xmm1, %xmm2
-; SSE-SLOW-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0]
-; SSE-SLOW-NEXT:    retq
-;
-; SSE-FAST-LABEL: PR34724_add_v4f32_012u:
-; SSE-FAST:       # %bb.0:
-; SSE-FAST-NEXT:    haddps %xmm1, %xmm0
-; SSE-FAST-NEXT:    retq
-;
-; AVX-SLOW-LABEL: PR34724_add_v4f32_012u:
-; AVX-SLOW:       # %bb.0:
-; AVX-SLOW-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
-; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; AVX-SLOW-NEXT:    vaddps %xmm1, %xmm2, %xmm1
-; AVX-SLOW-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX-SLOW-NEXT:    retq
+; SSE-LABEL: PR34724_add_v4f32_012u:
+; SSE:       # %bb.0:
+; SSE-NEXT:    haddps %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
-; AVX-FAST-LABEL: PR34724_add_v4f32_012u:
-; AVX-FAST:       # %bb.0:
-; AVX-FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
-; AVX-FAST-NEXT:    retq
+; AVX-LABEL: PR34724_add_v4f32_012u:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    retq
   %3 = shufflevector <4 x float> %0, <4 x float> undef, <2 x i32> <i32 0, i32 2>
   %4 = shufflevector <4 x float> %0, <4 x float> undef, <2 x i32> <i32 1, i32 3>
   %5 = fadd <2 x float> %3, %4