Revert r373172 "[X86] Add custom isel logic to match VPTERNLOG from 2 logic ops."

This seems to be causing some performance regresions that I'm
trying to investigate.

One thing that stands out is that this transform can increase
the live range of the operands of the earlier logic op. This
can be bad for register allocation. If there are two logic
op inputs we should really combine the one that is closest, but
SelectionDAG doesn't have a good way to do that. Maybe we need
to do this as a basic block transform in Machine IR.

llvm-svn: 373401

Author: topperc

llvm/lib/Target/X86/X86ISelDAGToDAG.cpp

@@ -514,7 +514,6 @@ namespace {
     bool tryShiftAmountMod(SDNode *N);
     bool combineIncDecVector(SDNode *Node);
     bool tryShrinkShlLogicImm(SDNode *N);
-    bool tryVPTERNLOG(SDNode *N);
     bool tryVPTESTM(SDNode *Root, SDValue Setcc, SDValue Mask);
     bool tryMatchBitSelect(SDNode *N);
 
@@ -3833,82 +3832,6 @@ bool X86DAGToDAGISel::tryShrinkShlLogicImm(SDNode *N) {
   return true;
 }
 
-// Try to match two logic ops to a VPTERNLOG.
-// FIXME: Handle inverted inputs?
-// FIXME: Handle more complex patterns that use an operand more than once?
-// FIXME: Support X86ISD::ANDNP
-bool X86DAGToDAGISel::tryVPTERNLOG(SDNode *N) {
-  MVT NVT = N->getSimpleValueType(0);
-
-  // Make sure we support VPTERNLOG.
-  if (!NVT.isVector() || !Subtarget->hasAVX512() ||
-      NVT.getVectorElementType() == MVT::i1)
-    return false;
-
-  // We need VLX for 128/256-bit.
-  if (!(Subtarget->hasVLX() || NVT.is512BitVector()))
-    return false;
-
-  unsigned Opc1 = N->getOpcode();
-  SDValue N0 = N->getOperand(0);
-  SDValue N1 = N->getOperand(1);
-
-  auto isLogicOp = [](unsigned Opc) {
-    return Opc == ISD::AND || Opc == ISD::OR || Opc == ISD::XOR;
-  };
-
-  SDValue A, B, C;
-  unsigned Opc2;
-  if (isLogicOp(N1.getOpcode()) && N1.hasOneUse()) {
-    Opc2 = N1.getOpcode();
-    A = N0;
-    B = N1.getOperand(0);
-    C = N1.getOperand(1);
-  } else if (isLogicOp(N0.getOpcode()) && N0.hasOneUse()) {
-    Opc2 = N0.getOpcode();
-    A = N1;
-    B = N0.getOperand(0);
-    C = N0.getOperand(1);
-  } else
-    return false;
-
-  uint64_t Imm;
-  switch (Opc1) {
-  default: llvm_unreachable("Unexpected opcode!");
-  case ISD::AND:
-    switch (Opc2) {
-    default: llvm_unreachable("Unexpected opcode!");
-    case ISD::AND: Imm = 0x80; break;
-    case ISD::OR:  Imm = 0xe0; break;
-    case ISD::XOR: Imm = 0x60; break;
-    }
-    break;
-  case ISD::OR:
-    switch (Opc2) {
-    default: llvm_unreachable("Unexpected opcode!");
-    case ISD::AND: Imm = 0xf8; break;
-    case ISD::OR:  Imm = 0xfe; break;
-    case ISD::XOR: Imm = 0xf6; break;
-    }
-    break;
-  case ISD::XOR:
-    switch (Opc2) {
-    default: llvm_unreachable("Unexpected opcode!");
-    case ISD::AND: Imm = 0x78; break;
-    case ISD::OR:  Imm = 0x1e; break;
-    case ISD::XOR: Imm = 0x96; break;
-    }
-    break;
-  }
-
-  SDLoc DL(N);
-  SDValue New = CurDAG->getNode(X86ISD::VPTERNLOG, DL, NVT, A, B, C,
-                                CurDAG->getTargetConstant(Imm, DL, MVT::i8));
-  ReplaceNode(N, New.getNode());
-  SelectCode(New.getNode());
-  return true;
-}
-
 /// Convert vector increment or decrement to sub/add with an all-ones constant:
 /// add X, <1, 1...> --> sub X, <-1, -1...>
 /// sub X, <1, 1...> --> add X, <-1, -1...>
@@ -4580,10 +4503,9 @@ void X86DAGToDAGISel::Select(SDNode *Node) {
   case ISD::XOR:
     if (tryShrinkShlLogicImm(Node))
       return;
+
     if (Opcode == ISD::OR && tryMatchBitSelect(Node))
       return;
-    if (tryVPTERNLOG(Node))
-      return;
 
     LLVM_FALLTHROUGH;
   case ISD::ADD:

llvm/test/CodeGen/X86/avx512-cvt.ll

@@ -331,8 +331,8 @@ define <4 x float> @ulto4f32(<4 x i64> %a) {
 define <8 x double> @ulto8f64(<8 x i64> %a) {
 ; NODQ-LABEL: ulto8f64:
 ; NODQ:       # %bb.0:
-; NODQ-NEXT:    vpbroadcastq {{.*#+}} zmm1 = [4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200]
-; NODQ-NEXT:    vpternlogq $248, {{.*}}(%rip){1to8}, %zmm0, %zmm1
+; NODQ-NEXT:    vpandq {{.*}}(%rip){1to8}, %zmm0, %zmm1
+; NODQ-NEXT:    vporq {{.*}}(%rip){1to8}, %zmm1, %zmm1
 ; NODQ-NEXT:    vpsrlq $32, %zmm0, %zmm0
 ; NODQ-NEXT:    vporq {{.*}}(%rip){1to8}, %zmm0, %zmm0
 ; NODQ-NEXT:    vsubpd {{.*}}(%rip){1to8}, %zmm0, %zmm0
@@ -356,20 +356,21 @@ define <16 x double> @ulto16f64(<16 x i64> %a) {
 ; NODQ-LABEL: ulto16f64:
 ; NODQ:       # %bb.0:
 ; NODQ-NEXT:    vpbroadcastq {{.*#+}} zmm2 = [4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,4294967295]
-; NODQ-NEXT:    vpbroadcastq {{.*#+}} zmm3 = [4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200]
-; NODQ-NEXT:    vmovdqa64 %zmm3, %zmm4
-; NODQ-NEXT:    vpternlogq $248, %zmm2, %zmm0, %zmm4
+; NODQ-NEXT:    vpandq %zmm2, %zmm0, %zmm3
+; NODQ-NEXT:    vpbroadcastq {{.*#+}} zmm4 = [4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200,4841369599423283200]
+; NODQ-NEXT:    vporq %zmm4, %zmm3, %zmm3
 ; NODQ-NEXT:    vpsrlq $32, %zmm0, %zmm0
 ; NODQ-NEXT:    vpbroadcastq {{.*#+}} zmm5 = [4985484787499139072,4985484787499139072,4985484787499139072,4985484787499139072,4985484787499139072,4985484787499139072,4985484787499139072,4985484787499139072]
 ; NODQ-NEXT:    vporq %zmm5, %zmm0, %zmm0
 ; NODQ-NEXT:    vbroadcastsd {{.*#+}} zmm6 = [1.9342813118337666E+25,1.9342813118337666E+25,1.9342813118337666E+25,1.9342813118337666E+25,1.9342813118337666E+25,1.9342813118337666E+25,1.9342813118337666E+25,1.9342813118337666E+25]
 ; NODQ-NEXT:    vsubpd %zmm6, %zmm0, %zmm0
-; NODQ-NEXT:    vaddpd %zmm0, %zmm4, %zmm0
-; NODQ-NEXT:    vpternlogq $248, %zmm2, %zmm1, %zmm3
+; NODQ-NEXT:    vaddpd %zmm0, %zmm3, %zmm0
+; NODQ-NEXT:    vpandq %zmm2, %zmm1, %zmm2
+; NODQ-NEXT:    vporq %zmm4, %zmm2, %zmm2
 ; NODQ-NEXT:    vpsrlq $32, %zmm1, %zmm1
 ; NODQ-NEXT:    vporq %zmm5, %zmm1, %zmm1
 ; NODQ-NEXT:    vsubpd %zmm6, %zmm1, %zmm1
-; NODQ-NEXT:    vaddpd %zmm1, %zmm3, %zmm1
+; NODQ-NEXT:    vaddpd %zmm1, %zmm2, %zmm1
 ; NODQ-NEXT:    retq
 ;
 ; VLDQ-LABEL: ulto16f64: