AVX-512: Fixed a bug in FSUB lowering

[delena]

Summary

The compiler failed to lower X86ISD::FXOR node for 512-bit data type on KNL because VXORPS-zmm is supported in DQ instruction set.
The solution is using integer XOR - VPXOR*.

(The X86ISD::FOR is the same but nobody generates X86ISD::FOR for 512-bit data types).

[igorb]

LGTM

../lib/Target/X86/X86ISelLowering.cpp

static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
   if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
        (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
       isHorizontalBinOp(LHS, RHS, false))
     return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
   return SDValue();
 }
 
 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
-static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
+static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
+                                 const X86Subtarget *Subtarget) {
   assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
 
   // F[X]OR(0.0, x) -> x
   if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
     if (C->getValueAPF().isPosZero())
       return N->getOperand(1);
 
   // F[X]OR(x, 0.0) -> x
   if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
     if (C->getValueAPF().isPosZero())
       return N->getOperand(0);
+
+  EVT VT = N->getValueType(0);
+  if (VT.is512BitVector() && !Subtarget->hasDQI()) {
+    SDLoc dl(N);
+    MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
+    MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
+
+    SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
+    SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
+    unsigned IntOpcode = (N->getOpcode() == X86ISD::FOR) ? ISD::OR : ISD::XOR;
+    SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
+    return  DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
+  }
   return SDValue();
 }
 
 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
   assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
 
   // Only perform optimizations if UnsafeMath is used.
SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::MLOAD:          return PerformMLOADCombine(N, DAG, DCI, Subtarget);
   case ISD::STORE:          return PerformSTORECombine(N, DAG, Subtarget);
   case ISD::MSTORE:         return PerformMSTORECombine(N, DAG, Subtarget);
   case ISD::SINT_TO_FP:     return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
   case ISD::UINT_TO_FP:     return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
   case ISD::FADD:           return PerformFADDCombine(N, DAG, Subtarget);
   case ISD::FSUB:           return PerformFSUBCombine(N, DAG, Subtarget);
   case X86ISD::FXOR:
-  case X86ISD::FOR:         return PerformFORCombine(N, DAG);
+  case X86ISD::FOR:         return PerformFORCombine(N, DAG, Subtarget);
   case X86ISD::FMIN:
   case X86ISD::FMAX:        return PerformFMinFMaxCombine(N, DAG);
   case X86ISD::FAND:        return PerformFANDCombine(N, DAG);
   case X86ISD::FANDN:       return PerformFANDNCombine(N, DAG);
   case X86ISD::BT:          return PerformBTCombine(N, DAG, DCI);
   case X86ISD::VZEXT_MOVL:  return PerformVZEXT_MOVLCombine(N, DAG);
   case ISD::ANY_EXTEND:
   case ISD::ZERO_EXTEND:    return PerformZExtCombine(N, DAG, DCI, Subtarget);

../test/CodeGen/X86/avx512-arith.ll

define <8 x double> @test_maskz_broadcast_vaddpd(<8 x double> %i, double* %j,
   %tmp = load double, double* %j
   %b = insertelement <8 x double> undef, double %tmp, i32 0
   %c = shufflevector <8 x double> %b, <8 x double> undef,
                      <8 x i32> zeroinitializer
   %x = fadd <8 x double> %c, %i
   %r = select <8 x i1> %mask, <8 x double> %x, <8 x double> zeroinitializer
   ret <8 x double> %r
 }
+
+; CHECK-LABEL: test_fxor
+; CHECK: vpxord
+; CHECK: ret
+define <16 x float>  @test_fxor(<16 x float> %a) {
+
+  %res = fsub <16 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %a
+  ret <16 x float>%res
+}
+