AVX-512: Fixed a bug in OR/XOR operations for 512-bit FP values on KNL.

KNL does not have VXORPS, VORPS for 512-bit values.
I use integer VPXOR, VPOR that actually do the same.

X86ISD::FXOR/FOR are generated as a result of FSUB combining.

Differential Revision: http://reviews.llvm.org/D12753

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@247523 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 4ac3b0f7ae5c12f9c39fb74c7ff947b0df319461..ee7c445d0b13db30200233486f02d76fd0613ee6 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -25152,7 +25152,8 @@ static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
 }
 
 /// 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
@@ -25164,6 +25165,19 @@ static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
   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();
 }
 
@@ -26027,7 +26041,7 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   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);

diff --git a/test/CodeGen/X86/avx512-arith.ll b/test/CodeGen/X86/avx512-arith.ll
index 1ecd1007905ae0f3eb59c3ecfa5e758a1db6667a..522abd261472682c5f23b54a85644c6fcabd37a8 100644 (file)
--- a/test/CodeGen/X86/avx512-arith.ll
+++ b/test/CodeGen/X86/avx512-arith.ll
@@ -652,3 +652,13 @@ define <8 x double> @test_maskz_broadcast_vaddpd(<8 x double> %i, double* %j,
   %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
+}
+