[llvm] r345564 - [X86] Stop changing f128 fand/for/fxor to v2i64.

[Craig Topper via llvm-commits]

Author: ctopper
Date: Mon Oct 29 20:27:11 2018
New Revision: 345564

URL: http://llvm.org/viewvc/llvm-project?rev=345564&view=rev
Log:
[X86] Stop changing f128 fand/for/fxor to v2i64.

The additional patterns don't cost us much and it seems better than changing element widths.

Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
    llvm/trunk/lib/Target/X86/X86InstrVecCompiler.td

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=345564&r1=345563&r2=345564&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Mon Oct 29 20:27:11 2018
@@ -37745,27 +37745,27 @@ static SDValue lowerX86FPLogicOp(SDNode
                                  const X86Subtarget &Subtarget) {
   MVT VT = N->getSimpleValueType(0);
   // If we have integer vector types available, use the integer opcodes.
-  if ((VT.isVector() || VT == MVT::f128) && Subtarget.hasSSE2()) {
-    SDLoc dl(N);
+  if (!VT.isVector() || !Subtarget.hasSSE2())
+    return SDValue();
 
-    unsigned IntBits = std::min(VT.getScalarSizeInBits(), 64U);
-    MVT IntSVT = MVT::getIntegerVT(IntBits);
-    MVT IntVT = MVT::getVectorVT(IntSVT, VT.getSizeInBits() / IntBits);
+  SDLoc dl(N);
 
-    SDValue Op0 = DAG.getBitcast(IntVT, N->getOperand(0));
-    SDValue Op1 = DAG.getBitcast(IntVT, N->getOperand(1));
-    unsigned IntOpcode;
-    switch (N->getOpcode()) {
-    default: llvm_unreachable("Unexpected FP logic op");
-    case X86ISD::FOR: IntOpcode = ISD::OR; break;
-    case X86ISD::FXOR: IntOpcode = ISD::XOR; break;
-    case X86ISD::FAND: IntOpcode = ISD::AND; break;
-    case X86ISD::FANDN: IntOpcode = X86ISD::ANDNP; break;
-    }
-    SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
-    return DAG.getBitcast(VT, IntOp);
+  unsigned IntBits = VT.getScalarSizeInBits();
+  MVT IntSVT = MVT::getIntegerVT(IntBits);
+  MVT IntVT = MVT::getVectorVT(IntSVT, VT.getSizeInBits() / IntBits);
+
+  SDValue Op0 = DAG.getBitcast(IntVT, N->getOperand(0));
+  SDValue Op1 = DAG.getBitcast(IntVT, N->getOperand(1));
+  unsigned IntOpcode;
+  switch (N->getOpcode()) {
+  default: llvm_unreachable("Unexpected FP logic op");
+  case X86ISD::FOR:   IntOpcode = ISD::OR; break;
+  case X86ISD::FXOR:  IntOpcode = ISD::XOR; break;
+  case X86ISD::FAND:  IntOpcode = ISD::AND; break;
+  case X86ISD::FANDN: IntOpcode = X86ISD::ANDNP; break;
   }
-  return SDValue();
+  SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
+  return DAG.getBitcast(VT, IntOp);
 }
 
 

Modified: llvm/trunk/lib/Target/X86/X86InstrVecCompiler.td
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86InstrVecCompiler.td?rev=345564&r1=345563&r2=345564&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86InstrVecCompiler.td (original)
+++ llvm/trunk/lib/Target/X86/X86InstrVecCompiler.td Mon Oct 29 20:27:11 2018
@@ -466,8 +466,6 @@ def : Pat<(loadf128 addr:$src),
           (VMOVUPSZ128rm addr:$src)>;
 }
 
-// With SSE2 the DAG combiner converts fp logic ops to integer logic ops to
-// reduce patterns.
 let Predicates = [UseSSE1] in {
 // andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2
 def : Pat<(f128 (X86fand VR128:$src1, (memopf128 addr:$src2))),
@@ -489,4 +487,23 @@ def : Pat<(f128 (X86fxor VR128:$src1, VR
           (XORPSrr VR128:$src1, VR128:$src2)>;
 }
 
+let Predicates = [HasAVX] in {
+// andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2
+def : Pat<(f128 (X86fand VR128:$src1, (loadf128 addr:$src2))),
+          (VANDPSrm VR128:$src1, f128mem:$src2)>;
+
+def : Pat<(f128 (X86fand VR128:$src1, VR128:$src2)),
+          (VANDPSrr VR128:$src1, VR128:$src2)>;
 
+def : Pat<(f128 (X86for VR128:$src1, (loadf128 addr:$src2))),
+          (VORPSrm VR128:$src1, f128mem:$src2)>;
+
+def : Pat<(f128 (X86for VR128:$src1, VR128:$src2)),
+          (VORPSrr VR128:$src1, VR128:$src2)>;
+
+def : Pat<(f128 (X86fxor VR128:$src1, (loadf128 addr:$src2))),
+          (VXORPSrm VR128:$src1, f128mem:$src2)>;
+
+def : Pat<(f128 (X86fxor VR128:$src1, VR128:$src2)),
+          (VXORPSrr VR128:$src1, VR128:$src2)>;
+}