Use x86 specific MOVSHDUP node and add more patterns to match it
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@112657 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 1b18eb6..d9207d9 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -2586,6 +2586,7 @@
case X86ISD::SHUFPS:
case X86ISD::MOVLHPS:
case X86ISD::MOVHLPS:
+ case X86ISD::MOVSHDUP:
case X86ISD::MOVSS:
case X86ISD::MOVSD:
case X86ISD::PUNPCKLDQ:
@@ -2595,6 +2596,17 @@
}
static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
+ SDValue V1, SelectionDAG &DAG) {
+ switch(Opc) {
+ default: llvm_unreachable("Unknown x86 shuffle node");
+ case X86ISD::MOVSHDUP:
+ return DAG.getNode(Opc, dl, VT, V1);
+ }
+
+ return SDValue();
+}
+
+static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
switch(Opc) {
default: llvm_unreachable("Unknown x86 shuffle node");
@@ -5053,6 +5065,7 @@
bool V1IsSplat = false;
bool V2IsSplat = false;
bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
+ bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
MachineFunction &MF = DAG.getMachineFunction();
bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
@@ -5152,8 +5165,10 @@
if (X86::isMOVHLPSMask(SVOp))
return getMOVHighToLow(Op, dl, DAG);
- if (X86::isMOVSHDUPMask(SVOp) ||
- X86::isMOVSLDUPMask(SVOp) ||
+ if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
+ return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
+
+ if (X86::isMOVSLDUPMask(SVOp) ||
X86::isMOVHLPSMask(SVOp) ||
X86::isMOVLPMask(SVOp))
return Op;