Fixed a bug which causes x86 be to incorrectly match

shuffle v, undef, <2, ?, 3, ?>
to movhlps
It should match to unpckhps instead.

Added proper matching code for
shuffle v, undef, <2, 3, 2, 3>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31519 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index ac7327f9e9531433058b6dc24d2c4f62c1906d84..55022dab247f09695e947fdd837ecdab9e7fa605 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -2582,6 +2582,22 @@ bool X86::isMOVHLPSMask(SDNode *N) {
          isUndefOrEqual(N->getOperand(3), 3);
 }
 
+/// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
+/// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
+/// <2, 3, 2, 3>
+bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
+  assert(N->getOpcode() == ISD::BUILD_VECTOR);
+
+  if (N->getNumOperands() != 4)
+    return false;
+
+  // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
+  return isUndefOrEqual(N->getOperand(0), 2) &&
+         isUndefOrEqual(N->getOperand(1), 3) &&
+         isUndefOrEqual(N->getOperand(2), 2) &&
+         isUndefOrEqual(N->getOperand(3), 3);
+}
+
 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
 bool X86::isMOVLPMask(SDNode *N) {
@@ -3724,7 +3740,7 @@ X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
     SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
                                  &IdxVec[0], IdxVec.size());
     Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
-                      Vec, Vec, Mask);
+                      Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
     return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
                        DAG.getConstant(0, getPointerTy()));
   } else if (MVT::getSizeInBits(VT) == 64) {

diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index dde123c02f083f5fe7a27fb2d25733d657da92c8..68088ad2654dda2d0a2e7bd38d696b53b93d436f 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -186,6 +186,11 @@ namespace llvm {
    /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
    bool isMOVHLPSMask(SDNode *N);
 
+   /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
+   /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
+   /// <2, 3, 2, 3>
+   bool isMOVHLPS_v_undef_Mask(SDNode *N);
+
    /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
    /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
    bool isMOVLPMask(SDNode *N);

diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index 8154d1730ad06823e10ccb6a6c4f2e0d68392e7b..b49beaf5b2f4523d7df0b49ec0d73d8808b8642c 100644 (file)
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -110,6 +110,10 @@ def MOVHLPS_shuffle_mask : PatLeaf<(build_vector), [{
   return X86::isMOVHLPSMask(N);
 }]>;
 
+def MOVHLPS_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
+  return X86::isMOVHLPS_v_undef_Mask(N);
+}]>;
+
 def MOVHP_shuffle_mask : PatLeaf<(build_vector), [{
   return X86::isMOVHPMask(N);
 }]>;
@@ -1987,16 +1991,16 @@ def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
                   MOVHLPS_shuffle_mask)),
           (MOVHLPSrr VR128:$src1, VR128:$src2)>;
 
-// vector_shuffle v1, undef <2, 3, ?, ?> using MOVHLPS
+// vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
-                  UNPCKH_shuffle_mask)),
+                  MOVHLPS_v_undef_shuffle_mask)),
           (MOVHLPSrr VR128:$src1, VR128:$src1)>;
 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
-                  UNPCKH_shuffle_mask)),
+                  MOVHLPS_v_undef_shuffle_mask)),
           (MOVHLPSrr VR128:$src1, VR128:$src1)>;
 }
 
-  let AddedComplexity = 20 in {
+let AddedComplexity = 20 in {
 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (loadv4f32 addr:$src2),