On Sandybridge loading unaligned 256bits using two XMM loads (vmovups and vinsertf128) is faster than using a single vmovups instruction.

llvm-svn: 172868

1 files changed, 31 insertions, 1 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 2b6ff3602a4..73a1d2e0075 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -16340,8 +16340,39 @@ static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
   EVT MemVT = Ld->getMemoryVT();
   DebugLoc dl = Ld->getDebugLoc();
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+  unsigned RegSz = RegVT.getSizeInBits();
 
   ISD::LoadExtType Ext = Ld->getExtensionType();
+  unsigned Alignment = Ld->getAlignment();
+
+  // On Sandybridge unaligned 256bit loads are inefficient.
+  if (RegVT.is256BitVector() && !Subtarget->hasInt256() &&
+      !DCI.isBeforeLegalizeOps() && Alignment < 32 &&
+      Ext == ISD::NON_EXTLOAD) {
+    unsigned NumElems = RegVT.getVectorNumElements();
+    SDValue Ptr = Ld->getBasePtr();
+    SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
+
+    EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
+                                  NumElems/2);
+    SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
+                                Ld->getPointerInfo(), Ld->isVolatile(),
+                                Ld->isNonTemporal(), Ld->isInvariant(),
+                                Alignment);
+    Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
+    SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
+                                Ld->getPointerInfo(), Ld->isVolatile(),
+                                Ld->isNonTemporal(), Ld->isInvariant(),
+                                Alignment);
+    SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+                             Load1.getValue(1),
+                             Load2.getValue(1));
+
+    SDValue NewVec = DAG.getUNDEF(RegVT);
+    NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
+    NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
+    return DCI.CombineTo(N, NewVec, TF, true);
+  }
 
   // If this is a vector EXT Load then attempt to optimize it using a
   // shuffle. If SSSE3 is not available we may emit an illegal shuffle but the
@@ -16356,7 +16387,6 @@ static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
     assert(MemVT.isVector() && "Must load a vector from memory");
 
     unsigned NumElems = RegVT.getVectorNumElements();
-    unsigned RegSz = RegVT.getSizeInBits();
     unsigned MemSz = MemVT.getSizeInBits();
     assert(RegSz > MemSz && "Register size must be greater than the mem size");