On Sandybridge split unaligned 256bit stores into two xmm-sized stores.

[oota-llvm.git] / lib / Target / X86 / X86ISelLowering.cpp

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index b6b10e2dcab4f7882acc272612aae40775ac9cdc..ca8cd741e78fd4de0ec4ee9db94ceca0e29dde16 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -16344,12 +16344,15 @@ static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
 
   ISD::LoadExtType Ext = Ld->getExtensionType();
   unsigned Alignment = Ld->getAlignment();
+  bool IsAligned = Alignment == 0 || Alignment == MemVT.getSizeInBits()/8;
 
   // On Sandybridge unaligned 256bit loads are inefficient.
   if (RegVT.is256BitVector() && !Subtarget->hasInt256() &&
-      !DCI.isBeforeLegalizeOps() && Alignment < 32 &&
-      Ext == ISD::NON_EXTLOAD) {
+      !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
     unsigned NumElems = RegVT.getVectorNumElements();
+    if (NumElems < 2)
+      return SDValue();
+
     SDValue Ptr = Ld->getBasePtr();
     SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
 
@@ -16363,7 +16366,7 @@ static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
     SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
                                 Ld->getPointerInfo(), Ld->isVolatile(),
                                 Ld->isNonTemporal(), Ld->isInvariant(),
-                                Alignment);
+                                std::max(Alignment/2U, 1U));
     SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
                              Load1.getValue(1),
                              Load2.getValue(1));
@@ -16536,16 +16539,21 @@ static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
   DebugLoc dl = St->getDebugLoc();
   SDValue StoredVal = St->getOperand(1);
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+  unsigned Alignment = St->getAlignment();
+  bool IsAligned = Alignment == 0 || Alignment == VT.getSizeInBits()/8;
 
   // If we are saving a concatenation of two XMM registers, perform two stores.
   // On Sandy Bridge, 256-bit memory operations are executed by two
   // 128-bit ports. However, on Haswell it is better to issue a single 256-bit
   // memory  operation.
   if (VT.is256BitVector() && !Subtarget->hasInt256() &&
-      StoredVal.getNode()->getOpcode() == ISD::CONCAT_VECTORS &&
-      StoredVal.getNumOperands() == 2) {
-    SDValue Value0 = StoredVal.getOperand(0);
-    SDValue Value1 = StoredVal.getOperand(1);
+      StVT == VT && !IsAligned) {
+    unsigned NumElems = VT.getVectorNumElements();
+    if (NumElems < 2)
+      return SDValue();
+
+    SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
+    SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
 
     SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
     SDValue Ptr0 = St->getBasePtr();
@@ -16553,10 +16561,11 @@ static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
 
     SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
                                 St->getPointerInfo(), St->isVolatile(),
-                                St->isNonTemporal(), St->getAlignment());
+                                St->isNonTemporal(), Alignment);
     SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
                                 St->getPointerInfo(), St->isVolatile(),
-                                St->isNonTemporal(), St->getAlignment());
+                                St->isNonTemporal(),
+                                std::max(Alignment/2U, 1U));
     return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
   }