On Haswell, perfer storing YMM registers using a single instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@157129 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 203c873..2810f42 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -14532,13 +14532,12 @@
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
 
   // If we are saving a concatenation of two XMM registers, perform two stores.
-  // This is better in Sandy Bridge cause one 256-bit mem op is done via two
-  // 128-bit ones. If in the future the cost becomes only one memory access the
-  // first version would be better.
-  if (VT.getSizeInBits() == 256 &&
+  // 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.getSizeInBits() == 256 && !Subtarget->hasAVX2() &&
       StoredVal.getNode()->getOpcode() == ISD::CONCAT_VECTORS &&
       StoredVal.getNumOperands() == 2) {
-
     SDValue Value0 = StoredVal.getOperand(0);
     SDValue Value1 = StoredVal.getOperand(1);