[AVX512] Fix VSQRT packed instructions internal names.

[oota-llvm.git] / lib / Target / X86 / X86RegisterInfo.cpp

diff --git a/lib/Target/X86/X86RegisterInfo.cpp b/lib/Target/X86/X86RegisterInfo.cpp
index 97f86e2760d4a893aa7ebf5ef3b2da77d3205aa6..3fea9b028d128cd69b24d98b7fe73870e1cc8cc1 100644 (file)
--- a/lib/Target/X86/X86RegisterInfo.cpp
+++ b/lib/Target/X86/X86RegisterInfo.cpp
@@ -443,10 +443,8 @@ bool X86RegisterInfo::canRealignStack(const MachineFunction &MF) const {
 bool X86RegisterInfo::needsStackRealignment(const MachineFunction &MF) const {
   const MachineFrameInfo *MFI = MF.getFrameInfo();
   const Function *F = MF.getFunction();
-  unsigned StackAlign = MF.getTarget()
-                            .getSubtargetImpl()
-                            ->getFrameLowering()
-                            ->getStackAlignment();
+  unsigned StackAlign =
+    MF.getSubtarget().getFrameLowering()->getStackAlignment();
   bool requiresRealignment =
     ((MFI->getMaxAlignment() > StackAlign) ||
      F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
@@ -489,6 +487,12 @@ X86RegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
   else
     BasePtr = (TFI->hasFP(MF) ? FramePtr : StackPtr);
 
+  // For LEA64_32r when BasePtr is 32-bits (X32) we can use full-size 64-bit
+  // register as source operand, semantic is the same and destination is
+  // 32-bits. It saves one byte per lea in code since 0x67 prefix is avoided.
+  if (Opc == X86::LEA64_32r && X86::GR32RegClass.contains(BasePtr))
+    BasePtr = getX86SubSuperRegister(BasePtr, MVT::i64, false);
+
   // This must be part of a four operand memory reference.  Replace the
   // FrameIndex with base register with EBP.  Add an offset to the offset.
   MI.getOperand(FIOperandNum).ChangeToRegister(BasePtr, false);