Merge remote-tracking branch 'upstream/master' into merge-20130502

Conflicts:
	lib/Support/Unix/Signals.inc
	unittests/Transforms/Utils/Cloning.cpp

Change-Id: I027581a4390ec3ce4cd8d33da8b5f4c0c7d372c8

1 files changed, 50 insertions, 1 deletions

diff --git a/lib/Target/R600/SIInstrInfo.cpp b/lib/Target/R600/SIInstrInfo.cpp
index de2373b..9a04c60 100644
--- a/lib/Target/R600/SIInstrInfo.cpp
+++ b/lib/Target/R600/SIInstrInfo.cpp
@@ -58,6 +58,10 @@ SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
     AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3, 0
   };
 
+  const int16_t Sub0_2[] = {
+    AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, 0
+  };
+
   const int16_t Sub0_1[] = {
     AMDGPU::sub0, AMDGPU::sub1, 0
   };
@@ -65,6 +69,26 @@ SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
   unsigned Opcode;
   const int16_t *SubIndices;
 
+  if (AMDGPU::M0 == DestReg) {
+    // Check if M0 isn't already set to this value
+    for (MachineBasicBlock::reverse_iterator E = MBB.rend(),
+      I = MachineBasicBlock::reverse_iterator(MI); I != E; ++I) {
+
+      if (!I->definesRegister(AMDGPU::M0))
+        continue;
+
+      unsigned Opc = I->getOpcode();
+      if (Opc != TargetOpcode::COPY && Opc != AMDGPU::S_MOV_B32)
+        break;
+
+      if (!I->readsRegister(SrcReg))
+        break;
+
+      // The copy isn't necessary
+      return;
+    }
+  }
+
   if (AMDGPU::SReg_32RegClass.contains(DestReg)) {
     assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
     BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
@@ -105,6 +129,11 @@ SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
     Opcode = AMDGPU::V_MOV_B32_e32;
     SubIndices = Sub0_1;
 
+  } else if (AMDGPU::VReg_96RegClass.contains(DestReg)) {
+    assert(AMDGPU::VReg_96RegClass.contains(SrcReg));
+    Opcode = AMDGPU::V_MOV_B32_e32;
+    SubIndices = Sub0_2;
+
   } else if (AMDGPU::VReg_128RegClass.contains(DestReg)) {
     assert(AMDGPU::VReg_128RegClass.contains(SrcReg) ||
 	   AMDGPU::SReg_128RegClass.contains(SrcReg));
@@ -138,6 +167,21 @@ SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
   }
 }
 
+unsigned SIInstrInfo::commuteOpcode(unsigned Opcode) const {
+
+  int NewOpc;
+
+  // Try to map original to commuted opcode
+  if ((NewOpc = AMDGPU::getCommuteRev(Opcode)) != -1)
+    return NewOpc;
+
+  // Try to map commuted to original opcode
+  if ((NewOpc = AMDGPU::getCommuteOrig(Opcode)) != -1)
+    return NewOpc;
+
+  return Opcode;
+}
+
 MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
                                               bool NewMI) const {
 
@@ -145,7 +189,12 @@ MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
       !MI->getOperand(2).isReg())
     return 0;
 
-  return TargetInstrInfo::commuteInstruction(MI, NewMI);
+  MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
+
+  if (MI)
+    MI->setDesc(get(commuteOpcode(MI->getOpcode())));
+
+  return MI;
 }
 
 MachineInstr * SIInstrInfo::getMovImmInstr(MachineFunction *MF, unsigned DstReg,