R600/SI: Allow SI_KILL for geometry shaders

[oota-llvm.git] / lib / Target / R600 / R600OptimizeVectorRegisters.cpp

diff --git a/lib/Target/R600/R600OptimizeVectorRegisters.cpp b/lib/Target/R600/R600OptimizeVectorRegisters.cpp
index 726cfbcdf3ee2fedad5f247976870f4e9c2df25e..49a90658a77db93c695c6356bdb86dc148869986 100644 (file)
--- a/lib/Target/R600/R600OptimizeVectorRegisters.cpp
+++ b/lib/Target/R600/R600OptimizeVectorRegisters.cpp
@@ -34,11 +34,11 @@
 #include "llvm/CodeGen/DFAPacketizer.h"
 #include "llvm/CodeGen/MachineDominators.h"
 #include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineLoopInfo.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/Passes.h"
-#include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/Support/raw_ostream.h"
-#include "llvm/CodeGen/MachineRegisterInfo.h"
 
 using namespace llvm;
 
@@ -50,6 +50,9 @@ isImplicitlyDef(MachineRegisterInfo &MRI, unsigned Reg) {
       E = MRI.def_end(); It != E; ++It) {
     return (*It).isImplicitDef();
   }
+  if (MRI.isReserved(Reg)) {
+    return false;
+  }
   llvm_unreachable("Reg without a def");
   return false;
 }
@@ -60,7 +63,7 @@ public:
   DenseMap<unsigned, unsigned> RegToChan;
   std::vector<unsigned> UndefReg;
   RegSeqInfo(MachineRegisterInfo &MRI, MachineInstr *MI) : Instr(MI) {
-    assert (MI->getOpcode() == AMDGPU::REG_SEQUENCE);
+    assert(MI->getOpcode() == AMDGPU::REG_SEQUENCE);
     for (unsigned i = 1, e = Instr->getNumOperands(); i < e; i+=2) {
       MachineOperand &MO = Instr->getOperand(i);
       unsigned Chan = Instr->getOperand(i + 1).getImm();
@@ -104,7 +107,7 @@ private:
 public:
   static char ID;
   R600VectorRegMerger(TargetMachine &tm) : MachineFunctionPass(ID),
-  TII (static_cast<const R600InstrInfo *>(tm.getInstrInfo())) { }
+  TII(0) { }
 
   void getAnalysisUsage(AnalysisUsage &AU) const {
     AU.setPreservesCFG();
@@ -183,10 +186,6 @@ MachineInstr *R600VectorRegMerger::RebuildVector(
   std::vector<unsigned> UpdatedUndef = BaseRSI->UndefReg;
   for (DenseMap<unsigned, unsigned>::iterator It = RSI->RegToChan.begin(),
       E = RSI->RegToChan.end(); It != E; ++It) {
-    if (BaseRSI->RegToChan.find((*It).first) != BaseRSI->RegToChan.end()) {
-      UpdatedRegToChan[(*It).first] = (*It).second;
-      continue;
-    }
     unsigned DstReg = MRI->createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
     unsigned SubReg = (*It).first;
     unsigned Swizzle = (*It).second;
@@ -314,6 +313,7 @@ void R600VectorRegMerger::trackRSI(const RegSeqInfo &RSI) {
 }
 
 bool R600VectorRegMerger::runOnMachineFunction(MachineFunction &Fn) {
+  TII = static_cast<const R600InstrInfo *>(Fn.getTarget().getInstrInfo());
   MRI = &(Fn.getRegInfo());
   for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
        MBB != MBBe; ++MBB) {
@@ -325,8 +325,17 @@ bool R600VectorRegMerger::runOnMachineFunction(MachineFunction &Fn) {
     for (MachineBasicBlock::iterator MII = MB->begin(), MIIE = MB->end();
          MII != MIIE; ++MII) {
       MachineInstr *MI = MII;
-      if (MI->getOpcode() != AMDGPU::REG_SEQUENCE)
+      if (MI->getOpcode() != AMDGPU::REG_SEQUENCE) {
+        if (TII->get(MI->getOpcode()).TSFlags & R600_InstFlag::TEX_INST) {
+          unsigned Reg = MI->getOperand(1).getReg();
+          for (MachineRegisterInfo::def_iterator It = MRI->def_begin(Reg),
+              E = MRI->def_end(); It != E; ++It) {
+            RemoveMI(&(*It));
+          }
+        }
         continue;
+      }
+
 
       RegSeqInfo RSI(*MRI, MI);