1 files changed, 47 insertions, 19 deletions

diff --git a/lib/Target/R600/R600ControlFlowFinalizer.cpp b/lib/Target/R600/R600ControlFlowFinalizer.cpp
index ab29d60..715be37 100644
--- a/lib/Target/R600/R600ControlFlowFinalizer.cpp
+++ b/lib/Target/R600/R600ControlFlowFinalizer.cpp
@@ -172,7 +172,7 @@ private:
   }
 
   void getLiteral(MachineInstr *MI, std::vector<int64_t> &Lits) const {
-    unsigned LiteralRegs[] = {
+    static const unsigned LiteralRegs[] = {
       AMDGPU::ALU_LITERAL_X,
       AMDGPU::ALU_LITERAL_Y,
       AMDGPU::ALU_LITERAL_Z,
@@ -256,6 +256,7 @@ private:
         ClauseContent.push_back(MILit);
       }
     }
+    assert(ClauseContent.size() < 128 && "ALU clause is too big");
     ClauseHead->getOperand(7).setImm(ClauseContent.size() - 1);
     return ClauseFile(ClauseHead, ClauseContent);
   }
@@ -276,6 +277,7 @@ private:
   void
   EmitALUClause(MachineBasicBlock::iterator InsertPos, ClauseFile &Clause,
       unsigned &CfCount) {
+    Clause.first->getOperand(0).setImm(0);
     CounterPropagateAddr(Clause.first, CfCount);
     MachineBasicBlock *BB = Clause.first->getParent();
     BuildMI(BB, InsertPos->getDebugLoc(), TII->get(AMDGPU::ALU_CLAUSE))
@@ -345,6 +347,9 @@ public:
         MaxStack = 1;
       }
       std::vector<ClauseFile> FetchClauses, AluClauses;
+      std::vector<MachineInstr *> LastAlu(1);
+      std::vector<MachineInstr *> ToPopAfter;
+      
       for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
           I != E;) {
         if (TII->usesTextureCache(I) || TII->usesVertexCache(I)) {
@@ -355,6 +360,10 @@ public:
         }
 
         MachineBasicBlock::iterator MI = I;
+        if (MI->getOpcode() != AMDGPU::ENDIF)
+          LastAlu.back() = 0;
+        if (MI->getOpcode() == AMDGPU::CF_ALU)
+          LastAlu.back() = MI;
         I++;
         switch (MI->getOpcode()) {
         case AMDGPU::CF_ALU_PUSH_BEFORE:
@@ -369,7 +378,10 @@ public:
         case AMDGPU::R600_ExportBuf:
         case AMDGPU::R600_ExportSwz:
         case AMDGPU::RAT_WRITE_CACHELESS_32_eg:
+        case AMDGPU::RAT_WRITE_CACHELESS_64_eg:
         case AMDGPU::RAT_WRITE_CACHELESS_128_eg:
+        case AMDGPU::RAT_STORE_DWORD32_cm:
+        case AMDGPU::RAT_STORE_DWORD64_cm:
           DEBUG(dbgs() << CfCount << ":"; MI->dump(););
           CfCount++;
           break;
@@ -400,6 +412,7 @@ public:
           break;
         }
         case AMDGPU::IF_PREDICATE_SET: {
+          LastAlu.push_back(0);
           MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
               getHWInstrDesc(CF_JUMP))
               .addImm(0)
@@ -417,7 +430,7 @@ public:
           MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
               getHWInstrDesc(CF_ELSE))
               .addImm(0)
-              .addImm(1);
+              .addImm(0);
           DEBUG(dbgs() << CfCount << ":"; MIb->dump(););
           IfThenElseStack.push_back(MIb);
           MI->eraseFromParent();
@@ -426,31 +439,31 @@ public:
         }
         case AMDGPU::ENDIF: {
           CurrentStack--;
+          if (LastAlu.back()) {
+            ToPopAfter.push_back(LastAlu.back());
+          } else {
+            MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
+                getHWInstrDesc(CF_POP))
+                .addImm(CfCount + 1)
+                .addImm(1);
+            (void)MIb;
+            DEBUG(dbgs() << CfCount << ":"; MIb->dump(););
+            CfCount++;
+          }
+          
           MachineInstr *IfOrElseInst = IfThenElseStack.back();
           IfThenElseStack.pop_back();
-          CounterPropagateAddr(IfOrElseInst, CfCount + 1);
-          MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
-              getHWInstrDesc(CF_POP))
-              .addImm(CfCount + 1)
-              .addImm(1);
-          (void)MIb;
-          DEBUG(dbgs() << CfCount << ":"; MIb->dump(););
+          CounterPropagateAddr(IfOrElseInst, CfCount);
+          IfOrElseInst->getOperand(1).setImm(1);
+          LastAlu.pop_back();
           MI->eraseFromParent();
-          CfCount++;
           break;
         }
-        case AMDGPU::PREDICATED_BREAK: {
-          CurrentStack--;
-          CfCount += 3;
-          BuildMI(MBB, MI, MBB.findDebugLoc(MI), getHWInstrDesc(CF_JUMP))
-              .addImm(CfCount)
-              .addImm(1);
+        case AMDGPU::BREAK: {
+          CfCount ++;
           MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
               getHWInstrDesc(CF_LOOP_BREAK))
               .addImm(0);
-          BuildMI(MBB, MI, MBB.findDebugLoc(MI), getHWInstrDesc(CF_POP))
-              .addImm(CfCount)
-              .addImm(1);
           LoopStack.back().second.insert(MIb);
           MI->eraseFromParent();
           break;
@@ -481,6 +494,21 @@ public:
           break;
         }
       }
+      for (unsigned i = 0, e = ToPopAfter.size(); i < e; ++i) {
+        MachineInstr *Alu = ToPopAfter[i];
+        BuildMI(MBB, Alu, MBB.findDebugLoc((MachineBasicBlock::iterator)Alu),
+            TII->get(AMDGPU::CF_ALU_POP_AFTER))
+            .addImm(Alu->getOperand(0).getImm())
+            .addImm(Alu->getOperand(1).getImm())
+            .addImm(Alu->getOperand(2).getImm())
+            .addImm(Alu->getOperand(3).getImm())
+            .addImm(Alu->getOperand(4).getImm())
+            .addImm(Alu->getOperand(5).getImm())
+            .addImm(Alu->getOperand(6).getImm())
+            .addImm(Alu->getOperand(7).getImm())
+            .addImm(Alu->getOperand(8).getImm());
+        Alu->eraseFromParent();
+      }
       MFI->StackSize = getHWStackSize(MaxStack, HasPush);
     }