On DataLayout, omit the default of p:64:64:64.

[oota-llvm.git] / lib / Target / R600 / R600ControlFlowFinalizer.cpp

diff --git a/lib/Target/R600/R600ControlFlowFinalizer.cpp b/lib/Target/R600/R600ControlFlowFinalizer.cpp
index 1cd0ac38a223f097123af204e6a71f975e0cf605..ec39e097ba951fab3c9fc2a6fc15f5cf04b00298 100644 (file)
--- a/lib/Target/R600/R600ControlFlowFinalizer.cpp
+++ b/lib/Target/R600/R600ControlFlowFinalizer.cpp
@@ -332,6 +332,7 @@ public:
 
     unsigned MaxStack = 0;
     unsigned CurrentStack = 0;
+    unsigned CurrentLoopDepth = 0;
     bool HasPush = false;
     for (MachineFunction::iterator MB = MF.begin(), ME = MF.end(); MB != ME;
         ++MB) {
@@ -347,6 +348,9 @@ 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)) {
@@ -357,27 +361,32 @@ 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:
           CurrentStack++;
           MaxStack = std::max(MaxStack, CurrentStack);
           HasPush = true;
+          if (ST.hasCaymanISA() && CurrentLoopDepth > 1) {
+            BuildMI(MBB, MI, MBB.findDebugLoc(MI), TII->get(AMDGPU::CF_PUSH_CM))
+                .addImm(CfCount + 1)
+                .addImm(1);
+            MI->setDesc(TII->get(AMDGPU::CF_ALU));
+            CfCount++;
+          }
         case AMDGPU::CF_ALU:
           I = MI;
           AluClauses.push_back(MakeALUClause(MBB, I));
-        case AMDGPU::EG_ExportBuf:
-        case AMDGPU::EG_ExportSwz:
-        case AMDGPU::R600_ExportBuf:
-        case AMDGPU::R600_ExportSwz:
-        case AMDGPU::RAT_WRITE_CACHELESS_32_eg:
-        case AMDGPU::RAT_WRITE_CACHELESS_128_eg:
-        case AMDGPU::RAT_STORE_DWORD_cm:
           DEBUG(dbgs() << CfCount << ":"; MI->dump(););
           CfCount++;
           break;
         case AMDGPU::WHILELOOP: {
           CurrentStack+=4;
+          CurrentLoopDepth++;
           MaxStack = std::max(MaxStack, CurrentStack);
           MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
               getHWInstrDesc(CF_WHILE_LOOP))
@@ -392,6 +401,7 @@ public:
         }
         case AMDGPU::ENDLOOP: {
           CurrentStack-=4;
+          CurrentLoopDepth--;
           std::pair<unsigned, std::set<MachineInstr *> > Pair =
               LoopStack.back();
           LoopStack.pop_back();
@@ -403,6 +413,7 @@ public:
           break;
         }
         case AMDGPU::IF_PREDICATE_SET: {
+          LastAlu.push_back(0);
           MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
               getHWInstrDesc(CF_JUMP))
               .addImm(0)
@@ -420,7 +431,7 @@ 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();
@@ -429,31 +440,31 @@ 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;
@@ -481,9 +492,28 @@ public:
             EmitALUClause(I, AluClauses[i], CfCount);
         }
         default:
+          if (TII->isExport(MI->getOpcode())) {
+            DEBUG(dbgs() << CfCount << ":"; MI->dump(););
+            CfCount++;
+          }
           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);
     }