R600: Fold remaining CONST_COPY after expand pseudo inst

Patch by: Vincent Lejeune

Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 174395

diff --git a/llvm/lib/Target/R600/AMDGPUTargetMachine.cpp b/llvm/lib/Target/R600/AMDGPUTargetMachine.cpp
index 7b069e7..2185be3 100644 (file)
--- a/llvm/lib/Target/R600/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/R600/AMDGPUTargetMachine.cpp
@@ -136,8 +136,8 @@ bool AMDGPUPassConfig::addPreEmitPass() {
     addPass(createAMDGPUCFGPreparationPass(*TM));
     addPass(createAMDGPUCFGStructurizerPass(*TM));
     addPass(createR600ExpandSpecialInstrsPass(*TM));
-    addPass(createR600LowerConstCopy(*TM));
     addPass(&FinalizeMachineBundlesID);
+    addPass(createR600LowerConstCopy(*TM));
   } else {
     addPass(createSILowerLiteralConstantsPass(*TM));
     addPass(createSILowerControlFlowPass(*TM));

diff --git a/llvm/lib/Target/R600/R600LowerConstCopy.cpp b/llvm/lib/Target/R600/R600LowerConstCopy.cpp
index 70a2b13..46f2aef 100644 (file)
--- a/llvm/lib/Target/R600/R600LowerConstCopy.cpp
+++ b/llvm/lib/Target/R600/R600LowerConstCopy.cpp
@@ -13,7 +13,6 @@
 /// fold them inside vector instruction, like DOT4 or Cube ; ISel emits
 /// ConstCopy instead. This pass (executed after ExpandingSpecialInstr) will try
 /// to fold them if possible or replace them by MOV otherwise.
-/// TODO : Implement the folding part, using Copy Propagation algorithm.
 //
 //===----------------------------------------------------------------------===//
 
@@ -30,6 +29,13 @@ class R600LowerConstCopy : public MachineFunctionPass {
 private:
   static char ID;
   const R600InstrInfo *TII;
+
+  struct ConstPairs {
+    unsigned XYPair;
+    unsigned ZWPair;
+  };
+
+  bool canFoldInBundle(ConstPairs &UsedConst, unsigned ReadConst) const;
 public:
   R600LowerConstCopy(TargetMachine &tm);
   virtual bool runOnMachineFunction(MachineFunction &MF);
@@ -39,27 +45,169 @@ public:
 
 char R600LowerConstCopy::ID = 0;
 
-
 R600LowerConstCopy::R600LowerConstCopy(TargetMachine &tm) :
     MachineFunctionPass(ID),
     TII (static_cast<const R600InstrInfo *>(tm.getInstrInfo()))
 {
 }
 
+bool R600LowerConstCopy::canFoldInBundle(ConstPairs &UsedConst,
+    unsigned ReadConst) const {
+  unsigned ReadConstChan = ReadConst & 3;
+  unsigned ReadConstIndex = ReadConst & (~3);
+  if (ReadConstChan < 2) {
+    if (!UsedConst.XYPair) {
+      UsedConst.XYPair = ReadConstIndex;
+    }
+    return UsedConst.XYPair == ReadConstIndex;
+  } else {
+    if (!UsedConst.ZWPair) {
+      UsedConst.ZWPair = ReadConstIndex;
+    }
+    return UsedConst.ZWPair == ReadConstIndex;
+  }
+}
+
+static bool isControlFlow(const MachineInstr &MI) {
+  return (MI.getOpcode() == AMDGPU::IF_PREDICATE_SET) ||
+  (MI.getOpcode() == AMDGPU::ENDIF) ||
+  (MI.getOpcode() == AMDGPU::ELSE) ||
+  (MI.getOpcode() == AMDGPU::WHILELOOP) ||
+  (MI.getOpcode() == AMDGPU::BREAK);
+}
+
 bool R600LowerConstCopy::runOnMachineFunction(MachineFunction &MF) {
+
   for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
                                                   BB != BB_E; ++BB) {
     MachineBasicBlock &MBB = *BB;
-    for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
-                                                      I != E;) {
-      MachineInstr &MI = *I;
-      I = llvm::next(I);
-      if (MI.getOpcode() != AMDGPU::CONST_COPY)
+    DenseMap<unsigned, MachineInstr *> RegToConstIndex;
+    for (MachineBasicBlock::instr_iterator I = MBB.instr_begin(),
+        E = MBB.instr_end(); I != E;) {
+
+      if (I->getOpcode() == AMDGPU::CONST_COPY) {
+        MachineInstr &MI = *I;
+        I = llvm::next(I);
+        unsigned DstReg = MI.getOperand(0).getReg();
+        DenseMap<unsigned, MachineInstr *>::iterator SrcMI =
+            RegToConstIndex.find(DstReg);
+        if (SrcMI != RegToConstIndex.end()) {
+          SrcMI->second->eraseFromParent();
+          RegToConstIndex.erase(SrcMI);
+        }
+        MachineInstr *NewMI = 
+            TII->buildDefaultInstruction(MBB, &MI, AMDGPU::MOV,
+            MI.getOperand(0).getReg(), AMDGPU::ALU_CONST);
+        TII->setImmOperand(NewMI, R600Operands::SRC0_SEL,
+            MI.getOperand(1).getImm());
+        RegToConstIndex[DstReg] = NewMI;
+        MI.eraseFromParent();
         continue;
-      MachineInstr *NewMI = TII->buildDefaultInstruction(MBB, I, AMDGPU::MOV,
-          MI.getOperand(0).getReg(), AMDGPU::ALU_CONST);
-      NewMI->getOperand(9).setImm(MI.getOperand(1).getImm());
-      MI.eraseFromParent();
+      }
+
+      std::vector<unsigned> Defs;
+      // We consider all Instructions as bundled because algorithm that  handle
+      // const read port limitations inside an IG is still valid with single
+      // instructions.
+      std::vector<MachineInstr *> Bundle;
+
+      if (I->isBundle()) {
+        unsigned BundleSize = I->getBundleSize();
+        for (unsigned i = 0; i < BundleSize; i++) {
+          I = llvm::next(I);
+          Bundle.push_back(I);
+        }
+      } else if (TII->isALUInstr(I->getOpcode())){
+        Bundle.push_back(I);
+      } else if (isControlFlow(*I)) {
+          RegToConstIndex.clear();
+          I = llvm::next(I);
+          continue;
+      } else {
+        MachineInstr &MI = *I;
+        for (MachineInstr::mop_iterator MOp = MI.operands_begin(),
+            MOpE = MI.operands_end(); MOp != MOpE; ++MOp) {
+          MachineOperand &MO = *MOp;
+          if (!MO.isReg())
+            continue;
+          if (MO.isDef()) {
+            Defs.push_back(MO.getReg());
+          } else {
+            // Either a TEX or an Export inst, prevent from erasing def of used
+            // operand
+            RegToConstIndex.erase(MO.getReg());
+            for (MCSubRegIterator SR(MO.getReg(), &TII->getRegisterInfo());
+                SR.isValid(); ++SR) {
+              RegToConstIndex.erase(*SR);
+            }
+          }
+        }
+      }
+
+
+      R600Operands::Ops OpTable[3][2] = {
+        {R600Operands::SRC0, R600Operands::SRC0_SEL},
+        {R600Operands::SRC1, R600Operands::SRC1_SEL},
+        {R600Operands::SRC2, R600Operands::SRC2_SEL},
+      };
+
+      for(std::vector<MachineInstr *>::iterator It = Bundle.begin(),
+          ItE = Bundle.end(); It != ItE; ++It) {
+        MachineInstr *MI = *It;
+        if (TII->isPredicated(MI)) {
+          // We don't want to erase previous assignment
+          RegToConstIndex.erase(MI->getOperand(0).getReg());
+        } else {
+          int WriteIDX = TII->getOperandIdx(MI->getOpcode(), R600Operands::WRITE);
+          if (WriteIDX < 0 || MI->getOperand(WriteIDX).getImm())
+            Defs.push_back(MI->getOperand(0).getReg());
+        }
+      }
+
+      ConstPairs CP = {0,0};
+      for (unsigned SrcOp = 0; SrcOp < 3; SrcOp++) {
+        for(std::vector<MachineInstr *>::iterator It = Bundle.begin(),
+            ItE = Bundle.end(); It != ItE; ++It) {
+          MachineInstr *MI = *It;
+          int SrcIdx = TII->getOperandIdx(MI->getOpcode(), OpTable[SrcOp][0]);
+          if (SrcIdx < 0)
+            continue;
+          MachineOperand &MO = MI->getOperand(SrcIdx);
+          DenseMap<unsigned, MachineInstr *>::iterator SrcMI =
+              RegToConstIndex.find(MO.getReg());
+          if (SrcMI != RegToConstIndex.end()) {
+            MachineInstr *CstMov = SrcMI->second;
+            int ConstMovSel =
+                TII->getOperandIdx(CstMov->getOpcode(), R600Operands::SRC0_SEL);
+            unsigned ConstIndex = CstMov->getOperand(ConstMovSel).getImm();
+            if (canFoldInBundle(CP, ConstIndex)) {
+              TII->setImmOperand(MI, OpTable[SrcOp][1], ConstIndex);
+              MI->getOperand(SrcIdx).setReg(AMDGPU::ALU_CONST);
+            } else {
+              RegToConstIndex.erase(SrcMI);
+            }
+          }
+        }
+      }
+
+      for (std::vector<unsigned>::iterator It = Defs.begin(), ItE = Defs.end();
+          It != ItE; ++It) {
+        DenseMap<unsigned, MachineInstr *>::iterator SrcMI =
+            RegToConstIndex.find(*It);
+        if (SrcMI != RegToConstIndex.end()) {
+          SrcMI->second->eraseFromParent();
+          RegToConstIndex.erase(SrcMI);
+        }
+      }
+      I = llvm::next(I);
+    }
+
+    if (MBB.succ_empty()) {
+      for (DenseMap<unsigned, MachineInstr *>::iterator
+          DI = RegToConstIndex.begin(), DE = RegToConstIndex.end();
+          DI != DE; ++DI) {
+        DI->second->eraseFromParent();
+      }
     }
   }
   return false;