projects / oota-llvm.git / blobdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | inline | side by side
Have MachineFunction cache a pointer to the subtarget to make lookups
[oota-llvm.git] / lib / Target / R600 / R600InstrInfo.cpp
diff --git a/lib/Target/R600/R600InstrInfo.cpp b/lib/Target/R600/R600InstrInfo.cpp
index a48afc9eaa10bfb757a2439857b7404147a7e5aa..ff1cedbb288975d067cae95dd5c3ceb4bfe40cf6 100644 (file)
--- a/lib/Target/R600/R600InstrInfo.cpp
+++ b/lib/Target/R600/R600InstrInfo.cpp
@@ -28,10 +28,9 @@ using namespace llvm;
 #define GET_INSTRINFO_CTOR_DTOR
 #include "AMDGPUGenDFAPacketizer.inc"
 
-R600InstrInfo::R600InstrInfo(AMDGPUTargetMachine &tm)
-  : AMDGPUInstrInfo(tm),
-    RI(tm),
-    ST(tm.getSubtarget<AMDGPUSubtarget>())
+R600InstrInfo::R600InstrInfo(const AMDGPUSubtarget &st)
+  : AMDGPUInstrInfo(st),
+    RI(st)
   { }
 
 const R600RegisterInfo &R600InstrInfo::getRegisterInfo() const {
@@ -52,11 +51,15 @@ R600InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
                            unsigned DestReg, unsigned SrcReg,
                            bool KillSrc) const {
   unsigned VectorComponents = 0;
-  if (AMDGPU::R600_Reg128RegClass.contains(DestReg) &&
-      AMDGPU::R600_Reg128RegClass.contains(SrcReg)) {
+  if ((AMDGPU::R600_Reg128RegClass.contains(DestReg) ||
+      AMDGPU::R600_Reg128VerticalRegClass.contains(DestReg)) &&
+      (AMDGPU::R600_Reg128RegClass.contains(SrcReg) ||
+       AMDGPU::R600_Reg128VerticalRegClass.contains(SrcReg))) {
     VectorComponents = 4;
-  } else if(AMDGPU::R600_Reg64RegClass.contains(DestReg) &&
-            AMDGPU::R600_Reg64RegClass.contains(SrcReg)) {
+  } else if((AMDGPU::R600_Reg64RegClass.contains(DestReg) ||
+            AMDGPU::R600_Reg64VerticalRegClass.contains(DestReg)) &&
+            (AMDGPU::R600_Reg64RegClass.contains(SrcReg) ||
+             AMDGPU::R600_Reg64VerticalRegClass.contains(SrcReg))) {
     VectorComponents = 2;
   }
 
@@ -89,10 +92,6 @@ bool R600InstrInfo::isLegalToSplitMBBAt(MachineBasicBlock &MBB,
   return true;
 }
 
-unsigned R600InstrInfo::getIEQOpcode() const {
-  return AMDGPU::SETE_INT;
-}
-
 bool R600InstrInfo::isMov(unsigned Opcode) const {
 
 
@@ -206,8 +205,10 @@ bool R600InstrInfo::usesVertexCache(unsigned Opcode) const {
 }
 
 bool R600InstrInfo::usesVertexCache(const MachineInstr *MI) const {
-  const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>();
-  return MFI->ShaderType != ShaderType::COMPUTE && usesVertexCache(MI->getOpcode());
+  const MachineFunction *MF = MI->getParent()->getParent();
+  const R600MachineFunctionInfo *MFI = MF->getInfo<R600MachineFunctionInfo>();
+  return MFI->getShaderType() != ShaderType::COMPUTE &&
+    usesVertexCache(MI->getOpcode());
 }
 
 bool R600InstrInfo::usesTextureCache(unsigned Opcode) const {
@@ -215,9 +216,11 @@ bool R600InstrInfo::usesTextureCache(unsigned Opcode) const {
 }
 
 bool R600InstrInfo::usesTextureCache(const MachineInstr *MI) const {
-  const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>();
-  return (MFI->ShaderType == ShaderType::COMPUTE && usesVertexCache(MI->getOpcode())) ||
-         usesTextureCache(MI->getOpcode());
+  const MachineFunction *MF = MI->getParent()->getParent();
+  const R600MachineFunctionInfo *MFI = MF->getInfo<R600MachineFunctionInfo>();
+  return (MFI->getShaderType() == ShaderType::COMPUTE &&
+          usesVertexCache(MI->getOpcode())) ||
+    usesTextureCache(MI->getOpcode());
 }
 
 bool R600InstrInfo::mustBeLastInClause(unsigned Opcode) const {
@@ -316,7 +319,7 @@ R600InstrInfo::getSrcs(MachineInstr *MI) const {
         Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
         continue;
       }
-      
+
     }
     return Result;
   }
@@ -653,7 +656,8 @@ R600InstrInfo::fitsConstReadLimitations(const std::vector<MachineInstr *> &MIs)
 
 DFAPacketizer *R600InstrInfo::CreateTargetScheduleState(const TargetMachine *TM,
     const ScheduleDAG *DAG) const {
-  const InstrItineraryData *II = TM->getInstrItineraryData();
+  const InstrItineraryData *II =
+      TM->getSubtargetImpl()->getInstrItineraryData();
   return TM->getSubtarget<AMDGPUSubtarget>().createDFAPacketizer(II);
 }
 
@@ -677,7 +681,7 @@ findFirstPredicateSetterFrom(MachineBasicBlock &MBB,
       return MI;
   }
 
-  return NULL;
+  return nullptr;
 }
 
 static
@@ -768,16 +772,6 @@ R600InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
   return true;
 }
 
-int R600InstrInfo::getBranchInstr(const MachineOperand &op) const {
-  const MachineInstr *MI = op.getParent();
-
-  switch (MI->getDesc().OpInfo->RegClass) {
-  default: // FIXME: fallthrough??
-  case AMDGPU::GPRI32RegClassID: return AMDGPU::BRANCH_COND_i32;
-  case AMDGPU::GPRF32RegClassID: return AMDGPU::BRANCH_COND_f32;
-  };
-}
-
 static
 MachineBasicBlock::iterator FindLastAluClause(MachineBasicBlock &MBB) {
   for (MachineBasicBlock::reverse_iterator It = MBB.rbegin(), E = MBB.rend();
@@ -797,7 +791,7 @@ R600InstrInfo::InsertBranch(MachineBasicBlock &MBB,
                             DebugLoc DL) const {
   assert(TBB && "InsertBranch must not be told to insert a fallthrough");
 
-  if (FBB == 0) {
+  if (!FBB) {
     if (Cond.empty()) {
       BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(TBB);
       return 1;
@@ -1064,10 +1058,33 @@ unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
   return 2;
 }
 
+bool R600InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
+
+  switch(MI->getOpcode()) {
+  default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
+  case AMDGPU::R600_EXTRACT_ELT_V2:
+  case AMDGPU::R600_EXTRACT_ELT_V4:
+    buildIndirectRead(MI->getParent(), MI, MI->getOperand(0).getReg(),
+                      RI.getHWRegIndex(MI->getOperand(1).getReg()), //  Address
+                      MI->getOperand(2).getReg(),
+                      RI.getHWRegChan(MI->getOperand(1).getReg()));
+    break;
+  case AMDGPU::R600_INSERT_ELT_V2:
+  case AMDGPU::R600_INSERT_ELT_V4:
+    buildIndirectWrite(MI->getParent(), MI, MI->getOperand(2).getReg(), // Value
+                       RI.getHWRegIndex(MI->getOperand(1).getReg()),  // Address
+                       MI->getOperand(3).getReg(),                    // Offset
+                       RI.getHWRegChan(MI->getOperand(1).getReg()));  // Channel
+    break;
+  }
+  MI->eraseFromParent();
+  return true;
+}
+
 void  R600InstrInfo::reserveIndirectRegisters(BitVector &Reserved,
                                              const MachineFunction &MF) const {
-  const AMDGPUFrameLowering *TFL =
-                 static_cast<const AMDGPUFrameLowering*>(TM.getFrameLowering());
+  const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering *>(
+      MF.getSubtarget().getFrameLowering());
 
   unsigned StackWidth = TFL->getStackWidth(MF);
   int End = getIndirectIndexEnd(MF);
@@ -1100,7 +1117,22 @@ MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
                                        MachineBasicBlock::iterator I,
                                        unsigned ValueReg, unsigned Address,
                                        unsigned OffsetReg) const {
-  unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
+  return buildIndirectWrite(MBB, I, ValueReg, Address, OffsetReg, 0);
+}
+
+MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
+                                       MachineBasicBlock::iterator I,
+                                       unsigned ValueReg, unsigned Address,
+                                       unsigned OffsetReg,
+                                       unsigned AddrChan) const {
+  unsigned AddrReg;
+  switch (AddrChan) {
+    default: llvm_unreachable("Invalid Channel");
+    case 0: AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address); break;
+    case 1: AddrReg = AMDGPU::R600_Addr_YRegClass.getRegister(Address); break;
+    case 2: AddrReg = AMDGPU::R600_Addr_ZRegClass.getRegister(Address); break;
+    case 3: AddrReg = AMDGPU::R600_Addr_WRegClass.getRegister(Address); break;
+  }
   MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
                                                AMDGPU::AR_X, OffsetReg);
   setImmOperand(MOVA, AMDGPU::OpName::write, 0);
@@ -1117,7 +1149,22 @@ MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
                                        MachineBasicBlock::iterator I,
                                        unsigned ValueReg, unsigned Address,
                                        unsigned OffsetReg) const {
-  unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
+  return buildIndirectRead(MBB, I, ValueReg, Address, OffsetReg, 0);
+}
+
+MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
+                                       MachineBasicBlock::iterator I,
+                                       unsigned ValueReg, unsigned Address,
+                                       unsigned OffsetReg,
+                                       unsigned AddrChan) const {
+  unsigned AddrReg;
+  switch (AddrChan) {
+    default: llvm_unreachable("Invalid Channel");
+    case 0: AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address); break;
+    case 1: AddrReg = AMDGPU::R600_Addr_YRegClass.getRegister(Address); break;
+    case 2: AddrReg = AMDGPU::R600_Addr_ZRegClass.getRegister(Address); break;
+    case 3: AddrReg = AMDGPU::R600_Addr_WRegClass.getRegister(Address); break;
+  }
   MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
                                                        AMDGPU::AR_X,
                                                        OffsetReg);
@@ -1220,7 +1267,6 @@ MachineInstr *R600InstrInfo::buildSlotOfVectorInstruction(
     const {
   assert (MI->getOpcode() == AMDGPU::DOT_4 && "Not Implemented");
   unsigned Opcode;
-  const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
   if (ST.getGeneration() <= AMDGPUSubtarget::R700)
     Opcode = AMDGPU::DOT4_r600;
   else
C/C++ LLVM-based compilers that forbids OOTA behaviors