#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
-#define GET_INSTRINFO_CTOR
-#include "AMDGPUGenDFAPacketizer.inc"
-
using namespace llvm;
-R600InstrInfo::R600InstrInfo(AMDGPUTargetMachine &tm)
- : AMDGPUInstrInfo(tm),
- RI(tm),
- ST(tm.getSubtarget<AMDGPUSubtarget>())
+#define GET_INSTRINFO_CTOR_DTOR
+#include "AMDGPUGenDFAPacketizer.inc"
+
+R600InstrInfo::R600InstrInfo(const AMDGPUSubtarget &st)
+ : AMDGPUInstrInfo(st),
+ RI(st)
{ }
const R600RegisterInfo &R600InstrInfo::getRegisterInfo() const {
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;
}
}
}
-unsigned R600InstrInfo::getIEQOpcode() const {
- return AMDGPU::SETE_INT;
+/// \returns true if \p MBBI can be moved into a new basic.
+bool R600InstrInfo::isLegalToSplitMBBAt(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator MBBI) const {
+ for (MachineInstr::const_mop_iterator I = MBBI->operands_begin(),
+ E = MBBI->operands_end(); I != E; ++I) {
+ if (I->isReg() && !TargetRegisterInfo::isVirtualRegister(I->getReg()) &&
+ I->isUse() && RI.isPhysRegLiveAcrossClauses(I->getReg()))
+ return false;
+ }
+ return true;
}
bool R600InstrInfo::isMov(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 {
}
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 {
Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
continue;
}
-
+
}
return Result;
}
if (!isLastAluTrans)
return FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps, TransBS);
- TransOps = IGSrcs.back();
+ TransOps = std::move(IGSrcs.back());
IGSrcs.pop_back();
ValidSwizzle.pop_back();
return fitsConstReadLimitations(Consts);
}
-DFAPacketizer *R600InstrInfo::CreateTargetScheduleState(const TargetMachine *TM,
- const ScheduleDAG *DAG) const {
- const InstrItineraryData *II = TM->getInstrItineraryData();
- return TM->getSubtarget<AMDGPUSubtarget>().createDFAPacketizer(II);
+DFAPacketizer *
+R600InstrInfo::CreateTargetScheduleState(const TargetSubtargetInfo &STI) const {
+ const InstrItineraryData *II = STI.getInstrItineraryData();
+ return static_cast<const AMDGPUSubtarget &>(STI).createDFAPacketizer(II);
}
static bool
return MI;
}
- return NULL;
+ return nullptr;
}
static
return false;
}
- // Get the last instruction in the block.
+ // Remove successive JUMP
+ while (I != MBB.begin() && std::prev(I)->getOpcode() == AMDGPU::JUMP) {
+ MachineBasicBlock::iterator PriorI = std::prev(I);
+ if (AllowModify)
+ I->removeFromParent();
+ I = PriorI;
+ }
MachineInstr *LastInst = I;
// If there is only one terminator instruction, process it.
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();
It != E; ++It) {
if (It->getOpcode() == AMDGPU::CF_ALU ||
It->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE)
- return llvm::prior(It.base());
+ return std::prev(It.base());
}
return MBB.end();
}
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;
return true;
}
+ if (MI->getOpcode() == AMDGPU::DOT_4) {
+ MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_X))
+ .setReg(Pred[2].getReg());
+ MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_Y))
+ .setReg(Pred[2].getReg());
+ MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_Z))
+ .setReg(Pred[2].getReg());
+ MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_W))
+ .setReg(Pred[2].getReg());
+ MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
+ MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
+ return true;
+ }
+
if (PIdx != -1) {
MachineOperand &PMO = MI->getOperand(PIdx);
PMO.setReg(Pred[2].getReg());
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);
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);
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);
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
AMDGPU::OpName::src1_sel,
};
+ MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
+ getSlotedOps(AMDGPU::OpName::pred_sel, Slot)));
+ MIB->getOperand(getOperandIdx(Opcode, AMDGPU::OpName::pred_sel))
+ .setReg(MO.getReg());
+
for (unsigned i = 0; i < 14; i++) {
MachineOperand &MO = MI->getOperand(
getOperandIdx(MI->getOpcode(), getSlotedOps(Operands[i], Slot)));