private:
const R600InstrInfo *TII;
const R600RegisterInfo &TRI;
+ bool VLIW5;
+ bool ConsideredInstUsesAlreadyWrittenVectorElement;
unsigned getSlot(const MachineInstr *MI) const {
return TRI.getHWRegChan(MI->getOperand(0).getReg());
}
/// \returns register to PV chan mapping for bundle/single instructions that
- /// immediatly precedes I.
+ /// immediately precedes I.
DenseMap<unsigned, unsigned> getPreviousVector(MachineBasicBlock::iterator I)
const {
DenseMap<unsigned, unsigned> Result;
MachineBasicBlock::instr_iterator BI = I.getInstrIterator();
if (I->isBundle())
BI++;
+ int LastDstChan = -1;
do {
+ bool isTrans = false;
+ int BISlot = getSlot(BI);
+ if (LastDstChan >= BISlot)
+ isTrans = true;
+ LastDstChan = BISlot;
if (TII->isPredicated(BI))
continue;
- if (TII->isTransOnly(BI))
- continue;
int OperandIdx = TII->getOperandIdx(BI->getOpcode(), AMDGPU::OpName::write);
if (OperandIdx > -1 && BI->getOperand(OperandIdx).getImm() == 0)
continue;
continue;
}
unsigned Dst = BI->getOperand(DstIdx).getReg();
+ if (isTrans || TII->isTransOnly(BI)) {
+ Result[Dst] = AMDGPU::PS;
+ continue;
+ }
if (BI->getOpcode() == AMDGPU::DOT4_r600 ||
BI->getOpcode() == AMDGPU::DOT4_eg) {
Result[Dst] = AMDGPU::PV_X;
continue;
}
+ if (Dst == AMDGPU::OQAP) {
+ continue;
+ }
unsigned PVReg = 0;
switch (TRI.getHWRegChan(Dst)) {
case 0:
MachineDominatorTree &MDT)
: VLIWPacketizerList(MF, MLI, MDT, true),
TII (static_cast<const R600InstrInfo *>(MF.getTarget().getInstrInfo())),
- TRI(TII->getRegisterInfo()) { }
+ TRI(TII->getRegisterInfo()) {
+ VLIW5 = !MF.getTarget().getSubtarget<AMDGPUSubtarget>().hasCaymanISA();
+ }
// initPacketizerState - initialize some internal flags.
- void initPacketizerState() { }
+ void initPacketizerState() {
+ ConsideredInstUsesAlreadyWrittenVectorElement = false;
+ }
// ignorePseudoInstruction - Ignore bundling of pseudo instructions.
bool ignorePseudoInstruction(MachineInstr *MI, MachineBasicBlock *MBB) {
return true;
if (!TII->isALUInstr(MI->getOpcode()))
return true;
- if (TII->get(MI->getOpcode()).TSFlags & R600_InstFlag::TRANS_ONLY)
- return true;
- if (TII->isTransOnly(MI))
- return true;
if (MI->getOpcode() == AMDGPU::GROUP_BARRIER)
return true;
+ // XXX: This can be removed once the packetizer properly handles all the
+ // LDS instruction group restrictions.
+ if (TII->isLDSInstr(MI->getOpcode()))
+ return true;
return false;
}
// together.
bool isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) {
MachineInstr *MII = SUI->getInstr(), *MIJ = SUJ->getInstr();
- if (getSlot(MII) <= getSlot(MIJ))
- return false;
+ if (getSlot(MII) == getSlot(MIJ))
+ ConsideredInstUsesAlreadyWrittenVectorElement = true;
// Does MII and MIJ share the same pred_sel ?
int OpI = TII->getOperandIdx(MII->getOpcode(), AMDGPU::OpName::pred_sel),
OpJ = TII->getOperandIdx(MIJ->getOpcode(), AMDGPU::OpName::pred_sel);
return false;
}
}
+
+ bool ARDef = TII->definesAddressRegister(MII) ||
+ TII->definesAddressRegister(MIJ);
+ bool ARUse = TII->usesAddressRegister(MII) ||
+ TII->usesAddressRegister(MIJ);
+ if (ARDef && ARUse)
+ return false;
+
return true;
}
MI->getOperand(LastOp).setImm(Bit);
}
- MachineBasicBlock::iterator addToPacket(MachineInstr *MI) {
+ bool isBundlableWithCurrentPMI(MachineInstr *MI,
+ const DenseMap<unsigned, unsigned> &PV,
+ std::vector<R600InstrInfo::BankSwizzle> &BS,
+ bool &isTransSlot) {
+ isTransSlot = TII->isTransOnly(MI);
+ assert (!isTransSlot || VLIW5);
+
+ // Is the dst reg sequence legal ?
+ if (!isTransSlot && !CurrentPacketMIs.empty()) {
+ if (getSlot(MI) <= getSlot(CurrentPacketMIs.back())) {
+ if (ConsideredInstUsesAlreadyWrittenVectorElement &&
+ !TII->isVectorOnly(MI) && VLIW5) {
+ isTransSlot = true;
+ DEBUG(dbgs() << "Considering as Trans Inst :"; MI->dump(););
+ }
+ else
+ return false;
+ }
+ }
+
+ // Are the Constants limitations met ?
CurrentPacketMIs.push_back(MI);
- bool FitsConstLimits = TII->canBundle(CurrentPacketMIs);
- DEBUG(
- if (!FitsConstLimits) {
+ if (!TII->fitsConstReadLimitations(CurrentPacketMIs)) {
+ DEBUG(
dbgs() << "Couldn't pack :\n";
MI->dump();
dbgs() << "with the following packets :\n";
dbgs() << "\n";
}
dbgs() << "because of Consts read limitations\n";
- });
- const DenseMap<unsigned, unsigned> &PV =
- getPreviousVector(CurrentPacketMIs.front());
- std::vector<R600InstrInfo::BankSwizzle> BS;
- bool FitsReadPortLimits =
- TII->fitsReadPortLimitations(CurrentPacketMIs, PV, BS);
- DEBUG(
- if (!FitsReadPortLimits) {
+ );
+ CurrentPacketMIs.pop_back();
+ return false;
+ }
+
+ // Is there a BankSwizzle set that meet Read Port limitations ?
+ if (!TII->fitsReadPortLimitations(CurrentPacketMIs,
+ PV, BS, isTransSlot)) {
+ DEBUG(
dbgs() << "Couldn't pack :\n";
MI->dump();
dbgs() << "with the following packets :\n";
dbgs() << "\n";
}
dbgs() << "because of Read port limitations\n";
- });
- bool isBundlable = FitsConstLimits && FitsReadPortLimits;
- if (isBundlable) {
+ );
+ CurrentPacketMIs.pop_back();
+ return false;
+ }
+
+ // We cannot read LDS source registrs from the Trans slot.
+ if (isTransSlot && TII->readsLDSSrcReg(MI))
+ return false;
+
+ CurrentPacketMIs.pop_back();
+ return true;
+ }
+
+ MachineBasicBlock::iterator addToPacket(MachineInstr *MI) {
+ MachineBasicBlock::iterator FirstInBundle =
+ CurrentPacketMIs.empty() ? MI : CurrentPacketMIs.front();
+ const DenseMap<unsigned, unsigned> &PV =
+ getPreviousVector(FirstInBundle);
+ std::vector<R600InstrInfo::BankSwizzle> BS;
+ bool isTransSlot;
+
+ if (isBundlableWithCurrentPMI(MI, PV, BS, isTransSlot)) {
for (unsigned i = 0, e = CurrentPacketMIs.size(); i < e; i++) {
MachineInstr *MI = CurrentPacketMIs[i];
- unsigned Op = TII->getOperandIdx(MI->getOpcode(),
- AMDGPU::OpName::bank_swizzle);
- MI->getOperand(Op).setImm(BS[i]);
+ unsigned Op = TII->getOperandIdx(MI->getOpcode(),
+ AMDGPU::OpName::bank_swizzle);
+ MI->getOperand(Op).setImm(BS[i]);
}
+ unsigned Op = TII->getOperandIdx(MI->getOpcode(),
+ AMDGPU::OpName::bank_swizzle);
+ MI->getOperand(Op).setImm(BS.back());
+ if (!CurrentPacketMIs.empty())
+ setIsLastBit(CurrentPacketMIs.back(), 0);
+ substitutePV(MI, PV);
+ MachineBasicBlock::iterator It = VLIWPacketizerList::addToPacket(MI);
+ if (isTransSlot) {
+ endPacket(llvm::next(It)->getParent(), llvm::next(It));
+ }
+ return It;
}
- CurrentPacketMIs.pop_back();
- if (!isBundlable) {
- endPacket(MI->getParent(), MI);
- substitutePV(MI, getPreviousVector(MI));
- return VLIWPacketizerList::addToPacket(MI);
- }
- if (!CurrentPacketMIs.empty())
- setIsLastBit(CurrentPacketMIs.back(), 0);
- substitutePV(MI, PV);
+ endPacket(MI->getParent(), MI);
+ if (TII->isTransOnly(MI))
+ return MI;
return VLIWPacketizerList::addToPacket(MI);
}
};
MachineBasicBlock::iterator End = MBB->end();
MachineBasicBlock::iterator MI = MBB->begin();
while (MI != End) {
- if (MI->isKill()) {
+ if (MI->isKill() || MI->getOpcode() == AMDGPU::IMPLICIT_DEF ||
+ (MI->getOpcode() == AMDGPU::CF_ALU && !MI->getOperand(8).getImm())) {
MachineBasicBlock::iterator DeleteMI = MI;
++MI;
MBB->erase(DeleteMI);