-/// CriticalPathStep - Return the next SUnit after SU on the bottom-up
-/// critical path.
-static SDep *CriticalPathStep(SUnit *SU) {
- SDep *Next = 0;
- unsigned NextDepth = 0;
- // Find the predecessor edge with the greatest depth.
- for (SUnit::pred_iterator P = SU->Preds.begin(), PE = SU->Preds.end();
- P != PE; ++P) {
- SUnit *PredSU = P->getSUnit();
- unsigned PredLatency = P->getLatency();
- unsigned PredTotalLatency = PredSU->getDepth() + PredLatency;
- // In the case of a latency tie, prefer an anti-dependency edge over
- // other types of edges.
- if (NextDepth < PredTotalLatency ||
- (NextDepth == PredTotalLatency && P->getKind() == SDep::Anti)) {
- NextDepth = PredTotalLatency;
- Next = &*P;
- }
- }
- return Next;
-}
-
-void SchedulePostRATDList::PrescanInstruction(MachineInstr *MI) {
- // Scan the register operands for this instruction and update
- // Classes and RegRefs.
- for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
- MachineOperand &MO = MI->getOperand(i);
- if (!MO.isReg()) continue;
- unsigned Reg = MO.getReg();
- if (Reg == 0) continue;
- const TargetRegisterClass *NewRC = 0;
-
- if (i < MI->getDesc().getNumOperands())
- NewRC = MI->getDesc().OpInfo[i].getRegClass(TRI);
-
- // For now, only allow the register to be changed if its register
- // class is consistent across all uses.
- if (!Classes[Reg] && NewRC)
- Classes[Reg] = NewRC;
- else if (!NewRC || Classes[Reg] != NewRC)
- Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
-
- // Now check for aliases.
- for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
- // If an alias of the reg is used during the live range, give up.
- // Note that this allows us to skip checking if AntiDepReg
- // overlaps with any of the aliases, among other things.
- unsigned AliasReg = *Alias;
- if (Classes[AliasReg]) {
- Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
- Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
- }
- }
-
- // If we're still willing to consider this register, note the reference.
- if (Classes[Reg] != reinterpret_cast<TargetRegisterClass *>(-1))
- RegRefs.insert(std::make_pair(Reg, &MO));
-
- // It's not safe to change register allocation for source operands of
- // that have special allocation requirements.
- if (MO.isUse() && MI->getDesc().hasExtraSrcRegAllocReq()) {
- if (KeepRegs.insert(Reg)) {
- for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
- *Subreg; ++Subreg)
- KeepRegs.insert(*Subreg);
- }
- }
- }
-}
-
-void SchedulePostRATDList::ScanInstruction(MachineInstr *MI,
- unsigned Count) {
- // Update liveness.
- // Proceding upwards, registers that are defed but not used in this
- // instruction are now dead.
- for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
- MachineOperand &MO = MI->getOperand(i);
- if (!MO.isReg()) continue;
- unsigned Reg = MO.getReg();
- if (Reg == 0) continue;
- if (!MO.isDef()) continue;
- // Ignore two-addr defs.
- if (MI->isRegTiedToUseOperand(i)) continue;
-
- DefIndices[Reg] = Count;
- KillIndices[Reg] = ~0u;
- assert(((KillIndices[Reg] == ~0u) !=
- (DefIndices[Reg] == ~0u)) &&
- "Kill and Def maps aren't consistent for Reg!");
- KeepRegs.erase(Reg);
- Classes[Reg] = 0;
- RegRefs.erase(Reg);
- // Repeat, for all subregs.
- for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
- *Subreg; ++Subreg) {
- unsigned SubregReg = *Subreg;
- DefIndices[SubregReg] = Count;
- KillIndices[SubregReg] = ~0u;
- KeepRegs.erase(SubregReg);
- Classes[SubregReg] = 0;
- RegRefs.erase(SubregReg);
- }
- // Conservatively mark super-registers as unusable.
- for (const unsigned *Super = TRI->getSuperRegisters(Reg);
- *Super; ++Super) {
- unsigned SuperReg = *Super;
- Classes[SuperReg] = reinterpret_cast<TargetRegisterClass *>(-1);
- }
- }
- for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
- MachineOperand &MO = MI->getOperand(i);
- if (!MO.isReg()) continue;
- unsigned Reg = MO.getReg();
- if (Reg == 0) continue;
- if (!MO.isUse()) continue;
-
- const TargetRegisterClass *NewRC = 0;
- if (i < MI->getDesc().getNumOperands())
- NewRC = MI->getDesc().OpInfo[i].getRegClass(TRI);
-
- // For now, only allow the register to be changed if its register
- // class is consistent across all uses.
- if (!Classes[Reg] && NewRC)
- Classes[Reg] = NewRC;
- else if (!NewRC || Classes[Reg] != NewRC)
- Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
-
- RegRefs.insert(std::make_pair(Reg, &MO));
-
- // It wasn't previously live but now it is, this is a kill.
- if (KillIndices[Reg] == ~0u) {
- KillIndices[Reg] = Count;
- DefIndices[Reg] = ~0u;
- assert(((KillIndices[Reg] == ~0u) !=
- (DefIndices[Reg] == ~0u)) &&
- "Kill and Def maps aren't consistent for Reg!");
- }
- // Repeat, for all aliases.
- for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
- unsigned AliasReg = *Alias;
- if (KillIndices[AliasReg] == ~0u) {
- KillIndices[AliasReg] = Count;
- DefIndices[AliasReg] = ~0u;
- }
- }
- }
-}
-
-unsigned
-SchedulePostRATDList::findSuitableFreeRegister(unsigned AntiDepReg,
- unsigned LastNewReg,
- const TargetRegisterClass *RC) {
- for (TargetRegisterClass::iterator R = RC->allocation_order_begin(MF),
- RE = RC->allocation_order_end(MF); R != RE; ++R) {
- unsigned NewReg = *R;
- // Don't replace a register with itself.
- if (NewReg == AntiDepReg) continue;
- // Don't replace a register with one that was recently used to repair
- // an anti-dependence with this AntiDepReg, because that would
- // re-introduce that anti-dependence.
- if (NewReg == LastNewReg) continue;
- // If NewReg is dead and NewReg's most recent def is not before
- // AntiDepReg's kill, it's safe to replace AntiDepReg with NewReg.
- assert(((KillIndices[AntiDepReg] == ~0u) != (DefIndices[AntiDepReg] == ~0u)) &&
- "Kill and Def maps aren't consistent for AntiDepReg!");
- assert(((KillIndices[NewReg] == ~0u) != (DefIndices[NewReg] == ~0u)) &&
- "Kill and Def maps aren't consistent for NewReg!");
- if (KillIndices[NewReg] != ~0u ||
- Classes[NewReg] == reinterpret_cast<TargetRegisterClass *>(-1) ||
- KillIndices[AntiDepReg] > DefIndices[NewReg])
- continue;
- return NewReg;
- }
-
- // No registers are free and available!
- return 0;
-}
-
-/// BreakAntiDependencies - Identifiy anti-dependencies along the critical path
-/// of the ScheduleDAG and break them by renaming registers.
-///
-bool SchedulePostRATDList::BreakAntiDependencies() {
- // The code below assumes that there is at least one instruction,
- // so just duck out immediately if the block is empty.
- if (SUnits.empty()) return false;
-
- // Find the node at the bottom of the critical path.
- SUnit *Max = 0;
- for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
- SUnit *SU = &SUnits[i];
- if (!Max || SU->getDepth() + SU->Latency > Max->getDepth() + Max->Latency)
- Max = SU;
- }
-
- DEBUG(errs() << "Critical path has total latency "
- << (Max->getDepth() + Max->Latency) << "\n");
-
- // Track progress along the critical path through the SUnit graph as we walk
- // the instructions.
- SUnit *CriticalPathSU = Max;
- MachineInstr *CriticalPathMI = CriticalPathSU->getInstr();
-
- // Consider this pattern:
- // A = ...
- // ... = A
- // A = ...
- // ... = A
- // A = ...
- // ... = A
- // A = ...
- // ... = A
- // There are three anti-dependencies here, and without special care,
- // we'd break all of them using the same register:
- // A = ...
- // ... = A
- // B = ...
- // ... = B
- // B = ...
- // ... = B
- // B = ...
- // ... = B
- // because at each anti-dependence, B is the first register that
- // isn't A which is free. This re-introduces anti-dependencies
- // at all but one of the original anti-dependencies that we were
- // trying to break. To avoid this, keep track of the most recent
- // register that each register was replaced with, avoid
- // using it to repair an anti-dependence on the same register.
- // This lets us produce this:
- // A = ...
- // ... = A
- // B = ...
- // ... = B
- // C = ...
- // ... = C
- // B = ...
- // ... = B
- // This still has an anti-dependence on B, but at least it isn't on the
- // original critical path.
- //
- // TODO: If we tracked more than one register here, we could potentially
- // fix that remaining critical edge too. This is a little more involved,
- // because unlike the most recent register, less recent registers should
- // still be considered, though only if no other registers are available.
- unsigned LastNewReg[TargetRegisterInfo::FirstVirtualRegister] = {};
-
- // Attempt to break anti-dependence edges on the critical path. Walk the
- // instructions from the bottom up, tracking information about liveness
- // as we go to help determine which registers are available.
- bool Changed = false;
- unsigned Count = InsertPosIndex - 1;
- for (MachineBasicBlock::iterator I = InsertPos, E = Begin;
- I != E; --Count) {
- MachineInstr *MI = --I;
-
- // Check if this instruction has a dependence on the critical path that
- // is an anti-dependence that we may be able to break. If it is, set
- // AntiDepReg to the non-zero register associated with the anti-dependence.
- //
- // We limit our attention to the critical path as a heuristic to avoid
- // breaking anti-dependence edges that aren't going to significantly
- // impact the overall schedule. There are a limited number of registers
- // and we want to save them for the important edges.
- //
- // TODO: Instructions with multiple defs could have multiple
- // anti-dependencies. The current code here only knows how to break one
- // edge per instruction. Note that we'd have to be able to break all of
- // the anti-dependencies in an instruction in order to be effective.
- unsigned AntiDepReg = 0;
- if (MI == CriticalPathMI) {
- if (SDep *Edge = CriticalPathStep(CriticalPathSU)) {
- SUnit *NextSU = Edge->getSUnit();
-
- // Only consider anti-dependence edges.
- if (Edge->getKind() == SDep::Anti) {
- AntiDepReg = Edge->getReg();
- assert(AntiDepReg != 0 && "Anti-dependence on reg0?");
- if (!AllocatableSet.test(AntiDepReg))
- // Don't break anti-dependencies on non-allocatable registers.
- AntiDepReg = 0;
- else if (KeepRegs.count(AntiDepReg))
- // Don't break anti-dependencies if an use down below requires
- // this exact register.
- AntiDepReg = 0;
- else {
- // If the SUnit has other dependencies on the SUnit that it
- // anti-depends on, don't bother breaking the anti-dependency
- // since those edges would prevent such units from being
- // scheduled past each other regardless.
- //
- // Also, if there are dependencies on other SUnits with the
- // same register as the anti-dependency, don't attempt to
- // break it.
- for (SUnit::pred_iterator P = CriticalPathSU->Preds.begin(),
- PE = CriticalPathSU->Preds.end(); P != PE; ++P)
- if (P->getSUnit() == NextSU ?
- (P->getKind() != SDep::Anti || P->getReg() != AntiDepReg) :
- (P->getKind() == SDep::Data && P->getReg() == AntiDepReg)) {
- AntiDepReg = 0;
- break;
- }
- }
- }
- CriticalPathSU = NextSU;
- CriticalPathMI = CriticalPathSU->getInstr();
- } else {
- // We've reached the end of the critical path.
- CriticalPathSU = 0;
- CriticalPathMI = 0;
- }
- }
-
- PrescanInstruction(MI);
-
- if (MI->getDesc().hasExtraDefRegAllocReq())
- // If this instruction's defs have special allocation requirement, don't
- // break this anti-dependency.
- AntiDepReg = 0;
- else if (AntiDepReg) {
- // If this instruction has a use of AntiDepReg, breaking it
- // is invalid.
- for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
- MachineOperand &MO = MI->getOperand(i);
- if (!MO.isReg()) continue;
- unsigned Reg = MO.getReg();
- if (Reg == 0) continue;
- if (MO.isUse() && AntiDepReg == Reg) {
- AntiDepReg = 0;
- break;
- }
- }
- }
-
- // Determine AntiDepReg's register class, if it is live and is
- // consistently used within a single class.
- const TargetRegisterClass *RC = AntiDepReg != 0 ? Classes[AntiDepReg] : 0;
- assert((AntiDepReg == 0 || RC != NULL) &&
- "Register should be live if it's causing an anti-dependence!");
- if (RC == reinterpret_cast<TargetRegisterClass *>(-1))
- AntiDepReg = 0;
-
- // Look for a suitable register to use to break the anti-depenence.
- //
- // TODO: Instead of picking the first free register, consider which might
- // be the best.
- if (AntiDepReg != 0) {
- if (unsigned NewReg = findSuitableFreeRegister(AntiDepReg,
- LastNewReg[AntiDepReg],
- RC)) {
- DEBUG(errs() << "Breaking anti-dependence edge on "
- << TRI->getName(AntiDepReg)
- << " with " << RegRefs.count(AntiDepReg) << " references"
- << " using " << TRI->getName(NewReg) << "!\n");
-
- // Update the references to the old register to refer to the new
- // register.
- std::pair<std::multimap<unsigned, MachineOperand *>::iterator,
- std::multimap<unsigned, MachineOperand *>::iterator>
- Range = RegRefs.equal_range(AntiDepReg);
- for (std::multimap<unsigned, MachineOperand *>::iterator
- Q = Range.first, QE = Range.second; Q != QE; ++Q)
- Q->second->setReg(NewReg);
-
- // We just went back in time and modified history; the
- // liveness information for the anti-depenence reg is now
- // inconsistent. Set the state as if it were dead.
- Classes[NewReg] = Classes[AntiDepReg];
- DefIndices[NewReg] = DefIndices[AntiDepReg];
- KillIndices[NewReg] = KillIndices[AntiDepReg];
- assert(((KillIndices[NewReg] == ~0u) !=
- (DefIndices[NewReg] == ~0u)) &&
- "Kill and Def maps aren't consistent for NewReg!");
-
- Classes[AntiDepReg] = 0;
- DefIndices[AntiDepReg] = KillIndices[AntiDepReg];
- KillIndices[AntiDepReg] = ~0u;
- assert(((KillIndices[AntiDepReg] == ~0u) !=
- (DefIndices[AntiDepReg] == ~0u)) &&
- "Kill and Def maps aren't consistent for AntiDepReg!");
-
- RegRefs.erase(AntiDepReg);
- Changed = true;
- LastNewReg[AntiDepReg] = NewReg;
- }
- }
-
- ScanInstruction(MI, Count);
- }
-
- return Changed;
-}
-