diff --git a/lib/CodeGen/MachineLICM.cpp b/lib/CodeGen/MachineLICM.cpp index b9bb5b4255a..8c562cc4454 100644 --- a/lib/CodeGen/MachineLICM.cpp +++ b/lib/CodeGen/MachineLICM.cpp @@ -205,7 +205,7 @@ namespace { /// CanCauseHighRegPressure - Visit BBs from header to current BB, /// check if hoisting an instruction of the given cost matrix can cause high /// register pressure. - bool CanCauseHighRegPressure(DenseMap &Cost); + bool CanCauseHighRegPressure(DenseMap &Cost, bool Cheap); /// UpdateBackTraceRegPressure - Traverse the back trace from header to /// the current block and update their register pressures to reflect the @@ -1067,7 +1067,8 @@ bool MachineLICM::IsCheapInstruction(MachineInstr &MI) const { /// CanCauseHighRegPressure - Visit BBs from header to current BB, check /// if hoisting an instruction of the given cost matrix can cause high /// register pressure. -bool MachineLICM::CanCauseHighRegPressure(DenseMap &Cost) { +bool MachineLICM::CanCauseHighRegPressure(DenseMap &Cost, + bool CheapInstr) { for (DenseMap::iterator CI = Cost.begin(), CE = Cost.end(); CI != CE; ++CI) { if (CI->second <= 0) @@ -1076,6 +1077,12 @@ bool MachineLICM::CanCauseHighRegPressure(DenseMap &Cost) { unsigned RCId = CI->first; unsigned Limit = RegLimit[RCId]; int Cost = CI->second; + + // Don't hoist cheap instructions if they would increase register pressure, + // even if we're under the limit. + if (CheapInstr) + return true; + for (unsigned i = BackTrace.size(); i != 0; --i) { SmallVector &RP = BackTrace[i-1]; if (RP[RCId] + Cost >= Limit) @@ -1138,83 +1145,96 @@ bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) { if (MI.isImplicitDef()) return true; - // If the instruction is cheap, only hoist if it is re-materilizable. LICM - // will increase register pressure. It's probably not worth it if the - // instruction is cheap. - // Also hoist loads from constant memory, e.g. load from stubs, GOT. Hoisting - // these tend to help performance in low register pressure situation. The - // trade off is it may cause spill in high pressure situation. It will end up - // adding a store in the loop preheader. But the reload is no more expensive. - // The side benefit is these loads are frequently CSE'ed. - if (IsCheapInstruction(MI)) { - if (!TII->isTriviallyReMaterializable(&MI, AA)) - return false; - } else { - // Estimate register pressure to determine whether to LICM the instruction. - // In low register pressure situation, we can be more aggressive about - // hoisting. Also, favors hoisting long latency instructions even in - // moderately high pressure situation. - // FIXME: If there are long latency loop-invariant instructions inside the - // loop at this point, why didn't the optimizer's LICM hoist them? - DenseMap Cost; - for (unsigned i = 0, e = MI.getDesc().getNumOperands(); i != e; ++i) { - const MachineOperand &MO = MI.getOperand(i); - if (!MO.isReg() || MO.isImplicit()) - continue; - unsigned Reg = MO.getReg(); - if (!TargetRegisterInfo::isVirtualRegister(Reg)) - continue; + // Besides removing computation from the loop, hoisting an instruction has + // these effects: + // + // - The value defined by the instruction becomes live across the entire + // loop. This increases register pressure in the loop. + // + // - If the value is used by a PHI in the loop, a copy will be required for + // lowering the PHI after extending the live range. + // + // - When hoisting the last use of a value in the loop, that value no longer + // needs to be live in the loop. This lowers register pressure in the loop. - unsigned RCId, RCCost; - getRegisterClassIDAndCost(&MI, Reg, i, RCId, RCCost); - if (MO.isDef()) { - if (HasHighOperandLatency(MI, i, Reg)) { - ++NumHighLatency; - return true; - } + bool CheapInstr = IsCheapInstruction(MI); + bool CreatesCopy = HasLoopPHIUse(&MI); - DenseMap::iterator CI = Cost.find(RCId); - if (CI != Cost.end()) - CI->second += RCCost; - else - Cost.insert(std::make_pair(RCId, RCCost)); - } else if (isOperandKill(MO, MRI)) { - // Is a virtual register use is a kill, hoisting it out of the loop - // may actually reduce register pressure or be register pressure - // neutral. - DenseMap::iterator CI = Cost.find(RCId); - if (CI != Cost.end()) - CI->second -= RCCost; - else - Cost.insert(std::make_pair(RCId, -RCCost)); - } - } - - // Visit BBs from header to current BB, if hoisting this doesn't cause - // high register pressure, then it's safe to proceed. - if (!CanCauseHighRegPressure(Cost)) { - ++NumLowRP; - return true; - } - - // Do not "speculate" in high register pressure situation. If an - // instruction is not guaranteed to be executed in the loop, it's best to be - // conservative. - if (AvoidSpeculation && - (!IsGuaranteedToExecute(MI.getParent()) && !MayCSE(&MI))) - return false; - - // High register pressure situation, only hoist if the instruction is going - // to be remat'ed. - if (!TII->isTriviallyReMaterializable(&MI, AA) && - !MI.isInvariantLoad(AA)) - return false; + // Don't hoist a cheap instruction if it would create a copy in the loop. + if (CheapInstr && CreatesCopy) { + DEBUG(dbgs() << "Won't hoist cheap instr with loop PHI use: " << MI); + return false; } - // If result(s) of this instruction is used by PHIs inside the loop, then - // don't hoist it because it will introduce an extra copy. - if (HasLoopPHIUse(&MI)) + // Rematerializable instructions should always be hoisted since the register + // allocator can just pull them down again when needed. + if (TII->isTriviallyReMaterializable(&MI, AA)) + return true; + + // Estimate register pressure to determine whether to LICM the instruction. + // In low register pressure situation, we can be more aggressive about + // hoisting. Also, favors hoisting long latency instructions even in + // moderately high pressure situation. + // Cheap instructions will only be hoisted if they don't increase register + // pressure at all. + // FIXME: If there are long latency loop-invariant instructions inside the + // loop at this point, why didn't the optimizer's LICM hoist them? + DenseMap Cost; + for (unsigned i = 0, e = MI.getDesc().getNumOperands(); i != e; ++i) { + const MachineOperand &MO = MI.getOperand(i); + if (!MO.isReg() || MO.isImplicit()) + continue; + unsigned Reg = MO.getReg(); + if (!TargetRegisterInfo::isVirtualRegister(Reg)) + continue; + + unsigned RCId, RCCost; + getRegisterClassIDAndCost(&MI, Reg, i, RCId, RCCost); + if (MO.isDef()) { + if (HasHighOperandLatency(MI, i, Reg)) { + DEBUG(dbgs() << "Hoist High Latency: " << MI); + ++NumHighLatency; + return true; + } + Cost[RCId] += RCCost; + } else if (isOperandKill(MO, MRI)) { + // Is a virtual register use is a kill, hoisting it out of the loop + // may actually reduce register pressure or be register pressure + // neutral. + Cost[RCId] -= RCCost; + } + } + + // Visit BBs from header to current BB, if hoisting this doesn't cause + // high register pressure, then it's safe to proceed. + if (!CanCauseHighRegPressure(Cost, CheapInstr)) { + DEBUG(dbgs() << "Hoist non-reg-pressure: " << MI); + ++NumLowRP; + return true; + } + + // Don't risk increasing register pressure if it would create copies. + if (CreatesCopy) { + DEBUG(dbgs() << "Won't hoist instr with loop PHI use: " << MI); return false; + } + + // Do not "speculate" in high register pressure situation. If an + // instruction is not guaranteed to be executed in the loop, it's best to be + // conservative. + if (AvoidSpeculation && + (!IsGuaranteedToExecute(MI.getParent()) && !MayCSE(&MI))) { + DEBUG(dbgs() << "Won't speculate: " << MI); + return false; + } + + // High register pressure situation, only hoist if the instruction is going + // to be remat'ed. + if (!TII->isTriviallyReMaterializable(&MI, AA) && + !MI.isInvariantLoad(AA)) { + DEBUG(dbgs() << "Can't remat / high reg-pressure: " << MI); + return false; + } return true; }