//===-- GCNSchedStrategy.cpp - GCN Scheduler Strategy ---------------------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// // /// \file /// This contains a MachineSchedStrategy implementation for maximizing wave /// occupancy on GCN hardware. /// /// This pass will apply multiple scheduling stages to the same function. /// Regions are first recorded in GCNScheduleDAGMILive::schedule. The actual /// entry point for the scheduling of those regions is /// GCNScheduleDAGMILive::runSchedStages. /// Generally, the reason for having multiple scheduling stages is to account /// for the kernel-wide effect of register usage on occupancy. Usually, only a /// few scheduling regions will have register pressure high enough to limit /// occupancy for the kernel, so constraints can be relaxed to improve ILP in /// other regions. /// //===----------------------------------------------------------------------===// #include "GCNSchedStrategy.h" #include "AMDGPUIGroupLP.h" #include "GCNRegPressure.h" #include "SIMachineFunctionInfo.h" #include "Utils/AMDGPUBaseInfo.h" #include "llvm/ADT/STLExtras.h" #include "llvm/CodeGen/RegisterClassInfo.h" #include "llvm/MC/LaneBitmask.h" #include "llvm/Support/ErrorHandling.h" #define DEBUG_TYPE "machine-scheduler" using namespace llvm; static cl::opt DisableUnclusterHighRP( "amdgpu-disable-unclustered-high-rp-reschedule", cl::Hidden, cl::desc("Disable unclustered high register pressure " "reduction scheduling stage."), cl::init(false)); static cl::opt DisableClusteredLowOccupancy( "amdgpu-disable-clustered-low-occupancy-reschedule", cl::Hidden, cl::desc("Disable clustered low occupancy " "rescheduling for ILP scheduling stage."), cl::init(false)); static cl::opt ScheduleMetricBias( "amdgpu-schedule-metric-bias", cl::Hidden, cl::desc( "Sets the bias which adds weight to occupancy vs latency. Set it to " "100 to chase the occupancy only."), cl::init(10)); static cl::opt RelaxedOcc("amdgpu-schedule-relaxed-occupancy", cl::Hidden, cl::desc("Relax occupancy targets for kernels which are memory " "bound (amdgpu-membound-threshold), or " "Wave Limited (amdgpu-limit-wave-threshold)."), cl::init(false)); static cl::opt GCNTrackers( "amdgpu-use-amdgpu-trackers", cl::Hidden, cl::desc("Use the AMDGPU specific RPTrackers during scheduling"), cl::init(false)); const unsigned ScheduleMetrics::ScaleFactor = 100; GCNSchedStrategy::GCNSchedStrategy(const MachineSchedContext *C) : GenericScheduler(C), TargetOccupancy(0), MF(nullptr), DownwardTracker(*C->LIS), UpwardTracker(*C->LIS), HasHighPressure(false) { } void GCNSchedStrategy::initialize(ScheduleDAGMI *DAG) { GenericScheduler::initialize(DAG); MF = &DAG->MF; const GCNSubtarget &ST = MF->getSubtarget(); SGPRExcessLimit = Context->RegClassInfo->getNumAllocatableRegs(&AMDGPU::SGPR_32RegClass); VGPRExcessLimit = Context->RegClassInfo->getNumAllocatableRegs(&AMDGPU::VGPR_32RegClass); SIMachineFunctionInfo &MFI = *MF->getInfo(); // Set the initial TargetOccupnacy to the maximum occupancy that we can // achieve for this function. This effectively sets a lower bound on the // 'Critical' register limits in the scheduler. // Allow for lower occupancy targets if kernel is wave limited or memory // bound, and using the relaxed occupancy feature. TargetOccupancy = RelaxedOcc ? MFI.getMinAllowedOccupancy() : MFI.getOccupancy(); SGPRCriticalLimit = std::min(ST.getMaxNumSGPRs(TargetOccupancy, true), SGPRExcessLimit); if (!KnownExcessRP) { VGPRCriticalLimit = std::min( ST.getMaxNumVGPRs(TargetOccupancy, MFI.getDynamicVGPRBlockSize()), VGPRExcessLimit); } else { // This is similar to ST.getMaxNumVGPRs(TargetOccupancy) result except // returns a reasonably small number for targets with lots of VGPRs, such // as GFX10 and GFX11. LLVM_DEBUG(dbgs() << "Region is known to spill, use alternative " "VGPRCriticalLimit calculation method.\n"); unsigned DynamicVGPRBlockSize = MFI.getDynamicVGPRBlockSize(); unsigned Granule = AMDGPU::IsaInfo::getVGPRAllocGranule(&ST, DynamicVGPRBlockSize); unsigned Addressable = AMDGPU::IsaInfo::getAddressableNumVGPRs(&ST, DynamicVGPRBlockSize); unsigned VGPRBudget = alignDown(Addressable / TargetOccupancy, Granule); VGPRBudget = std::max(VGPRBudget, Granule); VGPRCriticalLimit = std::min(VGPRBudget, VGPRExcessLimit); } // Subtract error margin and bias from register limits and avoid overflow. SGPRCriticalLimit -= std::min(SGPRLimitBias + ErrorMargin, SGPRCriticalLimit); VGPRCriticalLimit -= std::min(VGPRLimitBias + ErrorMargin, VGPRCriticalLimit); SGPRExcessLimit -= std::min(SGPRLimitBias + ErrorMargin, SGPRExcessLimit); VGPRExcessLimit -= std::min(VGPRLimitBias + ErrorMargin, VGPRExcessLimit); LLVM_DEBUG(dbgs() << "VGPRCriticalLimit = " << VGPRCriticalLimit << ", VGPRExcessLimit = " << VGPRExcessLimit << ", SGPRCriticalLimit = " << SGPRCriticalLimit << ", SGPRExcessLimit = " << SGPRExcessLimit << "\n\n"); } /// Checks whether \p SU can use the cached DAG pressure diffs to compute the /// current register pressure. /// /// This works for the common case, but it has a few exceptions that have been /// observed through trial and error: /// - Explicit physical register operands /// - Subregister definitions /// /// In both of those cases, PressureDiff doesn't represent the actual pressure, /// and querying LiveIntervals through the RegPressureTracker is needed to get /// an accurate value. /// /// We should eventually only use PressureDiff for maximum performance, but this /// already allows 80% of SUs to take the fast path without changing scheduling /// at all. Further changes would either change scheduling, or require a lot /// more logic to recover an accurate pressure estimate from the PressureDiffs. static bool canUsePressureDiffs(const SUnit &SU) { if (!SU.isInstr()) return false; // Cannot use pressure diffs for subregister defs or with physregs, it's // imprecise in both cases. for (const auto &Op : SU.getInstr()->operands()) { if (!Op.isReg() || Op.isImplicit()) continue; if (Op.getReg().isPhysical() || (Op.isDef() && Op.getSubReg() != AMDGPU::NoSubRegister)) return false; } return true; } static void getRegisterPressures( bool AtTop, const RegPressureTracker &RPTracker, SUnit *SU, std::vector &Pressure, std::vector &MaxPressure, GCNDownwardRPTracker &DownwardTracker, GCNUpwardRPTracker &UpwardTracker, ScheduleDAGMI *DAG, const SIRegisterInfo *SRI) { // getDownwardPressure() and getUpwardPressure() make temporary changes to // the tracker, so we need to pass those function a non-const copy. RegPressureTracker &TempTracker = const_cast(RPTracker); if (!GCNTrackers) { AtTop ? TempTracker.getDownwardPressure(SU->getInstr(), Pressure, MaxPressure) : TempTracker.getUpwardPressure(SU->getInstr(), Pressure, MaxPressure); return; } // GCNTrackers Pressure.resize(4, 0); MachineInstr *MI = SU->getInstr(); GCNRegPressure NewPressure; if (AtTop) { GCNDownwardRPTracker TempDownwardTracker(DownwardTracker); NewPressure = TempDownwardTracker.bumpDownwardPressure(MI, SRI); } else { GCNUpwardRPTracker TempUpwardTracker(UpwardTracker); TempUpwardTracker.recede(*MI); NewPressure = TempUpwardTracker.getPressure(); } Pressure[AMDGPU::RegisterPressureSets::SReg_32] = NewPressure.getSGPRNum(); Pressure[AMDGPU::RegisterPressureSets::VGPR_32] = NewPressure.getArchVGPRNum(); Pressure[AMDGPU::RegisterPressureSets::AGPR_32] = NewPressure.getAGPRNum(); } void GCNSchedStrategy::initCandidate(SchedCandidate &Cand, SUnit *SU, bool AtTop, const RegPressureTracker &RPTracker, const SIRegisterInfo *SRI, unsigned SGPRPressure, unsigned VGPRPressure, bool IsBottomUp) { Cand.SU = SU; Cand.AtTop = AtTop; if (!DAG->isTrackingPressure()) return; Pressure.clear(); MaxPressure.clear(); // We try to use the cached PressureDiffs in the ScheduleDAG whenever // possible over querying the RegPressureTracker. // // RegPressureTracker will make a lot of LIS queries which are very // expensive, it is considered a slow function in this context. // // PressureDiffs are precomputed and cached, and getPressureDiff is just a // trivial lookup into an array. It is pretty much free. // // In EXPENSIVE_CHECKS, we always query RPTracker to verify the results of // PressureDiffs. if (AtTop || !canUsePressureDiffs(*SU) || GCNTrackers) { getRegisterPressures(AtTop, RPTracker, SU, Pressure, MaxPressure, DownwardTracker, UpwardTracker, DAG, SRI); } else { // Reserve 4 slots. Pressure.resize(4, 0); Pressure[AMDGPU::RegisterPressureSets::SReg_32] = SGPRPressure; Pressure[AMDGPU::RegisterPressureSets::VGPR_32] = VGPRPressure; for (const auto &Diff : DAG->getPressureDiff(SU)) { if (!Diff.isValid()) continue; // PressureDiffs is always bottom-up so if we're working top-down we need // to invert its sign. Pressure[Diff.getPSet()] += (IsBottomUp ? Diff.getUnitInc() : -Diff.getUnitInc()); } #ifdef EXPENSIVE_CHECKS std::vector CheckPressure, CheckMaxPressure; getRegisterPressures(AtTop, RPTracker, SU, CheckPressure, CheckMaxPressure, DownwardTracker, UpwardTracker, DAG, SRI); if (Pressure[AMDGPU::RegisterPressureSets::SReg_32] != CheckPressure[AMDGPU::RegisterPressureSets::SReg_32] || Pressure[AMDGPU::RegisterPressureSets::VGPR_32] != CheckPressure[AMDGPU::RegisterPressureSets::VGPR_32]) { errs() << "Register Pressure is inaccurate when calculated through " "PressureDiff\n" << "SGPR got " << Pressure[AMDGPU::RegisterPressureSets::SReg_32] << ", expected " << CheckPressure[AMDGPU::RegisterPressureSets::SReg_32] << "\n" << "VGPR got " << Pressure[AMDGPU::RegisterPressureSets::VGPR_32] << ", expected " << CheckPressure[AMDGPU::RegisterPressureSets::VGPR_32] << "\n"; report_fatal_error("inaccurate register pressure calculation"); } #endif } unsigned NewSGPRPressure = Pressure[AMDGPU::RegisterPressureSets::SReg_32]; unsigned NewVGPRPressure = Pressure[AMDGPU::RegisterPressureSets::VGPR_32]; // If two instructions increase the pressure of different register sets // by the same amount, the generic scheduler will prefer to schedule the // instruction that increases the set with the least amount of registers, // which in our case would be SGPRs. This is rarely what we want, so // when we report excess/critical register pressure, we do it either // only for VGPRs or only for SGPRs. // FIXME: Better heuristics to determine whether to prefer SGPRs or VGPRs. const unsigned MaxVGPRPressureInc = 16; bool ShouldTrackVGPRs = VGPRPressure + MaxVGPRPressureInc >= VGPRExcessLimit; bool ShouldTrackSGPRs = !ShouldTrackVGPRs && SGPRPressure >= SGPRExcessLimit; // FIXME: We have to enter REG-EXCESS before we reach the actual threshold // to increase the likelihood we don't go over the limits. We should improve // the analysis to look through dependencies to find the path with the least // register pressure. // We only need to update the RPDelta for instructions that increase register // pressure. Instructions that decrease or keep reg pressure the same will be // marked as RegExcess in tryCandidate() when they are compared with // instructions that increase the register pressure. if (ShouldTrackVGPRs && NewVGPRPressure >= VGPRExcessLimit) { HasHighPressure = true; Cand.RPDelta.Excess = PressureChange(AMDGPU::RegisterPressureSets::VGPR_32); Cand.RPDelta.Excess.setUnitInc(NewVGPRPressure - VGPRExcessLimit); } if (ShouldTrackSGPRs && NewSGPRPressure >= SGPRExcessLimit) { HasHighPressure = true; Cand.RPDelta.Excess = PressureChange(AMDGPU::RegisterPressureSets::SReg_32); Cand.RPDelta.Excess.setUnitInc(NewSGPRPressure - SGPRExcessLimit); } // Register pressure is considered 'CRITICAL' if it is approaching a value // that would reduce the wave occupancy for the execution unit. When // register pressure is 'CRITICAL', increasing SGPR and VGPR pressure both // has the same cost, so we don't need to prefer one over the other. int SGPRDelta = NewSGPRPressure - SGPRCriticalLimit; int VGPRDelta = NewVGPRPressure - VGPRCriticalLimit; if (SGPRDelta >= 0 || VGPRDelta >= 0) { HasHighPressure = true; if (SGPRDelta > VGPRDelta) { Cand.RPDelta.CriticalMax = PressureChange(AMDGPU::RegisterPressureSets::SReg_32); Cand.RPDelta.CriticalMax.setUnitInc(SGPRDelta); } else { Cand.RPDelta.CriticalMax = PressureChange(AMDGPU::RegisterPressureSets::VGPR_32); Cand.RPDelta.CriticalMax.setUnitInc(VGPRDelta); } } } // This function is mostly cut and pasted from // GenericScheduler::pickNodeFromQueue() void GCNSchedStrategy::pickNodeFromQueue(SchedBoundary &Zone, const CandPolicy &ZonePolicy, const RegPressureTracker &RPTracker, SchedCandidate &Cand, bool IsBottomUp) { const SIRegisterInfo *SRI = static_cast(TRI); ArrayRef Pressure = RPTracker.getRegSetPressureAtPos(); unsigned SGPRPressure = 0; unsigned VGPRPressure = 0; if (DAG->isTrackingPressure()) { if (!GCNTrackers) { SGPRPressure = Pressure[AMDGPU::RegisterPressureSets::SReg_32]; VGPRPressure = Pressure[AMDGPU::RegisterPressureSets::VGPR_32]; } else { GCNRPTracker *T = IsBottomUp ? static_cast(&UpwardTracker) : static_cast(&DownwardTracker); SGPRPressure = T->getPressure().getSGPRNum(); VGPRPressure = T->getPressure().getArchVGPRNum(); } } ReadyQueue &Q = Zone.Available; for (SUnit *SU : Q) { SchedCandidate TryCand(ZonePolicy); initCandidate(TryCand, SU, Zone.isTop(), RPTracker, SRI, SGPRPressure, VGPRPressure, IsBottomUp); // Pass SchedBoundary only when comparing nodes from the same boundary. SchedBoundary *ZoneArg = Cand.AtTop == TryCand.AtTop ? &Zone : nullptr; tryCandidate(Cand, TryCand, ZoneArg); if (TryCand.Reason != NoCand) { // Initialize resource delta if needed in case future heuristics query it. if (TryCand.ResDelta == SchedResourceDelta()) TryCand.initResourceDelta(Zone.DAG, SchedModel); Cand.setBest(TryCand); LLVM_DEBUG(traceCandidate(Cand)); } } } // This function is mostly cut and pasted from // GenericScheduler::pickNodeBidirectional() SUnit *GCNSchedStrategy::pickNodeBidirectional(bool &IsTopNode) { // Schedule as far as possible in the direction of no choice. This is most // efficient, but also provides the best heuristics for CriticalPSets. if (SUnit *SU = Bot.pickOnlyChoice()) { IsTopNode = false; return SU; } if (SUnit *SU = Top.pickOnlyChoice()) { IsTopNode = true; return SU; } // Set the bottom-up policy based on the state of the current bottom zone and // the instructions outside the zone, including the top zone. CandPolicy BotPolicy; setPolicy(BotPolicy, /*IsPostRA=*/false, Bot, &Top); // Set the top-down policy based on the state of the current top zone and // the instructions outside the zone, including the bottom zone. CandPolicy TopPolicy; setPolicy(TopPolicy, /*IsPostRA=*/false, Top, &Bot); // See if BotCand is still valid (because we previously scheduled from Top). LLVM_DEBUG(dbgs() << "Picking from Bot:\n"); if (!BotCand.isValid() || BotCand.SU->isScheduled || BotCand.Policy != BotPolicy) { BotCand.reset(CandPolicy()); pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), BotCand, /*IsBottomUp=*/true); assert(BotCand.Reason != NoCand && "failed to find the first candidate"); } else { LLVM_DEBUG(traceCandidate(BotCand)); #ifndef NDEBUG if (VerifyScheduling) { SchedCandidate TCand; TCand.reset(CandPolicy()); pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), TCand, /*IsBottomUp=*/true); assert(TCand.SU == BotCand.SU && "Last pick result should correspond to re-picking right now"); } #endif } // Check if the top Q has a better candidate. LLVM_DEBUG(dbgs() << "Picking from Top:\n"); if (!TopCand.isValid() || TopCand.SU->isScheduled || TopCand.Policy != TopPolicy) { TopCand.reset(CandPolicy()); pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TopCand, /*IsBottomUp=*/false); assert(TopCand.Reason != NoCand && "failed to find the first candidate"); } else { LLVM_DEBUG(traceCandidate(TopCand)); #ifndef NDEBUG if (VerifyScheduling) { SchedCandidate TCand; TCand.reset(CandPolicy()); pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TCand, /*IsBottomUp=*/false); assert(TCand.SU == TopCand.SU && "Last pick result should correspond to re-picking right now"); } #endif } // Pick best from BotCand and TopCand. LLVM_DEBUG(dbgs() << "Top Cand: "; traceCandidate(TopCand); dbgs() << "Bot Cand: "; traceCandidate(BotCand);); SchedCandidate Cand = BotCand; TopCand.Reason = NoCand; tryCandidate(Cand, TopCand, nullptr); if (TopCand.Reason != NoCand) { Cand.setBest(TopCand); } LLVM_DEBUG(dbgs() << "Picking: "; traceCandidate(Cand);); IsTopNode = Cand.AtTop; return Cand.SU; } // This function is mostly cut and pasted from // GenericScheduler::pickNode() SUnit *GCNSchedStrategy::pickNode(bool &IsTopNode) { if (DAG->top() == DAG->bottom()) { assert(Top.Available.empty() && Top.Pending.empty() && Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage"); return nullptr; } SUnit *SU; do { if (RegionPolicy.OnlyTopDown) { SU = Top.pickOnlyChoice(); if (!SU) { CandPolicy NoPolicy; TopCand.reset(NoPolicy); pickNodeFromQueue(Top, NoPolicy, DAG->getTopRPTracker(), TopCand, /*IsBottomUp=*/false); assert(TopCand.Reason != NoCand && "failed to find a candidate"); SU = TopCand.SU; } IsTopNode = true; } else if (RegionPolicy.OnlyBottomUp) { SU = Bot.pickOnlyChoice(); if (!SU) { CandPolicy NoPolicy; BotCand.reset(NoPolicy); pickNodeFromQueue(Bot, NoPolicy, DAG->getBotRPTracker(), BotCand, /*IsBottomUp=*/true); assert(BotCand.Reason != NoCand && "failed to find a candidate"); SU = BotCand.SU; } IsTopNode = false; } else { SU = pickNodeBidirectional(IsTopNode); } } while (SU->isScheduled); if (SU->isTopReady()) Top.removeReady(SU); if (SU->isBottomReady()) Bot.removeReady(SU); LLVM_DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr()); return SU; } void GCNSchedStrategy::schedNode(SUnit *SU, bool IsTopNode) { if (GCNTrackers) { MachineInstr *MI = SU->getInstr(); IsTopNode ? (void)DownwardTracker.advance(MI, false) : UpwardTracker.recede(*MI); } return GenericScheduler::schedNode(SU, IsTopNode); } GCNSchedStageID GCNSchedStrategy::getCurrentStage() { assert(CurrentStage && CurrentStage != SchedStages.end()); return *CurrentStage; } bool GCNSchedStrategy::advanceStage() { assert(CurrentStage != SchedStages.end()); if (!CurrentStage) CurrentStage = SchedStages.begin(); else CurrentStage++; return CurrentStage != SchedStages.end(); } bool GCNSchedStrategy::hasNextStage() const { assert(CurrentStage); return std::next(CurrentStage) != SchedStages.end(); } GCNSchedStageID GCNSchedStrategy::getNextStage() const { assert(CurrentStage && std::next(CurrentStage) != SchedStages.end()); return *std::next(CurrentStage); } GCNMaxOccupancySchedStrategy::GCNMaxOccupancySchedStrategy( const MachineSchedContext *C, bool IsLegacyScheduler) : GCNSchedStrategy(C) { SchedStages.push_back(GCNSchedStageID::OccInitialSchedule); SchedStages.push_back(GCNSchedStageID::UnclusteredHighRPReschedule); SchedStages.push_back(GCNSchedStageID::ClusteredLowOccupancyReschedule); SchedStages.push_back(GCNSchedStageID::PreRARematerialize); GCNTrackers = GCNTrackers & !IsLegacyScheduler; } GCNMaxILPSchedStrategy::GCNMaxILPSchedStrategy(const MachineSchedContext *C) : GCNSchedStrategy(C) { SchedStages.push_back(GCNSchedStageID::ILPInitialSchedule); } bool GCNMaxILPSchedStrategy::tryCandidate(SchedCandidate &Cand, SchedCandidate &TryCand, SchedBoundary *Zone) const { // Initialize the candidate if needed. if (!Cand.isValid()) { TryCand.Reason = NodeOrder; return true; } // Avoid spilling by exceeding the register limit. if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.Excess, Cand.RPDelta.Excess, TryCand, Cand, RegExcess, TRI, DAG->MF)) return TryCand.Reason != NoCand; // Bias PhysReg Defs and copies to their uses and defined respectively. if (tryGreater(biasPhysReg(TryCand.SU, TryCand.AtTop), biasPhysReg(Cand.SU, Cand.AtTop), TryCand, Cand, PhysReg)) return TryCand.Reason != NoCand; bool SameBoundary = Zone != nullptr; if (SameBoundary) { // Prioritize instructions that read unbuffered resources by stall cycles. if (tryLess(Zone->getLatencyStallCycles(TryCand.SU), Zone->getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall)) return TryCand.Reason != NoCand; // Avoid critical resource consumption and balance the schedule. TryCand.initResourceDelta(DAG, SchedModel); if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources, TryCand, Cand, ResourceReduce)) return TryCand.Reason != NoCand; if (tryGreater(TryCand.ResDelta.DemandedResources, Cand.ResDelta.DemandedResources, TryCand, Cand, ResourceDemand)) return TryCand.Reason != NoCand; // Unconditionally try to reduce latency. if (tryLatency(TryCand, Cand, *Zone)) return TryCand.Reason != NoCand; // Weak edges are for clustering and other constraints. if (tryLess(getWeakLeft(TryCand.SU, TryCand.AtTop), getWeakLeft(Cand.SU, Cand.AtTop), TryCand, Cand, Weak)) return TryCand.Reason != NoCand; } // Keep clustered nodes together to encourage downstream peephole // optimizations which may reduce resource requirements. // // This is a best effort to set things up for a post-RA pass. Optimizations // like generating loads of multiple registers should ideally be done within // the scheduler pass by combining the loads during DAG postprocessing. const ClusterInfo *CandCluster = Cand.AtTop ? TopCluster : BotCluster; const ClusterInfo *TryCandCluster = TryCand.AtTop ? TopCluster : BotCluster; if (tryGreater(TryCandCluster && TryCandCluster->contains(TryCand.SU), CandCluster && CandCluster->contains(Cand.SU), TryCand, Cand, Cluster)) return TryCand.Reason != NoCand; // Avoid increasing the max critical pressure in the scheduled region. if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CriticalMax, Cand.RPDelta.CriticalMax, TryCand, Cand, RegCritical, TRI, DAG->MF)) return TryCand.Reason != NoCand; // Avoid increasing the max pressure of the entire region. if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CurrentMax, Cand.RPDelta.CurrentMax, TryCand, Cand, RegMax, TRI, DAG->MF)) return TryCand.Reason != NoCand; if (SameBoundary) { // Fall through to original instruction order. if ((Zone->isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum) || (!Zone->isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) { TryCand.Reason = NodeOrder; return true; } } return false; } GCNMaxMemoryClauseSchedStrategy::GCNMaxMemoryClauseSchedStrategy( const MachineSchedContext *C) : GCNSchedStrategy(C) { SchedStages.push_back(GCNSchedStageID::MemoryClauseInitialSchedule); } /// GCNMaxMemoryClauseSchedStrategy tries best to clause memory instructions as /// much as possible. This is achieved by: // 1. Prioritize clustered operations before stall latency heuristic. // 2. Prioritize long-latency-load before stall latency heuristic. /// /// \param Cand provides the policy and current best candidate. /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized. /// \param Zone describes the scheduled zone that we are extending, or nullptr /// if Cand is from a different zone than TryCand. /// \return \c true if TryCand is better than Cand (Reason is NOT NoCand) bool GCNMaxMemoryClauseSchedStrategy::tryCandidate(SchedCandidate &Cand, SchedCandidate &TryCand, SchedBoundary *Zone) const { // Initialize the candidate if needed. if (!Cand.isValid()) { TryCand.Reason = NodeOrder; return true; } // Bias PhysReg Defs and copies to their uses and defined respectively. if (tryGreater(biasPhysReg(TryCand.SU, TryCand.AtTop), biasPhysReg(Cand.SU, Cand.AtTop), TryCand, Cand, PhysReg)) return TryCand.Reason != NoCand; if (DAG->isTrackingPressure()) { // Avoid exceeding the target's limit. if (tryPressure(TryCand.RPDelta.Excess, Cand.RPDelta.Excess, TryCand, Cand, RegExcess, TRI, DAG->MF)) return TryCand.Reason != NoCand; // Avoid increasing the max critical pressure in the scheduled region. if (tryPressure(TryCand.RPDelta.CriticalMax, Cand.RPDelta.CriticalMax, TryCand, Cand, RegCritical, TRI, DAG->MF)) return TryCand.Reason != NoCand; } // MaxMemoryClause-specific: We prioritize clustered instructions as we would // get more benefit from clausing these memory instructions. const ClusterInfo *CandCluster = Cand.AtTop ? TopCluster : BotCluster; const ClusterInfo *TryCandCluster = TryCand.AtTop ? TopCluster : BotCluster; if (tryGreater(TryCandCluster && TryCandCluster->contains(TryCand.SU), CandCluster && CandCluster->contains(Cand.SU), TryCand, Cand, Cluster)) return TryCand.Reason != NoCand; // We only compare a subset of features when comparing nodes between // Top and Bottom boundary. Some properties are simply incomparable, in many // other instances we should only override the other boundary if something // is a clear good pick on one boundary. Skip heuristics that are more // "tie-breaking" in nature. bool SameBoundary = Zone != nullptr; if (SameBoundary) { // For loops that are acyclic path limited, aggressively schedule for // latency. Within an single cycle, whenever CurrMOps > 0, allow normal // heuristics to take precedence. if (Rem.IsAcyclicLatencyLimited && !Zone->getCurrMOps() && tryLatency(TryCand, Cand, *Zone)) return TryCand.Reason != NoCand; // MaxMemoryClause-specific: Prioritize long latency memory load // instructions in top-bottom order to hide more latency. The mayLoad check // is used to exclude store-like instructions, which we do not want to // scheduler them too early. bool TryMayLoad = TryCand.SU->isInstr() && TryCand.SU->getInstr()->mayLoad(); bool CandMayLoad = Cand.SU->isInstr() && Cand.SU->getInstr()->mayLoad(); if (TryMayLoad || CandMayLoad) { bool TryLongLatency = TryCand.SU->Latency > 10 * Cand.SU->Latency && TryMayLoad; bool CandLongLatency = 10 * TryCand.SU->Latency < Cand.SU->Latency && CandMayLoad; if (tryGreater(Zone->isTop() ? TryLongLatency : CandLongLatency, Zone->isTop() ? CandLongLatency : TryLongLatency, TryCand, Cand, Stall)) return TryCand.Reason != NoCand; } // Prioritize instructions that read unbuffered resources by stall cycles. if (tryLess(Zone->getLatencyStallCycles(TryCand.SU), Zone->getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall)) return TryCand.Reason != NoCand; } if (SameBoundary) { // Weak edges are for clustering and other constraints. if (tryLess(getWeakLeft(TryCand.SU, TryCand.AtTop), getWeakLeft(Cand.SU, Cand.AtTop), TryCand, Cand, Weak)) return TryCand.Reason != NoCand; } // Avoid increasing the max pressure of the entire region. if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CurrentMax, Cand.RPDelta.CurrentMax, TryCand, Cand, RegMax, TRI, DAG->MF)) return TryCand.Reason != NoCand; if (SameBoundary) { // Avoid critical resource consumption and balance the schedule. TryCand.initResourceDelta(DAG, SchedModel); if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources, TryCand, Cand, ResourceReduce)) return TryCand.Reason != NoCand; if (tryGreater(TryCand.ResDelta.DemandedResources, Cand.ResDelta.DemandedResources, TryCand, Cand, ResourceDemand)) return TryCand.Reason != NoCand; // Avoid serializing long latency dependence chains. // For acyclic path limited loops, latency was already checked above. if (!RegionPolicy.DisableLatencyHeuristic && TryCand.Policy.ReduceLatency && !Rem.IsAcyclicLatencyLimited && tryLatency(TryCand, Cand, *Zone)) return TryCand.Reason != NoCand; // Fall through to original instruction order. if (Zone->isTop() == (TryCand.SU->NodeNum < Cand.SU->NodeNum)) { assert(TryCand.SU->NodeNum != Cand.SU->NodeNum); TryCand.Reason = NodeOrder; return true; } } return false; } GCNScheduleDAGMILive::GCNScheduleDAGMILive( MachineSchedContext *C, std::unique_ptr S) : ScheduleDAGMILive(C, std::move(S)), ST(MF.getSubtarget()), MFI(*MF.getInfo()), StartingOccupancy(MFI.getOccupancy()), MinOccupancy(StartingOccupancy), RegionLiveOuts(this, /*IsLiveOut=*/true) { // We want regions with a single MI to be scheduled so that we can reason // about them correctly during scheduling stages that move MIs between regions // (e.g., rematerialization). ScheduleSingleMIRegions = true; LLVM_DEBUG(dbgs() << "Starting occupancy is " << StartingOccupancy << ".\n"); if (RelaxedOcc) { MinOccupancy = std::min(MFI.getMinAllowedOccupancy(), StartingOccupancy); if (MinOccupancy != StartingOccupancy) LLVM_DEBUG(dbgs() << "Allowing Occupancy drops to " << MinOccupancy << ".\n"); } } std::unique_ptr GCNScheduleDAGMILive::createSchedStage(GCNSchedStageID SchedStageID) { switch (SchedStageID) { case GCNSchedStageID::OccInitialSchedule: return std::make_unique(SchedStageID, *this); case GCNSchedStageID::UnclusteredHighRPReschedule: return std::make_unique(SchedStageID, *this); case GCNSchedStageID::ClusteredLowOccupancyReschedule: return std::make_unique(SchedStageID, *this); case GCNSchedStageID::PreRARematerialize: return std::make_unique(SchedStageID, *this); case GCNSchedStageID::ILPInitialSchedule: return std::make_unique(SchedStageID, *this); case GCNSchedStageID::MemoryClauseInitialSchedule: return std::make_unique(SchedStageID, *this); } llvm_unreachable("Unknown SchedStageID."); } void GCNScheduleDAGMILive::schedule() { // Collect all scheduling regions. The actual scheduling is performed in // GCNScheduleDAGMILive::finalizeSchedule. Regions.push_back(std::pair(RegionBegin, RegionEnd)); } GCNRegPressure GCNScheduleDAGMILive::getRealRegPressure(unsigned RegionIdx) const { GCNDownwardRPTracker RPTracker(*LIS); RPTracker.advance(begin(), end(), &LiveIns[RegionIdx]); return RPTracker.moveMaxPressure(); } static MachineInstr *getLastMIForRegion(MachineBasicBlock::iterator RegionBegin, MachineBasicBlock::iterator RegionEnd) { auto REnd = RegionEnd == RegionBegin->getParent()->end() ? std::prev(RegionEnd) : RegionEnd; return &*skipDebugInstructionsBackward(REnd, RegionBegin); } void GCNScheduleDAGMILive::computeBlockPressure(unsigned RegionIdx, const MachineBasicBlock *MBB) { GCNDownwardRPTracker RPTracker(*LIS); // If the block has the only successor then live-ins of that successor are // live-outs of the current block. We can reuse calculated live set if the // successor will be sent to scheduling past current block. // However, due to the bug in LiveInterval analysis it may happen that two // predecessors of the same successor block have different lane bitmasks for // a live-out register. Workaround that by sticking to one-to-one relationship // i.e. one predecessor with one successor block. const MachineBasicBlock *OnlySucc = nullptr; if (MBB->succ_size() == 1) { auto *Candidate = *MBB->succ_begin(); if (!Candidate->empty() && Candidate->pred_size() == 1) { SlotIndexes *Ind = LIS->getSlotIndexes(); if (Ind->getMBBStartIdx(MBB) < Ind->getMBBStartIdx(Candidate)) OnlySucc = Candidate; } } // Scheduler sends regions from the end of the block upwards. size_t CurRegion = RegionIdx; for (size_t E = Regions.size(); CurRegion != E; ++CurRegion) if (Regions[CurRegion].first->getParent() != MBB) break; --CurRegion; auto I = MBB->begin(); auto LiveInIt = MBBLiveIns.find(MBB); auto &Rgn = Regions[CurRegion]; auto *NonDbgMI = &*skipDebugInstructionsForward(Rgn.first, Rgn.second); if (LiveInIt != MBBLiveIns.end()) { auto LiveIn = std::move(LiveInIt->second); RPTracker.reset(*MBB->begin(), &LiveIn); MBBLiveIns.erase(LiveInIt); } else { I = Rgn.first; auto LRS = BBLiveInMap.lookup(NonDbgMI); #ifdef EXPENSIVE_CHECKS assert(isEqual(getLiveRegsBefore(*NonDbgMI, *LIS), LRS)); #endif RPTracker.reset(*I, &LRS); } for (;;) { I = RPTracker.getNext(); if (Regions[CurRegion].first == I || NonDbgMI == I) { LiveIns[CurRegion] = RPTracker.getLiveRegs(); RPTracker.clearMaxPressure(); } if (Regions[CurRegion].second == I) { Pressure[CurRegion] = RPTracker.moveMaxPressure(); if (CurRegion-- == RegionIdx) break; auto &Rgn = Regions[CurRegion]; NonDbgMI = &*skipDebugInstructionsForward(Rgn.first, Rgn.second); } RPTracker.advanceToNext(); RPTracker.advanceBeforeNext(); } if (OnlySucc) { if (I != MBB->end()) { RPTracker.advanceToNext(); RPTracker.advance(MBB->end()); } RPTracker.advanceBeforeNext(); MBBLiveIns[OnlySucc] = RPTracker.moveLiveRegs(); } } DenseMap GCNScheduleDAGMILive::getRegionLiveInMap() const { assert(!Regions.empty()); std::vector RegionFirstMIs; RegionFirstMIs.reserve(Regions.size()); auto I = Regions.rbegin(), E = Regions.rend(); do { const MachineBasicBlock *MBB = I->first->getParent(); auto *MI = &*skipDebugInstructionsForward(I->first, I->second); RegionFirstMIs.push_back(MI); do { ++I; } while (I != E && I->first->getParent() == MBB); } while (I != E); return getLiveRegMap(RegionFirstMIs, /*After=*/false, *LIS); } DenseMap GCNScheduleDAGMILive::getRegionLiveOutMap() const { assert(!Regions.empty()); std::vector RegionLastMIs; RegionLastMIs.reserve(Regions.size()); for (auto &[RegionBegin, RegionEnd] : reverse(Regions)) RegionLastMIs.push_back(getLastMIForRegion(RegionBegin, RegionEnd)); return getLiveRegMap(RegionLastMIs, /*After=*/true, *LIS); } void RegionPressureMap::buildLiveRegMap() { IdxToInstruction.clear(); RegionLiveRegMap = IsLiveOut ? DAG->getRegionLiveOutMap() : DAG->getRegionLiveInMap(); for (unsigned I = 0; I < DAG->Regions.size(); I++) { MachineInstr *RegionKey = IsLiveOut ? getLastMIForRegion(DAG->Regions[I].first, DAG->Regions[I].second) : &*DAG->Regions[I].first; IdxToInstruction[I] = RegionKey; } } void GCNScheduleDAGMILive::finalizeSchedule() { // Start actual scheduling here. This function is called by the base // MachineScheduler after all regions have been recorded by // GCNScheduleDAGMILive::schedule(). LiveIns.resize(Regions.size()); Pressure.resize(Regions.size()); RegionsWithHighRP.resize(Regions.size()); RegionsWithExcessRP.resize(Regions.size()); RegionsWithMinOcc.resize(Regions.size()); RegionsWithIGLPInstrs.resize(Regions.size()); RegionsWithHighRP.reset(); RegionsWithExcessRP.reset(); RegionsWithMinOcc.reset(); RegionsWithIGLPInstrs.reset(); runSchedStages(); } void GCNScheduleDAGMILive::runSchedStages() { LLVM_DEBUG(dbgs() << "All regions recorded, starting actual scheduling.\n"); if (!Regions.empty()) { BBLiveInMap = getRegionLiveInMap(); if (GCNTrackers) RegionLiveOuts.buildLiveRegMap(); } GCNSchedStrategy &S = static_cast(*SchedImpl); while (S.advanceStage()) { auto Stage = createSchedStage(S.getCurrentStage()); if (!Stage->initGCNSchedStage()) continue; for (auto Region : Regions) { RegionBegin = Region.first; RegionEnd = Region.second; // Setup for scheduling the region and check whether it should be skipped. if (!Stage->initGCNRegion()) { Stage->advanceRegion(); exitRegion(); continue; } if (GCNTrackers) { GCNDownwardRPTracker *DownwardTracker = S.getDownwardTracker(); GCNUpwardRPTracker *UpwardTracker = S.getUpwardTracker(); GCNRPTracker::LiveRegSet *RegionLiveIns = &LiveIns[Stage->getRegionIdx()]; reinterpret_cast(DownwardTracker) ->reset(MRI, *RegionLiveIns); reinterpret_cast(UpwardTracker) ->reset(MRI, RegionLiveOuts.getLiveRegsForRegionIdx( Stage->getRegionIdx())); } ScheduleDAGMILive::schedule(); Stage->finalizeGCNRegion(); } Stage->finalizeGCNSchedStage(); } } #ifndef NDEBUG raw_ostream &llvm::operator<<(raw_ostream &OS, const GCNSchedStageID &StageID) { switch (StageID) { case GCNSchedStageID::OccInitialSchedule: OS << "Max Occupancy Initial Schedule"; break; case GCNSchedStageID::UnclusteredHighRPReschedule: OS << "Unclustered High Register Pressure Reschedule"; break; case GCNSchedStageID::ClusteredLowOccupancyReschedule: OS << "Clustered Low Occupancy Reschedule"; break; case GCNSchedStageID::PreRARematerialize: OS << "Pre-RA Rematerialize"; break; case GCNSchedStageID::ILPInitialSchedule: OS << "Max ILP Initial Schedule"; break; case GCNSchedStageID::MemoryClauseInitialSchedule: OS << "Max memory clause Initial Schedule"; break; } return OS; } #endif GCNSchedStage::GCNSchedStage(GCNSchedStageID StageID, GCNScheduleDAGMILive &DAG) : DAG(DAG), S(static_cast(*DAG.SchedImpl)), MF(DAG.MF), MFI(DAG.MFI), ST(DAG.ST), StageID(StageID) {} bool GCNSchedStage::initGCNSchedStage() { if (!DAG.LIS) return false; LLVM_DEBUG(dbgs() << "Starting scheduling stage: " << StageID << "\n"); return true; } bool UnclusteredHighRPStage::initGCNSchedStage() { if (DisableUnclusterHighRP) return false; if (!GCNSchedStage::initGCNSchedStage()) return false; if (DAG.RegionsWithHighRP.none() && DAG.RegionsWithExcessRP.none()) return false; SavedMutations.swap(DAG.Mutations); DAG.addMutation( createIGroupLPDAGMutation(AMDGPU::SchedulingPhase::PreRAReentry)); InitialOccupancy = DAG.MinOccupancy; // Aggressivly try to reduce register pressure in the unclustered high RP // stage. Temporarily increase occupancy target in the region. S.SGPRLimitBias = S.HighRPSGPRBias; S.VGPRLimitBias = S.HighRPVGPRBias; if (MFI.getMaxWavesPerEU() > DAG.MinOccupancy) MFI.increaseOccupancy(MF, ++DAG.MinOccupancy); LLVM_DEBUG( dbgs() << "Retrying function scheduling without clustering. " "Aggressivly try to reduce register pressure to achieve occupancy " << DAG.MinOccupancy << ".\n"); return true; } bool ClusteredLowOccStage::initGCNSchedStage() { if (DisableClusteredLowOccupancy) return false; if (!GCNSchedStage::initGCNSchedStage()) return false; // Don't bother trying to improve ILP in lower RP regions if occupancy has not // been dropped. All regions will have already been scheduled with the ideal // occupancy targets. if (DAG.StartingOccupancy <= DAG.MinOccupancy) return false; LLVM_DEBUG( dbgs() << "Retrying function scheduling with lowest recorded occupancy " << DAG.MinOccupancy << ".\n"); return true; } /// Allows to easily filter for this stage's debug output. #define REMAT_DEBUG(X) LLVM_DEBUG(dbgs() << "[PreRARemat] "; X;) bool PreRARematStage::initGCNSchedStage() { // FIXME: This pass will invalidate cached BBLiveInMap and MBBLiveIns for // regions inbetween the defs and region we sinked the def to. Will need to be // fixed if there is another pass after this pass. assert(!S.hasNextStage()); if (!GCNSchedStage::initGCNSchedStage() || DAG.RegionsWithMinOcc.none() || DAG.Regions.size() == 1) return false; // Before performing any IR modification record the parent region of each MI // and the parent MBB of each region. const unsigned NumRegions = DAG.Regions.size(); RegionBB.reserve(NumRegions); for (unsigned I = 0; I < NumRegions; ++I) { RegionBoundaries Region = DAG.Regions[I]; for (auto MI = Region.first; MI != Region.second; ++MI) MIRegion.insert({&*MI, I}); RegionBB.push_back(Region.first->getParent()); } if (!canIncreaseOccupancyOrReduceSpill()) return false; // Rematerialize identified instructions and update scheduler's state. rematerialize(); if (GCNTrackers) DAG.RegionLiveOuts.buildLiveRegMap(); REMAT_DEBUG( dbgs() << "Retrying function scheduling with new min. occupancy of " << AchievedOcc << " from rematerializing (original was " << DAG.MinOccupancy << ", target was " << TargetOcc << ")\n"); if (AchievedOcc > DAG.MinOccupancy) { DAG.MinOccupancy = AchievedOcc; SIMachineFunctionInfo &MFI = *MF.getInfo(); MFI.increaseOccupancy(MF, DAG.MinOccupancy); } return true; } void GCNSchedStage::finalizeGCNSchedStage() { DAG.finishBlock(); LLVM_DEBUG(dbgs() << "Ending scheduling stage: " << StageID << "\n"); } void UnclusteredHighRPStage::finalizeGCNSchedStage() { SavedMutations.swap(DAG.Mutations); S.SGPRLimitBias = S.VGPRLimitBias = 0; if (DAG.MinOccupancy > InitialOccupancy) { for (unsigned IDX = 0; IDX < DAG.Pressure.size(); ++IDX) DAG.RegionsWithMinOcc[IDX] = DAG.Pressure[IDX].getOccupancy( DAG.ST, DAG.MFI.getDynamicVGPRBlockSize()) == DAG.MinOccupancy; LLVM_DEBUG(dbgs() << StageID << " stage successfully increased occupancy to " << DAG.MinOccupancy << '\n'); } GCNSchedStage::finalizeGCNSchedStage(); } bool GCNSchedStage::initGCNRegion() { // Check whether this new region is also a new block. if (DAG.RegionBegin->getParent() != CurrentMBB) setupNewBlock(); unsigned NumRegionInstrs = std::distance(DAG.begin(), DAG.end()); DAG.enterRegion(CurrentMBB, DAG.begin(), DAG.end(), NumRegionInstrs); // Skip empty scheduling regions (0 or 1 schedulable instructions). if (DAG.begin() == DAG.end() || DAG.begin() == std::prev(DAG.end())) return false; LLVM_DEBUG(dbgs() << "********** MI Scheduling **********\n"); LLVM_DEBUG(dbgs() << MF.getName() << ":" << printMBBReference(*CurrentMBB) << " " << CurrentMBB->getName() << "\n From: " << *DAG.begin() << " To: "; if (DAG.RegionEnd != CurrentMBB->end()) dbgs() << *DAG.RegionEnd; else dbgs() << "End"; dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n'); // Save original instruction order before scheduling for possible revert. Unsched.clear(); Unsched.reserve(DAG.NumRegionInstrs); if (StageID == GCNSchedStageID::OccInitialSchedule || StageID == GCNSchedStageID::ILPInitialSchedule) { const SIInstrInfo *SII = static_cast(DAG.TII); for (auto &I : DAG) { Unsched.push_back(&I); if (SII->isIGLPMutationOnly(I.getOpcode())) DAG.RegionsWithIGLPInstrs[RegionIdx] = true; } } else { for (auto &I : DAG) Unsched.push_back(&I); } PressureBefore = DAG.Pressure[RegionIdx]; LLVM_DEBUG( dbgs() << "Pressure before scheduling:\nRegion live-ins:" << print(DAG.LiveIns[RegionIdx], DAG.MRI) << "Region live-in pressure: " << print(llvm::getRegPressure(DAG.MRI, DAG.LiveIns[RegionIdx])) << "Region register pressure: " << print(PressureBefore)); S.HasHighPressure = false; S.KnownExcessRP = isRegionWithExcessRP(); if (DAG.RegionsWithIGLPInstrs[RegionIdx] && StageID != GCNSchedStageID::UnclusteredHighRPReschedule) { SavedMutations.clear(); SavedMutations.swap(DAG.Mutations); bool IsInitialStage = StageID == GCNSchedStageID::OccInitialSchedule || StageID == GCNSchedStageID::ILPInitialSchedule; DAG.addMutation(createIGroupLPDAGMutation( IsInitialStage ? AMDGPU::SchedulingPhase::Initial : AMDGPU::SchedulingPhase::PreRAReentry)); } return true; } bool UnclusteredHighRPStage::initGCNRegion() { // Only reschedule regions with the minimum occupancy or regions that may have // spilling (excess register pressure). if ((!DAG.RegionsWithMinOcc[RegionIdx] || DAG.MinOccupancy <= InitialOccupancy) && !DAG.RegionsWithExcessRP[RegionIdx]) return false; return GCNSchedStage::initGCNRegion(); } bool ClusteredLowOccStage::initGCNRegion() { // We may need to reschedule this region if it wasn't rescheduled in the last // stage, or if we found it was testing critical register pressure limits in // the unclustered reschedule stage. The later is because we may not have been // able to raise the min occupancy in the previous stage so the region may be // overly constrained even if it was already rescheduled. if (!DAG.RegionsWithHighRP[RegionIdx]) return false; return GCNSchedStage::initGCNRegion(); } bool PreRARematStage::initGCNRegion() { return RescheduleRegions[RegionIdx] && GCNSchedStage::initGCNRegion(); } void GCNSchedStage::setupNewBlock() { if (CurrentMBB) DAG.finishBlock(); CurrentMBB = DAG.RegionBegin->getParent(); DAG.startBlock(CurrentMBB); // Get real RP for the region if it hasn't be calculated before. After the // initial schedule stage real RP will be collected after scheduling. if (StageID == GCNSchedStageID::OccInitialSchedule || StageID == GCNSchedStageID::ILPInitialSchedule || StageID == GCNSchedStageID::MemoryClauseInitialSchedule) DAG.computeBlockPressure(RegionIdx, CurrentMBB); } void GCNSchedStage::finalizeGCNRegion() { DAG.Regions[RegionIdx] = std::pair(DAG.RegionBegin, DAG.RegionEnd); if (S.HasHighPressure) DAG.RegionsWithHighRP[RegionIdx] = true; // Revert scheduling if we have dropped occupancy or there is some other // reason that the original schedule is better. checkScheduling(); if (DAG.RegionsWithIGLPInstrs[RegionIdx] && StageID != GCNSchedStageID::UnclusteredHighRPReschedule) SavedMutations.swap(DAG.Mutations); DAG.exitRegion(); advanceRegion(); } void GCNSchedStage::checkScheduling() { // Check the results of scheduling. PressureAfter = DAG.getRealRegPressure(RegionIdx); LLVM_DEBUG(dbgs() << "Pressure after scheduling: " << print(PressureAfter)); LLVM_DEBUG(dbgs() << "Region: " << RegionIdx << ".\n"); unsigned DynamicVGPRBlockSize = DAG.MFI.getDynamicVGPRBlockSize(); if (PressureAfter.getSGPRNum() <= S.SGPRCriticalLimit && PressureAfter.getVGPRNum(ST.hasGFX90AInsts()) <= S.VGPRCriticalLimit) { DAG.Pressure[RegionIdx] = PressureAfter; DAG.RegionsWithMinOcc[RegionIdx] = PressureAfter.getOccupancy(ST, DynamicVGPRBlockSize) == DAG.MinOccupancy; // Early out if we have achieved the occupancy target. LLVM_DEBUG(dbgs() << "Pressure in desired limits, done.\n"); return; } unsigned TargetOccupancy = std::min( S.getTargetOccupancy(), ST.getOccupancyWithWorkGroupSizes(MF).second); unsigned WavesAfter = std::min( TargetOccupancy, PressureAfter.getOccupancy(ST, DynamicVGPRBlockSize)); unsigned WavesBefore = std::min( TargetOccupancy, PressureBefore.getOccupancy(ST, DynamicVGPRBlockSize)); LLVM_DEBUG(dbgs() << "Occupancy before scheduling: " << WavesBefore << ", after " << WavesAfter << ".\n"); // We may not be able to keep the current target occupancy because of the just // scheduled region. We might still be able to revert scheduling if the // occupancy before was higher, or if the current schedule has register // pressure higher than the excess limits which could lead to more spilling. unsigned NewOccupancy = std::max(WavesAfter, WavesBefore); // Allow memory bound functions to drop to 4 waves if not limited by an // attribute. if (WavesAfter < WavesBefore && WavesAfter < DAG.MinOccupancy && WavesAfter >= MFI.getMinAllowedOccupancy()) { LLVM_DEBUG(dbgs() << "Function is memory bound, allow occupancy drop up to " << MFI.getMinAllowedOccupancy() << " waves\n"); NewOccupancy = WavesAfter; } if (NewOccupancy < DAG.MinOccupancy) { DAG.MinOccupancy = NewOccupancy; MFI.limitOccupancy(DAG.MinOccupancy); DAG.RegionsWithMinOcc.reset(); LLVM_DEBUG(dbgs() << "Occupancy lowered for the function to " << DAG.MinOccupancy << ".\n"); } // The maximum number of arch VGPR on non-unified register file, or the // maximum VGPR + AGPR in the unified register file case. unsigned MaxVGPRs = ST.getMaxNumVGPRs(MF); // The maximum number of arch VGPR for both unified and non-unified register // file. unsigned MaxArchVGPRs = std::min(MaxVGPRs, ST.getAddressableNumArchVGPRs()); unsigned MaxSGPRs = ST.getMaxNumSGPRs(MF); if (PressureAfter.getVGPRNum(ST.hasGFX90AInsts()) > MaxVGPRs || PressureAfter.getArchVGPRNum() > MaxArchVGPRs || PressureAfter.getAGPRNum() > MaxArchVGPRs || PressureAfter.getSGPRNum() > MaxSGPRs) { DAG.RegionsWithHighRP[RegionIdx] = true; DAG.RegionsWithExcessRP[RegionIdx] = true; } // Revert if this region's schedule would cause a drop in occupancy or // spilling. if (shouldRevertScheduling(WavesAfter)) { revertScheduling(); } else { DAG.Pressure[RegionIdx] = PressureAfter; DAG.RegionsWithMinOcc[RegionIdx] = PressureAfter.getOccupancy(ST, DynamicVGPRBlockSize) == DAG.MinOccupancy; } } unsigned GCNSchedStage::computeSUnitReadyCycle(const SUnit &SU, unsigned CurrCycle, DenseMap &ReadyCycles, const TargetSchedModel &SM) { unsigned ReadyCycle = CurrCycle; for (auto &D : SU.Preds) { if (D.isAssignedRegDep()) { MachineInstr *DefMI = D.getSUnit()->getInstr(); unsigned Latency = SM.computeInstrLatency(DefMI); unsigned DefReady = ReadyCycles[DAG.getSUnit(DefMI)->NodeNum]; ReadyCycle = std::max(ReadyCycle, DefReady + Latency); } } ReadyCycles[SU.NodeNum] = ReadyCycle; return ReadyCycle; } #ifndef NDEBUG struct EarlierIssuingCycle { bool operator()(std::pair A, std::pair B) const { return A.second < B.second; } }; static void printScheduleModel(std::set, EarlierIssuingCycle> &ReadyCycles) { if (ReadyCycles.empty()) return; unsigned BBNum = ReadyCycles.begin()->first->getParent()->getNumber(); dbgs() << "\n################## Schedule time ReadyCycles for MBB : " << BBNum << " ##################\n# Cycle #\t\t\tInstruction " " " " \n"; unsigned IPrev = 1; for (auto &I : ReadyCycles) { if (I.second > IPrev + 1) dbgs() << "****************************** BUBBLE OF " << I.second - IPrev << " CYCLES DETECTED ******************************\n\n"; dbgs() << "[ " << I.second << " ] : " << *I.first << "\n"; IPrev = I.second; } } #endif ScheduleMetrics GCNSchedStage::getScheduleMetrics(const std::vector &InputSchedule) { #ifndef NDEBUG std::set, EarlierIssuingCycle> ReadyCyclesSorted; #endif const TargetSchedModel &SM = ST.getInstrInfo()->getSchedModel(); unsigned SumBubbles = 0; DenseMap ReadyCycles; unsigned CurrCycle = 0; for (auto &SU : InputSchedule) { unsigned ReadyCycle = computeSUnitReadyCycle(SU, CurrCycle, ReadyCycles, SM); SumBubbles += ReadyCycle - CurrCycle; #ifndef NDEBUG ReadyCyclesSorted.insert(std::make_pair(SU.getInstr(), ReadyCycle)); #endif CurrCycle = ++ReadyCycle; } #ifndef NDEBUG LLVM_DEBUG( printScheduleModel(ReadyCyclesSorted); dbgs() << "\n\t" << "Metric: " << (SumBubbles ? (SumBubbles * ScheduleMetrics::ScaleFactor) / CurrCycle : 1) << "\n\n"); #endif return ScheduleMetrics(CurrCycle, SumBubbles); } ScheduleMetrics GCNSchedStage::getScheduleMetrics(const GCNScheduleDAGMILive &DAG) { #ifndef NDEBUG std::set, EarlierIssuingCycle> ReadyCyclesSorted; #endif const TargetSchedModel &SM = ST.getInstrInfo()->getSchedModel(); unsigned SumBubbles = 0; DenseMap ReadyCycles; unsigned CurrCycle = 0; for (auto &MI : DAG) { SUnit *SU = DAG.getSUnit(&MI); if (!SU) continue; unsigned ReadyCycle = computeSUnitReadyCycle(*SU, CurrCycle, ReadyCycles, SM); SumBubbles += ReadyCycle - CurrCycle; #ifndef NDEBUG ReadyCyclesSorted.insert(std::make_pair(SU->getInstr(), ReadyCycle)); #endif CurrCycle = ++ReadyCycle; } #ifndef NDEBUG LLVM_DEBUG( printScheduleModel(ReadyCyclesSorted); dbgs() << "\n\t" << "Metric: " << (SumBubbles ? (SumBubbles * ScheduleMetrics::ScaleFactor) / CurrCycle : 1) << "\n\n"); #endif return ScheduleMetrics(CurrCycle, SumBubbles); } bool GCNSchedStage::shouldRevertScheduling(unsigned WavesAfter) { if (WavesAfter < DAG.MinOccupancy) return true; // For dynamic VGPR mode, we don't want to waste any VGPR blocks. if (DAG.MFI.isDynamicVGPREnabled()) { unsigned BlocksBefore = AMDGPU::IsaInfo::getAllocatedNumVGPRBlocks( &ST, DAG.MFI.getDynamicVGPRBlockSize(), PressureBefore.getVGPRNum(false)); unsigned BlocksAfter = AMDGPU::IsaInfo::getAllocatedNumVGPRBlocks( &ST, DAG.MFI.getDynamicVGPRBlockSize(), PressureAfter.getVGPRNum(false)); if (BlocksAfter > BlocksBefore) return true; } return false; } bool OccInitialScheduleStage::shouldRevertScheduling(unsigned WavesAfter) { if (PressureAfter == PressureBefore) return false; if (GCNSchedStage::shouldRevertScheduling(WavesAfter)) return true; if (mayCauseSpilling(WavesAfter)) return true; return false; } bool UnclusteredHighRPStage::shouldRevertScheduling(unsigned WavesAfter) { // If RP is not reduced in the unclustered reschedule stage, revert to the // old schedule. if ((WavesAfter <= PressureBefore.getOccupancy(ST, DAG.MFI.getDynamicVGPRBlockSize()) && mayCauseSpilling(WavesAfter)) || GCNSchedStage::shouldRevertScheduling(WavesAfter)) { LLVM_DEBUG(dbgs() << "Unclustered reschedule did not help.\n"); return true; } // Do not attempt to relax schedule even more if we are already spilling. if (isRegionWithExcessRP()) return false; LLVM_DEBUG( dbgs() << "\n\t *** In shouldRevertScheduling ***\n" << " *********** BEFORE UnclusteredHighRPStage ***********\n"); ScheduleMetrics MBefore = getScheduleMetrics(DAG.SUnits); LLVM_DEBUG( dbgs() << "\n *********** AFTER UnclusteredHighRPStage ***********\n"); ScheduleMetrics MAfter = getScheduleMetrics(DAG); unsigned OldMetric = MBefore.getMetric(); unsigned NewMetric = MAfter.getMetric(); unsigned WavesBefore = std::min( S.getTargetOccupancy(), PressureBefore.getOccupancy(ST, DAG.MFI.getDynamicVGPRBlockSize())); unsigned Profit = ((WavesAfter * ScheduleMetrics::ScaleFactor) / WavesBefore * ((OldMetric + ScheduleMetricBias) * ScheduleMetrics::ScaleFactor) / NewMetric) / ScheduleMetrics::ScaleFactor; LLVM_DEBUG(dbgs() << "\tMetric before " << MBefore << "\tMetric after " << MAfter << "Profit: " << Profit << "\n"); return Profit < ScheduleMetrics::ScaleFactor; } bool ClusteredLowOccStage::shouldRevertScheduling(unsigned WavesAfter) { if (PressureAfter == PressureBefore) return false; if (GCNSchedStage::shouldRevertScheduling(WavesAfter)) return true; if (mayCauseSpilling(WavesAfter)) return true; return false; } bool PreRARematStage::shouldRevertScheduling(unsigned WavesAfter) { return GCNSchedStage::shouldRevertScheduling(WavesAfter) || mayCauseSpilling(WavesAfter) || (IncreaseOccupancy && WavesAfter < TargetOcc); } bool ILPInitialScheduleStage::shouldRevertScheduling(unsigned WavesAfter) { if (mayCauseSpilling(WavesAfter)) return true; return false; } bool MemoryClauseInitialScheduleStage::shouldRevertScheduling( unsigned WavesAfter) { return mayCauseSpilling(WavesAfter); } bool GCNSchedStage::mayCauseSpilling(unsigned WavesAfter) { if (WavesAfter <= MFI.getMinWavesPerEU() && isRegionWithExcessRP() && !PressureAfter.less(MF, PressureBefore)) { LLVM_DEBUG(dbgs() << "New pressure will result in more spilling.\n"); return true; } return false; } void GCNSchedStage::revertScheduling() { DAG.RegionsWithMinOcc[RegionIdx] = PressureBefore.getOccupancy(ST, DAG.MFI.getDynamicVGPRBlockSize()) == DAG.MinOccupancy; LLVM_DEBUG(dbgs() << "Attempting to revert scheduling.\n"); DAG.RegionEnd = DAG.RegionBegin; int SkippedDebugInstr = 0; for (MachineInstr *MI : Unsched) { if (MI->isDebugInstr()) { ++SkippedDebugInstr; continue; } if (MI->getIterator() != DAG.RegionEnd) { DAG.BB->splice(DAG.RegionEnd, DAG.BB, MI); if (!MI->isDebugInstr()) DAG.LIS->handleMove(*MI, true); } // Reset read-undef flags and update them later. for (auto &Op : MI->all_defs()) Op.setIsUndef(false); RegisterOperands RegOpers; RegOpers.collect(*MI, *DAG.TRI, DAG.MRI, DAG.ShouldTrackLaneMasks, false); if (!MI->isDebugInstr()) { if (DAG.ShouldTrackLaneMasks) { // Adjust liveness and add missing dead+read-undef flags. SlotIndex SlotIdx = DAG.LIS->getInstructionIndex(*MI).getRegSlot(); RegOpers.adjustLaneLiveness(*DAG.LIS, DAG.MRI, SlotIdx, MI); } else { // Adjust for missing dead-def flags. RegOpers.detectDeadDefs(*MI, *DAG.LIS); } } DAG.RegionEnd = MI->getIterator(); ++DAG.RegionEnd; LLVM_DEBUG(dbgs() << "Scheduling " << *MI); } // After reverting schedule, debug instrs will now be at the end of the block // and RegionEnd will point to the first debug instr. Increment RegionEnd // pass debug instrs to the actual end of the scheduling region. while (SkippedDebugInstr-- > 0) ++DAG.RegionEnd; // If Unsched.front() instruction is a debug instruction, this will actually // shrink the region since we moved all debug instructions to the end of the // block. Find the first instruction that is not a debug instruction. DAG.RegionBegin = Unsched.front()->getIterator(); if (DAG.RegionBegin->isDebugInstr()) { for (MachineInstr *MI : Unsched) { if (MI->isDebugInstr()) continue; DAG.RegionBegin = MI->getIterator(); break; } } // Then move the debug instructions back into their correct place and set // RegionBegin and RegionEnd if needed. DAG.placeDebugValues(); DAG.Regions[RegionIdx] = std::pair(DAG.RegionBegin, DAG.RegionEnd); } bool PreRARematStage::allUsesAvailableAt(const MachineInstr *InstToRemat, SlotIndex OriginalIdx, SlotIndex RematIdx) const { LiveIntervals *LIS = DAG.LIS; MachineRegisterInfo &MRI = DAG.MRI; OriginalIdx = OriginalIdx.getRegSlot(true); RematIdx = std::max(RematIdx, RematIdx.getRegSlot(true)); for (const MachineOperand &MO : InstToRemat->operands()) { if (!MO.isReg() || !MO.getReg() || !MO.readsReg()) continue; if (!MO.getReg().isVirtual()) { // Do not attempt to reason about PhysRegs // TODO: better analysis of PhysReg livness if (!DAG.MRI.isConstantPhysReg(MO.getReg()) && !DAG.TII->isIgnorableUse(MO)) return false; // Constant PhysRegs and IgnorableUses are okay continue; } LiveInterval &LI = LIS->getInterval(MO.getReg()); const VNInfo *OVNI = LI.getVNInfoAt(OriginalIdx); assert(OVNI); // Don't allow rematerialization immediately after the original def. // It would be incorrect if InstToRemat redefines the register. // See PR14098. if (SlotIndex::isSameInstr(OriginalIdx, RematIdx)) return false; if (OVNI != LI.getVNInfoAt(RematIdx)) return false; // Check that subrange is live at RematIdx. if (LI.hasSubRanges()) { const TargetRegisterInfo *TRI = MRI.getTargetRegisterInfo(); unsigned SubReg = MO.getSubReg(); LaneBitmask LM = SubReg ? TRI->getSubRegIndexLaneMask(SubReg) : MRI.getMaxLaneMaskForVReg(MO.getReg()); for (LiveInterval::SubRange &SR : LI.subranges()) { if ((SR.LaneMask & LM).none()) continue; if (!SR.liveAt(RematIdx)) return false; // Early exit if all used lanes are checked. No need to continue. LM &= ~SR.LaneMask; if (LM.none()) break; } } } return true; } bool PreRARematStage::canIncreaseOccupancyOrReduceSpill() { REMAT_DEBUG({ dbgs() << "Collecting rematerializable instructions in "; MF.getFunction().printAsOperand(dbgs(), false); dbgs() << '\n'; }); // Maps optimizable regions (i.e., regions at minimum and register-limited // occupancy, or regions with spilling) to the target RP we would like to // reach. DenseMap OptRegions; const Function &F = MF.getFunction(); unsigned DynamicVGPRBlockSize = MF.getInfo()->getDynamicVGPRBlockSize(); std::pair WavesPerEU = ST.getWavesPerEU(F); const unsigned MaxSGPRsNoSpill = ST.getMaxNumSGPRs(F); const unsigned MaxVGPRsNoSpill = ST.getMaxNumVGPRs(F); const unsigned MaxSGPRsIncOcc = ST.getMaxNumSGPRs(DAG.MinOccupancy + 1, false); const unsigned MaxVGPRsIncOcc = ST.getMaxNumVGPRs(DAG.MinOccupancy + 1, DynamicVGPRBlockSize); IncreaseOccupancy = WavesPerEU.second > DAG.MinOccupancy; // Collect optimizable regions. If there is spilling in any region we will // just try to reduce spilling. Otherwise we will try to increase occupancy by // one in the whole function. for (unsigned I = 0, E = DAG.Regions.size(); I != E; ++I) { GCNRegPressure &RP = DAG.Pressure[I]; // We allow ArchVGPR or AGPR savings to count as savings of the other kind // of VGPR only when trying to eliminate spilling. We cannot do this when // trying to increase occupancy since VGPR class swaps only occur later in // the register allocator i.e., the scheduler will not be able to reason // about these savings and will not report an increase in the achievable // occupancy, triggering rollbacks. GCNRPTarget Target(MaxSGPRsNoSpill, MaxVGPRsNoSpill, MF, RP, /*CombineVGPRSavings=*/true); if (!Target.satisfied() && IncreaseOccupancy) { // There is spilling in the region and we were so far trying to increase // occupancy. Strop trying that and focus on reducing spilling. IncreaseOccupancy = false; OptRegions.clear(); } else if (IncreaseOccupancy) { // There is no spilling in the region, try to increase occupancy. Target = GCNRPTarget(MaxSGPRsIncOcc, MaxVGPRsIncOcc, MF, RP, /*CombineVGPRSavings=*/false); } if (!Target.satisfied()) OptRegions.insert({I, Target}); } if (OptRegions.empty()) return false; #ifndef NDEBUG if (IncreaseOccupancy) { REMAT_DEBUG(dbgs() << "Occupancy minimal (" << DAG.MinOccupancy << ") in regions:\n"); } else { REMAT_DEBUG(dbgs() << "Spilling w.r.t. minimum target occupancy (" << WavesPerEU.first << ") in regions:\n"); } for (unsigned I = 0, E = DAG.Regions.size(); I != E; ++I) { if (auto OptIt = OptRegions.find(I); OptIt != OptRegions.end()) REMAT_DEBUG(dbgs() << " [" << I << "] " << OptIt->getSecond() << '\n'); } #endif // When we are reducing spilling, the target is the minimum target number of // waves/EU determined by the subtarget. In cases where either one of // "amdgpu-num-sgpr" or "amdgpu-num-vgpr" are set on the function, the current // minimum region occupancy may be higher than the latter. TargetOcc = IncreaseOccupancy ? DAG.MinOccupancy + 1 : std::max(DAG.MinOccupancy, WavesPerEU.first); // Accounts for a reduction in RP in an optimizable region. Returns whether we // estimate that we have identified enough rematerialization opportunities to // achieve our goal, and sets Progress to true when this particular reduction // in pressure was helpful toward that goal. auto ReduceRPInRegion = [&](auto OptIt, Register Reg, LaneBitmask Mask, bool &Progress) -> bool { GCNRPTarget &Target = OptIt->getSecond(); if (!Target.isSaveBeneficial(Reg, DAG.MRI)) return false; Progress = true; Target.saveReg(Reg, Mask, DAG.MRI); if (Target.satisfied()) OptRegions.erase(OptIt->getFirst()); return OptRegions.empty(); }; // We need up-to-date live-out info. to query live-out register masks in // regions containing rematerializable instructions. DAG.RegionLiveOuts.buildLiveRegMap(); // Cache set of registers that are going to be rematerialized. DenseSet RematRegs; // Identify rematerializable instructions in the function. for (unsigned I = 0, E = DAG.Regions.size(); I != E; ++I) { auto Region = DAG.Regions[I]; for (auto MI = Region.first; MI != Region.second; ++MI) { // The instruction must be trivially rematerializable. MachineInstr &DefMI = *MI; if (!isTriviallyReMaterializable(DefMI)) continue; // We only support rematerializing virtual registers with one definition. Register Reg = DefMI.getOperand(0).getReg(); if (!Reg.isVirtual() || !DAG.MRI.hasOneDef(Reg)) continue; // We only care to rematerialize the instruction if it has a single // non-debug user in a different region. The using MI may not belong to a // region if it is a lone region terminator. MachineInstr *UseMI = DAG.MRI.getOneNonDBGUser(Reg); if (!UseMI) continue; auto UseRegion = MIRegion.find(UseMI); if (UseRegion != MIRegion.end() && UseRegion->second == I) continue; // Do not rematerialize an instruction if it uses or is used by an // instruction that we have designated for rematerialization. // FIXME: Allow for rematerialization chains: this requires 1. updating // remat points to account for uses that are rematerialized, and 2. either // rematerializing the candidates in careful ordering, or deferring the // MBB RP walk until the entire chain has been rematerialized. if (Rematerializations.contains(UseMI) || llvm::any_of(DefMI.operands(), [&RematRegs](MachineOperand &MO) { return MO.isReg() && RematRegs.contains(MO.getReg()); })) continue; // Do not rematerialize an instruction it it uses registers that aren't // available at its use. This ensures that we are not extending any live // range while rematerializing. SlotIndex DefIdx = DAG.LIS->getInstructionIndex(DefMI); SlotIndex UseIdx = DAG.LIS->getInstructionIndex(*UseMI).getRegSlot(true); if (!allUsesAvailableAt(&DefMI, DefIdx, UseIdx)) continue; REMAT_DEBUG(dbgs() << "Region " << I << ": remat instruction " << DefMI); RematInstruction &Remat = Rematerializations.try_emplace(&DefMI, UseMI).first->second; bool RematUseful = false; if (auto It = OptRegions.find(I); It != OptRegions.end()) { // Optimistically consider that moving the instruction out of its // defining region will reduce RP in the latter; this assumes that // maximum RP in the region is reached somewhere between the defining // instruction and the end of the region. REMAT_DEBUG(dbgs() << " Defining region is optimizable\n"); LaneBitmask Mask = DAG.RegionLiveOuts.getLiveRegsForRegionIdx(I)[Reg]; if (ReduceRPInRegion(It, Reg, Mask, RematUseful)) return true; } for (unsigned LIRegion = 0; LIRegion != E; ++LIRegion) { // We are only collecting regions in which the register is a live-in // (and may be live-through). auto It = DAG.LiveIns[LIRegion].find(Reg); if (It == DAG.LiveIns[LIRegion].end() || It->second.none()) continue; Remat.LiveInRegions.insert(LIRegion); // Account for the reduction in RP due to the rematerialization in an // optimizable region in which the defined register is a live-in. This // is exact for live-through region but optimistic in the using region, // where RP is actually reduced only if maximum RP is reached somewhere // between the beginning of the region and the rematerializable // instruction's use. if (auto It = OptRegions.find(LIRegion); It != OptRegions.end()) { REMAT_DEBUG(dbgs() << " Live-in in region " << LIRegion << '\n'); if (ReduceRPInRegion(It, Reg, DAG.LiveIns[LIRegion][Reg], RematUseful)) return true; } } // If the instruction is not a live-in or live-out in any optimizable // region then there is no point in rematerializing it. if (!RematUseful) { Rematerializations.pop_back(); REMAT_DEBUG(dbgs() << " No impact, not rematerializing instruction\n"); } else { RematRegs.insert(Reg); } } } if (IncreaseOccupancy) { // We were trying to increase occupancy but failed, abort the stage. REMAT_DEBUG(dbgs() << "Cannot increase occupancy\n"); Rematerializations.clear(); return false; } REMAT_DEBUG(dbgs() << "Can reduce but not eliminate spilling\n"); return !Rematerializations.empty(); } void PreRARematStage::rematerialize() { const SIInstrInfo *TII = MF.getSubtarget().getInstrInfo(); // Collect regions whose RP changes in unpredictable way; we will have to // fully recompute their RP after all rematerailizations. DenseSet RecomputeRP; // Rematerialize all instructions. for (auto &[DefMI, Remat] : Rematerializations) { MachineBasicBlock::iterator InsertPos(Remat.UseMI); Register Reg = DefMI->getOperand(0).getReg(); unsigned SubReg = DefMI->getOperand(0).getSubReg(); unsigned DefRegion = MIRegion.at(DefMI); // Rematerialize DefMI to its use block. TII->reMaterialize(*InsertPos->getParent(), InsertPos, Reg, SubReg, *DefMI, *DAG.TRI); Remat.RematMI = &*std::prev(InsertPos); Remat.RematMI->getOperand(0).setSubReg(SubReg); DAG.LIS->InsertMachineInstrInMaps(*Remat.RematMI); // Update region boundaries in regions we sinked from (remove defining MI) // and to (insert MI rematerialized in use block). Only then we can erase // the original MI. DAG.updateRegionBoundaries(DAG.Regions[DefRegion], DefMI, nullptr); auto UseRegion = MIRegion.find(Remat.UseMI); if (UseRegion != MIRegion.end()) { DAG.updateRegionBoundaries(DAG.Regions[UseRegion->second], InsertPos, Remat.RematMI); } DAG.LIS->RemoveMachineInstrFromMaps(*DefMI); DefMI->eraseFromParent(); // Collect all regions impacted by the rematerialization and update their // live-in/RP information. for (unsigned I : Remat.LiveInRegions) { ImpactedRegions.insert({I, DAG.Pressure[I]}); GCNRPTracker::LiveRegSet &RegionLiveIns = DAG.LiveIns[I]; #ifdef EXPENSIVE_CHECKS // All uses are known to be available / live at the remat point. Thus, the // uses should already be live in to the region. for (MachineOperand &MO : DefMI->operands()) { if (!MO.isReg() || !MO.getReg() || !MO.readsReg()) continue; Register UseReg = MO.getReg(); if (!UseReg.isVirtual()) continue; LiveInterval &LI = DAG.LIS->getInterval(UseReg); LaneBitmask LM = DAG.MRI.getMaxLaneMaskForVReg(MO.getReg()); if (LI.hasSubRanges() && MO.getSubReg()) LM = DAG.TRI->getSubRegIndexLaneMask(MO.getSubReg()); LaneBitmask LiveInMask = RegionLiveIns.at(UseReg); LaneBitmask UncoveredLanes = LM & ~(LiveInMask & LM); // If this register has lanes not covered by the LiveIns, be sure they // do not map to any subrange. ref: // machine-scheduler-sink-trivial-remats.mir::omitted_subrange if (UncoveredLanes.any()) { assert(LI.hasSubRanges()); for (LiveInterval::SubRange &SR : LI.subranges()) assert((SR.LaneMask & UncoveredLanes).none()); } } #endif // The register is no longer a live-in in all regions but the one that // contains the single use. In live-through regions, maximum register // pressure decreases predictably so we can directly update it. In the // using region, maximum RP may or may not decrease, so we will mark it // for re-computation after all materializations have taken place. LaneBitmask PrevMask = RegionLiveIns[Reg]; RegionLiveIns.erase(Reg); RegMasks.insert({{I, Remat.RematMI->getOperand(0).getReg()}, PrevMask}); if (Remat.UseMI->getParent() != DAG.Regions[I].first->getParent()) DAG.Pressure[I].inc(Reg, PrevMask, LaneBitmask::getNone(), DAG.MRI); else RecomputeRP.insert(I); } // RP in the region from which the instruction was rematerialized may or may // not decrease. ImpactedRegions.insert({DefRegion, DAG.Pressure[DefRegion]}); RecomputeRP.insert(DefRegion); // Recompute live interval to reflect the register's rematerialization. Register RematReg = Remat.RematMI->getOperand(0).getReg(); DAG.LIS->removeInterval(RematReg); DAG.LIS->createAndComputeVirtRegInterval(RematReg); } // All regions impacted by at least one rematerialization must be rescheduled. // Maximum pressure must also be recomputed for all regions where it changed // non-predictably and checked against the target occupancy. AchievedOcc = TargetOcc; for (auto &[I, OriginalRP] : ImpactedRegions) { bool IsEmptyRegion = DAG.Regions[I].first == DAG.Regions[I].second; RescheduleRegions[I] = !IsEmptyRegion; if (!RecomputeRP.contains(I)) continue; GCNRegPressure RP; if (IsEmptyRegion) { RP = getRegPressure(DAG.MRI, DAG.LiveIns[I]); } else { GCNDownwardRPTracker RPT(*DAG.LIS); auto *NonDbgMI = &*skipDebugInstructionsForward(DAG.Regions[I].first, DAG.Regions[I].second); if (NonDbgMI == DAG.Regions[I].second) { // Region is non-empty but contains only debug instructions. RP = getRegPressure(DAG.MRI, DAG.LiveIns[I]); } else { RPT.reset(*NonDbgMI, &DAG.LiveIns[I]); RPT.advance(DAG.Regions[I].second); RP = RPT.moveMaxPressure(); } } DAG.Pressure[I] = RP; AchievedOcc = std::min( AchievedOcc, RP.getOccupancy(ST, MF.getInfo() ->getDynamicVGPRBlockSize())); } REMAT_DEBUG(dbgs() << "Achieved occupancy " << AchievedOcc << "\n"); } // Copied from MachineLICM bool PreRARematStage::isTriviallyReMaterializable(const MachineInstr &MI) { if (!DAG.TII->isTriviallyReMaterializable(MI)) return false; for (const MachineOperand &MO : MI.all_uses()) { // We can't remat physreg uses, unless it is a constant or an ignorable // use (e.g. implicit exec use on VALU instructions) if (MO.getReg().isPhysical()) { if (DAG.MRI.isConstantPhysReg(MO.getReg()) || DAG.TII->isIgnorableUse(MO)) continue; return false; } } return true; } void PreRARematStage::finalizeGCNSchedStage() { // We consider that reducing spilling is always beneficial so we never // rollback rematerializations in such cases. It's also possible that // rescheduling lowers occupancy over the one achieved just through remats, in // which case we do not want to rollback either (the rescheduling was already // reverted in PreRARematStage::shouldRevertScheduling in such cases). unsigned MaxOcc = std::max(AchievedOcc, DAG.MinOccupancy); if (!IncreaseOccupancy || MaxOcc >= TargetOcc) return; REMAT_DEBUG(dbgs() << "Rolling back all rematerializations\n"); const SIInstrInfo *TII = MF.getSubtarget().getInstrInfo(); // Rollback the rematerializations. for (const auto &[DefMI, Remat] : Rematerializations) { MachineInstr &RematMI = *Remat.RematMI; unsigned DefRegion = MIRegion.at(DefMI); MachineBasicBlock::iterator InsertPos(DAG.Regions[DefRegion].second); MachineBasicBlock *MBB = RegionBB[DefRegion]; Register Reg = RematMI.getOperand(0).getReg(); unsigned SubReg = RematMI.getOperand(0).getSubReg(); // Re-rematerialize MI at the end of its original region. Note that it may // not be rematerialized exactly in the same position as originally within // the region, but it should not matter much. TII->reMaterialize(*MBB, InsertPos, Reg, SubReg, RematMI, *DAG.TRI); MachineInstr *NewMI = &*std::prev(InsertPos); NewMI->getOperand(0).setSubReg(SubReg); DAG.LIS->InsertMachineInstrInMaps(*NewMI); auto UseRegion = MIRegion.find(Remat.UseMI); if (UseRegion != MIRegion.end()) { DAG.updateRegionBoundaries(DAG.Regions[UseRegion->second], RematMI, nullptr); } DAG.updateRegionBoundaries(DAG.Regions[DefRegion], InsertPos, NewMI); // Erase rematerialized MI. DAG.LIS->RemoveMachineInstrFromMaps(RematMI); RematMI.eraseFromParent(); // Recompute live interval for the re-rematerialized register DAG.LIS->removeInterval(Reg); DAG.LIS->createAndComputeVirtRegInterval(Reg); // Re-add the register as a live-in in all regions it used to be one in. for (unsigned LIRegion : Remat.LiveInRegions) DAG.LiveIns[LIRegion].insert({Reg, RegMasks.at({LIRegion, Reg})}); } // Reset RP in all impacted regions. for (auto &[I, OriginalRP] : ImpactedRegions) DAG.Pressure[I] = OriginalRP; GCNSchedStage::finalizeGCNSchedStage(); } void GCNScheduleDAGMILive::updateRegionBoundaries( RegionBoundaries &RegionBounds, MachineBasicBlock::iterator MI, MachineInstr *NewMI) { assert((!NewMI || NewMI != RegionBounds.second) && "cannot remove at region end"); if (RegionBounds.first == RegionBounds.second) { assert(NewMI && "cannot remove from an empty region"); RegionBounds.first = NewMI; return; } // We only care for modifications at the beginning of a non-empty region since // the upper region boundary is exclusive. if (MI != RegionBounds.first) return; if (!NewMI) RegionBounds.first = std::next(MI); // Removal else RegionBounds.first = NewMI; // Insertion } static bool hasIGLPInstrs(ScheduleDAGInstrs *DAG) { const SIInstrInfo *SII = static_cast(DAG->TII); return any_of(*DAG, [SII](MachineBasicBlock::iterator MI) { return SII->isIGLPMutationOnly(MI->getOpcode()); }); } GCNPostScheduleDAGMILive::GCNPostScheduleDAGMILive( MachineSchedContext *C, std::unique_ptr S, bool RemoveKillFlags) : ScheduleDAGMI(C, std::move(S), RemoveKillFlags) {} void GCNPostScheduleDAGMILive::schedule() { HasIGLPInstrs = hasIGLPInstrs(this); if (HasIGLPInstrs) { SavedMutations.clear(); SavedMutations.swap(Mutations); addMutation(createIGroupLPDAGMutation(AMDGPU::SchedulingPhase::PostRA)); } ScheduleDAGMI::schedule(); } void GCNPostScheduleDAGMILive::finalizeSchedule() { if (HasIGLPInstrs) SavedMutations.swap(Mutations); ScheduleDAGMI::finalizeSchedule(); }