File: | lib/Target/AMDGPU/GCNSchedStrategy.cpp |
Warning: | line 531, column 7 Forming reference to null pointer |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===-- GCNSchedStrategy.cpp - GCN Scheduler Strategy ---------------------===// | |||
2 | // | |||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
4 | // See https://llvm.org/LICENSE.txt for license information. | |||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
6 | // | |||
7 | //===----------------------------------------------------------------------===// | |||
8 | // | |||
9 | /// \file | |||
10 | /// This contains a MachineSchedStrategy implementation for maximizing wave | |||
11 | /// occupancy on GCN hardware. | |||
12 | //===----------------------------------------------------------------------===// | |||
13 | ||||
14 | #include "GCNSchedStrategy.h" | |||
15 | #include "AMDGPUSubtarget.h" | |||
16 | #include "SIInstrInfo.h" | |||
17 | #include "SIMachineFunctionInfo.h" | |||
18 | #include "SIRegisterInfo.h" | |||
19 | #include "llvm/CodeGen/RegisterClassInfo.h" | |||
20 | #include "llvm/Support/MathExtras.h" | |||
21 | ||||
22 | #define DEBUG_TYPE"machine-scheduler" "machine-scheduler" | |||
23 | ||||
24 | using namespace llvm; | |||
25 | ||||
26 | GCNMaxOccupancySchedStrategy::GCNMaxOccupancySchedStrategy( | |||
27 | const MachineSchedContext *C) : | |||
28 | GenericScheduler(C), TargetOccupancy(0), MF(nullptr) { } | |||
29 | ||||
30 | void GCNMaxOccupancySchedStrategy::initialize(ScheduleDAGMI *DAG) { | |||
31 | GenericScheduler::initialize(DAG); | |||
32 | ||||
33 | const SIRegisterInfo *SRI = static_cast<const SIRegisterInfo*>(TRI); | |||
34 | ||||
35 | MF = &DAG->MF; | |||
36 | ||||
37 | const GCNSubtarget &ST = MF->getSubtarget<GCNSubtarget>(); | |||
38 | ||||
39 | // FIXME: This is also necessary, because some passes that run after | |||
40 | // scheduling and before regalloc increase register pressure. | |||
41 | const int ErrorMargin = 3; | |||
42 | ||||
43 | SGPRExcessLimit = Context->RegClassInfo | |||
44 | ->getNumAllocatableRegs(&AMDGPU::SGPR_32RegClass) - ErrorMargin; | |||
45 | VGPRExcessLimit = Context->RegClassInfo | |||
46 | ->getNumAllocatableRegs(&AMDGPU::VGPR_32RegClass) - ErrorMargin; | |||
47 | if (TargetOccupancy) { | |||
48 | SGPRCriticalLimit = ST.getMaxNumSGPRs(TargetOccupancy, true); | |||
49 | VGPRCriticalLimit = ST.getMaxNumVGPRs(TargetOccupancy); | |||
50 | } else { | |||
51 | SGPRCriticalLimit = SRI->getRegPressureSetLimit(DAG->MF, | |||
52 | SRI->getSGPRPressureSet()); | |||
53 | VGPRCriticalLimit = SRI->getRegPressureSetLimit(DAG->MF, | |||
54 | SRI->getVGPRPressureSet()); | |||
55 | } | |||
56 | ||||
57 | SGPRCriticalLimit -= ErrorMargin; | |||
58 | VGPRCriticalLimit -= ErrorMargin; | |||
59 | } | |||
60 | ||||
61 | void GCNMaxOccupancySchedStrategy::initCandidate(SchedCandidate &Cand, SUnit *SU, | |||
62 | bool AtTop, const RegPressureTracker &RPTracker, | |||
63 | const SIRegisterInfo *SRI, | |||
64 | unsigned SGPRPressure, | |||
65 | unsigned VGPRPressure) { | |||
66 | ||||
67 | Cand.SU = SU; | |||
68 | Cand.AtTop = AtTop; | |||
69 | ||||
70 | // getDownwardPressure() and getUpwardPressure() make temporary changes to | |||
71 | // the tracker, so we need to pass those function a non-const copy. | |||
72 | RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker); | |||
73 | ||||
74 | std::vector<unsigned> Pressure; | |||
75 | std::vector<unsigned> MaxPressure; | |||
76 | ||||
77 | if (AtTop) | |||
78 | TempTracker.getDownwardPressure(SU->getInstr(), Pressure, MaxPressure); | |||
79 | else { | |||
80 | // FIXME: I think for bottom up scheduling, the register pressure is cached | |||
81 | // and can be retrieved by DAG->getPressureDif(SU). | |||
82 | TempTracker.getUpwardPressure(SU->getInstr(), Pressure, MaxPressure); | |||
83 | } | |||
84 | ||||
85 | unsigned NewSGPRPressure = Pressure[SRI->getSGPRPressureSet()]; | |||
86 | unsigned NewVGPRPressure = Pressure[SRI->getVGPRPressureSet()]; | |||
87 | ||||
88 | // If two instructions increase the pressure of different register sets | |||
89 | // by the same amount, the generic scheduler will prefer to schedule the | |||
90 | // instruction that increases the set with the least amount of registers, | |||
91 | // which in our case would be SGPRs. This is rarely what we want, so | |||
92 | // when we report excess/critical register pressure, we do it either | |||
93 | // only for VGPRs or only for SGPRs. | |||
94 | ||||
95 | // FIXME: Better heuristics to determine whether to prefer SGPRs or VGPRs. | |||
96 | const unsigned MaxVGPRPressureInc = 16; | |||
97 | bool ShouldTrackVGPRs = VGPRPressure + MaxVGPRPressureInc >= VGPRExcessLimit; | |||
98 | bool ShouldTrackSGPRs = !ShouldTrackVGPRs && SGPRPressure >= SGPRExcessLimit; | |||
99 | ||||
100 | ||||
101 | // FIXME: We have to enter REG-EXCESS before we reach the actual threshold | |||
102 | // to increase the likelihood we don't go over the limits. We should improve | |||
103 | // the analysis to look through dependencies to find the path with the least | |||
104 | // register pressure. | |||
105 | ||||
106 | // We only need to update the RPDelata for instructions that increase | |||
107 | // register pressure. Instructions that decrease or keep reg pressure | |||
108 | // the same will be marked as RegExcess in tryCandidate() when they | |||
109 | // are compared with instructions that increase the register pressure. | |||
110 | if (ShouldTrackVGPRs && NewVGPRPressure >= VGPRExcessLimit) { | |||
111 | Cand.RPDelta.Excess = PressureChange(SRI->getVGPRPressureSet()); | |||
112 | Cand.RPDelta.Excess.setUnitInc(NewVGPRPressure - VGPRExcessLimit); | |||
113 | } | |||
114 | ||||
115 | if (ShouldTrackSGPRs && NewSGPRPressure >= SGPRExcessLimit) { | |||
116 | Cand.RPDelta.Excess = PressureChange(SRI->getSGPRPressureSet()); | |||
117 | Cand.RPDelta.Excess.setUnitInc(NewSGPRPressure - SGPRExcessLimit); | |||
118 | } | |||
119 | ||||
120 | // Register pressure is considered 'CRITICAL' if it is approaching a value | |||
121 | // that would reduce the wave occupancy for the execution unit. When | |||
122 | // register pressure is 'CRITICAL', increading SGPR and VGPR pressure both | |||
123 | // has the same cost, so we don't need to prefer one over the other. | |||
124 | ||||
125 | int SGPRDelta = NewSGPRPressure - SGPRCriticalLimit; | |||
126 | int VGPRDelta = NewVGPRPressure - VGPRCriticalLimit; | |||
127 | ||||
128 | if (SGPRDelta >= 0 || VGPRDelta >= 0) { | |||
129 | if (SGPRDelta > VGPRDelta) { | |||
130 | Cand.RPDelta.CriticalMax = PressureChange(SRI->getSGPRPressureSet()); | |||
131 | Cand.RPDelta.CriticalMax.setUnitInc(SGPRDelta); | |||
132 | } else { | |||
133 | Cand.RPDelta.CriticalMax = PressureChange(SRI->getVGPRPressureSet()); | |||
134 | Cand.RPDelta.CriticalMax.setUnitInc(VGPRDelta); | |||
135 | } | |||
136 | } | |||
137 | } | |||
138 | ||||
139 | // This function is mostly cut and pasted from | |||
140 | // GenericScheduler::pickNodeFromQueue() | |||
141 | void GCNMaxOccupancySchedStrategy::pickNodeFromQueue(SchedBoundary &Zone, | |||
142 | const CandPolicy &ZonePolicy, | |||
143 | const RegPressureTracker &RPTracker, | |||
144 | SchedCandidate &Cand) { | |||
145 | const SIRegisterInfo *SRI = static_cast<const SIRegisterInfo*>(TRI); | |||
146 | ArrayRef<unsigned> Pressure = RPTracker.getRegSetPressureAtPos(); | |||
147 | unsigned SGPRPressure = Pressure[SRI->getSGPRPressureSet()]; | |||
148 | unsigned VGPRPressure = Pressure[SRI->getVGPRPressureSet()]; | |||
149 | ReadyQueue &Q = Zone.Available; | |||
150 | for (SUnit *SU : Q) { | |||
151 | ||||
152 | SchedCandidate TryCand(ZonePolicy); | |||
153 | initCandidate(TryCand, SU, Zone.isTop(), RPTracker, SRI, | |||
154 | SGPRPressure, VGPRPressure); | |||
155 | // Pass SchedBoundary only when comparing nodes from the same boundary. | |||
156 | SchedBoundary *ZoneArg = Cand.AtTop == TryCand.AtTop ? &Zone : nullptr; | |||
157 | GenericScheduler::tryCandidate(Cand, TryCand, ZoneArg); | |||
158 | if (TryCand.Reason != NoCand) { | |||
159 | // Initialize resource delta if needed in case future heuristics query it. | |||
160 | if (TryCand.ResDelta == SchedResourceDelta()) | |||
161 | TryCand.initResourceDelta(Zone.DAG, SchedModel); | |||
162 | Cand.setBest(TryCand); | |||
163 | } | |||
164 | } | |||
165 | } | |||
166 | ||||
167 | // This function is mostly cut and pasted from | |||
168 | // GenericScheduler::pickNodeBidirectional() | |||
169 | SUnit *GCNMaxOccupancySchedStrategy::pickNodeBidirectional(bool &IsTopNode) { | |||
170 | // Schedule as far as possible in the direction of no choice. This is most | |||
171 | // efficient, but also provides the best heuristics for CriticalPSets. | |||
172 | if (SUnit *SU = Bot.pickOnlyChoice()) { | |||
173 | IsTopNode = false; | |||
174 | return SU; | |||
175 | } | |||
176 | if (SUnit *SU = Top.pickOnlyChoice()) { | |||
177 | IsTopNode = true; | |||
178 | return SU; | |||
179 | } | |||
180 | // Set the bottom-up policy based on the state of the current bottom zone and | |||
181 | // the instructions outside the zone, including the top zone. | |||
182 | CandPolicy BotPolicy; | |||
183 | setPolicy(BotPolicy, /*IsPostRA=*/false, Bot, &Top); | |||
184 | // Set the top-down policy based on the state of the current top zone and | |||
185 | // the instructions outside the zone, including the bottom zone. | |||
186 | CandPolicy TopPolicy; | |||
187 | setPolicy(TopPolicy, /*IsPostRA=*/false, Top, &Bot); | |||
188 | ||||
189 | // See if BotCand is still valid (because we previously scheduled from Top). | |||
190 | LLVM_DEBUG(dbgs() << "Picking from Bot:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Picking from Bot:\n" ; } } while (false); | |||
191 | if (!BotCand.isValid() || BotCand.SU->isScheduled || | |||
192 | BotCand.Policy != BotPolicy) { | |||
193 | BotCand.reset(CandPolicy()); | |||
194 | pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), BotCand); | |||
195 | assert(BotCand.Reason != NoCand && "failed to find the first candidate")((BotCand.Reason != NoCand && "failed to find the first candidate" ) ? static_cast<void> (0) : __assert_fail ("BotCand.Reason != NoCand && \"failed to find the first candidate\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/AMDGPU/GCNSchedStrategy.cpp" , 195, __PRETTY_FUNCTION__)); | |||
196 | } else { | |||
197 | LLVM_DEBUG(traceCandidate(BotCand))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { traceCandidate(BotCand); } } while ( false); | |||
198 | } | |||
199 | ||||
200 | // Check if the top Q has a better candidate. | |||
201 | LLVM_DEBUG(dbgs() << "Picking from Top:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Picking from Top:\n" ; } } while (false); | |||
202 | if (!TopCand.isValid() || TopCand.SU->isScheduled || | |||
203 | TopCand.Policy != TopPolicy) { | |||
204 | TopCand.reset(CandPolicy()); | |||
205 | pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TopCand); | |||
206 | assert(TopCand.Reason != NoCand && "failed to find the first candidate")((TopCand.Reason != NoCand && "failed to find the first candidate" ) ? static_cast<void> (0) : __assert_fail ("TopCand.Reason != NoCand && \"failed to find the first candidate\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/AMDGPU/GCNSchedStrategy.cpp" , 206, __PRETTY_FUNCTION__)); | |||
207 | } else { | |||
208 | LLVM_DEBUG(traceCandidate(TopCand))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { traceCandidate(TopCand); } } while ( false); | |||
209 | } | |||
210 | ||||
211 | // Pick best from BotCand and TopCand. | |||
212 | LLVM_DEBUG(dbgs() << "Top Cand: "; traceCandidate(TopCand);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Top Cand: "; traceCandidate (TopCand); dbgs() << "Bot Cand: "; traceCandidate(BotCand );; } } while (false) | |||
213 | dbgs() << "Bot Cand: "; traceCandidate(BotCand);)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Top Cand: "; traceCandidate (TopCand); dbgs() << "Bot Cand: "; traceCandidate(BotCand );; } } while (false); | |||
214 | SchedCandidate Cand; | |||
215 | if (TopCand.Reason == BotCand.Reason) { | |||
216 | Cand = BotCand; | |||
217 | GenericSchedulerBase::CandReason TopReason = TopCand.Reason; | |||
218 | TopCand.Reason = NoCand; | |||
219 | GenericScheduler::tryCandidate(Cand, TopCand, nullptr); | |||
220 | if (TopCand.Reason != NoCand) { | |||
221 | Cand.setBest(TopCand); | |||
222 | } else { | |||
223 | TopCand.Reason = TopReason; | |||
224 | } | |||
225 | } else { | |||
226 | if (TopCand.Reason == RegExcess && TopCand.RPDelta.Excess.getUnitInc() <= 0) { | |||
227 | Cand = TopCand; | |||
228 | } else if (BotCand.Reason == RegExcess && BotCand.RPDelta.Excess.getUnitInc() <= 0) { | |||
229 | Cand = BotCand; | |||
230 | } else if (TopCand.Reason == RegCritical && TopCand.RPDelta.CriticalMax.getUnitInc() <= 0) { | |||
231 | Cand = TopCand; | |||
232 | } else if (BotCand.Reason == RegCritical && BotCand.RPDelta.CriticalMax.getUnitInc() <= 0) { | |||
233 | Cand = BotCand; | |||
234 | } else { | |||
235 | if (BotCand.Reason > TopCand.Reason) { | |||
236 | Cand = TopCand; | |||
237 | } else { | |||
238 | Cand = BotCand; | |||
239 | } | |||
240 | } | |||
241 | } | |||
242 | LLVM_DEBUG(dbgs() << "Picking: "; traceCandidate(Cand);)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Picking: "; traceCandidate (Cand);; } } while (false); | |||
243 | ||||
244 | IsTopNode = Cand.AtTop; | |||
245 | return Cand.SU; | |||
246 | } | |||
247 | ||||
248 | // This function is mostly cut and pasted from | |||
249 | // GenericScheduler::pickNode() | |||
250 | SUnit *GCNMaxOccupancySchedStrategy::pickNode(bool &IsTopNode) { | |||
251 | if (DAG->top() == DAG->bottom()) { | |||
252 | assert(Top.Available.empty() && Top.Pending.empty() &&((Top.Available.empty() && Top.Pending.empty() && Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage") ? static_cast<void> (0) : __assert_fail ("Top.Available.empty() && Top.Pending.empty() && Bot.Available.empty() && Bot.Pending.empty() && \"ReadyQ garbage\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/AMDGPU/GCNSchedStrategy.cpp" , 253, __PRETTY_FUNCTION__)) | |||
253 | Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage")((Top.Available.empty() && Top.Pending.empty() && Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage") ? static_cast<void> (0) : __assert_fail ("Top.Available.empty() && Top.Pending.empty() && Bot.Available.empty() && Bot.Pending.empty() && \"ReadyQ garbage\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/AMDGPU/GCNSchedStrategy.cpp" , 253, __PRETTY_FUNCTION__)); | |||
254 | return nullptr; | |||
255 | } | |||
256 | SUnit *SU; | |||
257 | do { | |||
258 | if (RegionPolicy.OnlyTopDown) { | |||
259 | SU = Top.pickOnlyChoice(); | |||
260 | if (!SU) { | |||
261 | CandPolicy NoPolicy; | |||
262 | TopCand.reset(NoPolicy); | |||
263 | pickNodeFromQueue(Top, NoPolicy, DAG->getTopRPTracker(), TopCand); | |||
264 | assert(TopCand.Reason != NoCand && "failed to find a candidate")((TopCand.Reason != NoCand && "failed to find a candidate" ) ? static_cast<void> (0) : __assert_fail ("TopCand.Reason != NoCand && \"failed to find a candidate\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/AMDGPU/GCNSchedStrategy.cpp" , 264, __PRETTY_FUNCTION__)); | |||
265 | SU = TopCand.SU; | |||
266 | } | |||
267 | IsTopNode = true; | |||
268 | } else if (RegionPolicy.OnlyBottomUp) { | |||
269 | SU = Bot.pickOnlyChoice(); | |||
270 | if (!SU) { | |||
271 | CandPolicy NoPolicy; | |||
272 | BotCand.reset(NoPolicy); | |||
273 | pickNodeFromQueue(Bot, NoPolicy, DAG->getBotRPTracker(), BotCand); | |||
274 | assert(BotCand.Reason != NoCand && "failed to find a candidate")((BotCand.Reason != NoCand && "failed to find a candidate" ) ? static_cast<void> (0) : __assert_fail ("BotCand.Reason != NoCand && \"failed to find a candidate\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/AMDGPU/GCNSchedStrategy.cpp" , 274, __PRETTY_FUNCTION__)); | |||
275 | SU = BotCand.SU; | |||
276 | } | |||
277 | IsTopNode = false; | |||
278 | } else { | |||
279 | SU = pickNodeBidirectional(IsTopNode); | |||
280 | } | |||
281 | } while (SU->isScheduled); | |||
282 | ||||
283 | if (SU->isTopReady()) | |||
284 | Top.removeReady(SU); | |||
285 | if (SU->isBottomReady()) | |||
286 | Bot.removeReady(SU); | |||
287 | ||||
288 | LLVM_DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr(); } } while (false) | |||
289 | << *SU->getInstr())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr(); } } while (false); | |||
290 | return SU; | |||
291 | } | |||
292 | ||||
293 | GCNScheduleDAGMILive::GCNScheduleDAGMILive(MachineSchedContext *C, | |||
294 | std::unique_ptr<MachineSchedStrategy> S) : | |||
295 | ScheduleDAGMILive(C, std::move(S)), | |||
296 | ST(MF.getSubtarget<GCNSubtarget>()), | |||
297 | MFI(*MF.getInfo<SIMachineFunctionInfo>()), | |||
298 | StartingOccupancy(MFI.getOccupancy()), | |||
299 | MinOccupancy(StartingOccupancy), Stage(0), RegionIdx(0) { | |||
300 | ||||
301 | LLVM_DEBUG(dbgs() << "Starting occupancy is " << StartingOccupancy << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Starting occupancy is " << StartingOccupancy << ".\n"; } } while (false); | |||
302 | } | |||
303 | ||||
304 | void GCNScheduleDAGMILive::schedule() { | |||
305 | if (Stage == 0) { | |||
306 | // Just record regions at the first pass. | |||
307 | Regions.push_back(std::make_pair(RegionBegin, RegionEnd)); | |||
308 | return; | |||
309 | } | |||
310 | ||||
311 | std::vector<MachineInstr*> Unsched; | |||
312 | Unsched.reserve(NumRegionInstrs); | |||
313 | for (auto &I : *this) { | |||
314 | Unsched.push_back(&I); | |||
315 | } | |||
316 | ||||
317 | GCNRegPressure PressureBefore; | |||
318 | if (LIS) { | |||
319 | PressureBefore = Pressure[RegionIdx]; | |||
320 | ||||
321 | LLVM_DEBUG(dbgs() << "Pressure before scheduling:\nRegion live-ins:";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Pressure before scheduling:\nRegion live-ins:" ; GCNRPTracker::printLiveRegs(dbgs(), LiveIns[RegionIdx], MRI ); dbgs() << "Region live-in pressure: "; llvm::getRegPressure (MRI, LiveIns[RegionIdx]).print(dbgs()); dbgs() << "Region register pressure: " ; PressureBefore.print(dbgs()); } } while (false) | |||
322 | GCNRPTracker::printLiveRegs(dbgs(), LiveIns[RegionIdx], MRI);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Pressure before scheduling:\nRegion live-ins:" ; GCNRPTracker::printLiveRegs(dbgs(), LiveIns[RegionIdx], MRI ); dbgs() << "Region live-in pressure: "; llvm::getRegPressure (MRI, LiveIns[RegionIdx]).print(dbgs()); dbgs() << "Region register pressure: " ; PressureBefore.print(dbgs()); } } while (false) | |||
323 | dbgs() << "Region live-in pressure: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Pressure before scheduling:\nRegion live-ins:" ; GCNRPTracker::printLiveRegs(dbgs(), LiveIns[RegionIdx], MRI ); dbgs() << "Region live-in pressure: "; llvm::getRegPressure (MRI, LiveIns[RegionIdx]).print(dbgs()); dbgs() << "Region register pressure: " ; PressureBefore.print(dbgs()); } } while (false) | |||
324 | llvm::getRegPressure(MRI, LiveIns[RegionIdx]).print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Pressure before scheduling:\nRegion live-ins:" ; GCNRPTracker::printLiveRegs(dbgs(), LiveIns[RegionIdx], MRI ); dbgs() << "Region live-in pressure: "; llvm::getRegPressure (MRI, LiveIns[RegionIdx]).print(dbgs()); dbgs() << "Region register pressure: " ; PressureBefore.print(dbgs()); } } while (false) | |||
325 | dbgs() << "Region register pressure: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Pressure before scheduling:\nRegion live-ins:" ; GCNRPTracker::printLiveRegs(dbgs(), LiveIns[RegionIdx], MRI ); dbgs() << "Region live-in pressure: "; llvm::getRegPressure (MRI, LiveIns[RegionIdx]).print(dbgs()); dbgs() << "Region register pressure: " ; PressureBefore.print(dbgs()); } } while (false) | |||
326 | PressureBefore.print(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Pressure before scheduling:\nRegion live-ins:" ; GCNRPTracker::printLiveRegs(dbgs(), LiveIns[RegionIdx], MRI ); dbgs() << "Region live-in pressure: "; llvm::getRegPressure (MRI, LiveIns[RegionIdx]).print(dbgs()); dbgs() << "Region register pressure: " ; PressureBefore.print(dbgs()); } } while (false); | |||
327 | } | |||
328 | ||||
329 | ScheduleDAGMILive::schedule(); | |||
330 | Regions[RegionIdx] = std::make_pair(RegionBegin, RegionEnd); | |||
331 | ||||
332 | if (!LIS) | |||
333 | return; | |||
334 | ||||
335 | // Check the results of scheduling. | |||
336 | GCNMaxOccupancySchedStrategy &S = (GCNMaxOccupancySchedStrategy&)*SchedImpl; | |||
337 | auto PressureAfter = getRealRegPressure(); | |||
338 | ||||
339 | LLVM_DEBUG(dbgs() << "Pressure after scheduling: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Pressure after scheduling: " ; PressureAfter.print(dbgs()); } } while (false) | |||
340 | PressureAfter.print(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Pressure after scheduling: " ; PressureAfter.print(dbgs()); } } while (false); | |||
341 | ||||
342 | if (PressureAfter.getSGPRNum() <= S.SGPRCriticalLimit && | |||
343 | PressureAfter.getVGPRNum() <= S.VGPRCriticalLimit) { | |||
344 | Pressure[RegionIdx] = PressureAfter; | |||
345 | LLVM_DEBUG(dbgs() << "Pressure in desired limits, done.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Pressure in desired limits, done.\n" ; } } while (false); | |||
346 | return; | |||
347 | } | |||
348 | unsigned Occ = MFI.getOccupancy(); | |||
349 | unsigned WavesAfter = std::min(Occ, PressureAfter.getOccupancy(ST)); | |||
350 | unsigned WavesBefore = std::min(Occ, PressureBefore.getOccupancy(ST)); | |||
351 | LLVM_DEBUG(dbgs() << "Occupancy before scheduling: " << WavesBeforedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Occupancy before scheduling: " << WavesBefore << ", after " << WavesAfter << ".\n"; } } while (false) | |||
352 | << ", after " << WavesAfter << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Occupancy before scheduling: " << WavesBefore << ", after " << WavesAfter << ".\n"; } } while (false); | |||
353 | ||||
354 | // We could not keep current target occupancy because of the just scheduled | |||
355 | // region. Record new occupancy for next scheduling cycle. | |||
356 | unsigned NewOccupancy = std::max(WavesAfter, WavesBefore); | |||
357 | // Allow memory bound functions to drop to 4 waves if not limited by an | |||
358 | // attribute. | |||
359 | if (WavesAfter < WavesBefore && WavesAfter < MinOccupancy && | |||
360 | WavesAfter >= MFI.getMinAllowedOccupancy()) { | |||
361 | LLVM_DEBUG(dbgs() << "Function is memory bound, allow occupancy drop up to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Function is memory bound, allow occupancy drop up to " << MFI.getMinAllowedOccupancy() << " waves\n"; } } while (false) | |||
362 | << MFI.getMinAllowedOccupancy() << " waves\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Function is memory bound, allow occupancy drop up to " << MFI.getMinAllowedOccupancy() << " waves\n"; } } while (false); | |||
363 | NewOccupancy = WavesAfter; | |||
364 | } | |||
365 | if (NewOccupancy < MinOccupancy) { | |||
366 | MinOccupancy = NewOccupancy; | |||
367 | MFI.limitOccupancy(MinOccupancy); | |||
368 | LLVM_DEBUG(dbgs() << "Occupancy lowered for the function to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Occupancy lowered for the function to " << MinOccupancy << ".\n"; } } while (false) | |||
369 | << MinOccupancy << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Occupancy lowered for the function to " << MinOccupancy << ".\n"; } } while (false); | |||
370 | } | |||
371 | ||||
372 | if (WavesAfter >= MinOccupancy) { | |||
373 | Pressure[RegionIdx] = PressureAfter; | |||
374 | return; | |||
375 | } | |||
376 | ||||
377 | LLVM_DEBUG(dbgs() << "Attempting to revert scheduling.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Attempting to revert scheduling.\n" ; } } while (false); | |||
378 | RegionEnd = RegionBegin; | |||
379 | for (MachineInstr *MI : Unsched) { | |||
380 | if (MI->isDebugInstr()) | |||
381 | continue; | |||
382 | ||||
383 | if (MI->getIterator() != RegionEnd) { | |||
384 | BB->remove(MI); | |||
385 | BB->insert(RegionEnd, MI); | |||
386 | if (!MI->isDebugInstr()) | |||
387 | LIS->handleMove(*MI, true); | |||
388 | } | |||
389 | // Reset read-undef flags and update them later. | |||
390 | for (auto &Op : MI->operands()) | |||
391 | if (Op.isReg() && Op.isDef()) | |||
392 | Op.setIsUndef(false); | |||
393 | RegisterOperands RegOpers; | |||
394 | RegOpers.collect(*MI, *TRI, MRI, ShouldTrackLaneMasks, false); | |||
395 | if (!MI->isDebugInstr()) { | |||
396 | if (ShouldTrackLaneMasks) { | |||
397 | // Adjust liveness and add missing dead+read-undef flags. | |||
398 | SlotIndex SlotIdx = LIS->getInstructionIndex(*MI).getRegSlot(); | |||
399 | RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx, MI); | |||
400 | } else { | |||
401 | // Adjust for missing dead-def flags. | |||
402 | RegOpers.detectDeadDefs(*MI, *LIS); | |||
403 | } | |||
404 | } | |||
405 | RegionEnd = MI->getIterator(); | |||
406 | ++RegionEnd; | |||
407 | LLVM_DEBUG(dbgs() << "Scheduling " << *MI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Scheduling " << *MI; } } while (false); | |||
408 | } | |||
409 | RegionBegin = Unsched.front()->getIterator(); | |||
410 | Regions[RegionIdx] = std::make_pair(RegionBegin, RegionEnd); | |||
411 | ||||
412 | placeDebugValues(); | |||
413 | } | |||
414 | ||||
415 | GCNRegPressure GCNScheduleDAGMILive::getRealRegPressure() const { | |||
416 | GCNDownwardRPTracker RPTracker(*LIS); | |||
417 | RPTracker.advance(begin(), end(), &LiveIns[RegionIdx]); | |||
418 | return RPTracker.moveMaxPressure(); | |||
419 | } | |||
420 | ||||
421 | void GCNScheduleDAGMILive::computeBlockPressure(const MachineBasicBlock *MBB) { | |||
422 | GCNDownwardRPTracker RPTracker(*LIS); | |||
423 | ||||
424 | // If the block has the only successor then live-ins of that successor are | |||
425 | // live-outs of the current block. We can reuse calculated live set if the | |||
426 | // successor will be sent to scheduling past current block. | |||
427 | const MachineBasicBlock *OnlySucc = nullptr; | |||
428 | if (MBB->succ_size() == 1 && !(*MBB->succ_begin())->empty()) { | |||
429 | SlotIndexes *Ind = LIS->getSlotIndexes(); | |||
430 | if (Ind->getMBBStartIdx(MBB) < Ind->getMBBStartIdx(*MBB->succ_begin())) | |||
431 | OnlySucc = *MBB->succ_begin(); | |||
432 | } | |||
433 | ||||
434 | // Scheduler sends regions from the end of the block upwards. | |||
435 | size_t CurRegion = RegionIdx; | |||
436 | for (size_t E = Regions.size(); CurRegion != E; ++CurRegion) | |||
437 | if (Regions[CurRegion].first->getParent() != MBB) | |||
438 | break; | |||
439 | --CurRegion; | |||
440 | ||||
441 | auto I = MBB->begin(); | |||
442 | auto LiveInIt = MBBLiveIns.find(MBB); | |||
443 | if (LiveInIt != MBBLiveIns.end()) { | |||
444 | auto LiveIn = std::move(LiveInIt->second); | |||
445 | RPTracker.reset(*MBB->begin(), &LiveIn); | |||
446 | MBBLiveIns.erase(LiveInIt); | |||
447 | } else { | |||
448 | I = Regions[CurRegion].first; | |||
449 | RPTracker.reset(*I); | |||
450 | } | |||
451 | ||||
452 | for ( ; ; ) { | |||
453 | I = RPTracker.getNext(); | |||
454 | ||||
455 | if (Regions[CurRegion].first == I) { | |||
456 | LiveIns[CurRegion] = RPTracker.getLiveRegs(); | |||
457 | RPTracker.clearMaxPressure(); | |||
458 | } | |||
459 | ||||
460 | if (Regions[CurRegion].second == I) { | |||
461 | Pressure[CurRegion] = RPTracker.moveMaxPressure(); | |||
462 | if (CurRegion-- == RegionIdx) | |||
463 | break; | |||
464 | } | |||
465 | RPTracker.advanceToNext(); | |||
466 | RPTracker.advanceBeforeNext(); | |||
467 | } | |||
468 | ||||
469 | if (OnlySucc) { | |||
470 | if (I != MBB->end()) { | |||
471 | RPTracker.advanceToNext(); | |||
472 | RPTracker.advance(MBB->end()); | |||
473 | } | |||
474 | RPTracker.reset(*OnlySucc->begin(), &RPTracker.getLiveRegs()); | |||
475 | RPTracker.advanceBeforeNext(); | |||
476 | MBBLiveIns[OnlySucc] = RPTracker.moveLiveRegs(); | |||
477 | } | |||
478 | } | |||
479 | ||||
480 | void GCNScheduleDAGMILive::finalizeSchedule() { | |||
481 | GCNMaxOccupancySchedStrategy &S = (GCNMaxOccupancySchedStrategy&)*SchedImpl; | |||
482 | LLVM_DEBUG(dbgs() << "All regions recorded, starting actual scheduling.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "All regions recorded, starting actual scheduling.\n" ; } } while (false); | |||
| ||||
483 | ||||
484 | LiveIns.resize(Regions.size()); | |||
485 | Pressure.resize(Regions.size()); | |||
486 | ||||
487 | do { | |||
488 | Stage++; | |||
489 | RegionIdx = 0; | |||
490 | MachineBasicBlock *MBB = nullptr; | |||
491 | ||||
492 | if (Stage > 1) { | |||
493 | // Retry function scheduling if we found resulting occupancy and it is | |||
494 | // lower than used for first pass scheduling. This will give more freedom | |||
495 | // to schedule low register pressure blocks. | |||
496 | // Code is partially copied from MachineSchedulerBase::scheduleRegions(). | |||
497 | ||||
498 | if (!LIS || StartingOccupancy <= MinOccupancy) | |||
499 | break; | |||
500 | ||||
501 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Retrying function scheduling with lowest recorded occupancy " << MinOccupancy << ".\n"; } } while (false) | |||
502 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Retrying function scheduling with lowest recorded occupancy " << MinOccupancy << ".\n"; } } while (false) | |||
503 | << "Retrying function scheduling with lowest recorded occupancy "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Retrying function scheduling with lowest recorded occupancy " << MinOccupancy << ".\n"; } } while (false) | |||
504 | << MinOccupancy << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "Retrying function scheduling with lowest recorded occupancy " << MinOccupancy << ".\n"; } } while (false); | |||
505 | ||||
506 | S.setTargetOccupancy(MinOccupancy); | |||
507 | } | |||
508 | ||||
509 | for (auto Region : Regions) { | |||
510 | RegionBegin = Region.first; | |||
511 | RegionEnd = Region.second; | |||
512 | ||||
513 | if (RegionBegin->getParent() != MBB) { | |||
514 | if (MBB) finishBlock(); | |||
515 | MBB = RegionBegin->getParent(); | |||
516 | startBlock(MBB); | |||
517 | if (Stage == 1) | |||
518 | computeBlockPressure(MBB); | |||
519 | } | |||
520 | ||||
521 | unsigned NumRegionInstrs = std::distance(begin(), end()); | |||
522 | enterRegion(MBB, begin(), end(), NumRegionInstrs); | |||
523 | ||||
524 | // Skip empty scheduling regions (0 or 1 schedulable instructions). | |||
525 | if (begin() == end() || begin() == std::prev(end())) { | |||
526 | exitRegion(); | |||
527 | continue; | |||
528 | } | |||
529 | ||||
530 | LLVM_DEBUG(dbgs() << "********** MI Scheduling **********\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << "********** MI Scheduling **********\n" ; } } while (false); | |||
531 | LLVM_DEBUG(dbgs() << MF.getName() << ":" << printMBBReference(*MBB) << " "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << MF.getName() << ":" << printMBBReference(*MBB) << " " << MBB ->getName() << "\n From: " << *begin() << " To: "; if (RegionEnd != MBB->end()) dbgs() << * RegionEnd; else dbgs() << "End"; dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n'; } } while (false) | |||
| ||||
532 | << MBB->getName() << "\n From: " << *begin()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << MF.getName() << ":" << printMBBReference(*MBB) << " " << MBB ->getName() << "\n From: " << *begin() << " To: "; if (RegionEnd != MBB->end()) dbgs() << * RegionEnd; else dbgs() << "End"; dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n'; } } while (false) | |||
533 | << " To: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << MF.getName() << ":" << printMBBReference(*MBB) << " " << MBB ->getName() << "\n From: " << *begin() << " To: "; if (RegionEnd != MBB->end()) dbgs() << * RegionEnd; else dbgs() << "End"; dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n'; } } while (false) | |||
534 | if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << MF.getName() << ":" << printMBBReference(*MBB) << " " << MBB ->getName() << "\n From: " << *begin() << " To: "; if (RegionEnd != MBB->end()) dbgs() << * RegionEnd; else dbgs() << "End"; dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n'; } } while (false) | |||
535 | else dbgs() << "End";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << MF.getName() << ":" << printMBBReference(*MBB) << " " << MBB ->getName() << "\n From: " << *begin() << " To: "; if (RegionEnd != MBB->end()) dbgs() << * RegionEnd; else dbgs() << "End"; dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n'; } } while (false) | |||
536 | dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("machine-scheduler")) { dbgs() << MF.getName() << ":" << printMBBReference(*MBB) << " " << MBB ->getName() << "\n From: " << *begin() << " To: "; if (RegionEnd != MBB->end()) dbgs() << * RegionEnd; else dbgs() << "End"; dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n'; } } while (false); | |||
537 | ||||
538 | schedule(); | |||
539 | ||||
540 | exitRegion(); | |||
541 | ++RegionIdx; | |||
542 | } | |||
543 | finishBlock(); | |||
544 | ||||
545 | } while (Stage < 2); | |||
546 | } |