1 | //===- MachineScheduler.cpp - Machine Instruction Scheduler ---------------===// |
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 | // MachineScheduler schedules machine instructions after phi elimination. It |
10 | // preserves LiveIntervals so it can be invoked before register allocation. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "llvm/CodeGen/MachineScheduler.h" |
15 | #include "llvm/ADT/ArrayRef.h" |
16 | #include "llvm/ADT/BitVector.h" |
17 | #include "llvm/ADT/DenseMap.h" |
18 | #include "llvm/ADT/EquivalenceClasses.h" |
19 | #include "llvm/ADT/PriorityQueue.h" |
20 | #include "llvm/ADT/STLExtras.h" |
21 | #include "llvm/ADT/SmallVector.h" |
22 | #include "llvm/ADT/Statistic.h" |
23 | #include "llvm/ADT/iterator_range.h" |
24 | #include "llvm/Analysis/AliasAnalysis.h" |
25 | #include "llvm/CodeGen/LiveInterval.h" |
26 | #include "llvm/CodeGen/LiveIntervals.h" |
27 | #include "llvm/CodeGen/MachineBasicBlock.h" |
28 | #include "llvm/CodeGen/MachineDominators.h" |
29 | #include "llvm/CodeGen/MachineFunction.h" |
30 | #include "llvm/CodeGen/MachineFunctionPass.h" |
31 | #include "llvm/CodeGen/MachineInstr.h" |
32 | #include "llvm/CodeGen/MachineLoopInfo.h" |
33 | #include "llvm/CodeGen/MachineOperand.h" |
34 | #include "llvm/CodeGen/MachinePassRegistry.h" |
35 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
36 | #include "llvm/CodeGen/RegisterClassInfo.h" |
37 | #include "llvm/CodeGen/RegisterPressure.h" |
38 | #include "llvm/CodeGen/ScheduleDAG.h" |
39 | #include "llvm/CodeGen/ScheduleDAGInstrs.h" |
40 | #include "llvm/CodeGen/ScheduleDAGMutation.h" |
41 | #include "llvm/CodeGen/ScheduleDFS.h" |
42 | #include "llvm/CodeGen/ScheduleHazardRecognizer.h" |
43 | #include "llvm/CodeGen/SlotIndexes.h" |
44 | #include "llvm/CodeGen/TargetFrameLowering.h" |
45 | #include "llvm/CodeGen/TargetInstrInfo.h" |
46 | #include "llvm/CodeGen/TargetLowering.h" |
47 | #include "llvm/CodeGen/TargetPassConfig.h" |
48 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
49 | #include "llvm/CodeGen/TargetSchedule.h" |
50 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
51 | #include "llvm/CodeGenTypes/MachineValueType.h" |
52 | #include "llvm/Config/llvm-config.h" |
53 | #include "llvm/InitializePasses.h" |
54 | #include "llvm/MC/LaneBitmask.h" |
55 | #include "llvm/Pass.h" |
56 | #include "llvm/Support/CommandLine.h" |
57 | #include "llvm/Support/Compiler.h" |
58 | #include "llvm/Support/Debug.h" |
59 | #include "llvm/Support/ErrorHandling.h" |
60 | #include "llvm/Support/GraphWriter.h" |
61 | #include "llvm/Support/raw_ostream.h" |
62 | #include "llvm/Target/TargetMachine.h" |
63 | #include <algorithm> |
64 | #include <cassert> |
65 | #include <cstdint> |
66 | #include <iterator> |
67 | #include <limits> |
68 | #include <memory> |
69 | #include <string> |
70 | #include <tuple> |
71 | #include <utility> |
72 | #include <vector> |
73 | |
74 | using namespace llvm; |
75 | |
76 | #define DEBUG_TYPE "machine-scheduler" |
77 | |
78 | STATISTIC(NumInstrsInSourceOrderPreRA, |
79 | "Number of instructions in source order after pre-RA scheduling" ); |
80 | STATISTIC(NumInstrsInSourceOrderPostRA, |
81 | "Number of instructions in source order after post-RA scheduling" ); |
82 | STATISTIC(, |
83 | "Number of instructions scheduled by pre-RA scheduler" ); |
84 | STATISTIC(, |
85 | "Number of instructions scheduled by post-RA scheduler" ); |
86 | STATISTIC(NumClustered, "Number of load/store pairs clustered" ); |
87 | |
88 | STATISTIC(NumTopPreRA, |
89 | "Number of scheduling units chosen from top queue pre-RA" ); |
90 | STATISTIC(NumBotPreRA, |
91 | "Number of scheduling units chosen from bottom queue pre-RA" ); |
92 | STATISTIC(NumNoCandPreRA, |
93 | "Number of scheduling units chosen for NoCand heuristic pre-RA" ); |
94 | STATISTIC(NumOnly1PreRA, |
95 | "Number of scheduling units chosen for Only1 heuristic pre-RA" ); |
96 | STATISTIC(NumPhysRegPreRA, |
97 | "Number of scheduling units chosen for PhysReg heuristic pre-RA" ); |
98 | STATISTIC(NumRegExcessPreRA, |
99 | "Number of scheduling units chosen for RegExcess heuristic pre-RA" ); |
100 | STATISTIC(NumRegCriticalPreRA, |
101 | "Number of scheduling units chosen for RegCritical heuristic pre-RA" ); |
102 | STATISTIC(NumStallPreRA, |
103 | "Number of scheduling units chosen for Stall heuristic pre-RA" ); |
104 | STATISTIC(NumClusterPreRA, |
105 | "Number of scheduling units chosen for Cluster heuristic pre-RA" ); |
106 | STATISTIC(NumWeakPreRA, |
107 | "Number of scheduling units chosen for Weak heuristic pre-RA" ); |
108 | STATISTIC(NumRegMaxPreRA, |
109 | "Number of scheduling units chosen for RegMax heuristic pre-RA" ); |
110 | STATISTIC( |
111 | NumResourceReducePreRA, |
112 | "Number of scheduling units chosen for ResourceReduce heuristic pre-RA" ); |
113 | STATISTIC( |
114 | NumResourceDemandPreRA, |
115 | "Number of scheduling units chosen for ResourceDemand heuristic pre-RA" ); |
116 | STATISTIC( |
117 | NumTopDepthReducePreRA, |
118 | "Number of scheduling units chosen for TopDepthReduce heuristic pre-RA" ); |
119 | STATISTIC( |
120 | NumTopPathReducePreRA, |
121 | "Number of scheduling units chosen for TopPathReduce heuristic pre-RA" ); |
122 | STATISTIC( |
123 | NumBotHeightReducePreRA, |
124 | "Number of scheduling units chosen for BotHeightReduce heuristic pre-RA" ); |
125 | STATISTIC( |
126 | NumBotPathReducePreRA, |
127 | "Number of scheduling units chosen for BotPathReduce heuristic pre-RA" ); |
128 | STATISTIC(NumNodeOrderPreRA, |
129 | "Number of scheduling units chosen for NodeOrder heuristic pre-RA" ); |
130 | STATISTIC(NumFirstValidPreRA, |
131 | "Number of scheduling units chosen for FirstValid heuristic pre-RA" ); |
132 | |
133 | STATISTIC(NumTopPostRA, |
134 | "Number of scheduling units chosen from top queue post-RA" ); |
135 | STATISTIC(NumBotPostRA, |
136 | "Number of scheduling units chosen from bottom queue post-RA" ); |
137 | STATISTIC(NumNoCandPostRA, |
138 | "Number of scheduling units chosen for NoCand heuristic post-RA" ); |
139 | STATISTIC(NumOnly1PostRA, |
140 | "Number of scheduling units chosen for Only1 heuristic post-RA" ); |
141 | STATISTIC(NumPhysRegPostRA, |
142 | "Number of scheduling units chosen for PhysReg heuristic post-RA" ); |
143 | STATISTIC(NumRegExcessPostRA, |
144 | "Number of scheduling units chosen for RegExcess heuristic post-RA" ); |
145 | STATISTIC( |
146 | NumRegCriticalPostRA, |
147 | "Number of scheduling units chosen for RegCritical heuristic post-RA" ); |
148 | STATISTIC(NumStallPostRA, |
149 | "Number of scheduling units chosen for Stall heuristic post-RA" ); |
150 | STATISTIC(NumClusterPostRA, |
151 | "Number of scheduling units chosen for Cluster heuristic post-RA" ); |
152 | STATISTIC(NumWeakPostRA, |
153 | "Number of scheduling units chosen for Weak heuristic post-RA" ); |
154 | STATISTIC(NumRegMaxPostRA, |
155 | "Number of scheduling units chosen for RegMax heuristic post-RA" ); |
156 | STATISTIC( |
157 | NumResourceReducePostRA, |
158 | "Number of scheduling units chosen for ResourceReduce heuristic post-RA" ); |
159 | STATISTIC( |
160 | NumResourceDemandPostRA, |
161 | "Number of scheduling units chosen for ResourceDemand heuristic post-RA" ); |
162 | STATISTIC( |
163 | NumTopDepthReducePostRA, |
164 | "Number of scheduling units chosen for TopDepthReduce heuristic post-RA" ); |
165 | STATISTIC( |
166 | NumTopPathReducePostRA, |
167 | "Number of scheduling units chosen for TopPathReduce heuristic post-RA" ); |
168 | STATISTIC( |
169 | NumBotHeightReducePostRA, |
170 | "Number of scheduling units chosen for BotHeightReduce heuristic post-RA" ); |
171 | STATISTIC( |
172 | NumBotPathReducePostRA, |
173 | "Number of scheduling units chosen for BotPathReduce heuristic post-RA" ); |
174 | STATISTIC(NumNodeOrderPostRA, |
175 | "Number of scheduling units chosen for NodeOrder heuristic post-RA" ); |
176 | STATISTIC(NumFirstValidPostRA, |
177 | "Number of scheduling units chosen for FirstValid heuristic post-RA" ); |
178 | |
179 | namespace llvm { |
180 | |
181 | cl::opt<MISched::Direction> PreRADirection( |
182 | "misched-prera-direction" , cl::Hidden, |
183 | cl::desc("Pre reg-alloc list scheduling direction" ), |
184 | cl::init(Val: MISched::Unspecified), |
185 | cl::values( |
186 | clEnumValN(MISched::TopDown, "topdown" , |
187 | "Force top-down pre reg-alloc list scheduling" ), |
188 | clEnumValN(MISched::BottomUp, "bottomup" , |
189 | "Force bottom-up pre reg-alloc list scheduling" ), |
190 | clEnumValN(MISched::Bidirectional, "bidirectional" , |
191 | "Force bidirectional pre reg-alloc list scheduling" ))); |
192 | |
193 | static cl::opt<MISched::Direction> PostRADirection( |
194 | "misched-postra-direction" , cl::Hidden, |
195 | cl::desc("Post reg-alloc list scheduling direction" ), |
196 | cl::init(Val: MISched::Unspecified), |
197 | cl::values( |
198 | clEnumValN(MISched::TopDown, "topdown" , |
199 | "Force top-down post reg-alloc list scheduling" ), |
200 | clEnumValN(MISched::BottomUp, "bottomup" , |
201 | "Force bottom-up post reg-alloc list scheduling" ), |
202 | clEnumValN(MISched::Bidirectional, "bidirectional" , |
203 | "Force bidirectional post reg-alloc list scheduling" ))); |
204 | |
205 | static cl::opt<bool> |
206 | DumpCriticalPathLength("misched-dcpl" , cl::Hidden, |
207 | cl::desc("Print critical path length to stdout" )); |
208 | |
209 | cl::opt<bool> VerifyScheduling( |
210 | "verify-misched" , cl::Hidden, |
211 | cl::desc("Verify machine instrs before and after machine scheduling" )); |
212 | |
213 | #ifndef NDEBUG |
214 | cl::opt<bool> ViewMISchedDAGs( |
215 | "view-misched-dags" , cl::Hidden, |
216 | cl::desc("Pop up a window to show MISched dags after they are processed" )); |
217 | cl::opt<bool> PrintDAGs("misched-print-dags" , cl::Hidden, |
218 | cl::desc("Print schedule DAGs" )); |
219 | cl::opt<bool> MISchedDumpReservedCycles( |
220 | "misched-dump-reserved-cycles" , cl::Hidden, cl::init(false), |
221 | cl::desc("Dump resource usage at schedule boundary." )); |
222 | cl::opt<bool> MischedDetailResourceBooking( |
223 | "misched-detail-resource-booking" , cl::Hidden, cl::init(false), |
224 | cl::desc("Show details of invoking getNextResoufceCycle." )); |
225 | #else |
226 | const bool ViewMISchedDAGs = false; |
227 | const bool PrintDAGs = false; |
228 | const bool MischedDetailResourceBooking = false; |
229 | #ifdef LLVM_ENABLE_DUMP |
230 | const bool MISchedDumpReservedCycles = false; |
231 | #endif // LLVM_ENABLE_DUMP |
232 | #endif // NDEBUG |
233 | |
234 | } // end namespace llvm |
235 | |
236 | #ifndef NDEBUG |
237 | /// In some situations a few uninteresting nodes depend on nearly all other |
238 | /// nodes in the graph, provide a cutoff to hide them. |
239 | static cl::opt<unsigned> ViewMISchedCutoff("view-misched-cutoff" , cl::Hidden, |
240 | cl::desc("Hide nodes with more predecessor/successor than cutoff" )); |
241 | |
242 | static cl::opt<unsigned> MISchedCutoff("misched-cutoff" , cl::Hidden, |
243 | cl::desc("Stop scheduling after N instructions" ), cl::init(~0U)); |
244 | |
245 | static cl::opt<std::string> SchedOnlyFunc("misched-only-func" , cl::Hidden, |
246 | cl::desc("Only schedule this function" )); |
247 | static cl::opt<unsigned> SchedOnlyBlock("misched-only-block" , cl::Hidden, |
248 | cl::desc("Only schedule this MBB#" )); |
249 | #endif // NDEBUG |
250 | |
251 | /// Avoid quadratic complexity in unusually large basic blocks by limiting the |
252 | /// size of the ready lists. |
253 | static cl::opt<unsigned> ReadyListLimit("misched-limit" , cl::Hidden, |
254 | cl::desc("Limit ready list to N instructions" ), cl::init(Val: 256)); |
255 | |
256 | static cl::opt<bool> EnableRegPressure("misched-regpressure" , cl::Hidden, |
257 | cl::desc("Enable register pressure scheduling." ), cl::init(Val: true)); |
258 | |
259 | static cl::opt<bool> EnableCyclicPath("misched-cyclicpath" , cl::Hidden, |
260 | cl::desc("Enable cyclic critical path analysis." ), cl::init(Val: true)); |
261 | |
262 | static cl::opt<bool> EnableMemOpCluster("misched-cluster" , cl::Hidden, |
263 | cl::desc("Enable memop clustering." ), |
264 | cl::init(Val: true)); |
265 | static cl::opt<bool> |
266 | ForceFastCluster("force-fast-cluster" , cl::Hidden, |
267 | cl::desc("Switch to fast cluster algorithm with the lost " |
268 | "of some fusion opportunities" ), |
269 | cl::init(Val: false)); |
270 | static cl::opt<unsigned> |
271 | FastClusterThreshold("fast-cluster-threshold" , cl::Hidden, |
272 | cl::desc("The threshold for fast cluster" ), |
273 | cl::init(Val: 1000)); |
274 | |
275 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
276 | static cl::opt<bool> MISchedDumpScheduleTrace( |
277 | "misched-dump-schedule-trace" , cl::Hidden, cl::init(false), |
278 | cl::desc("Dump resource usage at schedule boundary." )); |
279 | static cl::opt<unsigned> |
280 | HeaderColWidth("misched-dump-schedule-trace-col-header-width" , cl::Hidden, |
281 | cl::desc("Set width of the columns with " |
282 | "the resources and schedule units" ), |
283 | cl::init(19)); |
284 | static cl::opt<unsigned> |
285 | ColWidth("misched-dump-schedule-trace-col-width" , cl::Hidden, |
286 | cl::desc("Set width of the columns showing resource booking." ), |
287 | cl::init(5)); |
288 | static cl::opt<bool> MISchedSortResourcesInTrace( |
289 | "misched-sort-resources-in-trace" , cl::Hidden, cl::init(true), |
290 | cl::desc("Sort the resources printed in the dump trace" )); |
291 | #endif |
292 | |
293 | static cl::opt<unsigned> |
294 | MIResourceCutOff("misched-resource-cutoff" , cl::Hidden, |
295 | cl::desc("Number of intervals to track" ), cl::init(Val: 10)); |
296 | |
297 | // DAG subtrees must have at least this many nodes. |
298 | static const unsigned MinSubtreeSize = 8; |
299 | |
300 | // Pin the vtables to this file. |
301 | void MachineSchedStrategy::anchor() {} |
302 | |
303 | void ScheduleDAGMutation::anchor() {} |
304 | |
305 | //===----------------------------------------------------------------------===// |
306 | // Machine Instruction Scheduling Pass and Registry |
307 | //===----------------------------------------------------------------------===// |
308 | |
309 | MachineSchedContext::MachineSchedContext() { |
310 | RegClassInfo = new RegisterClassInfo(); |
311 | } |
312 | |
313 | MachineSchedContext::~MachineSchedContext() { |
314 | delete RegClassInfo; |
315 | } |
316 | |
317 | namespace llvm { |
318 | namespace impl_detail { |
319 | |
320 | /// Base class for the machine scheduler classes. |
321 | class MachineSchedulerBase : public MachineSchedContext { |
322 | protected: |
323 | void scheduleRegions(ScheduleDAGInstrs &Scheduler, bool FixKillFlags); |
324 | }; |
325 | |
326 | /// Impl class for MachineScheduler. |
327 | class MachineSchedulerImpl : public MachineSchedulerBase { |
328 | // These are only for using MF.verify() |
329 | // remove when verify supports passing in all analyses |
330 | MachineFunctionPass *P = nullptr; |
331 | MachineFunctionAnalysisManager *MFAM = nullptr; |
332 | |
333 | public: |
334 | struct RequiredAnalyses { |
335 | MachineLoopInfo &MLI; |
336 | MachineDominatorTree &MDT; |
337 | AAResults &AA; |
338 | LiveIntervals &LIS; |
339 | }; |
340 | |
341 | MachineSchedulerImpl() {} |
342 | // Migration only |
343 | void setLegacyPass(MachineFunctionPass *P) { this->P = P; } |
344 | void setMFAM(MachineFunctionAnalysisManager *MFAM) { this->MFAM = MFAM; } |
345 | |
346 | bool run(MachineFunction &MF, const TargetMachine &TM, |
347 | const RequiredAnalyses &Analyses); |
348 | |
349 | protected: |
350 | ScheduleDAGInstrs *createMachineScheduler(); |
351 | }; |
352 | |
353 | /// Impl class for PostMachineScheduler. |
354 | class PostMachineSchedulerImpl : public MachineSchedulerBase { |
355 | // These are only for using MF.verify() |
356 | // remove when verify supports passing in all analyses |
357 | MachineFunctionPass *P = nullptr; |
358 | MachineFunctionAnalysisManager *MFAM = nullptr; |
359 | |
360 | public: |
361 | struct RequiredAnalyses { |
362 | MachineLoopInfo &MLI; |
363 | AAResults &AA; |
364 | }; |
365 | PostMachineSchedulerImpl() {} |
366 | // Migration only |
367 | void setLegacyPass(MachineFunctionPass *P) { this->P = P; } |
368 | void setMFAM(MachineFunctionAnalysisManager *MFAM) { this->MFAM = MFAM; } |
369 | |
370 | bool run(MachineFunction &Func, const TargetMachine &TM, |
371 | const RequiredAnalyses &Analyses); |
372 | |
373 | protected: |
374 | ScheduleDAGInstrs *createPostMachineScheduler(); |
375 | }; |
376 | |
377 | } // namespace impl_detail |
378 | } // namespace llvm |
379 | |
380 | using impl_detail::MachineSchedulerBase; |
381 | using impl_detail::MachineSchedulerImpl; |
382 | using impl_detail::PostMachineSchedulerImpl; |
383 | |
384 | namespace { |
385 | /// MachineScheduler runs after coalescing and before register allocation. |
386 | class MachineSchedulerLegacy : public MachineFunctionPass { |
387 | MachineSchedulerImpl Impl; |
388 | |
389 | public: |
390 | MachineSchedulerLegacy(); |
391 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
392 | bool runOnMachineFunction(MachineFunction&) override; |
393 | |
394 | static char ID; // Class identification, replacement for typeinfo |
395 | }; |
396 | |
397 | /// PostMachineScheduler runs after shortly before code emission. |
398 | class PostMachineSchedulerLegacy : public MachineFunctionPass { |
399 | PostMachineSchedulerImpl Impl; |
400 | |
401 | public: |
402 | PostMachineSchedulerLegacy(); |
403 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
404 | bool runOnMachineFunction(MachineFunction &) override; |
405 | |
406 | static char ID; // Class identification, replacement for typeinfo |
407 | }; |
408 | |
409 | } // end anonymous namespace |
410 | |
411 | char MachineSchedulerLegacy::ID = 0; |
412 | |
413 | char &llvm::MachineSchedulerID = MachineSchedulerLegacy::ID; |
414 | |
415 | INITIALIZE_PASS_BEGIN(MachineSchedulerLegacy, DEBUG_TYPE, |
416 | "Machine Instruction Scheduler" , false, false) |
417 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) |
418 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass) |
419 | INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass) |
420 | INITIALIZE_PASS_DEPENDENCY(SlotIndexesWrapperPass) |
421 | INITIALIZE_PASS_DEPENDENCY(LiveIntervalsWrapperPass) |
422 | INITIALIZE_PASS_END(MachineSchedulerLegacy, DEBUG_TYPE, |
423 | "Machine Instruction Scheduler" , false, false) |
424 | |
425 | MachineSchedulerLegacy::MachineSchedulerLegacy() : MachineFunctionPass(ID) { |
426 | initializeMachineSchedulerLegacyPass(Registry&: *PassRegistry::getPassRegistry()); |
427 | } |
428 | |
429 | void MachineSchedulerLegacy::getAnalysisUsage(AnalysisUsage &AU) const { |
430 | AU.setPreservesCFG(); |
431 | AU.addRequired<MachineDominatorTreeWrapperPass>(); |
432 | AU.addRequired<MachineLoopInfoWrapperPass>(); |
433 | AU.addRequired<AAResultsWrapperPass>(); |
434 | AU.addRequired<TargetPassConfig>(); |
435 | AU.addRequired<SlotIndexesWrapperPass>(); |
436 | AU.addPreserved<SlotIndexesWrapperPass>(); |
437 | AU.addRequired<LiveIntervalsWrapperPass>(); |
438 | AU.addPreserved<LiveIntervalsWrapperPass>(); |
439 | MachineFunctionPass::getAnalysisUsage(AU); |
440 | } |
441 | |
442 | char PostMachineSchedulerLegacy::ID = 0; |
443 | |
444 | char &llvm::PostMachineSchedulerID = PostMachineSchedulerLegacy::ID; |
445 | |
446 | INITIALIZE_PASS_BEGIN(PostMachineSchedulerLegacy, "postmisched" , |
447 | "PostRA Machine Instruction Scheduler" , false, false) |
448 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass) |
449 | INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass) |
450 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) |
451 | INITIALIZE_PASS_END(PostMachineSchedulerLegacy, "postmisched" , |
452 | "PostRA Machine Instruction Scheduler" , false, false) |
453 | |
454 | PostMachineSchedulerLegacy::PostMachineSchedulerLegacy() |
455 | : MachineFunctionPass(ID) { |
456 | initializePostMachineSchedulerLegacyPass(Registry&: *PassRegistry::getPassRegistry()); |
457 | } |
458 | |
459 | void PostMachineSchedulerLegacy::getAnalysisUsage(AnalysisUsage &AU) const { |
460 | AU.setPreservesCFG(); |
461 | AU.addRequired<MachineDominatorTreeWrapperPass>(); |
462 | AU.addRequired<MachineLoopInfoWrapperPass>(); |
463 | AU.addRequired<AAResultsWrapperPass>(); |
464 | AU.addRequired<TargetPassConfig>(); |
465 | MachineFunctionPass::getAnalysisUsage(AU); |
466 | } |
467 | |
468 | MachinePassRegistry<MachineSchedRegistry::ScheduleDAGCtor> |
469 | MachineSchedRegistry::Registry; |
470 | |
471 | /// A dummy default scheduler factory indicates whether the scheduler |
472 | /// is overridden on the command line. |
473 | static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) { |
474 | return nullptr; |
475 | } |
476 | |
477 | /// MachineSchedOpt allows command line selection of the scheduler. |
478 | static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false, |
479 | RegisterPassParser<MachineSchedRegistry>> |
480 | MachineSchedOpt("misched" , |
481 | cl::init(Val: &useDefaultMachineSched), cl::Hidden, |
482 | cl::desc("Machine instruction scheduler to use" )); |
483 | |
484 | static MachineSchedRegistry |
485 | DefaultSchedRegistry("default" , "Use the target's default scheduler choice." , |
486 | useDefaultMachineSched); |
487 | |
488 | static cl::opt<bool> EnableMachineSched( |
489 | "enable-misched" , |
490 | cl::desc("Enable the machine instruction scheduling pass." ), cl::init(Val: true), |
491 | cl::Hidden); |
492 | |
493 | static cl::opt<bool> EnablePostRAMachineSched( |
494 | "enable-post-misched" , |
495 | cl::desc("Enable the post-ra machine instruction scheduling pass." ), |
496 | cl::init(Val: true), cl::Hidden); |
497 | |
498 | /// Decrement this iterator until reaching the top or a non-debug instr. |
499 | static MachineBasicBlock::const_iterator |
500 | priorNonDebug(MachineBasicBlock::const_iterator I, |
501 | MachineBasicBlock::const_iterator Beg) { |
502 | assert(I != Beg && "reached the top of the region, cannot decrement" ); |
503 | while (--I != Beg) { |
504 | if (!I->isDebugOrPseudoInstr()) |
505 | break; |
506 | } |
507 | return I; |
508 | } |
509 | |
510 | /// Non-const version. |
511 | static MachineBasicBlock::iterator |
512 | priorNonDebug(MachineBasicBlock::iterator I, |
513 | MachineBasicBlock::const_iterator Beg) { |
514 | return priorNonDebug(I: MachineBasicBlock::const_iterator(I), Beg) |
515 | .getNonConstIterator(); |
516 | } |
517 | |
518 | /// If this iterator is a debug value, increment until reaching the End or a |
519 | /// non-debug instruction. |
520 | static MachineBasicBlock::const_iterator |
521 | nextIfDebug(MachineBasicBlock::const_iterator I, |
522 | MachineBasicBlock::const_iterator End) { |
523 | for(; I != End; ++I) { |
524 | if (!I->isDebugOrPseudoInstr()) |
525 | break; |
526 | } |
527 | return I; |
528 | } |
529 | |
530 | /// Non-const version. |
531 | static MachineBasicBlock::iterator |
532 | nextIfDebug(MachineBasicBlock::iterator I, |
533 | MachineBasicBlock::const_iterator End) { |
534 | return nextIfDebug(I: MachineBasicBlock::const_iterator(I), End) |
535 | .getNonConstIterator(); |
536 | } |
537 | |
538 | /// Instantiate a ScheduleDAGInstrs that will be owned by the caller. |
539 | ScheduleDAGInstrs *MachineSchedulerImpl::createMachineScheduler() { |
540 | // Select the scheduler, or set the default. |
541 | MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt; |
542 | if (Ctor != useDefaultMachineSched) |
543 | return Ctor(this); |
544 | |
545 | // Get the default scheduler set by the target for this function. |
546 | ScheduleDAGInstrs *Scheduler = TM->createMachineScheduler(C: this); |
547 | if (Scheduler) |
548 | return Scheduler; |
549 | |
550 | // Default to GenericScheduler. |
551 | return createSchedLive(C: this); |
552 | } |
553 | |
554 | bool MachineSchedulerImpl::run(MachineFunction &Func, const TargetMachine &TM, |
555 | const RequiredAnalyses &Analyses) { |
556 | MF = &Func; |
557 | MLI = &Analyses.MLI; |
558 | MDT = &Analyses.MDT; |
559 | this->TM = &TM; |
560 | AA = &Analyses.AA; |
561 | LIS = &Analyses.LIS; |
562 | |
563 | if (VerifyScheduling) { |
564 | LLVM_DEBUG(LIS->dump()); |
565 | const char *MSchedBanner = "Before machine scheduling." ; |
566 | if (P) |
567 | MF->verify(p: P, Banner: MSchedBanner, OS: &errs()); |
568 | else |
569 | MF->verify(MFAM&: *MFAM, Banner: MSchedBanner, OS: &errs()); |
570 | } |
571 | RegClassInfo->runOnMachineFunction(MF: *MF); |
572 | |
573 | // Instantiate the selected scheduler for this target, function, and |
574 | // optimization level. |
575 | std::unique_ptr<ScheduleDAGInstrs> Scheduler(createMachineScheduler()); |
576 | scheduleRegions(Scheduler&: *Scheduler, FixKillFlags: false); |
577 | |
578 | LLVM_DEBUG(LIS->dump()); |
579 | if (VerifyScheduling) { |
580 | const char *MSchedBanner = "After machine scheduling." ; |
581 | if (P) |
582 | MF->verify(p: P, Banner: MSchedBanner, OS: &errs()); |
583 | else |
584 | MF->verify(MFAM&: *MFAM, Banner: MSchedBanner, OS: &errs()); |
585 | } |
586 | return true; |
587 | } |
588 | |
589 | /// Instantiate a ScheduleDAGInstrs for PostRA scheduling that will be owned by |
590 | /// the caller. We don't have a command line option to override the postRA |
591 | /// scheduler. The Target must configure it. |
592 | ScheduleDAGInstrs *PostMachineSchedulerImpl::createPostMachineScheduler() { |
593 | // Get the postRA scheduler set by the target for this function. |
594 | ScheduleDAGInstrs *Scheduler = TM->createPostMachineScheduler(C: this); |
595 | if (Scheduler) |
596 | return Scheduler; |
597 | |
598 | // Default to GenericScheduler. |
599 | return createSchedPostRA(C: this); |
600 | } |
601 | |
602 | bool PostMachineSchedulerImpl::run(MachineFunction &Func, |
603 | const TargetMachine &TM, |
604 | const RequiredAnalyses &Analyses) { |
605 | MF = &Func; |
606 | MLI = &Analyses.MLI; |
607 | this->TM = &TM; |
608 | AA = &Analyses.AA; |
609 | |
610 | if (VerifyScheduling) { |
611 | const char *PostMSchedBanner = "Before post machine scheduling." ; |
612 | if (P) |
613 | MF->verify(p: P, Banner: PostMSchedBanner, OS: &errs()); |
614 | else |
615 | MF->verify(MFAM&: *MFAM, Banner: PostMSchedBanner, OS: &errs()); |
616 | } |
617 | |
618 | // Instantiate the selected scheduler for this target, function, and |
619 | // optimization level. |
620 | std::unique_ptr<ScheduleDAGInstrs> Scheduler(createPostMachineScheduler()); |
621 | scheduleRegions(Scheduler&: *Scheduler, FixKillFlags: true); |
622 | |
623 | if (VerifyScheduling) { |
624 | const char *PostMSchedBanner = "After post machine scheduling." ; |
625 | if (P) |
626 | MF->verify(p: P, Banner: PostMSchedBanner, OS: &errs()); |
627 | else |
628 | MF->verify(MFAM&: *MFAM, Banner: PostMSchedBanner, OS: &errs()); |
629 | } |
630 | return true; |
631 | } |
632 | |
633 | /// Top-level MachineScheduler pass driver. |
634 | /// |
635 | /// Visit blocks in function order. Divide each block into scheduling regions |
636 | /// and visit them bottom-up. Visiting regions bottom-up is not required, but is |
637 | /// consistent with the DAG builder, which traverses the interior of the |
638 | /// scheduling regions bottom-up. |
639 | /// |
640 | /// This design avoids exposing scheduling boundaries to the DAG builder, |
641 | /// simplifying the DAG builder's support for "special" target instructions. |
642 | /// At the same time the design allows target schedulers to operate across |
643 | /// scheduling boundaries, for example to bundle the boundary instructions |
644 | /// without reordering them. This creates complexity, because the target |
645 | /// scheduler must update the RegionBegin and RegionEnd positions cached by |
646 | /// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler |
647 | /// design would be to split blocks at scheduling boundaries, but LLVM has a |
648 | /// general bias against block splitting purely for implementation simplicity. |
649 | bool MachineSchedulerLegacy::runOnMachineFunction(MachineFunction &MF) { |
650 | if (skipFunction(F: MF.getFunction())) |
651 | return false; |
652 | |
653 | if (EnableMachineSched.getNumOccurrences()) { |
654 | if (!EnableMachineSched) |
655 | return false; |
656 | } else if (!MF.getSubtarget().enableMachineScheduler()) { |
657 | return false; |
658 | } |
659 | |
660 | LLVM_DEBUG(dbgs() << "Before MISched:\n" ; MF.print(dbgs())); |
661 | |
662 | auto &MLI = getAnalysis<MachineLoopInfoWrapperPass>().getLI(); |
663 | auto &MDT = getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree(); |
664 | auto &TM = getAnalysis<TargetPassConfig>().getTM<TargetMachine>(); |
665 | auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); |
666 | auto &LIS = getAnalysis<LiveIntervalsWrapperPass>().getLIS(); |
667 | Impl.setLegacyPass(this); |
668 | return Impl.run(Func&: MF, TM, Analyses: {.MLI: MLI, .MDT: MDT, .AA: AA, .LIS: LIS}); |
669 | } |
670 | |
671 | MachineSchedulerPass::MachineSchedulerPass(const TargetMachine *TM) |
672 | : Impl(std::make_unique<MachineSchedulerImpl>()), TM(TM) {} |
673 | MachineSchedulerPass::~MachineSchedulerPass() = default; |
674 | MachineSchedulerPass::MachineSchedulerPass(MachineSchedulerPass &&Other) = |
675 | default; |
676 | |
677 | PostMachineSchedulerPass::PostMachineSchedulerPass(const TargetMachine *TM) |
678 | : Impl(std::make_unique<PostMachineSchedulerImpl>()), TM(TM) {} |
679 | PostMachineSchedulerPass::PostMachineSchedulerPass( |
680 | PostMachineSchedulerPass &&Other) = default; |
681 | PostMachineSchedulerPass::~PostMachineSchedulerPass() = default; |
682 | |
683 | PreservedAnalyses |
684 | MachineSchedulerPass::run(MachineFunction &MF, |
685 | MachineFunctionAnalysisManager &MFAM) { |
686 | if (EnableMachineSched.getNumOccurrences()) { |
687 | if (!EnableMachineSched) |
688 | return PreservedAnalyses::all(); |
689 | } else if (!MF.getSubtarget().enableMachineScheduler()) { |
690 | return PreservedAnalyses::all(); |
691 | } |
692 | |
693 | LLVM_DEBUG(dbgs() << "Before MISched:\n" ; MF.print(dbgs())); |
694 | auto &MLI = MFAM.getResult<MachineLoopAnalysis>(IR&: MF); |
695 | auto &MDT = MFAM.getResult<MachineDominatorTreeAnalysis>(IR&: MF); |
696 | auto &FAM = MFAM.getResult<FunctionAnalysisManagerMachineFunctionProxy>(IR&: MF) |
697 | .getManager(); |
698 | auto &AA = FAM.getResult<AAManager>(IR&: MF.getFunction()); |
699 | auto &LIS = MFAM.getResult<LiveIntervalsAnalysis>(IR&: MF); |
700 | Impl->setMFAM(&MFAM); |
701 | bool Changed = Impl->run(Func&: MF, TM: *TM, Analyses: {.MLI: MLI, .MDT: MDT, .AA: AA, .LIS: LIS}); |
702 | if (!Changed) |
703 | return PreservedAnalyses::all(); |
704 | |
705 | return getMachineFunctionPassPreservedAnalyses() |
706 | .preserveSet<CFGAnalyses>() |
707 | .preserve<SlotIndexesAnalysis>() |
708 | .preserve<LiveIntervalsAnalysis>(); |
709 | } |
710 | |
711 | bool PostMachineSchedulerLegacy::runOnMachineFunction(MachineFunction &MF) { |
712 | if (skipFunction(F: MF.getFunction())) |
713 | return false; |
714 | |
715 | if (EnablePostRAMachineSched.getNumOccurrences()) { |
716 | if (!EnablePostRAMachineSched) |
717 | return false; |
718 | } else if (!MF.getSubtarget().enablePostRAMachineScheduler()) { |
719 | LLVM_DEBUG(dbgs() << "Subtarget disables post-MI-sched.\n" ); |
720 | return false; |
721 | } |
722 | LLVM_DEBUG(dbgs() << "Before post-MI-sched:\n" ; MF.print(dbgs())); |
723 | auto &MLI = getAnalysis<MachineLoopInfoWrapperPass>().getLI(); |
724 | auto &TM = getAnalysis<TargetPassConfig>().getTM<TargetMachine>(); |
725 | auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); |
726 | Impl.setLegacyPass(this); |
727 | return Impl.run(Func&: MF, TM, Analyses: {.MLI: MLI, .AA: AA}); |
728 | } |
729 | |
730 | PreservedAnalyses |
731 | PostMachineSchedulerPass::run(MachineFunction &MF, |
732 | MachineFunctionAnalysisManager &MFAM) { |
733 | if (EnablePostRAMachineSched.getNumOccurrences()) { |
734 | if (!EnablePostRAMachineSched) |
735 | return PreservedAnalyses::all(); |
736 | } else if (!MF.getSubtarget().enablePostRAMachineScheduler()) { |
737 | LLVM_DEBUG(dbgs() << "Subtarget disables post-MI-sched.\n" ); |
738 | return PreservedAnalyses::all(); |
739 | } |
740 | LLVM_DEBUG(dbgs() << "Before post-MI-sched:\n" ; MF.print(dbgs())); |
741 | auto &MLI = MFAM.getResult<MachineLoopAnalysis>(IR&: MF); |
742 | auto &FAM = MFAM.getResult<FunctionAnalysisManagerMachineFunctionProxy>(IR&: MF) |
743 | .getManager(); |
744 | auto &AA = FAM.getResult<AAManager>(IR&: MF.getFunction()); |
745 | |
746 | Impl->setMFAM(&MFAM); |
747 | bool Changed = Impl->run(Func&: MF, TM: *TM, Analyses: {.MLI: MLI, .AA: AA}); |
748 | if (!Changed) |
749 | return PreservedAnalyses::all(); |
750 | |
751 | PreservedAnalyses PA = getMachineFunctionPassPreservedAnalyses(); |
752 | PA.preserveSet<CFGAnalyses>(); |
753 | return PA; |
754 | } |
755 | |
756 | /// Return true of the given instruction should not be included in a scheduling |
757 | /// region. |
758 | /// |
759 | /// MachineScheduler does not currently support scheduling across calls. To |
760 | /// handle calls, the DAG builder needs to be modified to create register |
761 | /// anti/output dependencies on the registers clobbered by the call's regmask |
762 | /// operand. In PreRA scheduling, the stack pointer adjustment already prevents |
763 | /// scheduling across calls. In PostRA scheduling, we need the isCall to enforce |
764 | /// the boundary, but there would be no benefit to postRA scheduling across |
765 | /// calls this late anyway. |
766 | static bool isSchedBoundary(MachineBasicBlock::iterator MI, |
767 | MachineBasicBlock *MBB, |
768 | MachineFunction *MF, |
769 | const TargetInstrInfo *TII) { |
770 | return MI->isCall() || TII->isSchedulingBoundary(MI: *MI, MBB, MF: *MF) || |
771 | MI->isFakeUse(); |
772 | } |
773 | |
774 | /// A region of an MBB for scheduling. |
775 | namespace { |
776 | struct SchedRegion { |
777 | /// RegionBegin is the first instruction in the scheduling region, and |
778 | /// RegionEnd is either MBB->end() or the scheduling boundary after the |
779 | /// last instruction in the scheduling region. These iterators cannot refer |
780 | /// to instructions outside of the identified scheduling region because |
781 | /// those may be reordered before scheduling this region. |
782 | MachineBasicBlock::iterator RegionBegin; |
783 | MachineBasicBlock::iterator RegionEnd; |
784 | unsigned NumRegionInstrs; |
785 | |
786 | SchedRegion(MachineBasicBlock::iterator B, MachineBasicBlock::iterator E, |
787 | unsigned N) : |
788 | RegionBegin(B), RegionEnd(E), NumRegionInstrs(N) {} |
789 | }; |
790 | } // end anonymous namespace |
791 | |
792 | using MBBRegionsVector = SmallVector<SchedRegion, 16>; |
793 | |
794 | static void |
795 | getSchedRegions(MachineBasicBlock *MBB, |
796 | MBBRegionsVector &Regions, |
797 | bool RegionsTopDown) { |
798 | MachineFunction *MF = MBB->getParent(); |
799 | const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo(); |
800 | |
801 | MachineBasicBlock::iterator I = nullptr; |
802 | for(MachineBasicBlock::iterator RegionEnd = MBB->end(); |
803 | RegionEnd != MBB->begin(); RegionEnd = I) { |
804 | |
805 | // Avoid decrementing RegionEnd for blocks with no terminator. |
806 | if (RegionEnd != MBB->end() || |
807 | isSchedBoundary(MI: &*std::prev(x: RegionEnd), MBB: &*MBB, MF, TII)) { |
808 | --RegionEnd; |
809 | } |
810 | |
811 | // The next region starts above the previous region. Look backward in the |
812 | // instruction stream until we find the nearest boundary. |
813 | unsigned NumRegionInstrs = 0; |
814 | I = RegionEnd; |
815 | for (;I != MBB->begin(); --I) { |
816 | MachineInstr &MI = *std::prev(x: I); |
817 | if (isSchedBoundary(MI: &MI, MBB: &*MBB, MF, TII)) |
818 | break; |
819 | if (!MI.isDebugOrPseudoInstr()) { |
820 | // MBB::size() uses instr_iterator to count. Here we need a bundle to |
821 | // count as a single instruction. |
822 | ++NumRegionInstrs; |
823 | } |
824 | } |
825 | |
826 | // It's possible we found a scheduling region that only has debug |
827 | // instructions. Don't bother scheduling these. |
828 | if (NumRegionInstrs != 0) |
829 | Regions.push_back(Elt: SchedRegion(I, RegionEnd, NumRegionInstrs)); |
830 | } |
831 | |
832 | if (RegionsTopDown) |
833 | std::reverse(first: Regions.begin(), last: Regions.end()); |
834 | } |
835 | |
836 | /// Main driver for both MachineScheduler and PostMachineScheduler. |
837 | void MachineSchedulerBase::scheduleRegions(ScheduleDAGInstrs &Scheduler, |
838 | bool FixKillFlags) { |
839 | // Visit all machine basic blocks. |
840 | // |
841 | // TODO: Visit blocks in global postorder or postorder within the bottom-up |
842 | // loop tree. Then we can optionally compute global RegPressure. |
843 | for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end(); |
844 | MBB != MBBEnd; ++MBB) { |
845 | |
846 | Scheduler.startBlock(BB: &*MBB); |
847 | |
848 | #ifndef NDEBUG |
849 | if (SchedOnlyFunc.getNumOccurrences() && SchedOnlyFunc != MF->getName()) |
850 | continue; |
851 | if (SchedOnlyBlock.getNumOccurrences() |
852 | && (int)SchedOnlyBlock != MBB->getNumber()) |
853 | continue; |
854 | #endif |
855 | |
856 | // Break the block into scheduling regions [I, RegionEnd). RegionEnd |
857 | // points to the scheduling boundary at the bottom of the region. The DAG |
858 | // does not include RegionEnd, but the region does (i.e. the next |
859 | // RegionEnd is above the previous RegionBegin). If the current block has |
860 | // no terminator then RegionEnd == MBB->end() for the bottom region. |
861 | // |
862 | // All the regions of MBB are first found and stored in MBBRegions, which |
863 | // will be processed (MBB) top-down if initialized with true. |
864 | // |
865 | // The Scheduler may insert instructions during either schedule() or |
866 | // exitRegion(), even for empty regions. So the local iterators 'I' and |
867 | // 'RegionEnd' are invalid across these calls. Instructions must not be |
868 | // added to other regions than the current one without updating MBBRegions. |
869 | |
870 | MBBRegionsVector MBBRegions; |
871 | getSchedRegions(MBB: &*MBB, Regions&: MBBRegions, RegionsTopDown: Scheduler.doMBBSchedRegionsTopDown()); |
872 | bool ScheduleSingleMI = Scheduler.shouldScheduleSingleMIRegions(); |
873 | for (const SchedRegion &R : MBBRegions) { |
874 | MachineBasicBlock::iterator I = R.RegionBegin; |
875 | MachineBasicBlock::iterator RegionEnd = R.RegionEnd; |
876 | unsigned NumRegionInstrs = R.NumRegionInstrs; |
877 | |
878 | // Notify the scheduler of the region, even if we may skip scheduling |
879 | // it. Perhaps it still needs to be bundled. |
880 | Scheduler.enterRegion(bb: &*MBB, begin: I, end: RegionEnd, regioninstrs: NumRegionInstrs); |
881 | |
882 | // Skip empty scheduling regions and, conditionally, regions with a single |
883 | // MI. |
884 | if (I == RegionEnd || (!ScheduleSingleMI && I == std::prev(x: RegionEnd))) { |
885 | // Close the current region. Bundle the terminator if needed. |
886 | // This invalidates 'RegionEnd' and 'I'. |
887 | Scheduler.exitRegion(); |
888 | continue; |
889 | } |
890 | LLVM_DEBUG(dbgs() << "********** MI Scheduling **********\n" ); |
891 | LLVM_DEBUG(dbgs() << MF->getName() << ":" << printMBBReference(*MBB) |
892 | << " " << MBB->getName() << "\n From: " << *I |
893 | << " To: " ; |
894 | if (RegionEnd != MBB->end()) dbgs() << *RegionEnd; |
895 | else dbgs() << "End\n" ; |
896 | dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n'); |
897 | if (DumpCriticalPathLength) { |
898 | errs() << MF->getName(); |
899 | errs() << ":%bb. " << MBB->getNumber(); |
900 | errs() << " " << MBB->getName() << " \n" ; |
901 | } |
902 | |
903 | // Schedule a region: possibly reorder instructions. |
904 | // This invalidates the original region iterators. |
905 | Scheduler.schedule(); |
906 | |
907 | // Close the current region. |
908 | Scheduler.exitRegion(); |
909 | } |
910 | Scheduler.finishBlock(); |
911 | // FIXME: Ideally, no further passes should rely on kill flags. However, |
912 | // thumb2 size reduction is currently an exception, so the PostMIScheduler |
913 | // needs to do this. |
914 | if (FixKillFlags) |
915 | Scheduler.fixupKills(MBB&: *MBB); |
916 | } |
917 | Scheduler.finalizeSchedule(); |
918 | } |
919 | |
920 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
921 | LLVM_DUMP_METHOD void ReadyQueue::dump() const { |
922 | dbgs() << "Queue " << Name << ": " ; |
923 | for (const SUnit *SU : Queue) |
924 | dbgs() << SU->NodeNum << " " ; |
925 | dbgs() << "\n" ; |
926 | } |
927 | #endif |
928 | |
929 | //===----------------------------------------------------------------------===// |
930 | // ScheduleDAGMI - Basic machine instruction scheduling. This is |
931 | // independent of PreRA/PostRA scheduling and involves no extra book-keeping for |
932 | // virtual registers. |
933 | // ===----------------------------------------------------------------------===/ |
934 | |
935 | // Provide a vtable anchor. |
936 | ScheduleDAGMI::~ScheduleDAGMI() = default; |
937 | |
938 | /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When |
939 | /// NumPredsLeft reaches zero, release the successor node. |
940 | /// |
941 | /// FIXME: Adjust SuccSU height based on MinLatency. |
942 | void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) { |
943 | SUnit *SuccSU = SuccEdge->getSUnit(); |
944 | |
945 | if (SuccEdge->isWeak()) { |
946 | --SuccSU->WeakPredsLeft; |
947 | return; |
948 | } |
949 | #ifndef NDEBUG |
950 | if (SuccSU->NumPredsLeft == 0) { |
951 | dbgs() << "*** Scheduling failed! ***\n" ; |
952 | dumpNode(*SuccSU); |
953 | dbgs() << " has been released too many times!\n" ; |
954 | llvm_unreachable(nullptr); |
955 | } |
956 | #endif |
957 | // SU->TopReadyCycle was set to CurrCycle when it was scheduled. However, |
958 | // CurrCycle may have advanced since then. |
959 | if (SuccSU->TopReadyCycle < SU->TopReadyCycle + SuccEdge->getLatency()) |
960 | SuccSU->TopReadyCycle = SU->TopReadyCycle + SuccEdge->getLatency(); |
961 | |
962 | --SuccSU->NumPredsLeft; |
963 | if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU) |
964 | SchedImpl->releaseTopNode(SU: SuccSU); |
965 | } |
966 | |
967 | /// releaseSuccessors - Call releaseSucc on each of SU's successors. |
968 | void ScheduleDAGMI::releaseSuccessors(SUnit *SU) { |
969 | for (SDep &Succ : SU->Succs) |
970 | releaseSucc(SU, SuccEdge: &Succ); |
971 | } |
972 | |
973 | /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When |
974 | /// NumSuccsLeft reaches zero, release the predecessor node. |
975 | /// |
976 | /// FIXME: Adjust PredSU height based on MinLatency. |
977 | void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) { |
978 | SUnit *PredSU = PredEdge->getSUnit(); |
979 | |
980 | if (PredEdge->isWeak()) { |
981 | --PredSU->WeakSuccsLeft; |
982 | return; |
983 | } |
984 | #ifndef NDEBUG |
985 | if (PredSU->NumSuccsLeft == 0) { |
986 | dbgs() << "*** Scheduling failed! ***\n" ; |
987 | dumpNode(*PredSU); |
988 | dbgs() << " has been released too many times!\n" ; |
989 | llvm_unreachable(nullptr); |
990 | } |
991 | #endif |
992 | // SU->BotReadyCycle was set to CurrCycle when it was scheduled. However, |
993 | // CurrCycle may have advanced since then. |
994 | if (PredSU->BotReadyCycle < SU->BotReadyCycle + PredEdge->getLatency()) |
995 | PredSU->BotReadyCycle = SU->BotReadyCycle + PredEdge->getLatency(); |
996 | |
997 | --PredSU->NumSuccsLeft; |
998 | if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU) |
999 | SchedImpl->releaseBottomNode(SU: PredSU); |
1000 | } |
1001 | |
1002 | /// releasePredecessors - Call releasePred on each of SU's predecessors. |
1003 | void ScheduleDAGMI::releasePredecessors(SUnit *SU) { |
1004 | for (SDep &Pred : SU->Preds) |
1005 | releasePred(SU, PredEdge: &Pred); |
1006 | } |
1007 | |
1008 | void ScheduleDAGMI::startBlock(MachineBasicBlock *bb) { |
1009 | ScheduleDAGInstrs::startBlock(BB: bb); |
1010 | SchedImpl->enterMBB(MBB: bb); |
1011 | } |
1012 | |
1013 | void ScheduleDAGMI::finishBlock() { |
1014 | SchedImpl->leaveMBB(); |
1015 | ScheduleDAGInstrs::finishBlock(); |
1016 | } |
1017 | |
1018 | /// enterRegion - Called back from PostMachineScheduler::runOnMachineFunction |
1019 | /// after crossing a scheduling boundary. [begin, end) includes all instructions |
1020 | /// in the region, including the boundary itself and single-instruction regions |
1021 | /// that don't get scheduled. |
1022 | void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb, |
1023 | MachineBasicBlock::iterator begin, |
1024 | MachineBasicBlock::iterator end, |
1025 | unsigned regioninstrs) |
1026 | { |
1027 | ScheduleDAGInstrs::enterRegion(bb, begin, end, regioninstrs); |
1028 | |
1029 | SchedImpl->initPolicy(Begin: begin, End: end, NumRegionInstrs: regioninstrs); |
1030 | |
1031 | // Set dump direction after initializing sched policy. |
1032 | ScheduleDAGMI::DumpDirection D; |
1033 | if (SchedImpl->getPolicy().OnlyTopDown) |
1034 | D = ScheduleDAGMI::DumpDirection::TopDown; |
1035 | else if (SchedImpl->getPolicy().OnlyBottomUp) |
1036 | D = ScheduleDAGMI::DumpDirection::BottomUp; |
1037 | else |
1038 | D = ScheduleDAGMI::DumpDirection::Bidirectional; |
1039 | setDumpDirection(D); |
1040 | } |
1041 | |
1042 | /// This is normally called from the main scheduler loop but may also be invoked |
1043 | /// by the scheduling strategy to perform additional code motion. |
1044 | void ScheduleDAGMI::moveInstruction( |
1045 | MachineInstr *MI, MachineBasicBlock::iterator InsertPos) { |
1046 | // Advance RegionBegin if the first instruction moves down. |
1047 | if (&*RegionBegin == MI) |
1048 | ++RegionBegin; |
1049 | |
1050 | // Update the instruction stream. |
1051 | BB->splice(Where: InsertPos, Other: BB, From: MI); |
1052 | |
1053 | // Update LiveIntervals |
1054 | if (LIS) |
1055 | LIS->handleMove(MI&: *MI, /*UpdateFlags=*/true); |
1056 | |
1057 | // Recede RegionBegin if an instruction moves above the first. |
1058 | if (RegionBegin == InsertPos) |
1059 | RegionBegin = MI; |
1060 | } |
1061 | |
1062 | bool ScheduleDAGMI::checkSchedLimit() { |
1063 | #if LLVM_ENABLE_ABI_BREAKING_CHECKS && !defined(NDEBUG) |
1064 | if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) { |
1065 | CurrentTop = CurrentBottom; |
1066 | return false; |
1067 | } |
1068 | ++NumInstrsScheduled; |
1069 | #endif |
1070 | return true; |
1071 | } |
1072 | |
1073 | /// Per-region scheduling driver, called back from |
1074 | /// PostMachineScheduler::runOnMachineFunction. This is a simplified driver |
1075 | /// that does not consider liveness or register pressure. It is useful for |
1076 | /// PostRA scheduling and potentially other custom schedulers. |
1077 | void ScheduleDAGMI::schedule() { |
1078 | LLVM_DEBUG(dbgs() << "ScheduleDAGMI::schedule starting\n" ); |
1079 | LLVM_DEBUG(SchedImpl->dumpPolicy()); |
1080 | |
1081 | // Build the DAG. |
1082 | buildSchedGraph(AA); |
1083 | |
1084 | postProcessDAG(); |
1085 | |
1086 | SmallVector<SUnit*, 8> TopRoots, BotRoots; |
1087 | findRootsAndBiasEdges(TopRoots, BotRoots); |
1088 | |
1089 | LLVM_DEBUG(dump()); |
1090 | if (PrintDAGs) dump(); |
1091 | if (ViewMISchedDAGs) viewGraph(); |
1092 | |
1093 | // Initialize the strategy before modifying the DAG. |
1094 | // This may initialize a DFSResult to be used for queue priority. |
1095 | SchedImpl->initialize(DAG: this); |
1096 | |
1097 | // Initialize ready queues now that the DAG and priority data are finalized. |
1098 | initQueues(TopRoots, BotRoots); |
1099 | |
1100 | bool IsTopNode = false; |
1101 | while (true) { |
1102 | if (!checkSchedLimit()) |
1103 | break; |
1104 | |
1105 | LLVM_DEBUG(dbgs() << "** ScheduleDAGMI::schedule picking next node\n" ); |
1106 | SUnit *SU = SchedImpl->pickNode(IsTopNode); |
1107 | if (!SU) break; |
1108 | |
1109 | assert(!SU->isScheduled && "Node already scheduled" ); |
1110 | |
1111 | MachineInstr *MI = SU->getInstr(); |
1112 | if (IsTopNode) { |
1113 | assert(SU->isTopReady() && "node still has unscheduled dependencies" ); |
1114 | if (&*CurrentTop == MI) |
1115 | CurrentTop = nextIfDebug(I: ++CurrentTop, End: CurrentBottom); |
1116 | else |
1117 | moveInstruction(MI, InsertPos: CurrentTop); |
1118 | } else { |
1119 | assert(SU->isBottomReady() && "node still has unscheduled dependencies" ); |
1120 | MachineBasicBlock::iterator priorII = |
1121 | priorNonDebug(I: CurrentBottom, Beg: CurrentTop); |
1122 | if (&*priorII == MI) |
1123 | CurrentBottom = priorII; |
1124 | else { |
1125 | if (&*CurrentTop == MI) |
1126 | CurrentTop = nextIfDebug(I: ++CurrentTop, End: priorII); |
1127 | moveInstruction(MI, InsertPos: CurrentBottom); |
1128 | CurrentBottom = MI; |
1129 | } |
1130 | } |
1131 | // Notify the scheduling strategy before updating the DAG. |
1132 | // This sets the scheduled node's ReadyCycle to CurrCycle. When updateQueues |
1133 | // runs, it can then use the accurate ReadyCycle time to determine whether |
1134 | // newly released nodes can move to the readyQ. |
1135 | SchedImpl->schedNode(SU, IsTopNode); |
1136 | |
1137 | updateQueues(SU, IsTopNode); |
1138 | } |
1139 | assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone." ); |
1140 | |
1141 | placeDebugValues(); |
1142 | |
1143 | LLVM_DEBUG({ |
1144 | dbgs() << "*** Final schedule for " |
1145 | << printMBBReference(*begin()->getParent()) << " ***\n" ; |
1146 | dumpSchedule(); |
1147 | dbgs() << '\n'; |
1148 | }); |
1149 | } |
1150 | |
1151 | /// Apply each ScheduleDAGMutation step in order. |
1152 | void ScheduleDAGMI::postProcessDAG() { |
1153 | for (auto &m : Mutations) |
1154 | m->apply(DAG: this); |
1155 | } |
1156 | |
1157 | void ScheduleDAGMI:: |
1158 | findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots, |
1159 | SmallVectorImpl<SUnit*> &BotRoots) { |
1160 | for (SUnit &SU : SUnits) { |
1161 | assert(!SU.isBoundaryNode() && "Boundary node should not be in SUnits" ); |
1162 | |
1163 | // Order predecessors so DFSResult follows the critical path. |
1164 | SU.biasCriticalPath(); |
1165 | |
1166 | // A SUnit is ready to top schedule if it has no predecessors. |
1167 | if (!SU.NumPredsLeft) |
1168 | TopRoots.push_back(Elt: &SU); |
1169 | // A SUnit is ready to bottom schedule if it has no successors. |
1170 | if (!SU.NumSuccsLeft) |
1171 | BotRoots.push_back(Elt: &SU); |
1172 | } |
1173 | ExitSU.biasCriticalPath(); |
1174 | } |
1175 | |
1176 | /// Identify DAG roots and setup scheduler queues. |
1177 | void ScheduleDAGMI::initQueues(ArrayRef<SUnit *> TopRoots, |
1178 | ArrayRef<SUnit *> BotRoots) { |
1179 | // Release all DAG roots for scheduling, not including EntrySU/ExitSU. |
1180 | // |
1181 | // Nodes with unreleased weak edges can still be roots. |
1182 | // Release top roots in forward order. |
1183 | for (SUnit *SU : TopRoots) |
1184 | SchedImpl->releaseTopNode(SU); |
1185 | |
1186 | // Release bottom roots in reverse order so the higher priority nodes appear |
1187 | // first. This is more natural and slightly more efficient. |
1188 | for (SmallVectorImpl<SUnit*>::const_reverse_iterator |
1189 | I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I) { |
1190 | SchedImpl->releaseBottomNode(SU: *I); |
1191 | } |
1192 | |
1193 | releaseSuccessors(SU: &EntrySU); |
1194 | releasePredecessors(SU: &ExitSU); |
1195 | |
1196 | SchedImpl->registerRoots(); |
1197 | |
1198 | // Advance past initial DebugValues. |
1199 | CurrentTop = nextIfDebug(I: RegionBegin, End: RegionEnd); |
1200 | CurrentBottom = RegionEnd; |
1201 | } |
1202 | |
1203 | /// Update scheduler queues after scheduling an instruction. |
1204 | void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) { |
1205 | // Release dependent instructions for scheduling. |
1206 | if (IsTopNode) |
1207 | releaseSuccessors(SU); |
1208 | else |
1209 | releasePredecessors(SU); |
1210 | |
1211 | SU->isScheduled = true; |
1212 | } |
1213 | |
1214 | /// Reinsert any remaining debug_values, just like the PostRA scheduler. |
1215 | void ScheduleDAGMI::placeDebugValues() { |
1216 | // If first instruction was a DBG_VALUE then put it back. |
1217 | if (FirstDbgValue) { |
1218 | BB->splice(Where: RegionBegin, Other: BB, From: FirstDbgValue); |
1219 | RegionBegin = FirstDbgValue; |
1220 | } |
1221 | |
1222 | for (std::vector<std::pair<MachineInstr *, MachineInstr *>>::iterator |
1223 | DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) { |
1224 | std::pair<MachineInstr *, MachineInstr *> P = *std::prev(x: DI); |
1225 | MachineInstr *DbgValue = P.first; |
1226 | MachineBasicBlock::iterator OrigPrevMI = P.second; |
1227 | if (&*RegionBegin == DbgValue) |
1228 | ++RegionBegin; |
1229 | BB->splice(Where: std::next(x: OrigPrevMI), Other: BB, From: DbgValue); |
1230 | if (RegionEnd != BB->end() && OrigPrevMI == &*RegionEnd) |
1231 | RegionEnd = DbgValue; |
1232 | } |
1233 | } |
1234 | |
1235 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
1236 | static const char *scheduleTableLegend = " i: issue\n x: resource booked" ; |
1237 | |
1238 | LLVM_DUMP_METHOD void ScheduleDAGMI::dumpScheduleTraceTopDown() const { |
1239 | // Bail off when there is no schedule model to query. |
1240 | if (!SchedModel.hasInstrSchedModel()) |
1241 | return; |
1242 | |
1243 | // Nothing to show if there is no or just one instruction. |
1244 | if (BB->size() < 2) |
1245 | return; |
1246 | |
1247 | dbgs() << " * Schedule table (TopDown):\n" ; |
1248 | dbgs() << scheduleTableLegend << "\n" ; |
1249 | const unsigned FirstCycle = getSUnit(&*(std::begin(*this)))->TopReadyCycle; |
1250 | unsigned LastCycle = getSUnit(&*(std::prev(std::end(*this))))->TopReadyCycle; |
1251 | for (MachineInstr &MI : *this) { |
1252 | SUnit *SU = getSUnit(&MI); |
1253 | if (!SU) |
1254 | continue; |
1255 | const MCSchedClassDesc *SC = getSchedClass(SU); |
1256 | for (TargetSchedModel::ProcResIter PI = SchedModel.getWriteProcResBegin(SC), |
1257 | PE = SchedModel.getWriteProcResEnd(SC); |
1258 | PI != PE; ++PI) { |
1259 | if (SU->TopReadyCycle + PI->ReleaseAtCycle - 1 > LastCycle) |
1260 | LastCycle = SU->TopReadyCycle + PI->ReleaseAtCycle - 1; |
1261 | } |
1262 | } |
1263 | // Print the header with the cycles |
1264 | dbgs() << llvm::left_justify("Cycle" , HeaderColWidth); |
1265 | for (unsigned C = FirstCycle; C <= LastCycle; ++C) |
1266 | dbgs() << llvm::left_justify("| " + std::to_string(C), ColWidth); |
1267 | dbgs() << "|\n" ; |
1268 | |
1269 | for (MachineInstr &MI : *this) { |
1270 | SUnit *SU = getSUnit(&MI); |
1271 | if (!SU) { |
1272 | dbgs() << "Missing SUnit\n" ; |
1273 | continue; |
1274 | } |
1275 | std::string NodeName("SU(" ); |
1276 | NodeName += std::to_string(SU->NodeNum) + ")" ; |
1277 | dbgs() << llvm::left_justify(NodeName, HeaderColWidth); |
1278 | unsigned C = FirstCycle; |
1279 | for (; C <= LastCycle; ++C) { |
1280 | if (C == SU->TopReadyCycle) |
1281 | dbgs() << llvm::left_justify("| i" , ColWidth); |
1282 | else |
1283 | dbgs() << llvm::left_justify("|" , ColWidth); |
1284 | } |
1285 | dbgs() << "|\n" ; |
1286 | const MCSchedClassDesc *SC = getSchedClass(SU); |
1287 | |
1288 | SmallVector<MCWriteProcResEntry, 4> ResourcesIt( |
1289 | make_range(SchedModel.getWriteProcResBegin(SC), |
1290 | SchedModel.getWriteProcResEnd(SC))); |
1291 | |
1292 | if (MISchedSortResourcesInTrace) |
1293 | llvm::stable_sort( |
1294 | ResourcesIt, |
1295 | [](const MCWriteProcResEntry &LHS, |
1296 | const MCWriteProcResEntry &RHS) -> bool { |
1297 | return std::tie(LHS.AcquireAtCycle, LHS.ReleaseAtCycle) < |
1298 | std::tie(RHS.AcquireAtCycle, RHS.ReleaseAtCycle); |
1299 | }); |
1300 | for (const MCWriteProcResEntry &PI : ResourcesIt) { |
1301 | C = FirstCycle; |
1302 | const std::string ResName = |
1303 | SchedModel.getResourceName(PI.ProcResourceIdx); |
1304 | dbgs() << llvm::right_justify(ResName + " " , HeaderColWidth); |
1305 | for (; C < SU->TopReadyCycle + PI.AcquireAtCycle; ++C) { |
1306 | dbgs() << llvm::left_justify("|" , ColWidth); |
1307 | } |
1308 | for (unsigned I = 0, E = PI.ReleaseAtCycle - PI.AcquireAtCycle; I != E; |
1309 | ++I, ++C) |
1310 | dbgs() << llvm::left_justify("| x" , ColWidth); |
1311 | while (C++ <= LastCycle) |
1312 | dbgs() << llvm::left_justify("|" , ColWidth); |
1313 | // Place end char |
1314 | dbgs() << "| \n" ; |
1315 | } |
1316 | } |
1317 | } |
1318 | |
1319 | LLVM_DUMP_METHOD void ScheduleDAGMI::dumpScheduleTraceBottomUp() const { |
1320 | // Bail off when there is no schedule model to query. |
1321 | if (!SchedModel.hasInstrSchedModel()) |
1322 | return; |
1323 | |
1324 | // Nothing to show if there is no or just one instruction. |
1325 | if (BB->size() < 2) |
1326 | return; |
1327 | |
1328 | dbgs() << " * Schedule table (BottomUp):\n" ; |
1329 | dbgs() << scheduleTableLegend << "\n" ; |
1330 | |
1331 | const int FirstCycle = getSUnit(&*(std::begin(*this)))->BotReadyCycle; |
1332 | int LastCycle = getSUnit(&*(std::prev(std::end(*this))))->BotReadyCycle; |
1333 | for (MachineInstr &MI : *this) { |
1334 | SUnit *SU = getSUnit(&MI); |
1335 | if (!SU) |
1336 | continue; |
1337 | const MCSchedClassDesc *SC = getSchedClass(SU); |
1338 | for (TargetSchedModel::ProcResIter PI = SchedModel.getWriteProcResBegin(SC), |
1339 | PE = SchedModel.getWriteProcResEnd(SC); |
1340 | PI != PE; ++PI) { |
1341 | if ((int)SU->BotReadyCycle - PI->ReleaseAtCycle + 1 < LastCycle) |
1342 | LastCycle = (int)SU->BotReadyCycle - PI->ReleaseAtCycle + 1; |
1343 | } |
1344 | } |
1345 | // Print the header with the cycles |
1346 | dbgs() << llvm::left_justify("Cycle" , HeaderColWidth); |
1347 | for (int C = FirstCycle; C >= LastCycle; --C) |
1348 | dbgs() << llvm::left_justify("| " + std::to_string(C), ColWidth); |
1349 | dbgs() << "|\n" ; |
1350 | |
1351 | for (MachineInstr &MI : *this) { |
1352 | SUnit *SU = getSUnit(&MI); |
1353 | if (!SU) { |
1354 | dbgs() << "Missing SUnit\n" ; |
1355 | continue; |
1356 | } |
1357 | std::string NodeName("SU(" ); |
1358 | NodeName += std::to_string(SU->NodeNum) + ")" ; |
1359 | dbgs() << llvm::left_justify(NodeName, HeaderColWidth); |
1360 | int C = FirstCycle; |
1361 | for (; C >= LastCycle; --C) { |
1362 | if (C == (int)SU->BotReadyCycle) |
1363 | dbgs() << llvm::left_justify("| i" , ColWidth); |
1364 | else |
1365 | dbgs() << llvm::left_justify("|" , ColWidth); |
1366 | } |
1367 | dbgs() << "|\n" ; |
1368 | const MCSchedClassDesc *SC = getSchedClass(SU); |
1369 | SmallVector<MCWriteProcResEntry, 4> ResourcesIt( |
1370 | make_range(SchedModel.getWriteProcResBegin(SC), |
1371 | SchedModel.getWriteProcResEnd(SC))); |
1372 | |
1373 | if (MISchedSortResourcesInTrace) |
1374 | llvm::stable_sort( |
1375 | ResourcesIt, |
1376 | [](const MCWriteProcResEntry &LHS, |
1377 | const MCWriteProcResEntry &RHS) -> bool { |
1378 | return std::tie(LHS.AcquireAtCycle, LHS.ReleaseAtCycle) < |
1379 | std::tie(RHS.AcquireAtCycle, RHS.ReleaseAtCycle); |
1380 | }); |
1381 | for (const MCWriteProcResEntry &PI : ResourcesIt) { |
1382 | C = FirstCycle; |
1383 | const std::string ResName = |
1384 | SchedModel.getResourceName(PI.ProcResourceIdx); |
1385 | dbgs() << llvm::right_justify(ResName + " " , HeaderColWidth); |
1386 | for (; C > ((int)SU->BotReadyCycle - (int)PI.AcquireAtCycle); --C) { |
1387 | dbgs() << llvm::left_justify("|" , ColWidth); |
1388 | } |
1389 | for (unsigned I = 0, E = PI.ReleaseAtCycle - PI.AcquireAtCycle; I != E; |
1390 | ++I, --C) |
1391 | dbgs() << llvm::left_justify("| x" , ColWidth); |
1392 | while (C-- >= LastCycle) |
1393 | dbgs() << llvm::left_justify("|" , ColWidth); |
1394 | // Place end char |
1395 | dbgs() << "| \n" ; |
1396 | } |
1397 | } |
1398 | } |
1399 | #endif |
1400 | |
1401 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
1402 | LLVM_DUMP_METHOD void ScheduleDAGMI::dumpSchedule() const { |
1403 | if (MISchedDumpScheduleTrace) { |
1404 | if (DumpDir == DumpDirection::TopDown) |
1405 | dumpScheduleTraceTopDown(); |
1406 | else if (DumpDir == DumpDirection::BottomUp) |
1407 | dumpScheduleTraceBottomUp(); |
1408 | else if (DumpDir == DumpDirection::Bidirectional) { |
1409 | dbgs() << "* Schedule table (Bidirectional): not implemented\n" ; |
1410 | } else { |
1411 | dbgs() << "* Schedule table: DumpDirection not set.\n" ; |
1412 | } |
1413 | } |
1414 | |
1415 | for (MachineInstr &MI : *this) { |
1416 | if (SUnit *SU = getSUnit(&MI)) |
1417 | dumpNode(*SU); |
1418 | else |
1419 | dbgs() << "Missing SUnit\n" ; |
1420 | } |
1421 | } |
1422 | #endif |
1423 | |
1424 | //===----------------------------------------------------------------------===// |
1425 | // ScheduleDAGMILive - Base class for MachineInstr scheduling with LiveIntervals |
1426 | // preservation. |
1427 | //===----------------------------------------------------------------------===// |
1428 | |
1429 | ScheduleDAGMILive::~ScheduleDAGMILive() { |
1430 | delete DFSResult; |
1431 | } |
1432 | |
1433 | void ScheduleDAGMILive::collectVRegUses(SUnit &SU) { |
1434 | const MachineInstr &MI = *SU.getInstr(); |
1435 | for (const MachineOperand &MO : MI.operands()) { |
1436 | if (!MO.isReg()) |
1437 | continue; |
1438 | if (!MO.readsReg()) |
1439 | continue; |
1440 | if (TrackLaneMasks && !MO.isUse()) |
1441 | continue; |
1442 | |
1443 | Register Reg = MO.getReg(); |
1444 | if (!Reg.isVirtual()) |
1445 | continue; |
1446 | |
1447 | // Ignore re-defs. |
1448 | if (TrackLaneMasks) { |
1449 | bool FoundDef = false; |
1450 | for (const MachineOperand &MO2 : MI.all_defs()) { |
1451 | if (MO2.getReg() == Reg && !MO2.isDead()) { |
1452 | FoundDef = true; |
1453 | break; |
1454 | } |
1455 | } |
1456 | if (FoundDef) |
1457 | continue; |
1458 | } |
1459 | |
1460 | // Record this local VReg use. |
1461 | VReg2SUnitMultiMap::iterator UI = VRegUses.find(Key: Reg); |
1462 | for (; UI != VRegUses.end(); ++UI) { |
1463 | if (UI->SU == &SU) |
1464 | break; |
1465 | } |
1466 | if (UI == VRegUses.end()) |
1467 | VRegUses.insert(Val: VReg2SUnit(Reg, LaneBitmask::getNone(), &SU)); |
1468 | } |
1469 | } |
1470 | |
1471 | /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after |
1472 | /// crossing a scheduling boundary. [begin, end) includes all instructions in |
1473 | /// the region, including the boundary itself and single-instruction regions |
1474 | /// that don't get scheduled. |
1475 | void ScheduleDAGMILive::enterRegion(MachineBasicBlock *bb, |
1476 | MachineBasicBlock::iterator begin, |
1477 | MachineBasicBlock::iterator end, |
1478 | unsigned regioninstrs) |
1479 | { |
1480 | // ScheduleDAGMI initializes SchedImpl's per-region policy. |
1481 | ScheduleDAGMI::enterRegion(bb, begin, end, regioninstrs); |
1482 | |
1483 | // For convenience remember the end of the liveness region. |
1484 | LiveRegionEnd = (RegionEnd == bb->end()) ? RegionEnd : std::next(x: RegionEnd); |
1485 | |
1486 | SUPressureDiffs.clear(); |
1487 | |
1488 | ShouldTrackPressure = SchedImpl->shouldTrackPressure(); |
1489 | ShouldTrackLaneMasks = SchedImpl->shouldTrackLaneMasks(); |
1490 | |
1491 | assert((!ShouldTrackLaneMasks || ShouldTrackPressure) && |
1492 | "ShouldTrackLaneMasks requires ShouldTrackPressure" ); |
1493 | } |
1494 | |
1495 | // Setup the register pressure trackers for the top scheduled and bottom |
1496 | // scheduled regions. |
1497 | void ScheduleDAGMILive::initRegPressure() { |
1498 | VRegUses.clear(); |
1499 | VRegUses.setUniverse(MRI.getNumVirtRegs()); |
1500 | for (SUnit &SU : SUnits) |
1501 | collectVRegUses(SU); |
1502 | |
1503 | TopRPTracker.init(mf: &MF, rci: RegClassInfo, lis: LIS, mbb: BB, pos: RegionBegin, |
1504 | TrackLaneMasks: ShouldTrackLaneMasks, TrackUntiedDefs: false); |
1505 | BotRPTracker.init(mf: &MF, rci: RegClassInfo, lis: LIS, mbb: BB, pos: LiveRegionEnd, |
1506 | TrackLaneMasks: ShouldTrackLaneMasks, TrackUntiedDefs: false); |
1507 | |
1508 | // Close the RPTracker to finalize live ins. |
1509 | RPTracker.closeRegion(); |
1510 | |
1511 | LLVM_DEBUG(RPTracker.dump()); |
1512 | |
1513 | // Initialize the live ins and live outs. |
1514 | TopRPTracker.addLiveRegs(Regs: RPTracker.getPressure().LiveInRegs); |
1515 | BotRPTracker.addLiveRegs(Regs: RPTracker.getPressure().LiveOutRegs); |
1516 | |
1517 | // Close one end of the tracker so we can call |
1518 | // getMaxUpward/DownwardPressureDelta before advancing across any |
1519 | // instructions. This converts currently live regs into live ins/outs. |
1520 | TopRPTracker.closeTop(); |
1521 | BotRPTracker.closeBottom(); |
1522 | |
1523 | BotRPTracker.initLiveThru(RPTracker); |
1524 | if (!BotRPTracker.getLiveThru().empty()) { |
1525 | TopRPTracker.initLiveThru(PressureSet: BotRPTracker.getLiveThru()); |
1526 | LLVM_DEBUG(dbgs() << "Live Thru: " ; |
1527 | dumpRegSetPressure(BotRPTracker.getLiveThru(), TRI)); |
1528 | }; |
1529 | |
1530 | // For each live out vreg reduce the pressure change associated with other |
1531 | // uses of the same vreg below the live-out reaching def. |
1532 | updatePressureDiffs(LiveUses: RPTracker.getPressure().LiveOutRegs); |
1533 | |
1534 | // Account for liveness generated by the region boundary. |
1535 | if (LiveRegionEnd != RegionEnd) { |
1536 | SmallVector<VRegMaskOrUnit, 8> LiveUses; |
1537 | BotRPTracker.recede(LiveUses: &LiveUses); |
1538 | updatePressureDiffs(LiveUses); |
1539 | } |
1540 | |
1541 | LLVM_DEBUG(dbgs() << "Top Pressure: " ; |
1542 | dumpRegSetPressure(TopRPTracker.getRegSetPressureAtPos(), TRI); |
1543 | dbgs() << "Bottom Pressure: " ; |
1544 | dumpRegSetPressure(BotRPTracker.getRegSetPressureAtPos(), TRI);); |
1545 | |
1546 | assert((BotRPTracker.getPos() == RegionEnd || |
1547 | (RegionEnd->isDebugInstr() && |
1548 | BotRPTracker.getPos() == priorNonDebug(RegionEnd, RegionBegin))) && |
1549 | "Can't find the region bottom" ); |
1550 | |
1551 | // Cache the list of excess pressure sets in this region. This will also track |
1552 | // the max pressure in the scheduled code for these sets. |
1553 | RegionCriticalPSets.clear(); |
1554 | const std::vector<unsigned> &RegionPressure = |
1555 | RPTracker.getPressure().MaxSetPressure; |
1556 | for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) { |
1557 | unsigned Limit = RegClassInfo->getRegPressureSetLimit(Idx: i); |
1558 | if (RegionPressure[i] > Limit) { |
1559 | LLVM_DEBUG(dbgs() << TRI->getRegPressureSetName(i) << " Limit " << Limit |
1560 | << " Actual " << RegionPressure[i] << "\n" ); |
1561 | RegionCriticalPSets.push_back(x: PressureChange(i)); |
1562 | } |
1563 | } |
1564 | LLVM_DEBUG({ |
1565 | if (RegionCriticalPSets.size() > 0) { |
1566 | dbgs() << "Excess PSets: " ; |
1567 | for (const PressureChange &RCPS : RegionCriticalPSets) |
1568 | dbgs() << TRI->getRegPressureSetName(RCPS.getPSet()) << " " ; |
1569 | dbgs() << "\n" ; |
1570 | } |
1571 | }); |
1572 | } |
1573 | |
1574 | void ScheduleDAGMILive:: |
1575 | updateScheduledPressure(const SUnit *SU, |
1576 | const std::vector<unsigned> &NewMaxPressure) { |
1577 | const PressureDiff &PDiff = getPressureDiff(SU); |
1578 | unsigned CritIdx = 0, CritEnd = RegionCriticalPSets.size(); |
1579 | for (const PressureChange &PC : PDiff) { |
1580 | if (!PC.isValid()) |
1581 | break; |
1582 | unsigned ID = PC.getPSet(); |
1583 | while (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() < ID) |
1584 | ++CritIdx; |
1585 | if (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() == ID) { |
1586 | if ((int)NewMaxPressure[ID] > RegionCriticalPSets[CritIdx].getUnitInc() |
1587 | && NewMaxPressure[ID] <= (unsigned)std::numeric_limits<int16_t>::max()) |
1588 | RegionCriticalPSets[CritIdx].setUnitInc(NewMaxPressure[ID]); |
1589 | } |
1590 | unsigned Limit = RegClassInfo->getRegPressureSetLimit(Idx: ID); |
1591 | if (NewMaxPressure[ID] >= Limit - 2) { |
1592 | LLVM_DEBUG(dbgs() << " " << TRI->getRegPressureSetName(ID) << ": " |
1593 | << NewMaxPressure[ID] |
1594 | << ((NewMaxPressure[ID] > Limit) ? " > " : " <= " ) |
1595 | << Limit << "(+ " << BotRPTracker.getLiveThru()[ID] |
1596 | << " livethru)\n" ); |
1597 | } |
1598 | } |
1599 | } |
1600 | |
1601 | /// Update the PressureDiff array for liveness after scheduling this |
1602 | /// instruction. |
1603 | void ScheduleDAGMILive::updatePressureDiffs(ArrayRef<VRegMaskOrUnit> LiveUses) { |
1604 | for (const VRegMaskOrUnit &P : LiveUses) { |
1605 | Register Reg = P.RegUnit; |
1606 | /// FIXME: Currently assuming single-use physregs. |
1607 | if (!Reg.isVirtual()) |
1608 | continue; |
1609 | |
1610 | if (ShouldTrackLaneMasks) { |
1611 | // If the register has just become live then other uses won't change |
1612 | // this fact anymore => decrement pressure. |
1613 | // If the register has just become dead then other uses make it come |
1614 | // back to life => increment pressure. |
1615 | bool Decrement = P.LaneMask.any(); |
1616 | |
1617 | for (const VReg2SUnit &V2SU |
1618 | : make_range(x: VRegUses.find(Key: Reg), y: VRegUses.end())) { |
1619 | SUnit &SU = *V2SU.SU; |
1620 | if (SU.isScheduled || &SU == &ExitSU) |
1621 | continue; |
1622 | |
1623 | PressureDiff &PDiff = getPressureDiff(SU: &SU); |
1624 | PDiff.addPressureChange(RegUnit: Reg, IsDec: Decrement, MRI: &MRI); |
1625 | if (llvm::any_of(Range&: PDiff, P: [](const PressureChange &Change) { |
1626 | return Change.isValid(); |
1627 | })) |
1628 | LLVM_DEBUG(dbgs() |
1629 | << " UpdateRegPressure: SU(" << SU.NodeNum << ") " |
1630 | << printReg(Reg, TRI) << ':' |
1631 | << PrintLaneMask(P.LaneMask) << ' ' << *SU.getInstr(); |
1632 | dbgs() << " to " ; PDiff.dump(*TRI);); |
1633 | } |
1634 | } else { |
1635 | assert(P.LaneMask.any()); |
1636 | LLVM_DEBUG(dbgs() << " LiveReg: " << printVRegOrUnit(Reg, TRI) << "\n" ); |
1637 | // This may be called before CurrentBottom has been initialized. However, |
1638 | // BotRPTracker must have a valid position. We want the value live into the |
1639 | // instruction or live out of the block, so ask for the previous |
1640 | // instruction's live-out. |
1641 | const LiveInterval &LI = LIS->getInterval(Reg); |
1642 | VNInfo *VNI; |
1643 | MachineBasicBlock::const_iterator I = |
1644 | nextIfDebug(I: BotRPTracker.getPos(), End: BB->end()); |
1645 | if (I == BB->end()) |
1646 | VNI = LI.getVNInfoBefore(Idx: LIS->getMBBEndIdx(mbb: BB)); |
1647 | else { |
1648 | LiveQueryResult LRQ = LI.Query(Idx: LIS->getInstructionIndex(Instr: *I)); |
1649 | VNI = LRQ.valueIn(); |
1650 | } |
1651 | // RegisterPressureTracker guarantees that readsReg is true for LiveUses. |
1652 | assert(VNI && "No live value at use." ); |
1653 | for (const VReg2SUnit &V2SU |
1654 | : make_range(x: VRegUses.find(Key: Reg), y: VRegUses.end())) { |
1655 | SUnit *SU = V2SU.SU; |
1656 | // If this use comes before the reaching def, it cannot be a last use, |
1657 | // so decrease its pressure change. |
1658 | if (!SU->isScheduled && SU != &ExitSU) { |
1659 | LiveQueryResult LRQ = |
1660 | LI.Query(Idx: LIS->getInstructionIndex(Instr: *SU->getInstr())); |
1661 | if (LRQ.valueIn() == VNI) { |
1662 | PressureDiff &PDiff = getPressureDiff(SU); |
1663 | PDiff.addPressureChange(RegUnit: Reg, IsDec: true, MRI: &MRI); |
1664 | if (llvm::any_of(Range&: PDiff, P: [](const PressureChange &Change) { |
1665 | return Change.isValid(); |
1666 | })) |
1667 | LLVM_DEBUG(dbgs() << " UpdateRegPressure: SU(" << SU->NodeNum |
1668 | << ") " << *SU->getInstr(); |
1669 | dbgs() << " to " ; |
1670 | PDiff.dump(*TRI);); |
1671 | } |
1672 | } |
1673 | } |
1674 | } |
1675 | } |
1676 | } |
1677 | |
1678 | void ScheduleDAGMILive::dump() const { |
1679 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
1680 | if (EntrySU.getInstr() != nullptr) |
1681 | dumpNodeAll(EntrySU); |
1682 | for (const SUnit &SU : SUnits) { |
1683 | dumpNodeAll(SU); |
1684 | if (ShouldTrackPressure) { |
1685 | dbgs() << " Pressure Diff : " ; |
1686 | getPressureDiff(&SU).dump(*TRI); |
1687 | } |
1688 | dbgs() << " Single Issue : " ; |
1689 | if (SchedModel.mustBeginGroup(SU.getInstr()) && |
1690 | SchedModel.mustEndGroup(SU.getInstr())) |
1691 | dbgs() << "true;" ; |
1692 | else |
1693 | dbgs() << "false;" ; |
1694 | dbgs() << '\n'; |
1695 | } |
1696 | if (ExitSU.getInstr() != nullptr) |
1697 | dumpNodeAll(ExitSU); |
1698 | #endif |
1699 | } |
1700 | |
1701 | /// schedule - Called back from MachineScheduler::runOnMachineFunction |
1702 | /// after setting up the current scheduling region. [RegionBegin, RegionEnd) |
1703 | /// only includes instructions that have DAG nodes, not scheduling boundaries. |
1704 | /// |
1705 | /// This is a skeletal driver, with all the functionality pushed into helpers, |
1706 | /// so that it can be easily extended by experimental schedulers. Generally, |
1707 | /// implementing MachineSchedStrategy should be sufficient to implement a new |
1708 | /// scheduling algorithm. However, if a scheduler further subclasses |
1709 | /// ScheduleDAGMILive then it will want to override this virtual method in order |
1710 | /// to update any specialized state. |
1711 | void ScheduleDAGMILive::schedule() { |
1712 | LLVM_DEBUG(dbgs() << "ScheduleDAGMILive::schedule starting\n" ); |
1713 | LLVM_DEBUG(SchedImpl->dumpPolicy()); |
1714 | buildDAGWithRegPressure(); |
1715 | |
1716 | postProcessDAG(); |
1717 | |
1718 | SmallVector<SUnit*, 8> TopRoots, BotRoots; |
1719 | findRootsAndBiasEdges(TopRoots, BotRoots); |
1720 | |
1721 | // Initialize the strategy before modifying the DAG. |
1722 | // This may initialize a DFSResult to be used for queue priority. |
1723 | SchedImpl->initialize(DAG: this); |
1724 | |
1725 | LLVM_DEBUG(dump()); |
1726 | if (PrintDAGs) dump(); |
1727 | if (ViewMISchedDAGs) viewGraph(); |
1728 | |
1729 | // Initialize ready queues now that the DAG and priority data are finalized. |
1730 | initQueues(TopRoots, BotRoots); |
1731 | |
1732 | bool IsTopNode = false; |
1733 | while (true) { |
1734 | if (!checkSchedLimit()) |
1735 | break; |
1736 | |
1737 | LLVM_DEBUG(dbgs() << "** ScheduleDAGMILive::schedule picking next node\n" ); |
1738 | SUnit *SU = SchedImpl->pickNode(IsTopNode); |
1739 | if (!SU) break; |
1740 | |
1741 | assert(!SU->isScheduled && "Node already scheduled" ); |
1742 | |
1743 | scheduleMI(SU, IsTopNode); |
1744 | |
1745 | if (DFSResult) { |
1746 | unsigned SubtreeID = DFSResult->getSubtreeID(SU); |
1747 | if (!ScheduledTrees.test(Idx: SubtreeID)) { |
1748 | ScheduledTrees.set(SubtreeID); |
1749 | DFSResult->scheduleTree(SubtreeID); |
1750 | SchedImpl->scheduleTree(SubtreeID); |
1751 | } |
1752 | } |
1753 | |
1754 | // Notify the scheduling strategy after updating the DAG. |
1755 | SchedImpl->schedNode(SU, IsTopNode); |
1756 | |
1757 | updateQueues(SU, IsTopNode); |
1758 | } |
1759 | assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone." ); |
1760 | |
1761 | placeDebugValues(); |
1762 | |
1763 | LLVM_DEBUG({ |
1764 | dbgs() << "*** Final schedule for " |
1765 | << printMBBReference(*begin()->getParent()) << " ***\n" ; |
1766 | dumpSchedule(); |
1767 | dbgs() << '\n'; |
1768 | }); |
1769 | } |
1770 | |
1771 | /// Build the DAG and setup three register pressure trackers. |
1772 | void ScheduleDAGMILive::buildDAGWithRegPressure() { |
1773 | if (!ShouldTrackPressure) { |
1774 | RPTracker.reset(); |
1775 | RegionCriticalPSets.clear(); |
1776 | buildSchedGraph(AA); |
1777 | return; |
1778 | } |
1779 | |
1780 | // Initialize the register pressure tracker used by buildSchedGraph. |
1781 | RPTracker.init(mf: &MF, rci: RegClassInfo, lis: LIS, mbb: BB, pos: LiveRegionEnd, |
1782 | TrackLaneMasks: ShouldTrackLaneMasks, /*TrackUntiedDefs=*/true); |
1783 | |
1784 | // Account for liveness generate by the region boundary. |
1785 | if (LiveRegionEnd != RegionEnd) |
1786 | RPTracker.recede(); |
1787 | |
1788 | // Build the DAG, and compute current register pressure. |
1789 | buildSchedGraph(AA, RPTracker: &RPTracker, PDiffs: &SUPressureDiffs, LIS, TrackLaneMasks: ShouldTrackLaneMasks); |
1790 | |
1791 | // Initialize top/bottom trackers after computing region pressure. |
1792 | initRegPressure(); |
1793 | } |
1794 | |
1795 | void ScheduleDAGMILive::computeDFSResult() { |
1796 | if (!DFSResult) |
1797 | DFSResult = new SchedDFSResult(/*BottomU*/true, MinSubtreeSize); |
1798 | DFSResult->clear(); |
1799 | ScheduledTrees.clear(); |
1800 | DFSResult->resize(NumSUnits: SUnits.size()); |
1801 | DFSResult->compute(SUnits); |
1802 | ScheduledTrees.resize(N: DFSResult->getNumSubtrees()); |
1803 | } |
1804 | |
1805 | /// Compute the max cyclic critical path through the DAG. The scheduling DAG |
1806 | /// only provides the critical path for single block loops. To handle loops that |
1807 | /// span blocks, we could use the vreg path latencies provided by |
1808 | /// MachineTraceMetrics instead. However, MachineTraceMetrics is not currently |
1809 | /// available for use in the scheduler. |
1810 | /// |
1811 | /// The cyclic path estimation identifies a def-use pair that crosses the back |
1812 | /// edge and considers the depth and height of the nodes. For example, consider |
1813 | /// the following instruction sequence where each instruction has unit latency |
1814 | /// and defines an eponymous virtual register: |
1815 | /// |
1816 | /// a->b(a,c)->c(b)->d(c)->exit |
1817 | /// |
1818 | /// The cyclic critical path is a two cycles: b->c->b |
1819 | /// The acyclic critical path is four cycles: a->b->c->d->exit |
1820 | /// LiveOutHeight = height(c) = len(c->d->exit) = 2 |
1821 | /// LiveOutDepth = depth(c) + 1 = len(a->b->c) + 1 = 3 |
1822 | /// LiveInHeight = height(b) + 1 = len(b->c->d->exit) + 1 = 4 |
1823 | /// LiveInDepth = depth(b) = len(a->b) = 1 |
1824 | /// |
1825 | /// LiveOutDepth - LiveInDepth = 3 - 1 = 2 |
1826 | /// LiveInHeight - LiveOutHeight = 4 - 2 = 2 |
1827 | /// CyclicCriticalPath = min(2, 2) = 2 |
1828 | /// |
1829 | /// This could be relevant to PostRA scheduling, but is currently implemented |
1830 | /// assuming LiveIntervals. |
1831 | unsigned ScheduleDAGMILive::computeCyclicCriticalPath() { |
1832 | // This only applies to single block loop. |
1833 | if (!BB->isSuccessor(MBB: BB)) |
1834 | return 0; |
1835 | |
1836 | unsigned MaxCyclicLatency = 0; |
1837 | // Visit each live out vreg def to find def/use pairs that cross iterations. |
1838 | for (const VRegMaskOrUnit &P : RPTracker.getPressure().LiveOutRegs) { |
1839 | Register Reg = P.RegUnit; |
1840 | if (!Reg.isVirtual()) |
1841 | continue; |
1842 | const LiveInterval &LI = LIS->getInterval(Reg); |
1843 | const VNInfo *DefVNI = LI.getVNInfoBefore(Idx: LIS->getMBBEndIdx(mbb: BB)); |
1844 | if (!DefVNI) |
1845 | continue; |
1846 | |
1847 | MachineInstr *DefMI = LIS->getInstructionFromIndex(index: DefVNI->def); |
1848 | const SUnit *DefSU = getSUnit(MI: DefMI); |
1849 | if (!DefSU) |
1850 | continue; |
1851 | |
1852 | unsigned LiveOutHeight = DefSU->getHeight(); |
1853 | unsigned LiveOutDepth = DefSU->getDepth() + DefSU->Latency; |
1854 | // Visit all local users of the vreg def. |
1855 | for (const VReg2SUnit &V2SU |
1856 | : make_range(x: VRegUses.find(Key: Reg), y: VRegUses.end())) { |
1857 | SUnit *SU = V2SU.SU; |
1858 | if (SU == &ExitSU) |
1859 | continue; |
1860 | |
1861 | // Only consider uses of the phi. |
1862 | LiveQueryResult LRQ = LI.Query(Idx: LIS->getInstructionIndex(Instr: *SU->getInstr())); |
1863 | if (!LRQ.valueIn()->isPHIDef()) |
1864 | continue; |
1865 | |
1866 | // Assume that a path spanning two iterations is a cycle, which could |
1867 | // overestimate in strange cases. This allows cyclic latency to be |
1868 | // estimated as the minimum slack of the vreg's depth or height. |
1869 | unsigned CyclicLatency = 0; |
1870 | if (LiveOutDepth > SU->getDepth()) |
1871 | CyclicLatency = LiveOutDepth - SU->getDepth(); |
1872 | |
1873 | unsigned LiveInHeight = SU->getHeight() + DefSU->Latency; |
1874 | if (LiveInHeight > LiveOutHeight) { |
1875 | if (LiveInHeight - LiveOutHeight < CyclicLatency) |
1876 | CyclicLatency = LiveInHeight - LiveOutHeight; |
1877 | } else |
1878 | CyclicLatency = 0; |
1879 | |
1880 | LLVM_DEBUG(dbgs() << "Cyclic Path: SU(" << DefSU->NodeNum << ") -> SU(" |
1881 | << SU->NodeNum << ") = " << CyclicLatency << "c\n" ); |
1882 | if (CyclicLatency > MaxCyclicLatency) |
1883 | MaxCyclicLatency = CyclicLatency; |
1884 | } |
1885 | } |
1886 | LLVM_DEBUG(dbgs() << "Cyclic Critical Path: " << MaxCyclicLatency << "c\n" ); |
1887 | return MaxCyclicLatency; |
1888 | } |
1889 | |
1890 | /// Release ExitSU predecessors and setup scheduler queues. Re-position |
1891 | /// the Top RP tracker in case the region beginning has changed. |
1892 | void ScheduleDAGMILive::initQueues(ArrayRef<SUnit*> TopRoots, |
1893 | ArrayRef<SUnit*> BotRoots) { |
1894 | ScheduleDAGMI::initQueues(TopRoots, BotRoots); |
1895 | if (ShouldTrackPressure) { |
1896 | assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker" ); |
1897 | TopRPTracker.setPos(CurrentTop); |
1898 | } |
1899 | } |
1900 | |
1901 | /// Move an instruction and update register pressure. |
1902 | void ScheduleDAGMILive::scheduleMI(SUnit *SU, bool IsTopNode) { |
1903 | // Move the instruction to its new location in the instruction stream. |
1904 | MachineInstr *MI = SU->getInstr(); |
1905 | |
1906 | if (IsTopNode) { |
1907 | assert(SU->isTopReady() && "node still has unscheduled dependencies" ); |
1908 | if (&*CurrentTop == MI) |
1909 | CurrentTop = nextIfDebug(I: ++CurrentTop, End: CurrentBottom); |
1910 | else { |
1911 | moveInstruction(MI, InsertPos: CurrentTop); |
1912 | TopRPTracker.setPos(MI); |
1913 | } |
1914 | |
1915 | if (ShouldTrackPressure) { |
1916 | // Update top scheduled pressure. |
1917 | RegisterOperands RegOpers; |
1918 | RegOpers.collect(MI: *MI, TRI: *TRI, MRI, TrackLaneMasks: ShouldTrackLaneMasks, |
1919 | /*IgnoreDead=*/false); |
1920 | if (ShouldTrackLaneMasks) { |
1921 | // Adjust liveness and add missing dead+read-undef flags. |
1922 | SlotIndex SlotIdx = LIS->getInstructionIndex(Instr: *MI).getRegSlot(); |
1923 | RegOpers.adjustLaneLiveness(LIS: *LIS, MRI, Pos: SlotIdx, AddFlagsMI: MI); |
1924 | } else { |
1925 | // Adjust for missing dead-def flags. |
1926 | RegOpers.detectDeadDefs(MI: *MI, LIS: *LIS); |
1927 | } |
1928 | |
1929 | TopRPTracker.advance(RegOpers); |
1930 | assert(TopRPTracker.getPos() == CurrentTop && "out of sync" ); |
1931 | LLVM_DEBUG(dbgs() << "Top Pressure: " ; dumpRegSetPressure( |
1932 | TopRPTracker.getRegSetPressureAtPos(), TRI);); |
1933 | |
1934 | updateScheduledPressure(SU, NewMaxPressure: TopRPTracker.getPressure().MaxSetPressure); |
1935 | } |
1936 | } else { |
1937 | assert(SU->isBottomReady() && "node still has unscheduled dependencies" ); |
1938 | MachineBasicBlock::iterator priorII = |
1939 | priorNonDebug(I: CurrentBottom, Beg: CurrentTop); |
1940 | if (&*priorII == MI) |
1941 | CurrentBottom = priorII; |
1942 | else { |
1943 | if (&*CurrentTop == MI) { |
1944 | CurrentTop = nextIfDebug(I: ++CurrentTop, End: priorII); |
1945 | TopRPTracker.setPos(CurrentTop); |
1946 | } |
1947 | moveInstruction(MI, InsertPos: CurrentBottom); |
1948 | CurrentBottom = MI; |
1949 | BotRPTracker.setPos(CurrentBottom); |
1950 | } |
1951 | if (ShouldTrackPressure) { |
1952 | RegisterOperands RegOpers; |
1953 | RegOpers.collect(MI: *MI, TRI: *TRI, MRI, TrackLaneMasks: ShouldTrackLaneMasks, |
1954 | /*IgnoreDead=*/false); |
1955 | if (ShouldTrackLaneMasks) { |
1956 | // Adjust liveness and add missing dead+read-undef flags. |
1957 | SlotIndex SlotIdx = LIS->getInstructionIndex(Instr: *MI).getRegSlot(); |
1958 | RegOpers.adjustLaneLiveness(LIS: *LIS, MRI, Pos: SlotIdx, AddFlagsMI: MI); |
1959 | } else { |
1960 | // Adjust for missing dead-def flags. |
1961 | RegOpers.detectDeadDefs(MI: *MI, LIS: *LIS); |
1962 | } |
1963 | |
1964 | if (BotRPTracker.getPos() != CurrentBottom) |
1965 | BotRPTracker.recedeSkipDebugValues(); |
1966 | SmallVector<VRegMaskOrUnit, 8> LiveUses; |
1967 | BotRPTracker.recede(RegOpers, LiveUses: &LiveUses); |
1968 | assert(BotRPTracker.getPos() == CurrentBottom && "out of sync" ); |
1969 | LLVM_DEBUG(dbgs() << "Bottom Pressure: " ; dumpRegSetPressure( |
1970 | BotRPTracker.getRegSetPressureAtPos(), TRI);); |
1971 | |
1972 | updateScheduledPressure(SU, NewMaxPressure: BotRPTracker.getPressure().MaxSetPressure); |
1973 | updatePressureDiffs(LiveUses); |
1974 | } |
1975 | } |
1976 | } |
1977 | |
1978 | //===----------------------------------------------------------------------===// |
1979 | // BaseMemOpClusterMutation - DAG post-processing to cluster loads or stores. |
1980 | //===----------------------------------------------------------------------===// |
1981 | |
1982 | namespace { |
1983 | |
1984 | /// Post-process the DAG to create cluster edges between neighboring |
1985 | /// loads or between neighboring stores. |
1986 | class BaseMemOpClusterMutation : public ScheduleDAGMutation { |
1987 | struct MemOpInfo { |
1988 | SUnit *SU; |
1989 | SmallVector<const MachineOperand *, 4> BaseOps; |
1990 | int64_t Offset; |
1991 | LocationSize Width; |
1992 | bool OffsetIsScalable; |
1993 | |
1994 | MemOpInfo(SUnit *SU, ArrayRef<const MachineOperand *> BaseOps, |
1995 | int64_t Offset, bool OffsetIsScalable, LocationSize Width) |
1996 | : SU(SU), BaseOps(BaseOps), Offset(Offset), Width(Width), |
1997 | OffsetIsScalable(OffsetIsScalable) {} |
1998 | |
1999 | static bool Compare(const MachineOperand *const &A, |
2000 | const MachineOperand *const &B) { |
2001 | if (A->getType() != B->getType()) |
2002 | return A->getType() < B->getType(); |
2003 | if (A->isReg()) |
2004 | return A->getReg() < B->getReg(); |
2005 | if (A->isFI()) { |
2006 | const MachineFunction &MF = *A->getParent()->getParent()->getParent(); |
2007 | const TargetFrameLowering &TFI = *MF.getSubtarget().getFrameLowering(); |
2008 | bool StackGrowsDown = TFI.getStackGrowthDirection() == |
2009 | TargetFrameLowering::StackGrowsDown; |
2010 | return StackGrowsDown ? A->getIndex() > B->getIndex() |
2011 | : A->getIndex() < B->getIndex(); |
2012 | } |
2013 | |
2014 | llvm_unreachable("MemOpClusterMutation only supports register or frame " |
2015 | "index bases." ); |
2016 | } |
2017 | |
2018 | bool operator<(const MemOpInfo &RHS) const { |
2019 | // FIXME: Don't compare everything twice. Maybe use C++20 three way |
2020 | // comparison instead when it's available. |
2021 | if (std::lexicographical_compare(first1: BaseOps.begin(), last1: BaseOps.end(), |
2022 | first2: RHS.BaseOps.begin(), last2: RHS.BaseOps.end(), |
2023 | comp: Compare)) |
2024 | return true; |
2025 | if (std::lexicographical_compare(first1: RHS.BaseOps.begin(), last1: RHS.BaseOps.end(), |
2026 | first2: BaseOps.begin(), last2: BaseOps.end(), comp: Compare)) |
2027 | return false; |
2028 | if (Offset != RHS.Offset) |
2029 | return Offset < RHS.Offset; |
2030 | return SU->NodeNum < RHS.SU->NodeNum; |
2031 | } |
2032 | }; |
2033 | |
2034 | const TargetInstrInfo *TII; |
2035 | const TargetRegisterInfo *TRI; |
2036 | bool IsLoad; |
2037 | bool ReorderWhileClustering; |
2038 | |
2039 | public: |
2040 | BaseMemOpClusterMutation(const TargetInstrInfo *tii, |
2041 | const TargetRegisterInfo *tri, bool IsLoad, |
2042 | bool ReorderWhileClustering) |
2043 | : TII(tii), TRI(tri), IsLoad(IsLoad), |
2044 | ReorderWhileClustering(ReorderWhileClustering) {} |
2045 | |
2046 | void apply(ScheduleDAGInstrs *DAGInstrs) override; |
2047 | |
2048 | protected: |
2049 | void clusterNeighboringMemOps(ArrayRef<MemOpInfo> MemOps, bool FastCluster, |
2050 | ScheduleDAGInstrs *DAG); |
2051 | void collectMemOpRecords(std::vector<SUnit> &SUnits, |
2052 | SmallVectorImpl<MemOpInfo> &MemOpRecords); |
2053 | bool groupMemOps(ArrayRef<MemOpInfo> MemOps, ScheduleDAGInstrs *DAG, |
2054 | DenseMap<unsigned, SmallVector<MemOpInfo, 32>> &Groups); |
2055 | }; |
2056 | |
2057 | class StoreClusterMutation : public BaseMemOpClusterMutation { |
2058 | public: |
2059 | StoreClusterMutation(const TargetInstrInfo *tii, |
2060 | const TargetRegisterInfo *tri, |
2061 | bool ReorderWhileClustering) |
2062 | : BaseMemOpClusterMutation(tii, tri, false, ReorderWhileClustering) {} |
2063 | }; |
2064 | |
2065 | class LoadClusterMutation : public BaseMemOpClusterMutation { |
2066 | public: |
2067 | LoadClusterMutation(const TargetInstrInfo *tii, const TargetRegisterInfo *tri, |
2068 | bool ReorderWhileClustering) |
2069 | : BaseMemOpClusterMutation(tii, tri, true, ReorderWhileClustering) {} |
2070 | }; |
2071 | |
2072 | } // end anonymous namespace |
2073 | |
2074 | namespace llvm { |
2075 | |
2076 | std::unique_ptr<ScheduleDAGMutation> |
2077 | createLoadClusterDAGMutation(const TargetInstrInfo *TII, |
2078 | const TargetRegisterInfo *TRI, |
2079 | bool ReorderWhileClustering) { |
2080 | return EnableMemOpCluster ? std::make_unique<LoadClusterMutation>( |
2081 | args&: TII, args&: TRI, args&: ReorderWhileClustering) |
2082 | : nullptr; |
2083 | } |
2084 | |
2085 | std::unique_ptr<ScheduleDAGMutation> |
2086 | createStoreClusterDAGMutation(const TargetInstrInfo *TII, |
2087 | const TargetRegisterInfo *TRI, |
2088 | bool ReorderWhileClustering) { |
2089 | return EnableMemOpCluster ? std::make_unique<StoreClusterMutation>( |
2090 | args&: TII, args&: TRI, args&: ReorderWhileClustering) |
2091 | : nullptr; |
2092 | } |
2093 | |
2094 | } // end namespace llvm |
2095 | |
2096 | // Sorting all the loads/stores first, then for each load/store, checking the |
2097 | // following load/store one by one, until reach the first non-dependent one and |
2098 | // call target hook to see if they can cluster. |
2099 | // If FastCluster is enabled, we assume that, all the loads/stores have been |
2100 | // preprocessed and now, they didn't have dependencies on each other. |
2101 | void BaseMemOpClusterMutation::clusterNeighboringMemOps( |
2102 | ArrayRef<MemOpInfo> MemOpRecords, bool FastCluster, |
2103 | ScheduleDAGInstrs *DAG) { |
2104 | // Keep track of the current cluster length and bytes for each SUnit. |
2105 | DenseMap<unsigned, std::pair<unsigned, unsigned>> SUnit2ClusterInfo; |
2106 | EquivalenceClasses<SUnit *> Clusters; |
2107 | |
2108 | // At this point, `MemOpRecords` array must hold atleast two mem ops. Try to |
2109 | // cluster mem ops collected within `MemOpRecords` array. |
2110 | for (unsigned Idx = 0, End = MemOpRecords.size(); Idx < (End - 1); ++Idx) { |
2111 | // Decision to cluster mem ops is taken based on target dependent logic |
2112 | auto MemOpa = MemOpRecords[Idx]; |
2113 | |
2114 | // Seek for the next load/store to do the cluster. |
2115 | unsigned NextIdx = Idx + 1; |
2116 | for (; NextIdx < End; ++NextIdx) |
2117 | // Skip if MemOpb has been clustered already or has dependency with |
2118 | // MemOpa. |
2119 | if (!SUnit2ClusterInfo.count(Val: MemOpRecords[NextIdx].SU->NodeNum) && |
2120 | (FastCluster || |
2121 | (!DAG->IsReachable(SU: MemOpRecords[NextIdx].SU, TargetSU: MemOpa.SU) && |
2122 | !DAG->IsReachable(SU: MemOpa.SU, TargetSU: MemOpRecords[NextIdx].SU)))) |
2123 | break; |
2124 | if (NextIdx == End) |
2125 | continue; |
2126 | |
2127 | auto MemOpb = MemOpRecords[NextIdx]; |
2128 | unsigned ClusterLength = 2; |
2129 | unsigned CurrentClusterBytes = MemOpa.Width.getValue().getKnownMinValue() + |
2130 | MemOpb.Width.getValue().getKnownMinValue(); |
2131 | auto It = SUnit2ClusterInfo.find(Val: MemOpa.SU->NodeNum); |
2132 | if (It != SUnit2ClusterInfo.end()) { |
2133 | const auto &[Len, Bytes] = It->second; |
2134 | ClusterLength = Len + 1; |
2135 | CurrentClusterBytes = Bytes + MemOpb.Width.getValue().getKnownMinValue(); |
2136 | } |
2137 | |
2138 | if (!TII->shouldClusterMemOps(BaseOps1: MemOpa.BaseOps, Offset1: MemOpa.Offset, |
2139 | OffsetIsScalable1: MemOpa.OffsetIsScalable, BaseOps2: MemOpb.BaseOps, |
2140 | Offset2: MemOpb.Offset, OffsetIsScalable2: MemOpb.OffsetIsScalable, |
2141 | ClusterSize: ClusterLength, NumBytes: CurrentClusterBytes)) |
2142 | continue; |
2143 | |
2144 | SUnit *SUa = MemOpa.SU; |
2145 | SUnit *SUb = MemOpb.SU; |
2146 | |
2147 | if (!ReorderWhileClustering && SUa->NodeNum > SUb->NodeNum) |
2148 | std::swap(a&: SUa, b&: SUb); |
2149 | |
2150 | // FIXME: Is this check really required? |
2151 | if (!DAG->addEdge(SuccSU: SUb, PredDep: SDep(SUa, SDep::Cluster))) |
2152 | continue; |
2153 | |
2154 | Clusters.unionSets(V1: SUa, V2: SUb); |
2155 | LLVM_DEBUG(dbgs() << "Cluster ld/st SU(" << SUa->NodeNum << ") - SU(" |
2156 | << SUb->NodeNum << ")\n" ); |
2157 | ++NumClustered; |
2158 | |
2159 | if (IsLoad) { |
2160 | // Copy successor edges from SUa to SUb. Interleaving computation |
2161 | // dependent on SUa can prevent load combining due to register reuse. |
2162 | // Predecessor edges do not need to be copied from SUb to SUa since |
2163 | // nearby loads should have effectively the same inputs. |
2164 | for (const SDep &Succ : SUa->Succs) { |
2165 | if (Succ.getSUnit() == SUb) |
2166 | continue; |
2167 | LLVM_DEBUG(dbgs() << " Copy Succ SU(" << Succ.getSUnit()->NodeNum |
2168 | << ")\n" ); |
2169 | DAG->addEdge(SuccSU: Succ.getSUnit(), PredDep: SDep(SUb, SDep::Artificial)); |
2170 | } |
2171 | } else { |
2172 | // Copy predecessor edges from SUb to SUa to avoid the SUnits that |
2173 | // SUb dependent on scheduled in-between SUb and SUa. Successor edges |
2174 | // do not need to be copied from SUa to SUb since no one will depend |
2175 | // on stores. |
2176 | // Notice that, we don't need to care about the memory dependency as |
2177 | // we won't try to cluster them if they have any memory dependency. |
2178 | for (const SDep &Pred : SUb->Preds) { |
2179 | if (Pred.getSUnit() == SUa) |
2180 | continue; |
2181 | LLVM_DEBUG(dbgs() << " Copy Pred SU(" << Pred.getSUnit()->NodeNum |
2182 | << ")\n" ); |
2183 | DAG->addEdge(SuccSU: SUa, PredDep: SDep(Pred.getSUnit(), SDep::Artificial)); |
2184 | } |
2185 | } |
2186 | |
2187 | SUnit2ClusterInfo[MemOpb.SU->NodeNum] = {ClusterLength, |
2188 | CurrentClusterBytes}; |
2189 | |
2190 | LLVM_DEBUG(dbgs() << " Curr cluster length: " << ClusterLength |
2191 | << ", Curr cluster bytes: " << CurrentClusterBytes |
2192 | << "\n" ); |
2193 | } |
2194 | |
2195 | // Add cluster group information. |
2196 | // Iterate over all of the equivalence sets. |
2197 | auto &AllClusters = DAG->getClusters(); |
2198 | for (const EquivalenceClasses<SUnit *>::ECValue *I : Clusters) { |
2199 | if (!I->isLeader()) |
2200 | continue; |
2201 | ClusterInfo Group; |
2202 | unsigned ClusterIdx = AllClusters.size(); |
2203 | for (SUnit *MemberI : Clusters.members(ECV: *I)) { |
2204 | MemberI->ParentClusterIdx = ClusterIdx; |
2205 | Group.insert(Ptr: MemberI); |
2206 | } |
2207 | AllClusters.push_back(Elt: Group); |
2208 | } |
2209 | } |
2210 | |
2211 | void BaseMemOpClusterMutation::collectMemOpRecords( |
2212 | std::vector<SUnit> &SUnits, SmallVectorImpl<MemOpInfo> &MemOpRecords) { |
2213 | for (auto &SU : SUnits) { |
2214 | if ((IsLoad && !SU.getInstr()->mayLoad()) || |
2215 | (!IsLoad && !SU.getInstr()->mayStore())) |
2216 | continue; |
2217 | |
2218 | const MachineInstr &MI = *SU.getInstr(); |
2219 | SmallVector<const MachineOperand *, 4> BaseOps; |
2220 | int64_t Offset; |
2221 | bool OffsetIsScalable; |
2222 | LocationSize Width = LocationSize::precise(Value: 0); |
2223 | if (TII->getMemOperandsWithOffsetWidth(MI, BaseOps, Offset, |
2224 | OffsetIsScalable, Width, TRI)) { |
2225 | if (!Width.hasValue()) |
2226 | continue; |
2227 | |
2228 | MemOpRecords.push_back( |
2229 | Elt: MemOpInfo(&SU, BaseOps, Offset, OffsetIsScalable, Width)); |
2230 | |
2231 | LLVM_DEBUG(dbgs() << "Num BaseOps: " << BaseOps.size() << ", Offset: " |
2232 | << Offset << ", OffsetIsScalable: " << OffsetIsScalable |
2233 | << ", Width: " << Width << "\n" ); |
2234 | } |
2235 | #ifndef NDEBUG |
2236 | for (const auto *Op : BaseOps) |
2237 | assert(Op); |
2238 | #endif |
2239 | } |
2240 | } |
2241 | |
2242 | bool BaseMemOpClusterMutation::groupMemOps( |
2243 | ArrayRef<MemOpInfo> MemOps, ScheduleDAGInstrs *DAG, |
2244 | DenseMap<unsigned, SmallVector<MemOpInfo, 32>> &Groups) { |
2245 | bool FastCluster = |
2246 | ForceFastCluster || |
2247 | MemOps.size() * DAG->SUnits.size() / 1000 > FastClusterThreshold; |
2248 | |
2249 | for (const auto &MemOp : MemOps) { |
2250 | unsigned ChainPredID = DAG->SUnits.size(); |
2251 | if (FastCluster) { |
2252 | for (const SDep &Pred : MemOp.SU->Preds) { |
2253 | // We only want to cluster the mem ops that have the same ctrl(non-data) |
2254 | // pred so that they didn't have ctrl dependency for each other. But for |
2255 | // store instrs, we can still cluster them if the pred is load instr. |
2256 | if ((Pred.isCtrl() && |
2257 | (IsLoad || |
2258 | (Pred.getSUnit() && Pred.getSUnit()->getInstr()->mayStore()))) && |
2259 | !Pred.isArtificial()) { |
2260 | ChainPredID = Pred.getSUnit()->NodeNum; |
2261 | break; |
2262 | } |
2263 | } |
2264 | } else |
2265 | ChainPredID = 0; |
2266 | |
2267 | Groups[ChainPredID].push_back(Elt: MemOp); |
2268 | } |
2269 | return FastCluster; |
2270 | } |
2271 | |
2272 | /// Callback from DAG postProcessing to create cluster edges for loads/stores. |
2273 | void BaseMemOpClusterMutation::apply(ScheduleDAGInstrs *DAG) { |
2274 | // Collect all the clusterable loads/stores |
2275 | SmallVector<MemOpInfo, 32> MemOpRecords; |
2276 | collectMemOpRecords(SUnits&: DAG->SUnits, MemOpRecords); |
2277 | |
2278 | if (MemOpRecords.size() < 2) |
2279 | return; |
2280 | |
2281 | // Put the loads/stores without dependency into the same group with some |
2282 | // heuristic if the DAG is too complex to avoid compiling time blow up. |
2283 | // Notice that, some fusion pair could be lost with this. |
2284 | DenseMap<unsigned, SmallVector<MemOpInfo, 32>> Groups; |
2285 | bool FastCluster = groupMemOps(MemOps: MemOpRecords, DAG, Groups); |
2286 | |
2287 | for (auto &Group : Groups) { |
2288 | // Sorting the loads/stores, so that, we can stop the cluster as early as |
2289 | // possible. |
2290 | llvm::sort(C&: Group.second); |
2291 | |
2292 | // Trying to cluster all the neighboring loads/stores. |
2293 | clusterNeighboringMemOps(MemOpRecords: Group.second, FastCluster, DAG); |
2294 | } |
2295 | } |
2296 | |
2297 | //===----------------------------------------------------------------------===// |
2298 | // CopyConstrain - DAG post-processing to encourage copy elimination. |
2299 | //===----------------------------------------------------------------------===// |
2300 | |
2301 | namespace { |
2302 | |
2303 | /// Post-process the DAG to create weak edges from all uses of a copy to |
2304 | /// the one use that defines the copy's source vreg, most likely an induction |
2305 | /// variable increment. |
2306 | class CopyConstrain : public ScheduleDAGMutation { |
2307 | // Transient state. |
2308 | SlotIndex RegionBeginIdx; |
2309 | |
2310 | // RegionEndIdx is the slot index of the last non-debug instruction in the |
2311 | // scheduling region. So we may have RegionBeginIdx == RegionEndIdx. |
2312 | SlotIndex RegionEndIdx; |
2313 | |
2314 | public: |
2315 | CopyConstrain(const TargetInstrInfo *, const TargetRegisterInfo *) {} |
2316 | |
2317 | void apply(ScheduleDAGInstrs *DAGInstrs) override; |
2318 | |
2319 | protected: |
2320 | void constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG); |
2321 | }; |
2322 | |
2323 | } // end anonymous namespace |
2324 | |
2325 | namespace llvm { |
2326 | |
2327 | std::unique_ptr<ScheduleDAGMutation> |
2328 | createCopyConstrainDAGMutation(const TargetInstrInfo *TII, |
2329 | const TargetRegisterInfo *TRI) { |
2330 | return std::make_unique<CopyConstrain>(args&: TII, args&: TRI); |
2331 | } |
2332 | |
2333 | } // end namespace llvm |
2334 | |
2335 | /// constrainLocalCopy handles two possibilities: |
2336 | /// 1) Local src: |
2337 | /// I0: = dst |
2338 | /// I1: src = ... |
2339 | /// I2: = dst |
2340 | /// I3: dst = src (copy) |
2341 | /// (create pred->succ edges I0->I1, I2->I1) |
2342 | /// |
2343 | /// 2) Local copy: |
2344 | /// I0: dst = src (copy) |
2345 | /// I1: = dst |
2346 | /// I2: src = ... |
2347 | /// I3: = dst |
2348 | /// (create pred->succ edges I1->I2, I3->I2) |
2349 | /// |
2350 | /// Although the MachineScheduler is currently constrained to single blocks, |
2351 | /// this algorithm should handle extended blocks. An EBB is a set of |
2352 | /// contiguously numbered blocks such that the previous block in the EBB is |
2353 | /// always the single predecessor. |
2354 | void CopyConstrain::constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG) { |
2355 | LiveIntervals *LIS = DAG->getLIS(); |
2356 | MachineInstr *Copy = CopySU->getInstr(); |
2357 | |
2358 | // Check for pure vreg copies. |
2359 | const MachineOperand &SrcOp = Copy->getOperand(i: 1); |
2360 | Register SrcReg = SrcOp.getReg(); |
2361 | if (!SrcReg.isVirtual() || !SrcOp.readsReg()) |
2362 | return; |
2363 | |
2364 | const MachineOperand &DstOp = Copy->getOperand(i: 0); |
2365 | Register DstReg = DstOp.getReg(); |
2366 | if (!DstReg.isVirtual() || DstOp.isDead()) |
2367 | return; |
2368 | |
2369 | // Check if either the dest or source is local. If it's live across a back |
2370 | // edge, it's not local. Note that if both vregs are live across the back |
2371 | // edge, we cannot successfully contrain the copy without cyclic scheduling. |
2372 | // If both the copy's source and dest are local live intervals, then we |
2373 | // should treat the dest as the global for the purpose of adding |
2374 | // constraints. This adds edges from source's other uses to the copy. |
2375 | unsigned LocalReg = SrcReg; |
2376 | unsigned GlobalReg = DstReg; |
2377 | LiveInterval *LocalLI = &LIS->getInterval(Reg: LocalReg); |
2378 | if (!LocalLI->isLocal(Start: RegionBeginIdx, End: RegionEndIdx)) { |
2379 | LocalReg = DstReg; |
2380 | GlobalReg = SrcReg; |
2381 | LocalLI = &LIS->getInterval(Reg: LocalReg); |
2382 | if (!LocalLI->isLocal(Start: RegionBeginIdx, End: RegionEndIdx)) |
2383 | return; |
2384 | } |
2385 | LiveInterval *GlobalLI = &LIS->getInterval(Reg: GlobalReg); |
2386 | |
2387 | // Find the global segment after the start of the local LI. |
2388 | LiveInterval::iterator GlobalSegment = GlobalLI->find(Pos: LocalLI->beginIndex()); |
2389 | // If GlobalLI does not overlap LocalLI->start, then a copy directly feeds a |
2390 | // local live range. We could create edges from other global uses to the local |
2391 | // start, but the coalescer should have already eliminated these cases, so |
2392 | // don't bother dealing with it. |
2393 | if (GlobalSegment == GlobalLI->end()) |
2394 | return; |
2395 | |
2396 | // If GlobalSegment is killed at the LocalLI->start, the call to find() |
2397 | // returned the next global segment. But if GlobalSegment overlaps with |
2398 | // LocalLI->start, then advance to the next segment. If a hole in GlobalLI |
2399 | // exists in LocalLI's vicinity, GlobalSegment will be the end of the hole. |
2400 | if (GlobalSegment->contains(I: LocalLI->beginIndex())) |
2401 | ++GlobalSegment; |
2402 | |
2403 | if (GlobalSegment == GlobalLI->end()) |
2404 | return; |
2405 | |
2406 | // Check if GlobalLI contains a hole in the vicinity of LocalLI. |
2407 | if (GlobalSegment != GlobalLI->begin()) { |
2408 | // Two address defs have no hole. |
2409 | if (SlotIndex::isSameInstr(A: std::prev(x: GlobalSegment)->end, |
2410 | B: GlobalSegment->start)) { |
2411 | return; |
2412 | } |
2413 | // If the prior global segment may be defined by the same two-address |
2414 | // instruction that also defines LocalLI, then can't make a hole here. |
2415 | if (SlotIndex::isSameInstr(A: std::prev(x: GlobalSegment)->start, |
2416 | B: LocalLI->beginIndex())) { |
2417 | return; |
2418 | } |
2419 | // If GlobalLI has a prior segment, it must be live into the EBB. Otherwise |
2420 | // it would be a disconnected component in the live range. |
2421 | assert(std::prev(GlobalSegment)->start < LocalLI->beginIndex() && |
2422 | "Disconnected LRG within the scheduling region." ); |
2423 | } |
2424 | MachineInstr *GlobalDef = LIS->getInstructionFromIndex(index: GlobalSegment->start); |
2425 | if (!GlobalDef) |
2426 | return; |
2427 | |
2428 | SUnit *GlobalSU = DAG->getSUnit(MI: GlobalDef); |
2429 | if (!GlobalSU) |
2430 | return; |
2431 | |
2432 | // GlobalDef is the bottom of the GlobalLI hole. Open the hole by |
2433 | // constraining the uses of the last local def to precede GlobalDef. |
2434 | SmallVector<SUnit*,8> LocalUses; |
2435 | const VNInfo *LastLocalVN = LocalLI->getVNInfoBefore(Idx: LocalLI->endIndex()); |
2436 | MachineInstr *LastLocalDef = LIS->getInstructionFromIndex(index: LastLocalVN->def); |
2437 | SUnit *LastLocalSU = DAG->getSUnit(MI: LastLocalDef); |
2438 | for (const SDep &Succ : LastLocalSU->Succs) { |
2439 | if (Succ.getKind() != SDep::Data || Succ.getReg() != LocalReg) |
2440 | continue; |
2441 | if (Succ.getSUnit() == GlobalSU) |
2442 | continue; |
2443 | if (!DAG->canAddEdge(SuccSU: GlobalSU, PredSU: Succ.getSUnit())) |
2444 | return; |
2445 | LocalUses.push_back(Elt: Succ.getSUnit()); |
2446 | } |
2447 | // Open the top of the GlobalLI hole by constraining any earlier global uses |
2448 | // to precede the start of LocalLI. |
2449 | SmallVector<SUnit*,8> GlobalUses; |
2450 | MachineInstr *FirstLocalDef = |
2451 | LIS->getInstructionFromIndex(index: LocalLI->beginIndex()); |
2452 | SUnit *FirstLocalSU = DAG->getSUnit(MI: FirstLocalDef); |
2453 | for (const SDep &Pred : GlobalSU->Preds) { |
2454 | if (Pred.getKind() != SDep::Anti || Pred.getReg() != GlobalReg) |
2455 | continue; |
2456 | if (Pred.getSUnit() == FirstLocalSU) |
2457 | continue; |
2458 | if (!DAG->canAddEdge(SuccSU: FirstLocalSU, PredSU: Pred.getSUnit())) |
2459 | return; |
2460 | GlobalUses.push_back(Elt: Pred.getSUnit()); |
2461 | } |
2462 | LLVM_DEBUG(dbgs() << "Constraining copy SU(" << CopySU->NodeNum << ")\n" ); |
2463 | // Add the weak edges. |
2464 | for (SUnit *LU : LocalUses) { |
2465 | LLVM_DEBUG(dbgs() << " Local use SU(" << LU->NodeNum << ") -> SU(" |
2466 | << GlobalSU->NodeNum << ")\n" ); |
2467 | DAG->addEdge(SuccSU: GlobalSU, PredDep: SDep(LU, SDep::Weak)); |
2468 | } |
2469 | for (SUnit *GU : GlobalUses) { |
2470 | LLVM_DEBUG(dbgs() << " Global use SU(" << GU->NodeNum << ") -> SU(" |
2471 | << FirstLocalSU->NodeNum << ")\n" ); |
2472 | DAG->addEdge(SuccSU: FirstLocalSU, PredDep: SDep(GU, SDep::Weak)); |
2473 | } |
2474 | } |
2475 | |
2476 | /// Callback from DAG postProcessing to create weak edges to encourage |
2477 | /// copy elimination. |
2478 | void CopyConstrain::apply(ScheduleDAGInstrs *DAGInstrs) { |
2479 | ScheduleDAGMI *DAG = static_cast<ScheduleDAGMI*>(DAGInstrs); |
2480 | assert(DAG->hasVRegLiveness() && "Expect VRegs with LiveIntervals" ); |
2481 | |
2482 | MachineBasicBlock::iterator FirstPos = nextIfDebug(I: DAG->begin(), End: DAG->end()); |
2483 | if (FirstPos == DAG->end()) |
2484 | return; |
2485 | RegionBeginIdx = DAG->getLIS()->getInstructionIndex(Instr: *FirstPos); |
2486 | RegionEndIdx = DAG->getLIS()->getInstructionIndex( |
2487 | Instr: *priorNonDebug(I: DAG->end(), Beg: DAG->begin())); |
2488 | |
2489 | for (SUnit &SU : DAG->SUnits) { |
2490 | if (!SU.getInstr()->isCopy()) |
2491 | continue; |
2492 | |
2493 | constrainLocalCopy(CopySU: &SU, DAG: static_cast<ScheduleDAGMILive*>(DAG)); |
2494 | } |
2495 | } |
2496 | |
2497 | //===----------------------------------------------------------------------===// |
2498 | // MachineSchedStrategy helpers used by GenericScheduler, GenericPostScheduler |
2499 | // and possibly other custom schedulers. |
2500 | //===----------------------------------------------------------------------===// |
2501 | |
2502 | static const unsigned InvalidCycle = ~0U; |
2503 | |
2504 | SchedBoundary::~SchedBoundary() { delete HazardRec; } |
2505 | |
2506 | /// Given a Count of resource usage and a Latency value, return true if a |
2507 | /// SchedBoundary becomes resource limited. |
2508 | /// If we are checking after scheduling a node, we should return true when |
2509 | /// we just reach the resource limit. |
2510 | static bool checkResourceLimit(unsigned LFactor, unsigned Count, |
2511 | unsigned Latency, bool AfterSchedNode) { |
2512 | int ResCntFactor = (int)(Count - (Latency * LFactor)); |
2513 | if (AfterSchedNode) |
2514 | return ResCntFactor >= (int)LFactor; |
2515 | else |
2516 | return ResCntFactor > (int)LFactor; |
2517 | } |
2518 | |
2519 | void SchedBoundary::reset() { |
2520 | // A new HazardRec is created for each DAG and owned by SchedBoundary. |
2521 | // Destroying and reconstructing it is very expensive though. So keep |
2522 | // invalid, placeholder HazardRecs. |
2523 | if (HazardRec && HazardRec->isEnabled()) { |
2524 | delete HazardRec; |
2525 | HazardRec = nullptr; |
2526 | } |
2527 | Available.clear(); |
2528 | Pending.clear(); |
2529 | CheckPending = false; |
2530 | CurrCycle = 0; |
2531 | CurrMOps = 0; |
2532 | MinReadyCycle = std::numeric_limits<unsigned>::max(); |
2533 | ExpectedLatency = 0; |
2534 | DependentLatency = 0; |
2535 | RetiredMOps = 0; |
2536 | MaxExecutedResCount = 0; |
2537 | ZoneCritResIdx = 0; |
2538 | IsResourceLimited = false; |
2539 | ReservedCycles.clear(); |
2540 | ReservedResourceSegments.clear(); |
2541 | ReservedCyclesIndex.clear(); |
2542 | ResourceGroupSubUnitMasks.clear(); |
2543 | #if LLVM_ENABLE_ABI_BREAKING_CHECKS |
2544 | // Track the maximum number of stall cycles that could arise either from the |
2545 | // latency of a DAG edge or the number of cycles that a processor resource is |
2546 | // reserved (SchedBoundary::ReservedCycles). |
2547 | MaxObservedStall = 0; |
2548 | #endif |
2549 | // Reserve a zero-count for invalid CritResIdx. |
2550 | ExecutedResCounts.resize(N: 1); |
2551 | assert(!ExecutedResCounts[0] && "nonzero count for bad resource" ); |
2552 | } |
2553 | |
2554 | void SchedRemainder:: |
2555 | init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) { |
2556 | reset(); |
2557 | if (!SchedModel->hasInstrSchedModel()) |
2558 | return; |
2559 | RemainingCounts.resize(N: SchedModel->getNumProcResourceKinds()); |
2560 | for (SUnit &SU : DAG->SUnits) { |
2561 | const MCSchedClassDesc *SC = DAG->getSchedClass(SU: &SU); |
2562 | RemIssueCount += SchedModel->getNumMicroOps(MI: SU.getInstr(), SC) |
2563 | * SchedModel->getMicroOpFactor(); |
2564 | for (TargetSchedModel::ProcResIter |
2565 | PI = SchedModel->getWriteProcResBegin(SC), |
2566 | PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) { |
2567 | unsigned PIdx = PI->ProcResourceIdx; |
2568 | unsigned Factor = SchedModel->getResourceFactor(ResIdx: PIdx); |
2569 | assert(PI->ReleaseAtCycle >= PI->AcquireAtCycle); |
2570 | RemainingCounts[PIdx] += |
2571 | (Factor * (PI->ReleaseAtCycle - PI->AcquireAtCycle)); |
2572 | } |
2573 | } |
2574 | } |
2575 | |
2576 | void SchedBoundary:: |
2577 | init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) { |
2578 | reset(); |
2579 | DAG = dag; |
2580 | SchedModel = smodel; |
2581 | Rem = rem; |
2582 | if (SchedModel->hasInstrSchedModel()) { |
2583 | unsigned ResourceCount = SchedModel->getNumProcResourceKinds(); |
2584 | ReservedCyclesIndex.resize(N: ResourceCount); |
2585 | ExecutedResCounts.resize(N: ResourceCount); |
2586 | ResourceGroupSubUnitMasks.resize(N: ResourceCount, NV: APInt(ResourceCount, 0)); |
2587 | unsigned NumUnits = 0; |
2588 | |
2589 | for (unsigned i = 0; i < ResourceCount; ++i) { |
2590 | ReservedCyclesIndex[i] = NumUnits; |
2591 | NumUnits += SchedModel->getProcResource(PIdx: i)->NumUnits; |
2592 | if (isUnbufferedGroup(PIdx: i)) { |
2593 | auto SubUnits = SchedModel->getProcResource(PIdx: i)->SubUnitsIdxBegin; |
2594 | for (unsigned U = 0, UE = SchedModel->getProcResource(PIdx: i)->NumUnits; |
2595 | U != UE; ++U) |
2596 | ResourceGroupSubUnitMasks[i].setBit(SubUnits[U]); |
2597 | } |
2598 | } |
2599 | |
2600 | ReservedCycles.resize(new_size: NumUnits, x: InvalidCycle); |
2601 | } |
2602 | } |
2603 | |
2604 | /// Compute the stall cycles based on this SUnit's ready time. Heuristics treat |
2605 | /// these "soft stalls" differently than the hard stall cycles based on CPU |
2606 | /// resources and computed by checkHazard(). A fully in-order model |
2607 | /// (MicroOpBufferSize==0) will not make use of this since instructions are not |
2608 | /// available for scheduling until they are ready. However, a weaker in-order |
2609 | /// model may use this for heuristics. For example, if a processor has in-order |
2610 | /// behavior when reading certain resources, this may come into play. |
2611 | unsigned SchedBoundary::getLatencyStallCycles(SUnit *SU) { |
2612 | if (!SU->isUnbuffered) |
2613 | return 0; |
2614 | |
2615 | unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle); |
2616 | if (ReadyCycle > CurrCycle) |
2617 | return ReadyCycle - CurrCycle; |
2618 | return 0; |
2619 | } |
2620 | |
2621 | /// Compute the next cycle at which the given processor resource unit |
2622 | /// can be scheduled. |
2623 | unsigned SchedBoundary::getNextResourceCycleByInstance(unsigned InstanceIdx, |
2624 | unsigned ReleaseAtCycle, |
2625 | unsigned AcquireAtCycle) { |
2626 | if (SchedModel && SchedModel->enableIntervals()) { |
2627 | if (isTop()) |
2628 | return ReservedResourceSegments[InstanceIdx].getFirstAvailableAtFromTop( |
2629 | CurrCycle, AcquireAtCycle, ReleaseAtCycle); |
2630 | |
2631 | return ReservedResourceSegments[InstanceIdx].getFirstAvailableAtFromBottom( |
2632 | CurrCycle, AcquireAtCycle, ReleaseAtCycle); |
2633 | } |
2634 | |
2635 | unsigned NextUnreserved = ReservedCycles[InstanceIdx]; |
2636 | // If this resource has never been used, always return cycle zero. |
2637 | if (NextUnreserved == InvalidCycle) |
2638 | return CurrCycle; |
2639 | // For bottom-up scheduling add the cycles needed for the current operation. |
2640 | if (!isTop()) |
2641 | NextUnreserved = std::max(a: CurrCycle, b: NextUnreserved + ReleaseAtCycle); |
2642 | return NextUnreserved; |
2643 | } |
2644 | |
2645 | /// Compute the next cycle at which the given processor resource can be |
2646 | /// scheduled. Returns the next cycle and the index of the processor resource |
2647 | /// instance in the reserved cycles vector. |
2648 | std::pair<unsigned, unsigned> |
2649 | SchedBoundary::getNextResourceCycle(const MCSchedClassDesc *SC, unsigned PIdx, |
2650 | unsigned ReleaseAtCycle, |
2651 | unsigned AcquireAtCycle) { |
2652 | if (MischedDetailResourceBooking) { |
2653 | LLVM_DEBUG(dbgs() << " Resource booking (@" << CurrCycle << "c): \n" ); |
2654 | LLVM_DEBUG(dumpReservedCycles()); |
2655 | LLVM_DEBUG(dbgs() << " getNextResourceCycle (@" << CurrCycle << "c): \n" ); |
2656 | } |
2657 | unsigned MinNextUnreserved = InvalidCycle; |
2658 | unsigned InstanceIdx = 0; |
2659 | unsigned StartIndex = ReservedCyclesIndex[PIdx]; |
2660 | unsigned NumberOfInstances = SchedModel->getProcResource(PIdx)->NumUnits; |
2661 | assert(NumberOfInstances > 0 && |
2662 | "Cannot have zero instances of a ProcResource" ); |
2663 | |
2664 | if (isUnbufferedGroup(PIdx)) { |
2665 | // If any subunits are used by the instruction, report that the |
2666 | // subunits of the resource group are available at the first cycle |
2667 | // in which the unit is available, effectively removing the group |
2668 | // record from hazarding and basing the hazarding decisions on the |
2669 | // subunit records. Otherwise, choose the first available instance |
2670 | // from among the subunits. Specifications which assign cycles to |
2671 | // both the subunits and the group or which use an unbuffered |
2672 | // group with buffered subunits will appear to schedule |
2673 | // strangely. In the first case, the additional cycles for the |
2674 | // group will be ignored. In the second, the group will be |
2675 | // ignored entirely. |
2676 | for (const MCWriteProcResEntry &PE : |
2677 | make_range(x: SchedModel->getWriteProcResBegin(SC), |
2678 | y: SchedModel->getWriteProcResEnd(SC))) |
2679 | if (ResourceGroupSubUnitMasks[PIdx][PE.ProcResourceIdx]) |
2680 | return std::make_pair(x: getNextResourceCycleByInstance( |
2681 | InstanceIdx: StartIndex, ReleaseAtCycle, AcquireAtCycle), |
2682 | y&: StartIndex); |
2683 | |
2684 | auto SubUnits = SchedModel->getProcResource(PIdx)->SubUnitsIdxBegin; |
2685 | for (unsigned I = 0, End = NumberOfInstances; I < End; ++I) { |
2686 | unsigned NextUnreserved, NextInstanceIdx; |
2687 | std::tie(args&: NextUnreserved, args&: NextInstanceIdx) = |
2688 | getNextResourceCycle(SC, PIdx: SubUnits[I], ReleaseAtCycle, AcquireAtCycle); |
2689 | if (MinNextUnreserved > NextUnreserved) { |
2690 | InstanceIdx = NextInstanceIdx; |
2691 | MinNextUnreserved = NextUnreserved; |
2692 | } |
2693 | } |
2694 | return std::make_pair(x&: MinNextUnreserved, y&: InstanceIdx); |
2695 | } |
2696 | |
2697 | for (unsigned I = StartIndex, End = StartIndex + NumberOfInstances; I < End; |
2698 | ++I) { |
2699 | unsigned NextUnreserved = |
2700 | getNextResourceCycleByInstance(InstanceIdx: I, ReleaseAtCycle, AcquireAtCycle); |
2701 | if (MischedDetailResourceBooking) |
2702 | LLVM_DEBUG(dbgs() << " Instance " << I - StartIndex << " available @" |
2703 | << NextUnreserved << "c\n" ); |
2704 | if (MinNextUnreserved > NextUnreserved) { |
2705 | InstanceIdx = I; |
2706 | MinNextUnreserved = NextUnreserved; |
2707 | } |
2708 | } |
2709 | if (MischedDetailResourceBooking) |
2710 | LLVM_DEBUG(dbgs() << " selecting " << SchedModel->getResourceName(PIdx) |
2711 | << "[" << InstanceIdx - StartIndex << "]" |
2712 | << " available @" << MinNextUnreserved << "c" |
2713 | << "\n" ); |
2714 | return std::make_pair(x&: MinNextUnreserved, y&: InstanceIdx); |
2715 | } |
2716 | |
2717 | /// Does this SU have a hazard within the current instruction group. |
2718 | /// |
2719 | /// The scheduler supports two modes of hazard recognition. The first is the |
2720 | /// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that |
2721 | /// supports highly complicated in-order reservation tables |
2722 | /// (ScoreboardHazardRecognizer) and arbitrary target-specific logic. |
2723 | /// |
2724 | /// The second is a streamlined mechanism that checks for hazards based on |
2725 | /// simple counters that the scheduler itself maintains. It explicitly checks |
2726 | /// for instruction dispatch limitations, including the number of micro-ops that |
2727 | /// can dispatch per cycle. |
2728 | /// |
2729 | /// TODO: Also check whether the SU must start a new group. |
2730 | bool SchedBoundary::checkHazard(SUnit *SU) { |
2731 | if (HazardRec->isEnabled() |
2732 | && HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard) { |
2733 | LLVM_DEBUG(dbgs().indent(2) |
2734 | << "hazard: SU(" << SU->NodeNum << ") reported by HazardRec\n" ); |
2735 | return true; |
2736 | } |
2737 | |
2738 | unsigned uops = SchedModel->getNumMicroOps(MI: SU->getInstr()); |
2739 | if ((CurrMOps > 0) && (CurrMOps + uops > SchedModel->getIssueWidth())) { |
2740 | LLVM_DEBUG(dbgs().indent(2) << "hazard: SU(" << SU->NodeNum << ") uops=" |
2741 | << uops << ", CurrMOps = " << CurrMOps << ", " |
2742 | << "CurrMOps + uops > issue width of " |
2743 | << SchedModel->getIssueWidth() << "\n" ); |
2744 | return true; |
2745 | } |
2746 | |
2747 | if (CurrMOps > 0 && |
2748 | ((isTop() && SchedModel->mustBeginGroup(MI: SU->getInstr())) || |
2749 | (!isTop() && SchedModel->mustEndGroup(MI: SU->getInstr())))) { |
2750 | LLVM_DEBUG(dbgs().indent(2) << "hazard: SU(" << SU->NodeNum << ") must " |
2751 | << (isTop() ? "begin" : "end" ) << " group\n" ); |
2752 | return true; |
2753 | } |
2754 | |
2755 | if (SchedModel->hasInstrSchedModel() && SU->hasReservedResource) { |
2756 | const MCSchedClassDesc *SC = DAG->getSchedClass(SU); |
2757 | for (const MCWriteProcResEntry &PE : |
2758 | make_range(x: SchedModel->getWriteProcResBegin(SC), |
2759 | y: SchedModel->getWriteProcResEnd(SC))) { |
2760 | unsigned ResIdx = PE.ProcResourceIdx; |
2761 | unsigned ReleaseAtCycle = PE.ReleaseAtCycle; |
2762 | unsigned AcquireAtCycle = PE.AcquireAtCycle; |
2763 | unsigned NRCycle, InstanceIdx; |
2764 | std::tie(args&: NRCycle, args&: InstanceIdx) = |
2765 | getNextResourceCycle(SC, PIdx: ResIdx, ReleaseAtCycle, AcquireAtCycle); |
2766 | if (NRCycle > CurrCycle) { |
2767 | #if LLVM_ENABLE_ABI_BREAKING_CHECKS |
2768 | MaxObservedStall = std::max(ReleaseAtCycle, MaxObservedStall); |
2769 | #endif |
2770 | LLVM_DEBUG(dbgs().indent(2) |
2771 | << "hazard: SU(" << SU->NodeNum << ") " |
2772 | << SchedModel->getResourceName(ResIdx) << '[' |
2773 | << InstanceIdx - ReservedCyclesIndex[ResIdx] << ']' << "=" |
2774 | << NRCycle << "c, is later than " |
2775 | << "CurrCycle = " << CurrCycle << "c\n" ); |
2776 | return true; |
2777 | } |
2778 | } |
2779 | } |
2780 | return false; |
2781 | } |
2782 | |
2783 | // Find the unscheduled node in ReadySUs with the highest latency. |
2784 | unsigned SchedBoundary:: |
2785 | findMaxLatency(ArrayRef<SUnit*> ReadySUs) { |
2786 | SUnit *LateSU = nullptr; |
2787 | unsigned RemLatency = 0; |
2788 | for (SUnit *SU : ReadySUs) { |
2789 | unsigned L = getUnscheduledLatency(SU); |
2790 | if (L > RemLatency) { |
2791 | RemLatency = L; |
2792 | LateSU = SU; |
2793 | } |
2794 | } |
2795 | if (LateSU) { |
2796 | LLVM_DEBUG(dbgs() << Available.getName() << " RemLatency SU(" |
2797 | << LateSU->NodeNum << ") " << RemLatency << "c\n" ); |
2798 | } |
2799 | return RemLatency; |
2800 | } |
2801 | |
2802 | // Count resources in this zone and the remaining unscheduled |
2803 | // instruction. Return the max count, scaled. Set OtherCritIdx to the critical |
2804 | // resource index, or zero if the zone is issue limited. |
2805 | unsigned SchedBoundary:: |
2806 | getOtherResourceCount(unsigned &OtherCritIdx) { |
2807 | OtherCritIdx = 0; |
2808 | if (!SchedModel->hasInstrSchedModel()) |
2809 | return 0; |
2810 | |
2811 | unsigned OtherCritCount = Rem->RemIssueCount |
2812 | + (RetiredMOps * SchedModel->getMicroOpFactor()); |
2813 | LLVM_DEBUG(dbgs() << " " << Available.getName() << " + Remain MOps: " |
2814 | << OtherCritCount / SchedModel->getMicroOpFactor() << '\n'); |
2815 | for (unsigned PIdx = 1, PEnd = SchedModel->getNumProcResourceKinds(); |
2816 | PIdx != PEnd; ++PIdx) { |
2817 | unsigned OtherCount = getResourceCount(ResIdx: PIdx) + Rem->RemainingCounts[PIdx]; |
2818 | if (OtherCount > OtherCritCount) { |
2819 | OtherCritCount = OtherCount; |
2820 | OtherCritIdx = PIdx; |
2821 | } |
2822 | } |
2823 | if (OtherCritIdx) { |
2824 | LLVM_DEBUG( |
2825 | dbgs() << " " << Available.getName() << " + Remain CritRes: " |
2826 | << OtherCritCount / SchedModel->getResourceFactor(OtherCritIdx) |
2827 | << " " << SchedModel->getResourceName(OtherCritIdx) << "\n" ); |
2828 | } |
2829 | return OtherCritCount; |
2830 | } |
2831 | |
2832 | void SchedBoundary::releaseNode(SUnit *SU, unsigned ReadyCycle, bool InPQueue, |
2833 | unsigned Idx) { |
2834 | assert(SU->getInstr() && "Scheduled SUnit must have instr" ); |
2835 | |
2836 | #if LLVM_ENABLE_ABI_BREAKING_CHECKS |
2837 | // ReadyCycle was been bumped up to the CurrCycle when this node was |
2838 | // scheduled, but CurrCycle may have been eagerly advanced immediately after |
2839 | // scheduling, so may now be greater than ReadyCycle. |
2840 | if (ReadyCycle > CurrCycle) |
2841 | MaxObservedStall = std::max(ReadyCycle - CurrCycle, MaxObservedStall); |
2842 | #endif |
2843 | |
2844 | if (ReadyCycle < MinReadyCycle) |
2845 | MinReadyCycle = ReadyCycle; |
2846 | |
2847 | // Check for interlocks first. For the purpose of other heuristics, an |
2848 | // instruction that cannot issue appears as if it's not in the ReadyQueue. |
2849 | bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0; |
2850 | bool HazardDetected = !IsBuffered && ReadyCycle > CurrCycle; |
2851 | if (HazardDetected) |
2852 | LLVM_DEBUG(dbgs().indent(2) << "hazard: SU(" << SU->NodeNum |
2853 | << ") ReadyCycle = " << ReadyCycle |
2854 | << " is later than CurrCycle = " << CurrCycle |
2855 | << " on an unbuffered resource" << "\n" ); |
2856 | else |
2857 | HazardDetected = checkHazard(SU); |
2858 | |
2859 | if (!HazardDetected && Available.size() >= ReadyListLimit) { |
2860 | HazardDetected = true; |
2861 | LLVM_DEBUG(dbgs().indent(2) << "hazard: Available Q is full (size: " |
2862 | << Available.size() << ")\n" ); |
2863 | } |
2864 | |
2865 | if (!HazardDetected) { |
2866 | Available.push(SU); |
2867 | LLVM_DEBUG(dbgs().indent(2) |
2868 | << "Move SU(" << SU->NodeNum << ") into Available Q\n" ); |
2869 | |
2870 | if (InPQueue) |
2871 | Pending.remove(I: Pending.begin() + Idx); |
2872 | return; |
2873 | } |
2874 | |
2875 | if (!InPQueue) |
2876 | Pending.push(SU); |
2877 | } |
2878 | |
2879 | /// Move the boundary of scheduled code by one cycle. |
2880 | void SchedBoundary::bumpCycle(unsigned NextCycle) { |
2881 | if (SchedModel->getMicroOpBufferSize() == 0) { |
2882 | assert(MinReadyCycle < std::numeric_limits<unsigned>::max() && |
2883 | "MinReadyCycle uninitialized" ); |
2884 | if (MinReadyCycle > NextCycle) |
2885 | NextCycle = MinReadyCycle; |
2886 | } |
2887 | // Update the current micro-ops, which will issue in the next cycle. |
2888 | unsigned DecMOps = SchedModel->getIssueWidth() * (NextCycle - CurrCycle); |
2889 | CurrMOps = (CurrMOps <= DecMOps) ? 0 : CurrMOps - DecMOps; |
2890 | |
2891 | // Decrement DependentLatency based on the next cycle. |
2892 | if ((NextCycle - CurrCycle) > DependentLatency) |
2893 | DependentLatency = 0; |
2894 | else |
2895 | DependentLatency -= (NextCycle - CurrCycle); |
2896 | |
2897 | if (!HazardRec->isEnabled()) { |
2898 | // Bypass HazardRec virtual calls. |
2899 | CurrCycle = NextCycle; |
2900 | } else { |
2901 | // Bypass getHazardType calls in case of long latency. |
2902 | for (; CurrCycle != NextCycle; ++CurrCycle) { |
2903 | if (isTop()) |
2904 | HazardRec->AdvanceCycle(); |
2905 | else |
2906 | HazardRec->RecedeCycle(); |
2907 | } |
2908 | } |
2909 | CheckPending = true; |
2910 | IsResourceLimited = |
2911 | checkResourceLimit(LFactor: SchedModel->getLatencyFactor(), Count: getCriticalCount(), |
2912 | Latency: getScheduledLatency(), AfterSchedNode: true); |
2913 | |
2914 | LLVM_DEBUG(dbgs() << "Cycle: " << CurrCycle << ' ' << Available.getName() |
2915 | << '\n'); |
2916 | } |
2917 | |
2918 | void SchedBoundary::incExecutedResources(unsigned PIdx, unsigned Count) { |
2919 | ExecutedResCounts[PIdx] += Count; |
2920 | if (ExecutedResCounts[PIdx] > MaxExecutedResCount) |
2921 | MaxExecutedResCount = ExecutedResCounts[PIdx]; |
2922 | } |
2923 | |
2924 | /// Add the given processor resource to this scheduled zone. |
2925 | /// |
2926 | /// \param ReleaseAtCycle indicates the number of consecutive (non-pipelined) |
2927 | /// cycles during which this resource is released. |
2928 | /// |
2929 | /// \param AcquireAtCycle indicates the number of consecutive (non-pipelined) |
2930 | /// cycles at which the resource is aquired after issue (assuming no stalls). |
2931 | /// |
2932 | /// \return the next cycle at which the instruction may execute without |
2933 | /// oversubscribing resources. |
2934 | unsigned SchedBoundary::countResource(const MCSchedClassDesc *SC, unsigned PIdx, |
2935 | unsigned ReleaseAtCycle, |
2936 | unsigned NextCycle, |
2937 | unsigned AcquireAtCycle) { |
2938 | unsigned Factor = SchedModel->getResourceFactor(ResIdx: PIdx); |
2939 | unsigned Count = Factor * (ReleaseAtCycle- AcquireAtCycle); |
2940 | LLVM_DEBUG(dbgs() << " " << SchedModel->getResourceName(PIdx) << " +" |
2941 | << ReleaseAtCycle << "x" << Factor << "u\n" ); |
2942 | |
2943 | // Update Executed resources counts. |
2944 | incExecutedResources(PIdx, Count); |
2945 | assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted" ); |
2946 | Rem->RemainingCounts[PIdx] -= Count; |
2947 | |
2948 | // Check if this resource exceeds the current critical resource. If so, it |
2949 | // becomes the critical resource. |
2950 | if (ZoneCritResIdx != PIdx && (getResourceCount(ResIdx: PIdx) > getCriticalCount())) { |
2951 | ZoneCritResIdx = PIdx; |
2952 | LLVM_DEBUG(dbgs() << " *** Critical resource " |
2953 | << SchedModel->getResourceName(PIdx) << ": " |
2954 | << getResourceCount(PIdx) / SchedModel->getLatencyFactor() |
2955 | << "c\n" ); |
2956 | } |
2957 | // For reserved resources, record the highest cycle using the resource. |
2958 | unsigned NextAvailable, InstanceIdx; |
2959 | std::tie(args&: NextAvailable, args&: InstanceIdx) = |
2960 | getNextResourceCycle(SC, PIdx, ReleaseAtCycle, AcquireAtCycle); |
2961 | if (NextAvailable > CurrCycle) { |
2962 | LLVM_DEBUG(dbgs() << " Resource conflict: " |
2963 | << SchedModel->getResourceName(PIdx) |
2964 | << '[' << InstanceIdx - ReservedCyclesIndex[PIdx] << ']' |
2965 | << " reserved until @" << NextAvailable << "\n" ); |
2966 | } |
2967 | return NextAvailable; |
2968 | } |
2969 | |
2970 | /// Move the boundary of scheduled code by one SUnit. |
2971 | void SchedBoundary::bumpNode(SUnit *SU) { |
2972 | // Update the reservation table. |
2973 | if (HazardRec->isEnabled()) { |
2974 | if (!isTop() && SU->isCall) { |
2975 | // Calls are scheduled with their preceding instructions. For bottom-up |
2976 | // scheduling, clear the pipeline state before emitting. |
2977 | HazardRec->Reset(); |
2978 | } |
2979 | HazardRec->EmitInstruction(SU); |
2980 | // Scheduling an instruction may have made pending instructions available. |
2981 | CheckPending = true; |
2982 | } |
2983 | // checkHazard should prevent scheduling multiple instructions per cycle that |
2984 | // exceed the issue width. |
2985 | const MCSchedClassDesc *SC = DAG->getSchedClass(SU); |
2986 | unsigned IncMOps = SchedModel->getNumMicroOps(MI: SU->getInstr()); |
2987 | assert( |
2988 | (CurrMOps == 0 || (CurrMOps + IncMOps) <= SchedModel->getIssueWidth()) && |
2989 | "Cannot schedule this instruction's MicroOps in the current cycle." ); |
2990 | |
2991 | unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle); |
2992 | LLVM_DEBUG(dbgs() << " Ready @" << ReadyCycle << "c\n" ); |
2993 | |
2994 | unsigned NextCycle = CurrCycle; |
2995 | switch (SchedModel->getMicroOpBufferSize()) { |
2996 | case 0: |
2997 | assert(ReadyCycle <= CurrCycle && "Broken PendingQueue" ); |
2998 | break; |
2999 | case 1: |
3000 | if (ReadyCycle > NextCycle) { |
3001 | NextCycle = ReadyCycle; |
3002 | LLVM_DEBUG(dbgs() << " *** Stall until: " << ReadyCycle << "\n" ); |
3003 | } |
3004 | break; |
3005 | default: |
3006 | // We don't currently model the OOO reorder buffer, so consider all |
3007 | // scheduled MOps to be "retired". We do loosely model in-order resource |
3008 | // latency. If this instruction uses an in-order resource, account for any |
3009 | // likely stall cycles. |
3010 | if (SU->isUnbuffered && ReadyCycle > NextCycle) |
3011 | NextCycle = ReadyCycle; |
3012 | break; |
3013 | } |
3014 | RetiredMOps += IncMOps; |
3015 | |
3016 | // Update resource counts and critical resource. |
3017 | if (SchedModel->hasInstrSchedModel()) { |
3018 | unsigned DecRemIssue = IncMOps * SchedModel->getMicroOpFactor(); |
3019 | assert(Rem->RemIssueCount >= DecRemIssue && "MOps double counted" ); |
3020 | Rem->RemIssueCount -= DecRemIssue; |
3021 | if (ZoneCritResIdx) { |
3022 | // Scale scheduled micro-ops for comparing with the critical resource. |
3023 | unsigned ScaledMOps = |
3024 | RetiredMOps * SchedModel->getMicroOpFactor(); |
3025 | |
3026 | // If scaled micro-ops are now more than the previous critical resource by |
3027 | // a full cycle, then micro-ops issue becomes critical. |
3028 | if ((int)(ScaledMOps - getResourceCount(ResIdx: ZoneCritResIdx)) |
3029 | >= (int)SchedModel->getLatencyFactor()) { |
3030 | ZoneCritResIdx = 0; |
3031 | LLVM_DEBUG(dbgs() << " *** Critical resource NumMicroOps: " |
3032 | << ScaledMOps / SchedModel->getLatencyFactor() |
3033 | << "c\n" ); |
3034 | } |
3035 | } |
3036 | for (TargetSchedModel::ProcResIter |
3037 | PI = SchedModel->getWriteProcResBegin(SC), |
3038 | PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) { |
3039 | unsigned RCycle = |
3040 | countResource(SC, PIdx: PI->ProcResourceIdx, ReleaseAtCycle: PI->ReleaseAtCycle, NextCycle, |
3041 | AcquireAtCycle: PI->AcquireAtCycle); |
3042 | if (RCycle > NextCycle) |
3043 | NextCycle = RCycle; |
3044 | } |
3045 | if (SU->hasReservedResource) { |
3046 | // For reserved resources, record the highest cycle using the resource. |
3047 | // For top-down scheduling, this is the cycle in which we schedule this |
3048 | // instruction plus the number of cycles the operations reserves the |
3049 | // resource. For bottom-up is it simply the instruction's cycle. |
3050 | for (TargetSchedModel::ProcResIter |
3051 | PI = SchedModel->getWriteProcResBegin(SC), |
3052 | PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) { |
3053 | unsigned PIdx = PI->ProcResourceIdx; |
3054 | if (SchedModel->getProcResource(PIdx)->BufferSize == 0) { |
3055 | |
3056 | if (SchedModel && SchedModel->enableIntervals()) { |
3057 | unsigned ReservedUntil, InstanceIdx; |
3058 | std::tie(args&: ReservedUntil, args&: InstanceIdx) = getNextResourceCycle( |
3059 | SC, PIdx, ReleaseAtCycle: PI->ReleaseAtCycle, AcquireAtCycle: PI->AcquireAtCycle); |
3060 | if (isTop()) { |
3061 | ReservedResourceSegments[InstanceIdx].add( |
3062 | A: ResourceSegments::getResourceIntervalTop( |
3063 | C: NextCycle, AcquireAtCycle: PI->AcquireAtCycle, ReleaseAtCycle: PI->ReleaseAtCycle), |
3064 | CutOff: MIResourceCutOff); |
3065 | } else { |
3066 | ReservedResourceSegments[InstanceIdx].add( |
3067 | A: ResourceSegments::getResourceIntervalBottom( |
3068 | C: NextCycle, AcquireAtCycle: PI->AcquireAtCycle, ReleaseAtCycle: PI->ReleaseAtCycle), |
3069 | CutOff: MIResourceCutOff); |
3070 | } |
3071 | } else { |
3072 | |
3073 | unsigned ReservedUntil, InstanceIdx; |
3074 | std::tie(args&: ReservedUntil, args&: InstanceIdx) = getNextResourceCycle( |
3075 | SC, PIdx, ReleaseAtCycle: PI->ReleaseAtCycle, AcquireAtCycle: PI->AcquireAtCycle); |
3076 | if (isTop()) { |
3077 | ReservedCycles[InstanceIdx] = |
3078 | std::max(a: ReservedUntil, b: NextCycle + PI->ReleaseAtCycle); |
3079 | } else |
3080 | ReservedCycles[InstanceIdx] = NextCycle; |
3081 | } |
3082 | } |
3083 | } |
3084 | } |
3085 | } |
3086 | // Update ExpectedLatency and DependentLatency. |
3087 | unsigned &TopLatency = isTop() ? ExpectedLatency : DependentLatency; |
3088 | unsigned &BotLatency = isTop() ? DependentLatency : ExpectedLatency; |
3089 | if (SU->getDepth() > TopLatency) { |
3090 | TopLatency = SU->getDepth(); |
3091 | LLVM_DEBUG(dbgs() << " " << Available.getName() << " TopLatency SU(" |
3092 | << SU->NodeNum << ") " << TopLatency << "c\n" ); |
3093 | } |
3094 | if (SU->getHeight() > BotLatency) { |
3095 | BotLatency = SU->getHeight(); |
3096 | LLVM_DEBUG(dbgs() << " " << Available.getName() << " BotLatency SU(" |
3097 | << SU->NodeNum << ") " << BotLatency << "c\n" ); |
3098 | } |
3099 | // If we stall for any reason, bump the cycle. |
3100 | if (NextCycle > CurrCycle) |
3101 | bumpCycle(NextCycle); |
3102 | else |
3103 | // After updating ZoneCritResIdx and ExpectedLatency, check if we're |
3104 | // resource limited. If a stall occurred, bumpCycle does this. |
3105 | IsResourceLimited = |
3106 | checkResourceLimit(LFactor: SchedModel->getLatencyFactor(), Count: getCriticalCount(), |
3107 | Latency: getScheduledLatency(), AfterSchedNode: true); |
3108 | |
3109 | // Update CurrMOps after calling bumpCycle to handle stalls, since bumpCycle |
3110 | // resets CurrMOps. Loop to handle instructions with more MOps than issue in |
3111 | // one cycle. Since we commonly reach the max MOps here, opportunistically |
3112 | // bump the cycle to avoid uselessly checking everything in the readyQ. |
3113 | CurrMOps += IncMOps; |
3114 | |
3115 | // Bump the cycle count for issue group constraints. |
3116 | // This must be done after NextCycle has been adjust for all other stalls. |
3117 | // Calling bumpCycle(X) will reduce CurrMOps by one issue group and set |
3118 | // currCycle to X. |
3119 | if ((isTop() && SchedModel->mustEndGroup(MI: SU->getInstr())) || |
3120 | (!isTop() && SchedModel->mustBeginGroup(MI: SU->getInstr()))) { |
3121 | LLVM_DEBUG(dbgs() << " Bump cycle to " << (isTop() ? "end" : "begin" ) |
3122 | << " group\n" ); |
3123 | bumpCycle(NextCycle: ++NextCycle); |
3124 | } |
3125 | |
3126 | while (CurrMOps >= SchedModel->getIssueWidth()) { |
3127 | LLVM_DEBUG(dbgs() << " *** Max MOps " << CurrMOps << " at cycle " |
3128 | << CurrCycle << '\n'); |
3129 | bumpCycle(NextCycle: ++NextCycle); |
3130 | } |
3131 | LLVM_DEBUG(dumpScheduledState()); |
3132 | } |
3133 | |
3134 | /// Release pending ready nodes in to the available queue. This makes them |
3135 | /// visible to heuristics. |
3136 | void SchedBoundary::releasePending() { |
3137 | // If the available queue is empty, it is safe to reset MinReadyCycle. |
3138 | if (Available.empty()) |
3139 | MinReadyCycle = std::numeric_limits<unsigned>::max(); |
3140 | |
3141 | // Check to see if any of the pending instructions are ready to issue. If |
3142 | // so, add them to the available queue. |
3143 | for (unsigned I = 0, E = Pending.size(); I < E; ++I) { |
3144 | SUnit *SU = *(Pending.begin() + I); |
3145 | unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle; |
3146 | |
3147 | LLVM_DEBUG(dbgs() << "Checking pending node SU(" << SU->NodeNum << ")\n" ); |
3148 | |
3149 | if (ReadyCycle < MinReadyCycle) |
3150 | MinReadyCycle = ReadyCycle; |
3151 | |
3152 | if (Available.size() >= ReadyListLimit) |
3153 | break; |
3154 | |
3155 | releaseNode(SU, ReadyCycle, InPQueue: true, Idx: I); |
3156 | if (E != Pending.size()) { |
3157 | --I; |
3158 | --E; |
3159 | } |
3160 | } |
3161 | CheckPending = false; |
3162 | } |
3163 | |
3164 | /// Remove SU from the ready set for this boundary. |
3165 | void SchedBoundary::removeReady(SUnit *SU) { |
3166 | if (Available.isInQueue(SU)) |
3167 | Available.remove(I: Available.find(SU)); |
3168 | else { |
3169 | assert(Pending.isInQueue(SU) && "bad ready count" ); |
3170 | Pending.remove(I: Pending.find(SU)); |
3171 | } |
3172 | } |
3173 | |
3174 | /// If this queue only has one ready candidate, return it. As a side effect, |
3175 | /// defer any nodes that now hit a hazard, and advance the cycle until at least |
3176 | /// one node is ready. If multiple instructions are ready, return NULL. |
3177 | SUnit *SchedBoundary::pickOnlyChoice() { |
3178 | if (CheckPending) |
3179 | releasePending(); |
3180 | |
3181 | // Defer any ready instrs that now have a hazard. |
3182 | for (ReadyQueue::iterator I = Available.begin(); I != Available.end();) { |
3183 | if (checkHazard(SU: *I)) { |
3184 | Pending.push(SU: *I); |
3185 | I = Available.remove(I); |
3186 | continue; |
3187 | } |
3188 | ++I; |
3189 | } |
3190 | for (unsigned i = 0; Available.empty(); ++i) { |
3191 | // FIXME: Re-enable assert once PR20057 is resolved. |
3192 | // assert(i <= (HazardRec->getMaxLookAhead() + MaxObservedStall) && |
3193 | // "permanent hazard"); |
3194 | (void)i; |
3195 | bumpCycle(NextCycle: CurrCycle + 1); |
3196 | releasePending(); |
3197 | } |
3198 | |
3199 | LLVM_DEBUG(Pending.dump()); |
3200 | LLVM_DEBUG(Available.dump()); |
3201 | |
3202 | if (Available.size() == 1) |
3203 | return *Available.begin(); |
3204 | return nullptr; |
3205 | } |
3206 | |
3207 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
3208 | |
3209 | /// Dump the content of the \ref ReservedCycles vector for the |
3210 | /// resources that are used in the basic block. |
3211 | /// |
3212 | LLVM_DUMP_METHOD void SchedBoundary::dumpReservedCycles() const { |
3213 | if (!SchedModel->hasInstrSchedModel()) |
3214 | return; |
3215 | |
3216 | unsigned ResourceCount = SchedModel->getNumProcResourceKinds(); |
3217 | unsigned StartIdx = 0; |
3218 | |
3219 | for (unsigned ResIdx = 0; ResIdx < ResourceCount; ++ResIdx) { |
3220 | const unsigned NumUnits = SchedModel->getProcResource(ResIdx)->NumUnits; |
3221 | std::string ResName = SchedModel->getResourceName(ResIdx); |
3222 | for (unsigned UnitIdx = 0; UnitIdx < NumUnits; ++UnitIdx) { |
3223 | dbgs() << ResName << "(" << UnitIdx << ") = " ; |
3224 | if (SchedModel && SchedModel->enableIntervals()) { |
3225 | if (ReservedResourceSegments.count(StartIdx + UnitIdx)) |
3226 | dbgs() << ReservedResourceSegments.at(StartIdx + UnitIdx); |
3227 | else |
3228 | dbgs() << "{ }\n" ; |
3229 | } else |
3230 | dbgs() << ReservedCycles[StartIdx + UnitIdx] << "\n" ; |
3231 | } |
3232 | StartIdx += NumUnits; |
3233 | } |
3234 | } |
3235 | |
3236 | // This is useful information to dump after bumpNode. |
3237 | // Note that the Queue contents are more useful before pickNodeFromQueue. |
3238 | LLVM_DUMP_METHOD void SchedBoundary::dumpScheduledState() const { |
3239 | unsigned ResFactor; |
3240 | unsigned ResCount; |
3241 | if (ZoneCritResIdx) { |
3242 | ResFactor = SchedModel->getResourceFactor(ZoneCritResIdx); |
3243 | ResCount = getResourceCount(ZoneCritResIdx); |
3244 | } else { |
3245 | ResFactor = SchedModel->getMicroOpFactor(); |
3246 | ResCount = RetiredMOps * ResFactor; |
3247 | } |
3248 | unsigned LFactor = SchedModel->getLatencyFactor(); |
3249 | dbgs() << Available.getName() << " @" << CurrCycle << "c\n" |
3250 | << " Retired: " << RetiredMOps; |
3251 | dbgs() << "\n Executed: " << getExecutedCount() / LFactor << "c" ; |
3252 | dbgs() << "\n Critical: " << ResCount / LFactor << "c, " |
3253 | << ResCount / ResFactor << " " |
3254 | << SchedModel->getResourceName(ZoneCritResIdx) |
3255 | << "\n ExpectedLatency: " << ExpectedLatency << "c\n" |
3256 | << (IsResourceLimited ? " - Resource" : " - Latency" ) |
3257 | << " limited.\n" ; |
3258 | if (MISchedDumpReservedCycles) |
3259 | dumpReservedCycles(); |
3260 | } |
3261 | #endif |
3262 | |
3263 | //===----------------------------------------------------------------------===// |
3264 | // GenericScheduler - Generic implementation of MachineSchedStrategy. |
3265 | //===----------------------------------------------------------------------===// |
3266 | |
3267 | void GenericSchedulerBase::SchedCandidate:: |
3268 | initResourceDelta(const ScheduleDAGMI *DAG, |
3269 | const TargetSchedModel *SchedModel) { |
3270 | if (!Policy.ReduceResIdx && !Policy.DemandResIdx) |
3271 | return; |
3272 | |
3273 | const MCSchedClassDesc *SC = DAG->getSchedClass(SU); |
3274 | for (TargetSchedModel::ProcResIter |
3275 | PI = SchedModel->getWriteProcResBegin(SC), |
3276 | PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) { |
3277 | if (PI->ProcResourceIdx == Policy.ReduceResIdx) |
3278 | ResDelta.CritResources += PI->ReleaseAtCycle; |
3279 | if (PI->ProcResourceIdx == Policy.DemandResIdx) |
3280 | ResDelta.DemandedResources += PI->ReleaseAtCycle; |
3281 | } |
3282 | } |
3283 | |
3284 | /// Compute remaining latency. We need this both to determine whether the |
3285 | /// overall schedule has become latency-limited and whether the instructions |
3286 | /// outside this zone are resource or latency limited. |
3287 | /// |
3288 | /// The "dependent" latency is updated incrementally during scheduling as the |
3289 | /// max height/depth of scheduled nodes minus the cycles since it was |
3290 | /// scheduled: |
3291 | /// DLat = max (N.depth - (CurrCycle - N.ReadyCycle) for N in Zone |
3292 | /// |
3293 | /// The "independent" latency is the max ready queue depth: |
3294 | /// ILat = max N.depth for N in Available|Pending |
3295 | /// |
3296 | /// RemainingLatency is the greater of independent and dependent latency. |
3297 | /// |
3298 | /// These computations are expensive, especially in DAGs with many edges, so |
3299 | /// only do them if necessary. |
3300 | static unsigned computeRemLatency(SchedBoundary &CurrZone) { |
3301 | unsigned RemLatency = CurrZone.getDependentLatency(); |
3302 | RemLatency = std::max(a: RemLatency, |
3303 | b: CurrZone.findMaxLatency(ReadySUs: CurrZone.Available.elements())); |
3304 | RemLatency = std::max(a: RemLatency, |
3305 | b: CurrZone.findMaxLatency(ReadySUs: CurrZone.Pending.elements())); |
3306 | return RemLatency; |
3307 | } |
3308 | |
3309 | /// Returns true if the current cycle plus remaning latency is greater than |
3310 | /// the critical path in the scheduling region. |
3311 | bool GenericSchedulerBase::shouldReduceLatency(const CandPolicy &Policy, |
3312 | SchedBoundary &CurrZone, |
3313 | bool ComputeRemLatency, |
3314 | unsigned &RemLatency) const { |
3315 | // The current cycle is already greater than the critical path, so we are |
3316 | // already latency limited and don't need to compute the remaining latency. |
3317 | if (CurrZone.getCurrCycle() > Rem.CriticalPath) |
3318 | return true; |
3319 | |
3320 | // If we haven't scheduled anything yet, then we aren't latency limited. |
3321 | if (CurrZone.getCurrCycle() == 0) |
3322 | return false; |
3323 | |
3324 | if (ComputeRemLatency) |
3325 | RemLatency = computeRemLatency(CurrZone); |
3326 | |
3327 | return RemLatency + CurrZone.getCurrCycle() > Rem.CriticalPath; |
3328 | } |
3329 | |
3330 | /// Set the CandPolicy given a scheduling zone given the current resources and |
3331 | /// latencies inside and outside the zone. |
3332 | void GenericSchedulerBase::setPolicy(CandPolicy &Policy, bool IsPostRA, |
3333 | SchedBoundary &CurrZone, |
3334 | SchedBoundary *OtherZone) { |
3335 | // Apply preemptive heuristics based on the total latency and resources |
3336 | // inside and outside this zone. Potential stalls should be considered before |
3337 | // following this policy. |
3338 | |
3339 | // Compute the critical resource outside the zone. |
3340 | unsigned OtherCritIdx = 0; |
3341 | unsigned OtherCount = |
3342 | OtherZone ? OtherZone->getOtherResourceCount(OtherCritIdx) : 0; |
3343 | |
3344 | bool OtherResLimited = false; |
3345 | unsigned RemLatency = 0; |
3346 | bool RemLatencyComputed = false; |
3347 | if (SchedModel->hasInstrSchedModel() && OtherCount != 0) { |
3348 | RemLatency = computeRemLatency(CurrZone); |
3349 | RemLatencyComputed = true; |
3350 | OtherResLimited = checkResourceLimit(LFactor: SchedModel->getLatencyFactor(), |
3351 | Count: OtherCount, Latency: RemLatency, AfterSchedNode: false); |
3352 | } |
3353 | |
3354 | // Schedule aggressively for latency in PostRA mode. We don't check for |
3355 | // acyclic latency during PostRA, and highly out-of-order processors will |
3356 | // skip PostRA scheduling. |
3357 | if (!OtherResLimited && |
3358 | (IsPostRA || shouldReduceLatency(Policy, CurrZone, ComputeRemLatency: !RemLatencyComputed, |
3359 | RemLatency))) { |
3360 | Policy.ReduceLatency |= true; |
3361 | LLVM_DEBUG(dbgs() << " " << CurrZone.Available.getName() |
3362 | << " RemainingLatency " << RemLatency << " + " |
3363 | << CurrZone.getCurrCycle() << "c > CritPath " |
3364 | << Rem.CriticalPath << "\n" ); |
3365 | } |
3366 | // If the same resource is limiting inside and outside the zone, do nothing. |
3367 | if (CurrZone.getZoneCritResIdx() == OtherCritIdx) |
3368 | return; |
3369 | |
3370 | LLVM_DEBUG(if (CurrZone.isResourceLimited()) { |
3371 | dbgs() << " " << CurrZone.Available.getName() << " ResourceLimited: " |
3372 | << SchedModel->getResourceName(CurrZone.getZoneCritResIdx()) << "\n" ; |
3373 | } if (OtherResLimited) dbgs() |
3374 | << " RemainingLimit: " |
3375 | << SchedModel->getResourceName(OtherCritIdx) << "\n" ; |
3376 | if (!CurrZone.isResourceLimited() && !OtherResLimited) dbgs() |
3377 | << " Latency limited both directions.\n" ); |
3378 | |
3379 | if (CurrZone.isResourceLimited() && !Policy.ReduceResIdx) |
3380 | Policy.ReduceResIdx = CurrZone.getZoneCritResIdx(); |
3381 | |
3382 | if (OtherResLimited) |
3383 | Policy.DemandResIdx = OtherCritIdx; |
3384 | } |
3385 | |
3386 | #ifndef NDEBUG |
3387 | const char *GenericSchedulerBase::getReasonStr( |
3388 | GenericSchedulerBase::CandReason Reason) { |
3389 | // clang-format off |
3390 | switch (Reason) { |
3391 | case NoCand: return "NOCAND " ; |
3392 | case Only1: return "ONLY1 " ; |
3393 | case PhysReg: return "PHYS-REG " ; |
3394 | case RegExcess: return "REG-EXCESS" ; |
3395 | case RegCritical: return "REG-CRIT " ; |
3396 | case Stall: return "STALL " ; |
3397 | case Cluster: return "CLUSTER " ; |
3398 | case Weak: return "WEAK " ; |
3399 | case RegMax: return "REG-MAX " ; |
3400 | case ResourceReduce: return "RES-REDUCE" ; |
3401 | case ResourceDemand: return "RES-DEMAND" ; |
3402 | case TopDepthReduce: return "TOP-DEPTH " ; |
3403 | case TopPathReduce: return "TOP-PATH " ; |
3404 | case BotHeightReduce:return "BOT-HEIGHT" ; |
3405 | case BotPathReduce: return "BOT-PATH " ; |
3406 | case NodeOrder: return "ORDER " ; |
3407 | case FirstValid: return "FIRST " ; |
3408 | }; |
3409 | // clang-format on |
3410 | llvm_unreachable("Unknown reason!" ); |
3411 | } |
3412 | |
3413 | void GenericSchedulerBase::traceCandidate(const SchedCandidate &Cand) { |
3414 | PressureChange P; |
3415 | unsigned ResIdx = 0; |
3416 | unsigned Latency = 0; |
3417 | switch (Cand.Reason) { |
3418 | default: |
3419 | break; |
3420 | case RegExcess: |
3421 | P = Cand.RPDelta.Excess; |
3422 | break; |
3423 | case RegCritical: |
3424 | P = Cand.RPDelta.CriticalMax; |
3425 | break; |
3426 | case RegMax: |
3427 | P = Cand.RPDelta.CurrentMax; |
3428 | break; |
3429 | case ResourceReduce: |
3430 | ResIdx = Cand.Policy.ReduceResIdx; |
3431 | break; |
3432 | case ResourceDemand: |
3433 | ResIdx = Cand.Policy.DemandResIdx; |
3434 | break; |
3435 | case TopDepthReduce: |
3436 | Latency = Cand.SU->getDepth(); |
3437 | break; |
3438 | case TopPathReduce: |
3439 | Latency = Cand.SU->getHeight(); |
3440 | break; |
3441 | case BotHeightReduce: |
3442 | Latency = Cand.SU->getHeight(); |
3443 | break; |
3444 | case BotPathReduce: |
3445 | Latency = Cand.SU->getDepth(); |
3446 | break; |
3447 | } |
3448 | dbgs() << " Cand SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason); |
3449 | if (P.isValid()) |
3450 | dbgs() << " " << TRI->getRegPressureSetName(P.getPSet()) |
3451 | << ":" << P.getUnitInc() << " " ; |
3452 | else |
3453 | dbgs() << " " ; |
3454 | if (ResIdx) |
3455 | dbgs() << " " << SchedModel->getProcResource(ResIdx)->Name << " " ; |
3456 | else |
3457 | dbgs() << " " ; |
3458 | if (Latency) |
3459 | dbgs() << " " << Latency << " cycles " ; |
3460 | else |
3461 | dbgs() << " " ; |
3462 | dbgs() << '\n'; |
3463 | } |
3464 | #endif |
3465 | |
3466 | namespace llvm { |
3467 | /// Return true if this heuristic determines order. |
3468 | /// TODO: Consider refactor return type of these functions as integer or enum, |
3469 | /// as we may need to differentiate whether TryCand is better than Cand. |
3470 | bool tryLess(int TryVal, int CandVal, |
3471 | GenericSchedulerBase::SchedCandidate &TryCand, |
3472 | GenericSchedulerBase::SchedCandidate &Cand, |
3473 | GenericSchedulerBase::CandReason Reason) { |
3474 | if (TryVal < CandVal) { |
3475 | TryCand.Reason = Reason; |
3476 | return true; |
3477 | } |
3478 | if (TryVal > CandVal) { |
3479 | if (Cand.Reason > Reason) |
3480 | Cand.Reason = Reason; |
3481 | return true; |
3482 | } |
3483 | return false; |
3484 | } |
3485 | |
3486 | bool tryGreater(int TryVal, int CandVal, |
3487 | GenericSchedulerBase::SchedCandidate &TryCand, |
3488 | GenericSchedulerBase::SchedCandidate &Cand, |
3489 | GenericSchedulerBase::CandReason Reason) { |
3490 | if (TryVal > CandVal) { |
3491 | TryCand.Reason = Reason; |
3492 | return true; |
3493 | } |
3494 | if (TryVal < CandVal) { |
3495 | if (Cand.Reason > Reason) |
3496 | Cand.Reason = Reason; |
3497 | return true; |
3498 | } |
3499 | return false; |
3500 | } |
3501 | |
3502 | bool tryLatency(GenericSchedulerBase::SchedCandidate &TryCand, |
3503 | GenericSchedulerBase::SchedCandidate &Cand, |
3504 | SchedBoundary &Zone) { |
3505 | if (Zone.isTop()) { |
3506 | // Prefer the candidate with the lesser depth, but only if one of them has |
3507 | // depth greater than the total latency scheduled so far, otherwise either |
3508 | // of them could be scheduled now with no stall. |
3509 | if (std::max(a: TryCand.SU->getDepth(), b: Cand.SU->getDepth()) > |
3510 | Zone.getScheduledLatency()) { |
3511 | if (tryLess(TryVal: TryCand.SU->getDepth(), CandVal: Cand.SU->getDepth(), |
3512 | TryCand, Cand, Reason: GenericSchedulerBase::TopDepthReduce)) |
3513 | return true; |
3514 | } |
3515 | if (tryGreater(TryVal: TryCand.SU->getHeight(), CandVal: Cand.SU->getHeight(), |
3516 | TryCand, Cand, Reason: GenericSchedulerBase::TopPathReduce)) |
3517 | return true; |
3518 | } else { |
3519 | // Prefer the candidate with the lesser height, but only if one of them has |
3520 | // height greater than the total latency scheduled so far, otherwise either |
3521 | // of them could be scheduled now with no stall. |
3522 | if (std::max(a: TryCand.SU->getHeight(), b: Cand.SU->getHeight()) > |
3523 | Zone.getScheduledLatency()) { |
3524 | if (tryLess(TryVal: TryCand.SU->getHeight(), CandVal: Cand.SU->getHeight(), |
3525 | TryCand, Cand, Reason: GenericSchedulerBase::BotHeightReduce)) |
3526 | return true; |
3527 | } |
3528 | if (tryGreater(TryVal: TryCand.SU->getDepth(), CandVal: Cand.SU->getDepth(), |
3529 | TryCand, Cand, Reason: GenericSchedulerBase::BotPathReduce)) |
3530 | return true; |
3531 | } |
3532 | return false; |
3533 | } |
3534 | } // end namespace llvm |
3535 | |
3536 | static void tracePick(GenericSchedulerBase::CandReason Reason, bool IsTop, |
3537 | bool IsPostRA = false) { |
3538 | LLVM_DEBUG(dbgs() << "Pick " << (IsTop ? "Top " : "Bot " ) |
3539 | << GenericSchedulerBase::getReasonStr(Reason) << " [" |
3540 | << (IsPostRA ? "post-RA" : "pre-RA" ) << "]\n" ); |
3541 | |
3542 | if (IsPostRA) { |
3543 | if (IsTop) |
3544 | NumTopPostRA++; |
3545 | else |
3546 | NumBotPostRA++; |
3547 | |
3548 | switch (Reason) { |
3549 | case GenericScheduler::NoCand: |
3550 | NumNoCandPostRA++; |
3551 | return; |
3552 | case GenericScheduler::Only1: |
3553 | NumOnly1PostRA++; |
3554 | return; |
3555 | case GenericScheduler::PhysReg: |
3556 | NumPhysRegPostRA++; |
3557 | return; |
3558 | case GenericScheduler::RegExcess: |
3559 | NumRegExcessPostRA++; |
3560 | return; |
3561 | case GenericScheduler::RegCritical: |
3562 | NumRegCriticalPostRA++; |
3563 | return; |
3564 | case GenericScheduler::Stall: |
3565 | NumStallPostRA++; |
3566 | return; |
3567 | case GenericScheduler::Cluster: |
3568 | NumClusterPostRA++; |
3569 | return; |
3570 | case GenericScheduler::Weak: |
3571 | NumWeakPostRA++; |
3572 | return; |
3573 | case GenericScheduler::RegMax: |
3574 | NumRegMaxPostRA++; |
3575 | return; |
3576 | case GenericScheduler::ResourceReduce: |
3577 | NumResourceReducePostRA++; |
3578 | return; |
3579 | case GenericScheduler::ResourceDemand: |
3580 | NumResourceDemandPostRA++; |
3581 | return; |
3582 | case GenericScheduler::TopDepthReduce: |
3583 | NumTopDepthReducePostRA++; |
3584 | return; |
3585 | case GenericScheduler::TopPathReduce: |
3586 | NumTopPathReducePostRA++; |
3587 | return; |
3588 | case GenericScheduler::BotHeightReduce: |
3589 | NumBotHeightReducePostRA++; |
3590 | return; |
3591 | case GenericScheduler::BotPathReduce: |
3592 | NumBotPathReducePostRA++; |
3593 | return; |
3594 | case GenericScheduler::NodeOrder: |
3595 | NumNodeOrderPostRA++; |
3596 | return; |
3597 | case GenericScheduler::FirstValid: |
3598 | NumFirstValidPostRA++; |
3599 | return; |
3600 | }; |
3601 | } else { |
3602 | if (IsTop) |
3603 | NumTopPreRA++; |
3604 | else |
3605 | NumBotPreRA++; |
3606 | |
3607 | switch (Reason) { |
3608 | case GenericScheduler::NoCand: |
3609 | NumNoCandPreRA++; |
3610 | return; |
3611 | case GenericScheduler::Only1: |
3612 | NumOnly1PreRA++; |
3613 | return; |
3614 | case GenericScheduler::PhysReg: |
3615 | NumPhysRegPreRA++; |
3616 | return; |
3617 | case GenericScheduler::RegExcess: |
3618 | NumRegExcessPreRA++; |
3619 | return; |
3620 | case GenericScheduler::RegCritical: |
3621 | NumRegCriticalPreRA++; |
3622 | return; |
3623 | case GenericScheduler::Stall: |
3624 | NumStallPreRA++; |
3625 | return; |
3626 | case GenericScheduler::Cluster: |
3627 | NumClusterPreRA++; |
3628 | return; |
3629 | case GenericScheduler::Weak: |
3630 | NumWeakPreRA++; |
3631 | return; |
3632 | case GenericScheduler::RegMax: |
3633 | NumRegMaxPreRA++; |
3634 | return; |
3635 | case GenericScheduler::ResourceReduce: |
3636 | NumResourceReducePreRA++; |
3637 | return; |
3638 | case GenericScheduler::ResourceDemand: |
3639 | NumResourceDemandPreRA++; |
3640 | return; |
3641 | case GenericScheduler::TopDepthReduce: |
3642 | NumTopDepthReducePreRA++; |
3643 | return; |
3644 | case GenericScheduler::TopPathReduce: |
3645 | NumTopPathReducePreRA++; |
3646 | return; |
3647 | case GenericScheduler::BotHeightReduce: |
3648 | NumBotHeightReducePreRA++; |
3649 | return; |
3650 | case GenericScheduler::BotPathReduce: |
3651 | NumBotPathReducePreRA++; |
3652 | return; |
3653 | case GenericScheduler::NodeOrder: |
3654 | NumNodeOrderPreRA++; |
3655 | return; |
3656 | case GenericScheduler::FirstValid: |
3657 | NumFirstValidPreRA++; |
3658 | return; |
3659 | }; |
3660 | } |
3661 | llvm_unreachable("Unknown reason!" ); |
3662 | } |
3663 | |
3664 | static void tracePick(const GenericSchedulerBase::SchedCandidate &Cand, |
3665 | bool IsPostRA = false) { |
3666 | tracePick(Reason: Cand.Reason, IsTop: Cand.AtTop, IsPostRA); |
3667 | } |
3668 | |
3669 | void GenericScheduler::initialize(ScheduleDAGMI *dag) { |
3670 | assert(dag->hasVRegLiveness() && |
3671 | "(PreRA)GenericScheduler needs vreg liveness" ); |
3672 | DAG = static_cast<ScheduleDAGMILive*>(dag); |
3673 | SchedModel = DAG->getSchedModel(); |
3674 | TRI = DAG->TRI; |
3675 | |
3676 | if (RegionPolicy.ComputeDFSResult) |
3677 | DAG->computeDFSResult(); |
3678 | |
3679 | Rem.init(DAG, SchedModel); |
3680 | Top.init(dag: DAG, smodel: SchedModel, rem: &Rem); |
3681 | Bot.init(dag: DAG, smodel: SchedModel, rem: &Rem); |
3682 | |
3683 | // Initialize resource counts. |
3684 | |
3685 | // Initialize the HazardRecognizers. If itineraries don't exist, are empty, or |
3686 | // are disabled, then these HazardRecs will be disabled. |
3687 | const InstrItineraryData *Itin = SchedModel->getInstrItineraries(); |
3688 | if (!Top.HazardRec) { |
3689 | Top.HazardRec = DAG->TII->CreateTargetMIHazardRecognizer(Itin, DAG); |
3690 | } |
3691 | if (!Bot.HazardRec) { |
3692 | Bot.HazardRec = DAG->TII->CreateTargetMIHazardRecognizer(Itin, DAG); |
3693 | } |
3694 | TopCand.SU = nullptr; |
3695 | BotCand.SU = nullptr; |
3696 | |
3697 | TopCluster = nullptr; |
3698 | BotCluster = nullptr; |
3699 | } |
3700 | |
3701 | /// Initialize the per-region scheduling policy. |
3702 | void GenericScheduler::initPolicy(MachineBasicBlock::iterator Begin, |
3703 | MachineBasicBlock::iterator End, |
3704 | unsigned NumRegionInstrs) { |
3705 | const MachineFunction &MF = *Begin->getMF(); |
3706 | const TargetLowering *TLI = MF.getSubtarget().getTargetLowering(); |
3707 | |
3708 | // Avoid setting up the register pressure tracker for small regions to save |
3709 | // compile time. As a rough heuristic, only track pressure when the number of |
3710 | // schedulable instructions exceeds half the allocatable integer register file |
3711 | // that is the largest legal integer regiser type. |
3712 | RegionPolicy.ShouldTrackPressure = true; |
3713 | for (unsigned VT = MVT::i64; VT > (unsigned)MVT::i1; --VT) { |
3714 | MVT::SimpleValueType LegalIntVT = (MVT::SimpleValueType)VT; |
3715 | if (TLI->isTypeLegal(VT: LegalIntVT)) { |
3716 | unsigned NIntRegs = Context->RegClassInfo->getNumAllocatableRegs( |
3717 | RC: TLI->getRegClassFor(VT: LegalIntVT)); |
3718 | RegionPolicy.ShouldTrackPressure = NumRegionInstrs > (NIntRegs / 2); |
3719 | break; |
3720 | } |
3721 | } |
3722 | |
3723 | // For generic targets, we default to bottom-up, because it's simpler and more |
3724 | // compile-time optimizations have been implemented in that direction. |
3725 | RegionPolicy.OnlyBottomUp = true; |
3726 | |
3727 | // Allow the subtarget to override default policy. |
3728 | MF.getSubtarget().overrideSchedPolicy(Policy&: RegionPolicy, NumRegionInstrs); |
3729 | |
3730 | // After subtarget overrides, apply command line options. |
3731 | if (!EnableRegPressure) { |
3732 | RegionPolicy.ShouldTrackPressure = false; |
3733 | RegionPolicy.ShouldTrackLaneMasks = false; |
3734 | } |
3735 | |
3736 | if (PreRADirection == MISched::TopDown) { |
3737 | RegionPolicy.OnlyTopDown = true; |
3738 | RegionPolicy.OnlyBottomUp = false; |
3739 | } else if (PreRADirection == MISched::BottomUp) { |
3740 | RegionPolicy.OnlyTopDown = false; |
3741 | RegionPolicy.OnlyBottomUp = true; |
3742 | } else if (PreRADirection == MISched::Bidirectional) { |
3743 | RegionPolicy.OnlyBottomUp = false; |
3744 | RegionPolicy.OnlyTopDown = false; |
3745 | } |
3746 | |
3747 | BotIdx = NumRegionInstrs - 1; |
3748 | this->NumRegionInstrs = NumRegionInstrs; |
3749 | } |
3750 | |
3751 | void GenericScheduler::dumpPolicy() const { |
3752 | // Cannot completely remove virtual function even in release mode. |
3753 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
3754 | dbgs() << "GenericScheduler RegionPolicy: " |
3755 | << " ShouldTrackPressure=" << RegionPolicy.ShouldTrackPressure |
3756 | << " OnlyTopDown=" << RegionPolicy.OnlyTopDown |
3757 | << " OnlyBottomUp=" << RegionPolicy.OnlyBottomUp |
3758 | << "\n" ; |
3759 | #endif |
3760 | } |
3761 | |
3762 | /// Set IsAcyclicLatencyLimited if the acyclic path is longer than the cyclic |
3763 | /// critical path by more cycles than it takes to drain the instruction buffer. |
3764 | /// We estimate an upper bounds on in-flight instructions as: |
3765 | /// |
3766 | /// CyclesPerIteration = max( CyclicPath, Loop-Resource-Height ) |
3767 | /// InFlightIterations = AcyclicPath / CyclesPerIteration |
3768 | /// InFlightResources = InFlightIterations * LoopResources |
3769 | /// |
3770 | /// TODO: Check execution resources in addition to IssueCount. |
3771 | void GenericScheduler::checkAcyclicLatency() { |
3772 | if (Rem.CyclicCritPath == 0 || Rem.CyclicCritPath >= Rem.CriticalPath) |
3773 | return; |
3774 | |
3775 | // Scaled number of cycles per loop iteration. |
3776 | unsigned IterCount = |
3777 | std::max(a: Rem.CyclicCritPath * SchedModel->getLatencyFactor(), |
3778 | b: Rem.RemIssueCount); |
3779 | // Scaled acyclic critical path. |
3780 | unsigned AcyclicCount = Rem.CriticalPath * SchedModel->getLatencyFactor(); |
3781 | // InFlightCount = (AcyclicPath / IterCycles) * InstrPerLoop |
3782 | unsigned InFlightCount = |
3783 | (AcyclicCount * Rem.RemIssueCount + IterCount-1) / IterCount; |
3784 | unsigned BufferLimit = |
3785 | SchedModel->getMicroOpBufferSize() * SchedModel->getMicroOpFactor(); |
3786 | |
3787 | Rem.IsAcyclicLatencyLimited = InFlightCount > BufferLimit; |
3788 | |
3789 | LLVM_DEBUG( |
3790 | dbgs() << "IssueCycles=" |
3791 | << Rem.RemIssueCount / SchedModel->getLatencyFactor() << "c " |
3792 | << "IterCycles=" << IterCount / SchedModel->getLatencyFactor() |
3793 | << "c NumIters=" << (AcyclicCount + IterCount - 1) / IterCount |
3794 | << " InFlight=" << InFlightCount / SchedModel->getMicroOpFactor() |
3795 | << "m BufferLim=" << SchedModel->getMicroOpBufferSize() << "m\n" ; |
3796 | if (Rem.IsAcyclicLatencyLimited) dbgs() << " ACYCLIC LATENCY LIMIT\n" ); |
3797 | } |
3798 | |
3799 | void GenericScheduler::registerRoots() { |
3800 | Rem.CriticalPath = DAG->ExitSU.getDepth(); |
3801 | |
3802 | // Some roots may not feed into ExitSU. Check all of them in case. |
3803 | for (const SUnit *SU : Bot.Available) { |
3804 | if (SU->getDepth() > Rem.CriticalPath) |
3805 | Rem.CriticalPath = SU->getDepth(); |
3806 | } |
3807 | LLVM_DEBUG(dbgs() << "Critical Path(GS-RR ): " << Rem.CriticalPath << '\n'); |
3808 | if (DumpCriticalPathLength) { |
3809 | errs() << "Critical Path(GS-RR ): " << Rem.CriticalPath << " \n" ; |
3810 | } |
3811 | |
3812 | if (EnableCyclicPath && SchedModel->getMicroOpBufferSize() > 0) { |
3813 | Rem.CyclicCritPath = DAG->computeCyclicCriticalPath(); |
3814 | checkAcyclicLatency(); |
3815 | } |
3816 | } |
3817 | |
3818 | namespace llvm { |
3819 | bool tryPressure(const PressureChange &TryP, |
3820 | const PressureChange &CandP, |
3821 | GenericSchedulerBase::SchedCandidate &TryCand, |
3822 | GenericSchedulerBase::SchedCandidate &Cand, |
3823 | GenericSchedulerBase::CandReason Reason, |
3824 | const TargetRegisterInfo *TRI, |
3825 | const MachineFunction &MF) { |
3826 | // If one candidate decreases and the other increases, go with it. |
3827 | // Invalid candidates have UnitInc==0. |
3828 | if (tryGreater(TryVal: TryP.getUnitInc() < 0, CandVal: CandP.getUnitInc() < 0, TryCand, Cand, |
3829 | Reason)) { |
3830 | return true; |
3831 | } |
3832 | // Do not compare the magnitude of pressure changes between top and bottom |
3833 | // boundary. |
3834 | if (Cand.AtTop != TryCand.AtTop) |
3835 | return false; |
3836 | |
3837 | // If both candidates affect the same set in the same boundary, go with the |
3838 | // smallest increase. |
3839 | unsigned TryPSet = TryP.getPSetOrMax(); |
3840 | unsigned CandPSet = CandP.getPSetOrMax(); |
3841 | if (TryPSet == CandPSet) { |
3842 | return tryLess(TryVal: TryP.getUnitInc(), CandVal: CandP.getUnitInc(), TryCand, Cand, |
3843 | Reason); |
3844 | } |
3845 | |
3846 | int TryRank = TryP.isValid() ? TRI->getRegPressureSetScore(MF, PSetID: TryPSet) : |
3847 | std::numeric_limits<int>::max(); |
3848 | |
3849 | int CandRank = CandP.isValid() ? TRI->getRegPressureSetScore(MF, PSetID: CandPSet) : |
3850 | std::numeric_limits<int>::max(); |
3851 | |
3852 | // If the candidates are decreasing pressure, reverse priority. |
3853 | if (TryP.getUnitInc() < 0) |
3854 | std::swap(a&: TryRank, b&: CandRank); |
3855 | return tryGreater(TryVal: TryRank, CandVal: CandRank, TryCand, Cand, Reason); |
3856 | } |
3857 | |
3858 | unsigned getWeakLeft(const SUnit *SU, bool isTop) { |
3859 | return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft; |
3860 | } |
3861 | |
3862 | /// Minimize physical register live ranges. Regalloc wants them adjacent to |
3863 | /// their physreg def/use. |
3864 | /// |
3865 | /// FIXME: This is an unnecessary check on the critical path. Most are root/leaf |
3866 | /// copies which can be prescheduled. The rest (e.g. x86 MUL) could be bundled |
3867 | /// with the operation that produces or consumes the physreg. We'll do this when |
3868 | /// regalloc has support for parallel copies. |
3869 | int biasPhysReg(const SUnit *SU, bool isTop) { |
3870 | const MachineInstr *MI = SU->getInstr(); |
3871 | |
3872 | if (MI->isCopy()) { |
3873 | unsigned ScheduledOper = isTop ? 1 : 0; |
3874 | unsigned UnscheduledOper = isTop ? 0 : 1; |
3875 | // If we have already scheduled the physreg produce/consumer, immediately |
3876 | // schedule the copy. |
3877 | if (MI->getOperand(i: ScheduledOper).getReg().isPhysical()) |
3878 | return 1; |
3879 | // If the physreg is at the boundary, defer it. Otherwise schedule it |
3880 | // immediately to free the dependent. We can hoist the copy later. |
3881 | bool AtBoundary = isTop ? !SU->NumSuccsLeft : !SU->NumPredsLeft; |
3882 | if (MI->getOperand(i: UnscheduledOper).getReg().isPhysical()) |
3883 | return AtBoundary ? -1 : 1; |
3884 | } |
3885 | |
3886 | if (MI->isMoveImmediate()) { |
3887 | // If we have a move immediate and all successors have been assigned, bias |
3888 | // towards scheduling this later. Make sure all register defs are to |
3889 | // physical registers. |
3890 | bool DoBias = true; |
3891 | for (const MachineOperand &Op : MI->defs()) { |
3892 | if (Op.isReg() && !Op.getReg().isPhysical()) { |
3893 | DoBias = false; |
3894 | break; |
3895 | } |
3896 | } |
3897 | |
3898 | if (DoBias) |
3899 | return isTop ? -1 : 1; |
3900 | } |
3901 | |
3902 | return 0; |
3903 | } |
3904 | } // end namespace llvm |
3905 | |
3906 | void GenericScheduler::initCandidate(SchedCandidate &Cand, SUnit *SU, |
3907 | bool AtTop, |
3908 | const RegPressureTracker &RPTracker, |
3909 | RegPressureTracker &TempTracker) { |
3910 | Cand.SU = SU; |
3911 | Cand.AtTop = AtTop; |
3912 | if (DAG->isTrackingPressure()) { |
3913 | if (AtTop) { |
3914 | TempTracker.getMaxDownwardPressureDelta( |
3915 | MI: Cand.SU->getInstr(), |
3916 | Delta&: Cand.RPDelta, |
3917 | CriticalPSets: DAG->getRegionCriticalPSets(), |
3918 | MaxPressureLimit: DAG->getRegPressure().MaxSetPressure); |
3919 | } else { |
3920 | if (VerifyScheduling) { |
3921 | TempTracker.getMaxUpwardPressureDelta( |
3922 | MI: Cand.SU->getInstr(), |
3923 | PDiff: &DAG->getPressureDiff(SU: Cand.SU), |
3924 | Delta&: Cand.RPDelta, |
3925 | CriticalPSets: DAG->getRegionCriticalPSets(), |
3926 | MaxPressureLimit: DAG->getRegPressure().MaxSetPressure); |
3927 | } else { |
3928 | RPTracker.getUpwardPressureDelta( |
3929 | MI: Cand.SU->getInstr(), |
3930 | PDiff&: DAG->getPressureDiff(SU: Cand.SU), |
3931 | Delta&: Cand.RPDelta, |
3932 | CriticalPSets: DAG->getRegionCriticalPSets(), |
3933 | MaxPressureLimit: DAG->getRegPressure().MaxSetPressure); |
3934 | } |
3935 | } |
3936 | } |
3937 | LLVM_DEBUG(if (Cand.RPDelta.Excess.isValid()) dbgs() |
3938 | << " Try SU(" << Cand.SU->NodeNum << ") " |
3939 | << TRI->getRegPressureSetName(Cand.RPDelta.Excess.getPSet()) << ":" |
3940 | << Cand.RPDelta.Excess.getUnitInc() << "\n" ); |
3941 | } |
3942 | |
3943 | /// Apply a set of heuristics to a new candidate. Heuristics are currently |
3944 | /// hierarchical. This may be more efficient than a graduated cost model because |
3945 | /// we don't need to evaluate all aspects of the model for each node in the |
3946 | /// queue. But it's really done to make the heuristics easier to debug and |
3947 | /// statistically analyze. |
3948 | /// |
3949 | /// \param Cand provides the policy and current best candidate. |
3950 | /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized. |
3951 | /// \param Zone describes the scheduled zone that we are extending, or nullptr |
3952 | /// if Cand is from a different zone than TryCand. |
3953 | /// \return \c true if TryCand is better than Cand (Reason is NOT NoCand) |
3954 | bool GenericScheduler::tryCandidate(SchedCandidate &Cand, |
3955 | SchedCandidate &TryCand, |
3956 | SchedBoundary *Zone) const { |
3957 | // Initialize the candidate if needed. |
3958 | if (!Cand.isValid()) { |
3959 | TryCand.Reason = FirstValid; |
3960 | return true; |
3961 | } |
3962 | |
3963 | // Bias PhysReg Defs and copies to their uses and defined respectively. |
3964 | if (tryGreater(TryVal: biasPhysReg(SU: TryCand.SU, isTop: TryCand.AtTop), |
3965 | CandVal: biasPhysReg(SU: Cand.SU, isTop: Cand.AtTop), TryCand, Cand, Reason: PhysReg)) |
3966 | return TryCand.Reason != NoCand; |
3967 | |
3968 | // Avoid exceeding the target's limit. |
3969 | if (DAG->isTrackingPressure() && tryPressure(TryP: TryCand.RPDelta.Excess, |
3970 | CandP: Cand.RPDelta.Excess, |
3971 | TryCand, Cand, Reason: RegExcess, TRI, |
3972 | MF: DAG->MF)) |
3973 | return TryCand.Reason != NoCand; |
3974 | |
3975 | // Avoid increasing the max critical pressure in the scheduled region. |
3976 | if (DAG->isTrackingPressure() && tryPressure(TryP: TryCand.RPDelta.CriticalMax, |
3977 | CandP: Cand.RPDelta.CriticalMax, |
3978 | TryCand, Cand, Reason: RegCritical, TRI, |
3979 | MF: DAG->MF)) |
3980 | return TryCand.Reason != NoCand; |
3981 | |
3982 | // We only compare a subset of features when comparing nodes between |
3983 | // Top and Bottom boundary. Some properties are simply incomparable, in many |
3984 | // other instances we should only override the other boundary if something |
3985 | // is a clear good pick on one boundary. Skip heuristics that are more |
3986 | // "tie-breaking" in nature. |
3987 | bool SameBoundary = Zone != nullptr; |
3988 | if (SameBoundary) { |
3989 | // For loops that are acyclic path limited, aggressively schedule for |
3990 | // latency. Within an single cycle, whenever CurrMOps > 0, allow normal |
3991 | // heuristics to take precedence. |
3992 | if (Rem.IsAcyclicLatencyLimited && !Zone->getCurrMOps() && |
3993 | tryLatency(TryCand, Cand, Zone&: *Zone)) |
3994 | return TryCand.Reason != NoCand; |
3995 | |
3996 | // Prioritize instructions that read unbuffered resources by stall cycles. |
3997 | if (tryLess(TryVal: Zone->getLatencyStallCycles(SU: TryCand.SU), |
3998 | CandVal: Zone->getLatencyStallCycles(SU: Cand.SU), TryCand, Cand, Reason: Stall)) |
3999 | return TryCand.Reason != NoCand; |
4000 | } |
4001 | |
4002 | // Keep clustered nodes together to encourage downstream peephole |
4003 | // optimizations which may reduce resource requirements. |
4004 | // |
4005 | // This is a best effort to set things up for a post-RA pass. Optimizations |
4006 | // like generating loads of multiple registers should ideally be done within |
4007 | // the scheduler pass by combining the loads during DAG postprocessing. |
4008 | const ClusterInfo *CandCluster = Cand.AtTop ? TopCluster : BotCluster; |
4009 | const ClusterInfo *TryCandCluster = TryCand.AtTop ? TopCluster : BotCluster; |
4010 | if (tryGreater(TryVal: TryCandCluster && TryCandCluster->contains(Ptr: TryCand.SU), |
4011 | CandVal: CandCluster && CandCluster->contains(Ptr: Cand.SU), TryCand, Cand, |
4012 | Reason: Cluster)) |
4013 | return TryCand.Reason != NoCand; |
4014 | |
4015 | if (SameBoundary) { |
4016 | // Weak edges are for clustering and other constraints. |
4017 | if (tryLess(TryVal: getWeakLeft(SU: TryCand.SU, isTop: TryCand.AtTop), |
4018 | CandVal: getWeakLeft(SU: Cand.SU, isTop: Cand.AtTop), |
4019 | TryCand, Cand, Reason: Weak)) |
4020 | return TryCand.Reason != NoCand; |
4021 | } |
4022 | |
4023 | // Avoid increasing the max pressure of the entire region. |
4024 | if (DAG->isTrackingPressure() && tryPressure(TryP: TryCand.RPDelta.CurrentMax, |
4025 | CandP: Cand.RPDelta.CurrentMax, |
4026 | TryCand, Cand, Reason: RegMax, TRI, |
4027 | MF: DAG->MF)) |
4028 | return TryCand.Reason != NoCand; |
4029 | |
4030 | if (SameBoundary) { |
4031 | // Avoid critical resource consumption and balance the schedule. |
4032 | TryCand.initResourceDelta(DAG, SchedModel); |
4033 | if (tryLess(TryVal: TryCand.ResDelta.CritResources, CandVal: Cand.ResDelta.CritResources, |
4034 | TryCand, Cand, Reason: ResourceReduce)) |
4035 | return TryCand.Reason != NoCand; |
4036 | if (tryGreater(TryVal: TryCand.ResDelta.DemandedResources, |
4037 | CandVal: Cand.ResDelta.DemandedResources, |
4038 | TryCand, Cand, Reason: ResourceDemand)) |
4039 | return TryCand.Reason != NoCand; |
4040 | |
4041 | // Avoid serializing long latency dependence chains. |
4042 | // For acyclic path limited loops, latency was already checked above. |
4043 | if (!RegionPolicy.DisableLatencyHeuristic && TryCand.Policy.ReduceLatency && |
4044 | !Rem.IsAcyclicLatencyLimited && tryLatency(TryCand, Cand, Zone&: *Zone)) |
4045 | return TryCand.Reason != NoCand; |
4046 | |
4047 | // Fall through to original instruction order. |
4048 | if ((Zone->isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum) |
4049 | || (!Zone->isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) { |
4050 | TryCand.Reason = NodeOrder; |
4051 | return true; |
4052 | } |
4053 | } |
4054 | |
4055 | return false; |
4056 | } |
4057 | |
4058 | /// Pick the best candidate from the queue. |
4059 | /// |
4060 | /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during |
4061 | /// DAG building. To adjust for the current scheduling location we need to |
4062 | /// maintain the number of vreg uses remaining to be top-scheduled. |
4063 | void GenericScheduler::pickNodeFromQueue(SchedBoundary &Zone, |
4064 | const CandPolicy &ZonePolicy, |
4065 | const RegPressureTracker &RPTracker, |
4066 | SchedCandidate &Cand) { |
4067 | // getMaxPressureDelta temporarily modifies the tracker. |
4068 | RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker); |
4069 | |
4070 | ReadyQueue &Q = Zone.Available; |
4071 | for (SUnit *SU : Q) { |
4072 | |
4073 | SchedCandidate TryCand(ZonePolicy); |
4074 | initCandidate(Cand&: TryCand, SU, AtTop: Zone.isTop(), RPTracker, TempTracker); |
4075 | // Pass SchedBoundary only when comparing nodes from the same boundary. |
4076 | SchedBoundary *ZoneArg = Cand.AtTop == TryCand.AtTop ? &Zone : nullptr; |
4077 | if (tryCandidate(Cand, TryCand, Zone: ZoneArg)) { |
4078 | // Initialize resource delta if needed in case future heuristics query it. |
4079 | if (TryCand.ResDelta == SchedResourceDelta()) |
4080 | TryCand.initResourceDelta(DAG, SchedModel); |
4081 | Cand.setBest(TryCand); |
4082 | LLVM_DEBUG(traceCandidate(Cand)); |
4083 | } |
4084 | } |
4085 | } |
4086 | |
4087 | /// Pick the best candidate node from either the top or bottom queue. |
4088 | SUnit *GenericScheduler::pickNodeBidirectional(bool &IsTopNode) { |
4089 | // Schedule as far as possible in the direction of no choice. This is most |
4090 | // efficient, but also provides the best heuristics for CriticalPSets. |
4091 | if (SUnit *SU = Bot.pickOnlyChoice()) { |
4092 | IsTopNode = false; |
4093 | tracePick(Reason: Only1, /*IsTopNode=*/IsTop: false); |
4094 | return SU; |
4095 | } |
4096 | if (SUnit *SU = Top.pickOnlyChoice()) { |
4097 | IsTopNode = true; |
4098 | tracePick(Reason: Only1, /*IsTopNode=*/IsTop: true); |
4099 | return SU; |
4100 | } |
4101 | // Set the bottom-up policy based on the state of the current bottom zone and |
4102 | // the instructions outside the zone, including the top zone. |
4103 | CandPolicy BotPolicy; |
4104 | setPolicy(Policy&: BotPolicy, /*IsPostRA=*/false, CurrZone&: Bot, OtherZone: &Top); |
4105 | // Set the top-down policy based on the state of the current top zone and |
4106 | // the instructions outside the zone, including the bottom zone. |
4107 | CandPolicy TopPolicy; |
4108 | setPolicy(Policy&: TopPolicy, /*IsPostRA=*/false, CurrZone&: Top, OtherZone: &Bot); |
4109 | |
4110 | // See if BotCand is still valid (because we previously scheduled from Top). |
4111 | LLVM_DEBUG(dbgs() << "Picking from Bot:\n" ); |
4112 | if (!BotCand.isValid() || BotCand.SU->isScheduled || |
4113 | BotCand.Policy != BotPolicy) { |
4114 | BotCand.reset(NewPolicy: CandPolicy()); |
4115 | pickNodeFromQueue(Zone&: Bot, ZonePolicy: BotPolicy, RPTracker: DAG->getBotRPTracker(), Cand&: BotCand); |
4116 | assert(BotCand.Reason != NoCand && "failed to find the first candidate" ); |
4117 | } else { |
4118 | LLVM_DEBUG(traceCandidate(BotCand)); |
4119 | #ifndef NDEBUG |
4120 | if (VerifyScheduling) { |
4121 | SchedCandidate TCand; |
4122 | TCand.reset(CandPolicy()); |
4123 | pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), TCand); |
4124 | assert(TCand.SU == BotCand.SU && |
4125 | "Last pick result should correspond to re-picking right now" ); |
4126 | } |
4127 | #endif |
4128 | } |
4129 | |
4130 | // Check if the top Q has a better candidate. |
4131 | LLVM_DEBUG(dbgs() << "Picking from Top:\n" ); |
4132 | if (!TopCand.isValid() || TopCand.SU->isScheduled || |
4133 | TopCand.Policy != TopPolicy) { |
4134 | TopCand.reset(NewPolicy: CandPolicy()); |
4135 | pickNodeFromQueue(Zone&: Top, ZonePolicy: TopPolicy, RPTracker: DAG->getTopRPTracker(), Cand&: TopCand); |
4136 | assert(TopCand.Reason != NoCand && "failed to find the first candidate" ); |
4137 | } else { |
4138 | LLVM_DEBUG(traceCandidate(TopCand)); |
4139 | #ifndef NDEBUG |
4140 | if (VerifyScheduling) { |
4141 | SchedCandidate TCand; |
4142 | TCand.reset(CandPolicy()); |
4143 | pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TCand); |
4144 | assert(TCand.SU == TopCand.SU && |
4145 | "Last pick result should correspond to re-picking right now" ); |
4146 | } |
4147 | #endif |
4148 | } |
4149 | |
4150 | // Pick best from BotCand and TopCand. |
4151 | assert(BotCand.isValid()); |
4152 | assert(TopCand.isValid()); |
4153 | SchedCandidate Cand = BotCand; |
4154 | TopCand.Reason = NoCand; |
4155 | if (tryCandidate(Cand, TryCand&: TopCand, Zone: nullptr)) { |
4156 | Cand.setBest(TopCand); |
4157 | LLVM_DEBUG(traceCandidate(Cand)); |
4158 | } |
4159 | |
4160 | IsTopNode = Cand.AtTop; |
4161 | tracePick(Cand); |
4162 | return Cand.SU; |
4163 | } |
4164 | |
4165 | /// Pick the best node to balance the schedule. Implements MachineSchedStrategy. |
4166 | SUnit *GenericScheduler::pickNode(bool &IsTopNode) { |
4167 | if (DAG->top() == DAG->bottom()) { |
4168 | assert(Top.Available.empty() && Top.Pending.empty() && |
4169 | Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage" ); |
4170 | return nullptr; |
4171 | } |
4172 | SUnit *SU; |
4173 | do { |
4174 | if (RegionPolicy.OnlyTopDown) { |
4175 | SU = Top.pickOnlyChoice(); |
4176 | if (!SU) { |
4177 | CandPolicy NoPolicy; |
4178 | TopCand.reset(NewPolicy: NoPolicy); |
4179 | pickNodeFromQueue(Zone&: Top, ZonePolicy: NoPolicy, RPTracker: DAG->getTopRPTracker(), Cand&: TopCand); |
4180 | assert(TopCand.Reason != NoCand && "failed to find a candidate" ); |
4181 | tracePick(Cand: TopCand); |
4182 | SU = TopCand.SU; |
4183 | } |
4184 | IsTopNode = true; |
4185 | } else if (RegionPolicy.OnlyBottomUp) { |
4186 | SU = Bot.pickOnlyChoice(); |
4187 | if (!SU) { |
4188 | CandPolicy NoPolicy; |
4189 | BotCand.reset(NewPolicy: NoPolicy); |
4190 | pickNodeFromQueue(Zone&: Bot, ZonePolicy: NoPolicy, RPTracker: DAG->getBotRPTracker(), Cand&: BotCand); |
4191 | assert(BotCand.Reason != NoCand && "failed to find a candidate" ); |
4192 | tracePick(Cand: BotCand); |
4193 | SU = BotCand.SU; |
4194 | } |
4195 | IsTopNode = false; |
4196 | } else { |
4197 | SU = pickNodeBidirectional(IsTopNode); |
4198 | } |
4199 | } while (SU->isScheduled); |
4200 | |
4201 | // If IsTopNode, then SU is in Top.Available and must be removed. Otherwise, |
4202 | // if isTopReady(), then SU is in either Top.Available or Top.Pending. |
4203 | // If !IsTopNode, then SU is in Bot.Available and must be removed. Otherwise, |
4204 | // if isBottomReady(), then SU is in either Bot.Available or Bot.Pending. |
4205 | // |
4206 | // It is coincidental when !IsTopNode && isTopReady or when IsTopNode && |
4207 | // isBottomReady. That is, it didn't factor into the decision to choose SU |
4208 | // because it isTopReady or isBottomReady, respectively. In fact, if the |
4209 | // RegionPolicy is OnlyTopDown or OnlyBottomUp, then the Bot queues and Top |
4210 | // queues respectivley contain the original roots and don't get updated when |
4211 | // picking a node. So if SU isTopReady on a OnlyBottomUp pick, then it was |
4212 | // because we schduled everything but the top roots. Conversley, if SU |
4213 | // isBottomReady on OnlyTopDown, then it was because we scheduled everything |
4214 | // but the bottom roots. If its in a queue even coincidentally, it should be |
4215 | // removed so it does not get re-picked in a subsequent pickNode call. |
4216 | if (SU->isTopReady()) |
4217 | Top.removeReady(SU); |
4218 | if (SU->isBottomReady()) |
4219 | Bot.removeReady(SU); |
4220 | |
4221 | LLVM_DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " |
4222 | << *SU->getInstr()); |
4223 | |
4224 | if (IsTopNode) { |
4225 | if (SU->NodeNum == TopIdx++) |
4226 | ++NumInstrsInSourceOrderPreRA; |
4227 | } else { |
4228 | assert(BotIdx < NumRegionInstrs && "out of bounds" ); |
4229 | if (SU->NodeNum == BotIdx--) |
4230 | ++NumInstrsInSourceOrderPreRA; |
4231 | } |
4232 | |
4233 | NumInstrsScheduledPreRA += 1; |
4234 | |
4235 | return SU; |
4236 | } |
4237 | |
4238 | void GenericScheduler::reschedulePhysReg(SUnit *SU, bool isTop) { |
4239 | MachineBasicBlock::iterator InsertPos = SU->getInstr(); |
4240 | if (!isTop) |
4241 | ++InsertPos; |
4242 | SmallVectorImpl<SDep> &Deps = isTop ? SU->Preds : SU->Succs; |
4243 | |
4244 | // Find already scheduled copies with a single physreg dependence and move |
4245 | // them just above the scheduled instruction. |
4246 | for (SDep &Dep : Deps) { |
4247 | if (Dep.getKind() != SDep::Data || !Dep.getReg().isPhysical()) |
4248 | continue; |
4249 | SUnit *DepSU = Dep.getSUnit(); |
4250 | if (isTop ? DepSU->Succs.size() > 1 : DepSU->Preds.size() > 1) |
4251 | continue; |
4252 | MachineInstr *Copy = DepSU->getInstr(); |
4253 | if (!Copy->isCopy() && !Copy->isMoveImmediate()) |
4254 | continue; |
4255 | LLVM_DEBUG(dbgs() << " Rescheduling physreg copy " ; |
4256 | DAG->dumpNode(*Dep.getSUnit())); |
4257 | DAG->moveInstruction(MI: Copy, InsertPos); |
4258 | } |
4259 | } |
4260 | |
4261 | /// Update the scheduler's state after scheduling a node. This is the same node |
4262 | /// that was just returned by pickNode(). However, ScheduleDAGMILive needs to |
4263 | /// update it's state based on the current cycle before MachineSchedStrategy |
4264 | /// does. |
4265 | /// |
4266 | /// FIXME: Eventually, we may bundle physreg copies rather than rescheduling |
4267 | /// them here. See comments in biasPhysReg. |
4268 | void GenericScheduler::schedNode(SUnit *SU, bool IsTopNode) { |
4269 | if (IsTopNode) { |
4270 | SU->TopReadyCycle = std::max(a: SU->TopReadyCycle, b: Top.getCurrCycle()); |
4271 | TopCluster = DAG->getCluster(Idx: SU->ParentClusterIdx); |
4272 | LLVM_DEBUG(if (TopCluster) { |
4273 | dbgs() << " Top Cluster: " ; |
4274 | for (auto *N : *TopCluster) |
4275 | dbgs() << N->NodeNum << '\t'; |
4276 | dbgs() << '\n'; |
4277 | }); |
4278 | Top.bumpNode(SU); |
4279 | if (SU->hasPhysRegUses) |
4280 | reschedulePhysReg(SU, isTop: true); |
4281 | } else { |
4282 | SU->BotReadyCycle = std::max(a: SU->BotReadyCycle, b: Bot.getCurrCycle()); |
4283 | BotCluster = DAG->getCluster(Idx: SU->ParentClusterIdx); |
4284 | LLVM_DEBUG(if (BotCluster) { |
4285 | dbgs() << " Bot Cluster: " ; |
4286 | for (auto *N : *BotCluster) |
4287 | dbgs() << N->NodeNum << '\t'; |
4288 | dbgs() << '\n'; |
4289 | }); |
4290 | Bot.bumpNode(SU); |
4291 | if (SU->hasPhysRegDefs) |
4292 | reschedulePhysReg(SU, isTop: false); |
4293 | } |
4294 | } |
4295 | |
4296 | static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C) { |
4297 | return createSchedLive(C); |
4298 | } |
4299 | |
4300 | static MachineSchedRegistry |
4301 | GenericSchedRegistry("converge" , "Standard converging scheduler." , |
4302 | createConvergingSched); |
4303 | |
4304 | //===----------------------------------------------------------------------===// |
4305 | // PostGenericScheduler - Generic PostRA implementation of MachineSchedStrategy. |
4306 | //===----------------------------------------------------------------------===// |
4307 | |
4308 | void PostGenericScheduler::initialize(ScheduleDAGMI *Dag) { |
4309 | DAG = Dag; |
4310 | SchedModel = DAG->getSchedModel(); |
4311 | TRI = DAG->TRI; |
4312 | |
4313 | Rem.init(DAG, SchedModel); |
4314 | Top.init(dag: DAG, smodel: SchedModel, rem: &Rem); |
4315 | Bot.init(dag: DAG, smodel: SchedModel, rem: &Rem); |
4316 | |
4317 | // Initialize the HazardRecognizers. If itineraries don't exist, are empty, |
4318 | // or are disabled, then these HazardRecs will be disabled. |
4319 | const InstrItineraryData *Itin = SchedModel->getInstrItineraries(); |
4320 | if (!Top.HazardRec) { |
4321 | Top.HazardRec = DAG->TII->CreateTargetMIHazardRecognizer(Itin, DAG); |
4322 | } |
4323 | if (!Bot.HazardRec) { |
4324 | Bot.HazardRec = DAG->TII->CreateTargetMIHazardRecognizer(Itin, DAG); |
4325 | } |
4326 | TopCluster = nullptr; |
4327 | BotCluster = nullptr; |
4328 | } |
4329 | |
4330 | void PostGenericScheduler::initPolicy(MachineBasicBlock::iterator Begin, |
4331 | MachineBasicBlock::iterator End, |
4332 | unsigned NumRegionInstrs) { |
4333 | const MachineFunction &MF = *Begin->getMF(); |
4334 | |
4335 | // Default to top-down because it was implemented first and existing targets |
4336 | // expect that behavior by default. |
4337 | RegionPolicy.OnlyTopDown = true; |
4338 | RegionPolicy.OnlyBottomUp = false; |
4339 | |
4340 | // Allow the subtarget to override default policy. |
4341 | MF.getSubtarget().overridePostRASchedPolicy(Policy&: RegionPolicy, NumRegionInstrs); |
4342 | |
4343 | // After subtarget overrides, apply command line options. |
4344 | if (PostRADirection == MISched::TopDown) { |
4345 | RegionPolicy.OnlyTopDown = true; |
4346 | RegionPolicy.OnlyBottomUp = false; |
4347 | } else if (PostRADirection == MISched::BottomUp) { |
4348 | RegionPolicy.OnlyTopDown = false; |
4349 | RegionPolicy.OnlyBottomUp = true; |
4350 | } else if (PostRADirection == MISched::Bidirectional) { |
4351 | RegionPolicy.OnlyBottomUp = false; |
4352 | RegionPolicy.OnlyTopDown = false; |
4353 | } |
4354 | |
4355 | BotIdx = NumRegionInstrs - 1; |
4356 | this->NumRegionInstrs = NumRegionInstrs; |
4357 | } |
4358 | |
4359 | void PostGenericScheduler::registerRoots() { |
4360 | Rem.CriticalPath = DAG->ExitSU.getDepth(); |
4361 | |
4362 | // Some roots may not feed into ExitSU. Check all of them in case. |
4363 | for (const SUnit *SU : Bot.Available) { |
4364 | if (SU->getDepth() > Rem.CriticalPath) |
4365 | Rem.CriticalPath = SU->getDepth(); |
4366 | } |
4367 | LLVM_DEBUG(dbgs() << "Critical Path: (PGS-RR) " << Rem.CriticalPath << '\n'); |
4368 | if (DumpCriticalPathLength) { |
4369 | errs() << "Critical Path(PGS-RR ): " << Rem.CriticalPath << " \n" ; |
4370 | } |
4371 | } |
4372 | |
4373 | /// Apply a set of heuristics to a new candidate for PostRA scheduling. |
4374 | /// |
4375 | /// \param Cand provides the policy and current best candidate. |
4376 | /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized. |
4377 | /// \return \c true if TryCand is better than Cand (Reason is NOT NoCand) |
4378 | bool PostGenericScheduler::tryCandidate(SchedCandidate &Cand, |
4379 | SchedCandidate &TryCand) { |
4380 | // Initialize the candidate if needed. |
4381 | if (!Cand.isValid()) { |
4382 | TryCand.Reason = FirstValid; |
4383 | return true; |
4384 | } |
4385 | |
4386 | // Prioritize instructions that read unbuffered resources by stall cycles. |
4387 | if (tryLess(TryVal: Top.getLatencyStallCycles(SU: TryCand.SU), |
4388 | CandVal: Top.getLatencyStallCycles(SU: Cand.SU), TryCand, Cand, Reason: Stall)) |
4389 | return TryCand.Reason != NoCand; |
4390 | |
4391 | // Keep clustered nodes together. |
4392 | const ClusterInfo *CandCluster = Cand.AtTop ? TopCluster : BotCluster; |
4393 | const ClusterInfo *TryCandCluster = TryCand.AtTop ? TopCluster : BotCluster; |
4394 | if (tryGreater(TryVal: TryCandCluster && TryCandCluster->contains(Ptr: TryCand.SU), |
4395 | CandVal: CandCluster && CandCluster->contains(Ptr: Cand.SU), TryCand, Cand, |
4396 | Reason: Cluster)) |
4397 | return TryCand.Reason != NoCand; |
4398 | // Avoid critical resource consumption and balance the schedule. |
4399 | if (tryLess(TryVal: TryCand.ResDelta.CritResources, CandVal: Cand.ResDelta.CritResources, |
4400 | TryCand, Cand, Reason: ResourceReduce)) |
4401 | return TryCand.Reason != NoCand; |
4402 | if (tryGreater(TryVal: TryCand.ResDelta.DemandedResources, |
4403 | CandVal: Cand.ResDelta.DemandedResources, |
4404 | TryCand, Cand, Reason: ResourceDemand)) |
4405 | return TryCand.Reason != NoCand; |
4406 | |
4407 | // We only compare a subset of features when comparing nodes between |
4408 | // Top and Bottom boundary. |
4409 | if (Cand.AtTop == TryCand.AtTop) { |
4410 | // Avoid serializing long latency dependence chains. |
4411 | if (Cand.Policy.ReduceLatency && |
4412 | tryLatency(TryCand, Cand, Zone&: Cand.AtTop ? Top : Bot)) |
4413 | return TryCand.Reason != NoCand; |
4414 | } |
4415 | |
4416 | // Fall through to original instruction order. |
4417 | if (TryCand.SU->NodeNum < Cand.SU->NodeNum) { |
4418 | TryCand.Reason = NodeOrder; |
4419 | return true; |
4420 | } |
4421 | |
4422 | return false; |
4423 | } |
4424 | |
4425 | void PostGenericScheduler::pickNodeFromQueue(SchedBoundary &Zone, |
4426 | SchedCandidate &Cand) { |
4427 | ReadyQueue &Q = Zone.Available; |
4428 | for (SUnit *SU : Q) { |
4429 | SchedCandidate TryCand(Cand.Policy); |
4430 | TryCand.SU = SU; |
4431 | TryCand.AtTop = Zone.isTop(); |
4432 | TryCand.initResourceDelta(DAG, SchedModel); |
4433 | if (tryCandidate(Cand, TryCand)) { |
4434 | Cand.setBest(TryCand); |
4435 | LLVM_DEBUG(traceCandidate(Cand)); |
4436 | } |
4437 | } |
4438 | } |
4439 | |
4440 | /// Pick the best candidate node from either the top or bottom queue. |
4441 | SUnit *PostGenericScheduler::pickNodeBidirectional(bool &IsTopNode) { |
4442 | // FIXME: This is similiar to GenericScheduler::pickNodeBidirectional. Factor |
4443 | // out common parts. |
4444 | |
4445 | // Schedule as far as possible in the direction of no choice. This is most |
4446 | // efficient, but also provides the best heuristics for CriticalPSets. |
4447 | if (SUnit *SU = Bot.pickOnlyChoice()) { |
4448 | IsTopNode = false; |
4449 | tracePick(Reason: Only1, /*IsTopNode=*/IsTop: false, /*IsPostRA=*/true); |
4450 | return SU; |
4451 | } |
4452 | if (SUnit *SU = Top.pickOnlyChoice()) { |
4453 | IsTopNode = true; |
4454 | tracePick(Reason: Only1, /*IsTopNode=*/IsTop: true, /*IsPostRA=*/true); |
4455 | return SU; |
4456 | } |
4457 | // Set the bottom-up policy based on the state of the current bottom zone and |
4458 | // the instructions outside the zone, including the top zone. |
4459 | CandPolicy BotPolicy; |
4460 | setPolicy(Policy&: BotPolicy, /*IsPostRA=*/true, CurrZone&: Bot, OtherZone: &Top); |
4461 | // Set the top-down policy based on the state of the current top zone and |
4462 | // the instructions outside the zone, including the bottom zone. |
4463 | CandPolicy TopPolicy; |
4464 | setPolicy(Policy&: TopPolicy, /*IsPostRA=*/true, CurrZone&: Top, OtherZone: &Bot); |
4465 | |
4466 | // See if BotCand is still valid (because we previously scheduled from Top). |
4467 | LLVM_DEBUG(dbgs() << "Picking from Bot:\n" ); |
4468 | if (!BotCand.isValid() || BotCand.SU->isScheduled || |
4469 | BotCand.Policy != BotPolicy) { |
4470 | BotCand.reset(NewPolicy: CandPolicy()); |
4471 | pickNodeFromQueue(Zone&: Bot, Cand&: BotCand); |
4472 | assert(BotCand.Reason != NoCand && "failed to find the first candidate" ); |
4473 | } else { |
4474 | LLVM_DEBUG(traceCandidate(BotCand)); |
4475 | #ifndef NDEBUG |
4476 | if (VerifyScheduling) { |
4477 | SchedCandidate TCand; |
4478 | TCand.reset(CandPolicy()); |
4479 | pickNodeFromQueue(Bot, BotCand); |
4480 | assert(TCand.SU == BotCand.SU && |
4481 | "Last pick result should correspond to re-picking right now" ); |
4482 | } |
4483 | #endif |
4484 | } |
4485 | |
4486 | // Check if the top Q has a better candidate. |
4487 | LLVM_DEBUG(dbgs() << "Picking from Top:\n" ); |
4488 | if (!TopCand.isValid() || TopCand.SU->isScheduled || |
4489 | TopCand.Policy != TopPolicy) { |
4490 | TopCand.reset(NewPolicy: CandPolicy()); |
4491 | pickNodeFromQueue(Zone&: Top, Cand&: TopCand); |
4492 | assert(TopCand.Reason != NoCand && "failed to find the first candidate" ); |
4493 | } else { |
4494 | LLVM_DEBUG(traceCandidate(TopCand)); |
4495 | #ifndef NDEBUG |
4496 | if (VerifyScheduling) { |
4497 | SchedCandidate TCand; |
4498 | TCand.reset(CandPolicy()); |
4499 | pickNodeFromQueue(Top, TopCand); |
4500 | assert(TCand.SU == TopCand.SU && |
4501 | "Last pick result should correspond to re-picking right now" ); |
4502 | } |
4503 | #endif |
4504 | } |
4505 | |
4506 | // Pick best from BotCand and TopCand. |
4507 | assert(BotCand.isValid()); |
4508 | assert(TopCand.isValid()); |
4509 | SchedCandidate Cand = BotCand; |
4510 | TopCand.Reason = NoCand; |
4511 | if (tryCandidate(Cand, TryCand&: TopCand)) { |
4512 | Cand.setBest(TopCand); |
4513 | LLVM_DEBUG(traceCandidate(Cand)); |
4514 | } |
4515 | |
4516 | IsTopNode = Cand.AtTop; |
4517 | tracePick(Cand, /*IsPostRA=*/true); |
4518 | return Cand.SU; |
4519 | } |
4520 | |
4521 | /// Pick the next node to schedule. |
4522 | SUnit *PostGenericScheduler::pickNode(bool &IsTopNode) { |
4523 | if (DAG->top() == DAG->bottom()) { |
4524 | assert(Top.Available.empty() && Top.Pending.empty() && |
4525 | Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage" ); |
4526 | return nullptr; |
4527 | } |
4528 | SUnit *SU; |
4529 | do { |
4530 | if (RegionPolicy.OnlyBottomUp) { |
4531 | SU = Bot.pickOnlyChoice(); |
4532 | if (SU) { |
4533 | tracePick(Reason: Only1, /*IsTopNode=*/IsTop: true, /*IsPostRA=*/true); |
4534 | } else { |
4535 | CandPolicy NoPolicy; |
4536 | BotCand.reset(NewPolicy: NoPolicy); |
4537 | // Set the bottom-up policy based on the state of the current bottom |
4538 | // zone and the instructions outside the zone, including the top zone. |
4539 | setPolicy(Policy&: BotCand.Policy, /*IsPostRA=*/true, CurrZone&: Bot, OtherZone: nullptr); |
4540 | pickNodeFromQueue(Zone&: Bot, Cand&: BotCand); |
4541 | assert(BotCand.Reason != NoCand && "failed to find a candidate" ); |
4542 | tracePick(Cand: BotCand, /*IsPostRA=*/true); |
4543 | SU = BotCand.SU; |
4544 | } |
4545 | IsTopNode = false; |
4546 | } else if (RegionPolicy.OnlyTopDown) { |
4547 | SU = Top.pickOnlyChoice(); |
4548 | if (SU) { |
4549 | tracePick(Reason: Only1, /*IsTopNode=*/IsTop: true, /*IsPostRA=*/true); |
4550 | } else { |
4551 | CandPolicy NoPolicy; |
4552 | TopCand.reset(NewPolicy: NoPolicy); |
4553 | // Set the top-down policy based on the state of the current top zone |
4554 | // and the instructions outside the zone, including the bottom zone. |
4555 | setPolicy(Policy&: TopCand.Policy, /*IsPostRA=*/true, CurrZone&: Top, OtherZone: nullptr); |
4556 | pickNodeFromQueue(Zone&: Top, Cand&: TopCand); |
4557 | assert(TopCand.Reason != NoCand && "failed to find a candidate" ); |
4558 | tracePick(Cand: TopCand, /*IsPostRA=*/true); |
4559 | SU = TopCand.SU; |
4560 | } |
4561 | IsTopNode = true; |
4562 | } else { |
4563 | SU = pickNodeBidirectional(IsTopNode); |
4564 | } |
4565 | } while (SU->isScheduled); |
4566 | |
4567 | if (SU->isTopReady()) |
4568 | Top.removeReady(SU); |
4569 | if (SU->isBottomReady()) |
4570 | Bot.removeReady(SU); |
4571 | |
4572 | LLVM_DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " |
4573 | << *SU->getInstr()); |
4574 | |
4575 | if (IsTopNode) { |
4576 | if (SU->NodeNum == TopIdx++) |
4577 | ++NumInstrsInSourceOrderPostRA; |
4578 | } else { |
4579 | assert(BotIdx < NumRegionInstrs && "out of bounds" ); |
4580 | if (SU->NodeNum == BotIdx--) |
4581 | ++NumInstrsInSourceOrderPostRA; |
4582 | } |
4583 | |
4584 | NumInstrsScheduledPostRA += 1; |
4585 | |
4586 | return SU; |
4587 | } |
4588 | |
4589 | /// Called after ScheduleDAGMI has scheduled an instruction and updated |
4590 | /// scheduled/remaining flags in the DAG nodes. |
4591 | void PostGenericScheduler::schedNode(SUnit *SU, bool IsTopNode) { |
4592 | if (IsTopNode) { |
4593 | SU->TopReadyCycle = std::max(a: SU->TopReadyCycle, b: Top.getCurrCycle()); |
4594 | TopCluster = DAG->getCluster(Idx: SU->ParentClusterIdx); |
4595 | Top.bumpNode(SU); |
4596 | } else { |
4597 | SU->BotReadyCycle = std::max(a: SU->BotReadyCycle, b: Bot.getCurrCycle()); |
4598 | BotCluster = DAG->getCluster(Idx: SU->ParentClusterIdx); |
4599 | Bot.bumpNode(SU); |
4600 | } |
4601 | } |
4602 | |
4603 | //===----------------------------------------------------------------------===// |
4604 | // ILP Scheduler. Currently for experimental analysis of heuristics. |
4605 | //===----------------------------------------------------------------------===// |
4606 | |
4607 | namespace { |
4608 | |
4609 | /// Order nodes by the ILP metric. |
4610 | struct ILPOrder { |
4611 | const SchedDFSResult *DFSResult = nullptr; |
4612 | const BitVector *ScheduledTrees = nullptr; |
4613 | bool MaximizeILP; |
4614 | |
4615 | ILPOrder(bool MaxILP) : MaximizeILP(MaxILP) {} |
4616 | |
4617 | /// Apply a less-than relation on node priority. |
4618 | /// |
4619 | /// (Return true if A comes after B in the Q.) |
4620 | bool operator()(const SUnit *A, const SUnit *B) const { |
4621 | unsigned SchedTreeA = DFSResult->getSubtreeID(SU: A); |
4622 | unsigned SchedTreeB = DFSResult->getSubtreeID(SU: B); |
4623 | if (SchedTreeA != SchedTreeB) { |
4624 | // Unscheduled trees have lower priority. |
4625 | if (ScheduledTrees->test(Idx: SchedTreeA) != ScheduledTrees->test(Idx: SchedTreeB)) |
4626 | return ScheduledTrees->test(Idx: SchedTreeB); |
4627 | |
4628 | // Trees with shallower connections have lower priority. |
4629 | if (DFSResult->getSubtreeLevel(SubtreeID: SchedTreeA) |
4630 | != DFSResult->getSubtreeLevel(SubtreeID: SchedTreeB)) { |
4631 | return DFSResult->getSubtreeLevel(SubtreeID: SchedTreeA) |
4632 | < DFSResult->getSubtreeLevel(SubtreeID: SchedTreeB); |
4633 | } |
4634 | } |
4635 | if (MaximizeILP) |
4636 | return DFSResult->getILP(SU: A) < DFSResult->getILP(SU: B); |
4637 | else |
4638 | return DFSResult->getILP(SU: A) > DFSResult->getILP(SU: B); |
4639 | } |
4640 | }; |
4641 | |
4642 | /// Schedule based on the ILP metric. |
4643 | class ILPScheduler : public MachineSchedStrategy { |
4644 | ScheduleDAGMILive *DAG = nullptr; |
4645 | ILPOrder Cmp; |
4646 | |
4647 | std::vector<SUnit*> ReadyQ; |
4648 | |
4649 | public: |
4650 | ILPScheduler(bool MaximizeILP) : Cmp(MaximizeILP) {} |
4651 | |
4652 | void initialize(ScheduleDAGMI *dag) override { |
4653 | assert(dag->hasVRegLiveness() && "ILPScheduler needs vreg liveness" ); |
4654 | DAG = static_cast<ScheduleDAGMILive*>(dag); |
4655 | DAG->computeDFSResult(); |
4656 | Cmp.DFSResult = DAG->getDFSResult(); |
4657 | Cmp.ScheduledTrees = &DAG->getScheduledTrees(); |
4658 | ReadyQ.clear(); |
4659 | } |
4660 | |
4661 | void registerRoots() override { |
4662 | // Restore the heap in ReadyQ with the updated DFS results. |
4663 | std::make_heap(first: ReadyQ.begin(), last: ReadyQ.end(), comp: Cmp); |
4664 | } |
4665 | |
4666 | /// Implement MachineSchedStrategy interface. |
4667 | /// ----------------------------------------- |
4668 | |
4669 | /// Callback to select the highest priority node from the ready Q. |
4670 | SUnit *pickNode(bool &IsTopNode) override { |
4671 | if (ReadyQ.empty()) return nullptr; |
4672 | std::pop_heap(first: ReadyQ.begin(), last: ReadyQ.end(), comp: Cmp); |
4673 | SUnit *SU = ReadyQ.back(); |
4674 | ReadyQ.pop_back(); |
4675 | IsTopNode = false; |
4676 | LLVM_DEBUG(dbgs() << "Pick node " |
4677 | << "SU(" << SU->NodeNum << ") " |
4678 | << " ILP: " << DAG->getDFSResult()->getILP(SU) |
4679 | << " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) |
4680 | << " @" |
4681 | << DAG->getDFSResult()->getSubtreeLevel( |
4682 | DAG->getDFSResult()->getSubtreeID(SU)) |
4683 | << '\n' |
4684 | << "Scheduling " << *SU->getInstr()); |
4685 | return SU; |
4686 | } |
4687 | |
4688 | /// Scheduler callback to notify that a new subtree is scheduled. |
4689 | void scheduleTree(unsigned SubtreeID) override { |
4690 | std::make_heap(first: ReadyQ.begin(), last: ReadyQ.end(), comp: Cmp); |
4691 | } |
4692 | |
4693 | /// Callback after a node is scheduled. Mark a newly scheduled tree, notify |
4694 | /// DFSResults, and resort the priority Q. |
4695 | void schedNode(SUnit *SU, bool IsTopNode) override { |
4696 | assert(!IsTopNode && "SchedDFSResult needs bottom-up" ); |
4697 | } |
4698 | |
4699 | void releaseTopNode(SUnit *) override { /*only called for top roots*/ } |
4700 | |
4701 | void releaseBottomNode(SUnit *SU) override { |
4702 | ReadyQ.push_back(x: SU); |
4703 | std::push_heap(first: ReadyQ.begin(), last: ReadyQ.end(), comp: Cmp); |
4704 | } |
4705 | }; |
4706 | |
4707 | } // end anonymous namespace |
4708 | |
4709 | static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) { |
4710 | return new ScheduleDAGMILive(C, std::make_unique<ILPScheduler>(args: true)); |
4711 | } |
4712 | static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) { |
4713 | return new ScheduleDAGMILive(C, std::make_unique<ILPScheduler>(args: false)); |
4714 | } |
4715 | |
4716 | static MachineSchedRegistry ILPMaxRegistry( |
4717 | "ilpmax" , "Schedule bottom-up for max ILP" , createILPMaxScheduler); |
4718 | static MachineSchedRegistry ILPMinRegistry( |
4719 | "ilpmin" , "Schedule bottom-up for min ILP" , createILPMinScheduler); |
4720 | |
4721 | //===----------------------------------------------------------------------===// |
4722 | // Machine Instruction Shuffler for Correctness Testing |
4723 | //===----------------------------------------------------------------------===// |
4724 | |
4725 | #ifndef NDEBUG |
4726 | namespace { |
4727 | |
4728 | /// Apply a less-than relation on the node order, which corresponds to the |
4729 | /// instruction order prior to scheduling. IsReverse implements greater-than. |
4730 | template<bool IsReverse> |
4731 | struct SUnitOrder { |
4732 | bool operator()(SUnit *A, SUnit *B) const { |
4733 | if (IsReverse) |
4734 | return A->NodeNum > B->NodeNum; |
4735 | else |
4736 | return A->NodeNum < B->NodeNum; |
4737 | } |
4738 | }; |
4739 | |
4740 | /// Reorder instructions as much as possible. |
4741 | class InstructionShuffler : public MachineSchedStrategy { |
4742 | bool IsAlternating; |
4743 | bool IsTopDown; |
4744 | |
4745 | // Using a less-than relation (SUnitOrder<false>) for the TopQ priority |
4746 | // gives nodes with a higher number higher priority causing the latest |
4747 | // instructions to be scheduled first. |
4748 | PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false>> |
4749 | TopQ; |
4750 | |
4751 | // When scheduling bottom-up, use greater-than as the queue priority. |
4752 | PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true>> |
4753 | BottomQ; |
4754 | |
4755 | public: |
4756 | InstructionShuffler(bool alternate, bool topdown) |
4757 | : IsAlternating(alternate), IsTopDown(topdown) {} |
4758 | |
4759 | void initialize(ScheduleDAGMI*) override { |
4760 | TopQ.clear(); |
4761 | BottomQ.clear(); |
4762 | } |
4763 | |
4764 | /// Implement MachineSchedStrategy interface. |
4765 | /// ----------------------------------------- |
4766 | |
4767 | SUnit *pickNode(bool &IsTopNode) override { |
4768 | SUnit *SU; |
4769 | if (IsTopDown) { |
4770 | do { |
4771 | if (TopQ.empty()) return nullptr; |
4772 | SU = TopQ.top(); |
4773 | TopQ.pop(); |
4774 | } while (SU->isScheduled); |
4775 | IsTopNode = true; |
4776 | } else { |
4777 | do { |
4778 | if (BottomQ.empty()) return nullptr; |
4779 | SU = BottomQ.top(); |
4780 | BottomQ.pop(); |
4781 | } while (SU->isScheduled); |
4782 | IsTopNode = false; |
4783 | } |
4784 | if (IsAlternating) |
4785 | IsTopDown = !IsTopDown; |
4786 | return SU; |
4787 | } |
4788 | |
4789 | void schedNode(SUnit *SU, bool IsTopNode) override {} |
4790 | |
4791 | void releaseTopNode(SUnit *SU) override { |
4792 | TopQ.push(SU); |
4793 | } |
4794 | void releaseBottomNode(SUnit *SU) override { |
4795 | BottomQ.push(SU); |
4796 | } |
4797 | }; |
4798 | |
4799 | } // end anonymous namespace |
4800 | |
4801 | static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) { |
4802 | bool Alternate = |
4803 | PreRADirection != MISched::TopDown && PreRADirection != MISched::BottomUp; |
4804 | bool TopDown = PreRADirection != MISched::BottomUp; |
4805 | return new ScheduleDAGMILive( |
4806 | C, std::make_unique<InstructionShuffler>(Alternate, TopDown)); |
4807 | } |
4808 | |
4809 | static MachineSchedRegistry ShufflerRegistry( |
4810 | "shuffle" , "Shuffle machine instructions alternating directions" , |
4811 | createInstructionShuffler); |
4812 | #endif // !NDEBUG |
4813 | |
4814 | //===----------------------------------------------------------------------===// |
4815 | // GraphWriter support for ScheduleDAGMILive. |
4816 | //===----------------------------------------------------------------------===// |
4817 | |
4818 | #ifndef NDEBUG |
4819 | namespace llvm { |
4820 | |
4821 | template<> struct GraphTraits< |
4822 | ScheduleDAGMI*> : public GraphTraits<ScheduleDAG*> {}; |
4823 | |
4824 | template<> |
4825 | struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits { |
4826 | DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {} |
4827 | |
4828 | static std::string getGraphName(const ScheduleDAG *G) { |
4829 | return std::string(G->MF.getName()); |
4830 | } |
4831 | |
4832 | static bool renderGraphFromBottomUp() { |
4833 | return true; |
4834 | } |
4835 | |
4836 | static bool isNodeHidden(const SUnit *Node, const ScheduleDAG *G) { |
4837 | if (ViewMISchedCutoff == 0) |
4838 | return false; |
4839 | return (Node->Preds.size() > ViewMISchedCutoff |
4840 | || Node->Succs.size() > ViewMISchedCutoff); |
4841 | } |
4842 | |
4843 | /// If you want to override the dot attributes printed for a particular |
4844 | /// edge, override this method. |
4845 | static std::string getEdgeAttributes(const SUnit *Node, |
4846 | SUnitIterator EI, |
4847 | const ScheduleDAG *Graph) { |
4848 | if (EI.isArtificialDep()) |
4849 | return "color=cyan,style=dashed" ; |
4850 | if (EI.isCtrlDep()) |
4851 | return "color=blue,style=dashed" ; |
4852 | return "" ; |
4853 | } |
4854 | |
4855 | static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) { |
4856 | std::string Str; |
4857 | raw_string_ostream SS(Str); |
4858 | const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G); |
4859 | const SchedDFSResult *DFS = DAG->hasVRegLiveness() ? |
4860 | static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : nullptr; |
4861 | SS << "SU:" << SU->NodeNum; |
4862 | if (DFS) |
4863 | SS << " I:" << DFS->getNumInstrs(SU); |
4864 | return Str; |
4865 | } |
4866 | |
4867 | static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) { |
4868 | return G->getGraphNodeLabel(SU); |
4869 | } |
4870 | |
4871 | static std::string getNodeAttributes(const SUnit *N, const ScheduleDAG *G) { |
4872 | std::string Str("shape=Mrecord" ); |
4873 | const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G); |
4874 | const SchedDFSResult *DFS = DAG->hasVRegLiveness() ? |
4875 | static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : nullptr; |
4876 | if (DFS) { |
4877 | Str += ",style=filled,fillcolor=\"#" ; |
4878 | Str += DOT::getColorString(DFS->getSubtreeID(N)); |
4879 | Str += '"'; |
4880 | } |
4881 | return Str; |
4882 | } |
4883 | }; |
4884 | |
4885 | } // end namespace llvm |
4886 | #endif // NDEBUG |
4887 | |
4888 | /// viewGraph - Pop up a ghostview window with the reachable parts of the DAG |
4889 | /// rendered using 'dot'. |
4890 | void ScheduleDAGMI::viewGraph(const Twine &Name, const Twine &Title) { |
4891 | #ifndef NDEBUG |
4892 | ViewGraph(this, Name, false, Title); |
4893 | #else |
4894 | errs() << "ScheduleDAGMI::viewGraph is only available in debug builds on " |
4895 | << "systems with Graphviz or gv!\n" ; |
4896 | #endif // NDEBUG |
4897 | } |
4898 | |
4899 | /// Out-of-line implementation with no arguments is handy for gdb. |
4900 | void ScheduleDAGMI::viewGraph() { |
4901 | viewGraph(Name: getDAGName(), Title: "Scheduling-Units Graph for " + getDAGName()); |
4902 | } |
4903 | |
4904 | /// Sort predicate for the intervals stored in an instance of |
4905 | /// ResourceSegments. Intervals are always disjoint (no intersection |
4906 | /// for any pairs of intervals), therefore we can sort the totality of |
4907 | /// the intervals by looking only at the left boundary. |
4908 | static bool sortIntervals(const ResourceSegments::IntervalTy &A, |
4909 | const ResourceSegments::IntervalTy &B) { |
4910 | return A.first < B.first; |
4911 | } |
4912 | |
4913 | unsigned ResourceSegments::getFirstAvailableAt( |
4914 | unsigned CurrCycle, unsigned AcquireAtCycle, unsigned ReleaseAtCycle, |
4915 | std::function<ResourceSegments::IntervalTy(unsigned, unsigned, unsigned)> |
4916 | IntervalBuilder) const { |
4917 | assert(llvm::is_sorted(_Intervals, sortIntervals) && |
4918 | "Cannot execute on an un-sorted set of intervals." ); |
4919 | |
4920 | // Zero resource usage is allowed by TargetSchedule.td but we do not construct |
4921 | // a ResourceSegment interval for that situation. |
4922 | if (AcquireAtCycle == ReleaseAtCycle) |
4923 | return CurrCycle; |
4924 | |
4925 | unsigned RetCycle = CurrCycle; |
4926 | ResourceSegments::IntervalTy NewInterval = |
4927 | IntervalBuilder(RetCycle, AcquireAtCycle, ReleaseAtCycle); |
4928 | for (auto &Interval : _Intervals) { |
4929 | if (!intersects(A: NewInterval, B: Interval)) |
4930 | continue; |
4931 | |
4932 | // Move the interval right next to the top of the one it |
4933 | // intersects. |
4934 | assert(Interval.second > NewInterval.first && |
4935 | "Invalid intervals configuration." ); |
4936 | RetCycle += (unsigned)Interval.second - (unsigned)NewInterval.first; |
4937 | NewInterval = IntervalBuilder(RetCycle, AcquireAtCycle, ReleaseAtCycle); |
4938 | } |
4939 | return RetCycle; |
4940 | } |
4941 | |
4942 | void ResourceSegments::add(ResourceSegments::IntervalTy A, |
4943 | const unsigned CutOff) { |
4944 | assert(A.first <= A.second && "Cannot add negative resource usage" ); |
4945 | assert(CutOff > 0 && "0-size interval history has no use." ); |
4946 | // Zero resource usage is allowed by TargetSchedule.td, in the case that the |
4947 | // instruction needed the resource to be available but does not use it. |
4948 | // However, ResourceSegment represents an interval that is closed on the left |
4949 | // and open on the right. It is impossible to represent an empty interval when |
4950 | // the left is closed. Do not add it to Intervals. |
4951 | if (A.first == A.second) |
4952 | return; |
4953 | |
4954 | assert(all_of(_Intervals, |
4955 | [&A](const ResourceSegments::IntervalTy &Interval) -> bool { |
4956 | return !intersects(A, Interval); |
4957 | }) && |
4958 | "A resource is being overwritten" ); |
4959 | _Intervals.push_back(x: A); |
4960 | |
4961 | sortAndMerge(); |
4962 | |
4963 | // Do not keep the full history of the intervals, just the |
4964 | // latest #CutOff. |
4965 | while (_Intervals.size() > CutOff) |
4966 | _Intervals.pop_front(); |
4967 | } |
4968 | |
4969 | bool ResourceSegments::intersects(ResourceSegments::IntervalTy A, |
4970 | ResourceSegments::IntervalTy B) { |
4971 | assert(A.first <= A.second && "Invalid interval" ); |
4972 | assert(B.first <= B.second && "Invalid interval" ); |
4973 | |
4974 | // Share one boundary. |
4975 | if ((A.first == B.first) || (A.second == B.second)) |
4976 | return true; |
4977 | |
4978 | // full intersersect: [ *** ) B |
4979 | // [***) A |
4980 | if ((A.first > B.first) && (A.second < B.second)) |
4981 | return true; |
4982 | |
4983 | // right intersect: [ ***) B |
4984 | // [*** ) A |
4985 | if ((A.first > B.first) && (A.first < B.second) && (A.second > B.second)) |
4986 | return true; |
4987 | |
4988 | // left intersect: [*** ) B |
4989 | // [ ***) A |
4990 | if ((A.first < B.first) && (B.first < A.second) && (B.second > B.first)) |
4991 | return true; |
4992 | |
4993 | return false; |
4994 | } |
4995 | |
4996 | void ResourceSegments::sortAndMerge() { |
4997 | if (_Intervals.size() <= 1) |
4998 | return; |
4999 | |
5000 | // First sort the collection. |
5001 | _Intervals.sort(comp: sortIntervals); |
5002 | |
5003 | // can use next because I have at least 2 elements in the list |
5004 | auto next = std::next(x: std::begin(cont&: _Intervals)); |
5005 | auto E = std::end(cont&: _Intervals); |
5006 | for (; next != E; ++next) { |
5007 | if (std::prev(x: next)->second >= next->first) { |
5008 | next->first = std::prev(x: next)->first; |
5009 | _Intervals.erase(position: std::prev(x: next)); |
5010 | continue; |
5011 | } |
5012 | } |
5013 | } |
5014 | |