1 | //===- MLRegAllocEvictAdvisor.cpp - ML eviction advisor -------------------===// |
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 | // Implementation of the ML eviction advisor and reward injection pass |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "AllocationOrder.h" |
14 | #include "RegAllocEvictionAdvisor.h" |
15 | #include "RegAllocGreedy.h" |
16 | #include "llvm/Analysis/InteractiveModelRunner.h" |
17 | #include "llvm/Analysis/MLModelRunner.h" |
18 | #include "llvm/Analysis/TensorSpec.h" |
19 | #if defined(LLVM_HAVE_TF_AOT_REGALLOCEVICTMODEL) || defined(LLVM_HAVE_TFLITE) |
20 | #include "llvm/Analysis/ModelUnderTrainingRunner.h" |
21 | #include "llvm/Analysis/NoInferenceModelRunner.h" |
22 | #include "llvm/Analysis/Utils/TrainingLogger.h" |
23 | #endif |
24 | #include "MLRegAllocEvictAdvisor.h" |
25 | #include "llvm/Analysis/ReleaseModeModelRunner.h" |
26 | #include "llvm/CodeGen/CalcSpillWeights.h" |
27 | #include "llvm/CodeGen/LiveRegMatrix.h" |
28 | #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" |
29 | #include "llvm/CodeGen/MachineFunction.h" |
30 | #include "llvm/CodeGen/MachineLoopInfo.h" |
31 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
32 | #include "llvm/CodeGen/Passes.h" |
33 | #include "llvm/CodeGen/RegisterClassInfo.h" |
34 | #include "llvm/CodeGen/VirtRegMap.h" |
35 | #include "llvm/IR/Module.h" |
36 | #include "llvm/InitializePasses.h" |
37 | #include "llvm/Pass.h" |
38 | #include "llvm/PassRegistry.h" |
39 | #include "llvm/Support/CommandLine.h" |
40 | #include "llvm/Support/ErrorHandling.h" |
41 | |
42 | #include <array> |
43 | #include <bitset> |
44 | #include <memory> |
45 | |
46 | using namespace llvm; |
47 | |
48 | #define DEBUG_TYPE "ml-regalloc" |
49 | |
50 | // Generated header in release (AOT) mode |
51 | #if defined(LLVM_HAVE_TF_AOT_REGALLOCEVICTMODEL) |
52 | #include "RegAllocEvictModel.h" |
53 | using CompiledModelType = RegAllocEvictModel; |
54 | #else |
55 | using CompiledModelType = NoopSavedModelImpl; |
56 | #endif |
57 | |
58 | static cl::opt<std::string> InteractiveChannelBaseName( |
59 | "regalloc-evict-interactive-channel-base" , cl::Hidden, |
60 | cl::desc( |
61 | "Base file path for the interactive mode. The incoming filename should " |
62 | "have the name <regalloc-evict-interactive-channel-base>.in, while the " |
63 | "outgoing name should be " |
64 | "<regalloc-evict-interactive-channel-base>.out" )); |
65 | |
66 | // Options that only make sense in development mode |
67 | #ifdef LLVM_HAVE_TFLITE |
68 | #include "RegAllocScore.h" |
69 | #include "llvm/Analysis/Utils/TFUtils.h" |
70 | |
71 | static cl::opt<std::string> TrainingLog( |
72 | "regalloc-training-log" , cl::Hidden, |
73 | cl::desc("Training log for the register allocator eviction model" )); |
74 | |
75 | static cl::opt<std::string> ModelUnderTraining( |
76 | "regalloc-model" , cl::Hidden, |
77 | cl::desc("The model being trained for register allocation eviction" )); |
78 | |
79 | static cl::opt<bool> EnableDevelopmentFeatures( |
80 | "regalloc-enable-development-features" , cl::Hidden, |
81 | cl::desc("Whether or not to enable features under development for the ML " |
82 | "regalloc advisor" )); |
83 | |
84 | #else |
85 | static const bool EnableDevelopmentFeatures = false; |
86 | #endif // #ifdef LLVM_HAVE_TFLITE |
87 | |
88 | /// The score injection pass. |
89 | /// This pass calculates the score for a function and inserts it in the log, but |
90 | /// this happens only in development mode. It's a no-op otherwise. |
91 | namespace llvm { |
92 | extern cl::opt<unsigned> EvictInterferenceCutoff; |
93 | |
94 | class RegAllocScoring : public MachineFunctionPass { |
95 | public: |
96 | static char ID; |
97 | |
98 | RegAllocScoring() : MachineFunctionPass(ID) { |
99 | initializeRegAllocScoringPass(*PassRegistry::getPassRegistry()); |
100 | } |
101 | |
102 | ~RegAllocScoring() override = default; |
103 | |
104 | StringRef getPassName() const override { |
105 | return "Register Allocation Pass Scoring" ; |
106 | } |
107 | |
108 | /// RegAllocReward analysis usage. |
109 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
110 | AU.setPreservesAll(); |
111 | AU.addRequired<RegAllocEvictionAdvisorAnalysis>(); |
112 | AU.addRequired<RegAllocPriorityAdvisorAnalysis>(); |
113 | AU.addRequired<MachineBlockFrequencyInfoWrapperPass>(); |
114 | MachineFunctionPass::getAnalysisUsage(AU); |
115 | } |
116 | |
117 | /// Performs this pass |
118 | bool runOnMachineFunction(MachineFunction &) override; |
119 | }; |
120 | |
121 | char RegAllocScoring::ID = 0; |
122 | FunctionPass *createRegAllocScoringPass() { return new RegAllocScoring(); } |
123 | |
124 | } // namespace llvm |
125 | |
126 | INITIALIZE_PASS(RegAllocScoring, "regallocscoringpass" , |
127 | "Register Allocation Scoring Pass" , false, false) |
128 | |
129 | // =================================== |
130 | // Common ML Advisor declarations |
131 | // =================================== |
132 | namespace { |
133 | // The model can only accept a specified number of opcodes and will error it if |
134 | // fed an opcode it hasn't seen before. This constant sets the current cutoff. |
135 | static const int OpcodeValueCutoff = 17716; |
136 | |
137 | // Most features are as described above, so we'll reuse this vector in defining |
138 | // them. |
139 | static const std::vector<int64_t> PerLiveRangeShape{1, NumberOfInterferences}; |
140 | |
141 | // -------------- |
142 | // Features table |
143 | // -------------- |
144 | // For each interfering live range (incl. the candidate) we collect a number of |
145 | // features. However, because the features are of different types (and because |
146 | // of ML best practices), we organize the tensors per feature, not per |
147 | // candidate. Each such tensor has a scalar value corresponding to the |
148 | // interferring live range at that position, in the order in AllocationOrder. |
149 | // The last position corresponds to the virt reg seeking allocation. |
150 | // Exception to all that is the progression feature, which is just a scalar (see |
151 | // its documentation for details). |
152 | // Note on naming: the "_by_max" are normalized using the largest value of that |
153 | // tensor, as observed in the current decision making stage (i.e. for the |
154 | // current call to the advisor's tryFindEvictionCandidate) |
155 | // |
156 | // The feature list format: type, name, shape, documentation. |
157 | // Note: we can really just use int64 and float, hence the modeling of some |
158 | // bools as int64 values. |
159 | #define RA_EVICT_FEATURES_LIST(M) \ |
160 | M(int64_t, mask, PerLiveRangeShape, \ |
161 | "boolean values, 0 for unavailable candidates (i.e. if a position is 0, " \ |
162 | "it " \ |
163 | "can't be evicted)") \ |
164 | M(int64_t, is_free, PerLiveRangeShape, \ |
165 | "boolean values, 1 if this phys reg is actually free (no interferences)") \ |
166 | M(float, nr_urgent, PerLiveRangeShape, \ |
167 | "number of 'urgent' intervals, normalized. Urgent are those that are OK " \ |
168 | "to break cascades") \ |
169 | M(float, nr_broken_hints, PerLiveRangeShape, \ |
170 | "if this position were evicted, how many broken hints would there be") \ |
171 | M(int64_t, is_hint, PerLiveRangeShape, \ |
172 | "is this a preferred phys reg for the candidate") \ |
173 | M(int64_t, is_local, PerLiveRangeShape, \ |
174 | "is this live range local to a basic block") \ |
175 | M(float, nr_rematerializable, PerLiveRangeShape, \ |
176 | "nr rematerializable ranges") \ |
177 | M(float, nr_defs_and_uses, PerLiveRangeShape, \ |
178 | "bb freq - weighed nr defs and uses") \ |
179 | M(float, weighed_reads_by_max, PerLiveRangeShape, \ |
180 | "bb freq - weighed nr of reads, normalized") \ |
181 | M(float, weighed_writes_by_max, PerLiveRangeShape, \ |
182 | "bb feq - weighed nr of writes, normalized") \ |
183 | M(float, weighed_read_writes_by_max, PerLiveRangeShape, \ |
184 | "bb freq - weighed nr of uses that are both read and writes, normalized") \ |
185 | M(float, weighed_indvars_by_max, PerLiveRangeShape, \ |
186 | "bb freq - weighed nr of uses that are indvars, normalized") \ |
187 | M(float, hint_weights_by_max, PerLiveRangeShape, \ |
188 | "bb freq - weighed nr of uses that are hints, normalized") \ |
189 | M(float, start_bb_freq_by_max, PerLiveRangeShape, \ |
190 | "the freq in the start block, normalized") \ |
191 | M(float, end_bb_freq_by_max, PerLiveRangeShape, \ |
192 | "freq of end block, normalized") \ |
193 | M(float, hottest_bb_freq_by_max, PerLiveRangeShape, \ |
194 | "hottest BB freq, normalized") \ |
195 | M(float, liverange_size, PerLiveRangeShape, \ |
196 | "size (instr index diff) of the LR") \ |
197 | M(float, use_def_density, PerLiveRangeShape, \ |
198 | "the max weight, as computed by the manual heuristic") \ |
199 | M(int64_t, max_stage, PerLiveRangeShape, \ |
200 | "largest stage of an interval in this LR") \ |
201 | M(int64_t, min_stage, PerLiveRangeShape, \ |
202 | "lowest stage of an interval in this LR") \ |
203 | M(float, progress, {1}, "ratio of current queue size to initial size") |
204 | |
205 | #ifdef LLVM_HAVE_TFLITE |
206 | #define RA_EVICT_FIRST_DEVELOPMENT_FEATURE(M) \ |
207 | M(int64_t, instructions, InstructionsShape, \ |
208 | "Opcodes of the instructions covered by the eviction problem") |
209 | |
210 | #define RA_EVICT_REST_DEVELOPMENT_FEATURES(M) \ |
211 | M(int64_t, instructions_mapping, InstructionsMappingShape, \ |
212 | "A binary matrix mapping LRs to instruction opcodes") \ |
213 | M(float, mbb_frequencies, MBBFrequencyShape, \ |
214 | "A vector of machine basic block frequencies") \ |
215 | M(int64_t, mbb_mapping, InstructionsShape, \ |
216 | "A vector of indices mapping instructions to MBBs") |
217 | #else |
218 | #define RA_EVICT_FIRST_DEVELOPMENT_FEATURE(M) |
219 | #define RA_EVICT_REST_DEVELOPMENT_FEATURES(M) |
220 | #endif |
221 | |
222 | // The model learns to pick one of the mask == 1 interferences. This is the |
223 | // name of the output tensor. The contract with the model is that the output |
224 | // will be guaranteed to be to a mask == 1 position. Using a macro here to |
225 | // avoid 'not used' warnings (and keep cond compilation to a minimum) |
226 | #define DecisionName "index_to_evict" |
227 | static const TensorSpec DecisionSpec = |
228 | TensorSpec::createSpec<int64_t>(DecisionName, Shape: {1}); |
229 | |
230 | // Named features index. |
231 | enum FeatureIDs { |
232 | #define _FEATURE_IDX_SIMPLE(_, name, __, ___) name |
233 | #define _FEATURE_IDX(A, B, C, D) _FEATURE_IDX_SIMPLE(A, B, C, D), |
234 | RA_EVICT_FEATURES_LIST(_FEATURE_IDX) FeatureCount, |
235 | #ifdef LLVM_HAVE_TFLITE |
236 | RA_EVICT_FIRST_DEVELOPMENT_FEATURE(_FEATURE_IDX_SIMPLE) = FeatureCount, |
237 | #else |
238 | RA_EVICT_FIRST_DEVELOPMENT_FEATURE(_FEATURE_IDX) |
239 | #endif // #ifdef LLVM_HAVE_TFLITE |
240 | RA_EVICT_REST_DEVELOPMENT_FEATURES(_FEATURE_IDX) FeaturesWithDevelopmentCount |
241 | #undef _FEATURE_IDX |
242 | #undef _FEATURE_IDX_SIMPLE |
243 | }; |
244 | |
245 | // The ML advisor will typically have a sparse input to the evaluator, because |
246 | // various phys regs won't be available. It's easier (maintenance-wise) to |
247 | // bulk-reset the state of the evaluator each time we are about to use it |
248 | // again. |
249 | template <typename T> size_t getTotalSize(const std::vector<int64_t> &Shape) { |
250 | size_t Ret = sizeof(T); |
251 | for (const auto V : Shape) |
252 | Ret *= V; |
253 | return Ret; |
254 | } |
255 | |
256 | void resetInputs(MLModelRunner &Runner) { |
257 | #define _RESET(TYPE, NAME, SHAPE, __) \ |
258 | std::memset(Runner.getTensorUntyped(FeatureIDs::NAME), 0, \ |
259 | getTotalSize<TYPE>(SHAPE)); |
260 | RA_EVICT_FEATURES_LIST(_RESET) |
261 | if (EnableDevelopmentFeatures) { |
262 | RA_EVICT_FIRST_DEVELOPMENT_FEATURE(_RESET) |
263 | RA_EVICT_REST_DEVELOPMENT_FEATURES(_RESET) |
264 | #undef _RESET |
265 | } |
266 | } |
267 | |
268 | // Per-live interval components that get aggregated into the feature values |
269 | // that will be passed to the evaluator. |
270 | struct LIFeatureComponents { |
271 | double R = 0; |
272 | double W = 0; |
273 | double RW = 0; |
274 | double IndVarUpdates = 0; |
275 | double HintWeights = 0.0; |
276 | int64_t NrDefsAndUses = 0; |
277 | float HottestBlockFreq = 0.0; |
278 | bool IsRemat = false; |
279 | }; |
280 | |
281 | using CandidateRegList = |
282 | std::array<std::pair<MCRegister, bool>, NumberOfInterferences>; |
283 | using FeaturesListNormalizer = |
284 | llvm::SmallVector<float, FeatureIDs::FeatureCount>; |
285 | |
286 | /// The ML evictor (commonalities between release and development mode) |
287 | class MLEvictAdvisor : public RegAllocEvictionAdvisor { |
288 | public: |
289 | MLEvictAdvisor(const MachineFunction &MF, const RAGreedy &RA, |
290 | MLModelRunner *Runner, const MachineBlockFrequencyInfo &MBFI, |
291 | const MachineLoopInfo &Loops); |
292 | |
293 | protected: |
294 | const RegAllocEvictionAdvisor &getDefaultAdvisor() const { |
295 | return static_cast<const RegAllocEvictionAdvisor &>(DefaultAdvisor); |
296 | } |
297 | |
298 | // The assumption is that if the Runner could not be constructed, we emit-ed |
299 | // error, and we shouldn't be asking for it here. |
300 | const MLModelRunner &getRunner() const { return *Runner; } |
301 | |
302 | /// This just calls Evaluate on the Runner, but in the development mode |
303 | /// case, if we're just capturing the log of the default advisor, it needs |
304 | /// to call the latter instead, so we need to pass all the necessary |
305 | /// parameters for it. In the development case, it will also log. |
306 | virtual int64_t |
307 | tryFindEvictionCandidatePosition(const LiveInterval &VirtReg, |
308 | const AllocationOrder &Order, |
309 | unsigned OrderLimit, uint8_t CostPerUseLimit, |
310 | const SmallVirtRegSet &FixedRegisters) const; |
311 | |
312 | /// Load the features of the given VirtReg (allocated or not) at column Pos, |
313 | /// but if that can't be evicted, return false instead. |
314 | bool |
315 | loadInterferenceFeatures(const LiveInterval &VirtReg, MCRegister PhysReg, |
316 | bool IsHint, const SmallVirtRegSet &FixedRegisters, |
317 | llvm::SmallVectorImpl<float> &Largest, size_t Pos, |
318 | SmallVectorImpl<LRStartEndInfo> &LRPosInfo) const; |
319 | |
320 | private: |
321 | static float getInitialQueueSize(const MachineFunction &MF); |
322 | |
323 | MCRegister tryFindEvictionCandidate( |
324 | const LiveInterval &VirtReg, const AllocationOrder &Order, |
325 | uint8_t CostPerUseLimit, |
326 | const SmallVirtRegSet &FixedRegisters) const override; |
327 | |
328 | void extractFeatures(const SmallVectorImpl<const LiveInterval *> &Intervals, |
329 | llvm::SmallVectorImpl<float> &Largest, size_t Pos, |
330 | int64_t IsHint, int64_t LocalIntfsCount, float NrUrgent, |
331 | SmallVectorImpl<LRStartEndInfo> &LRPosInfo) const; |
332 | |
333 | // Point-in-time: we didn't learn this, so we always delegate to the |
334 | // default. |
335 | bool canEvictHintInterference( |
336 | const LiveInterval &VirtReg, MCRegister PhysReg, |
337 | const SmallVirtRegSet &FixedRegisters) const override { |
338 | return getDefaultAdvisor().canEvictHintInterference(VirtReg, PhysReg, |
339 | FixedRegisters); |
340 | } |
341 | |
342 | const LIFeatureComponents & |
343 | getLIFeatureComponents(const LiveInterval &LI) const; |
344 | |
345 | // Hold on to a default advisor for: |
346 | // 1) the implementation of canEvictHintInterference, because we didn't |
347 | // learn that nuance yet; 2) for bootstrapping (logging) in the development |
348 | // mode case. |
349 | const DefaultEvictionAdvisor DefaultAdvisor; |
350 | MLModelRunner *const Runner; |
351 | const MachineBlockFrequencyInfo &MBFI; |
352 | const MachineLoopInfo &Loops; |
353 | |
354 | // Indices of those features we don't want to normalize. |
355 | // This could be static and shared, but its initialization is non-trivial. |
356 | std::bitset<FeatureIDs::FeatureCount> DoNotNormalize; |
357 | const float InitialQSize; |
358 | |
359 | using RegID = unsigned; |
360 | mutable DenseMap<RegID, LIFeatureComponents> CachedFeatures; |
361 | }; |
362 | |
363 | #define _DECL_FEATURES(type, name, shape, _) \ |
364 | TensorSpec::createSpec<type>(#name, shape), |
365 | |
366 | // =================================== |
367 | // Release (AOT) - specifics |
368 | // =================================== |
369 | class ReleaseModeEvictionAdvisorAnalysis final |
370 | : public RegAllocEvictionAdvisorAnalysis { |
371 | public: |
372 | ReleaseModeEvictionAdvisorAnalysis() |
373 | : RegAllocEvictionAdvisorAnalysis(AdvisorMode::Release) { |
374 | if (EnableDevelopmentFeatures) { |
375 | InputFeatures = {RA_EVICT_FEATURES_LIST( |
376 | _DECL_FEATURES) RA_EVICT_FIRST_DEVELOPMENT_FEATURE(_DECL_FEATURES) |
377 | RA_EVICT_REST_DEVELOPMENT_FEATURES(_DECL_FEATURES)}; |
378 | } else { |
379 | InputFeatures = {RA_EVICT_FEATURES_LIST(_DECL_FEATURES)}; |
380 | } |
381 | } |
382 | // support for isa<> and dyn_cast. |
383 | static bool classof(const RegAllocEvictionAdvisorAnalysis *R) { |
384 | return R->getAdvisorMode() == AdvisorMode::Release; |
385 | } |
386 | |
387 | private: |
388 | std::vector<TensorSpec> InputFeatures; |
389 | |
390 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
391 | AU.addRequired<MachineBlockFrequencyInfoWrapperPass>(); |
392 | AU.addRequired<MachineLoopInfoWrapperPass>(); |
393 | RegAllocEvictionAdvisorAnalysis::getAnalysisUsage(AU); |
394 | } |
395 | |
396 | std::unique_ptr<RegAllocEvictionAdvisor> |
397 | getAdvisor(const MachineFunction &MF, const RAGreedy &RA) override { |
398 | if (!Runner) { |
399 | if (InteractiveChannelBaseName.empty()) |
400 | Runner = std::make_unique<ReleaseModeModelRunner<CompiledModelType>>( |
401 | args&: MF.getFunction().getContext(), args&: InputFeatures, DecisionName); |
402 | else |
403 | Runner = std::make_unique<InteractiveModelRunner>( |
404 | args&: MF.getFunction().getContext(), args&: InputFeatures, args: DecisionSpec, |
405 | args: InteractiveChannelBaseName + ".out" , |
406 | args: InteractiveChannelBaseName + ".in" ); |
407 | } |
408 | return std::make_unique<MLEvictAdvisor>( |
409 | args: MF, args: RA, args: Runner.get(), |
410 | args&: getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI(), |
411 | args&: getAnalysis<MachineLoopInfoWrapperPass>().getLI()); |
412 | } |
413 | std::unique_ptr<MLModelRunner> Runner; |
414 | }; |
415 | |
416 | // =================================== |
417 | // Development mode-specifics |
418 | // =================================== |
419 | // |
420 | // Features we log |
421 | #ifdef LLVM_HAVE_TFLITE |
422 | static const TensorSpec Reward = TensorSpec::createSpec<float>("reward" , {1}); |
423 | |
424 | // Features we bind on the model. The tensor names have a prefix, and we also |
425 | // need to include some tensors that are expected to be present by the |
426 | // training algo. |
427 | // TODO: can we just get rid of these? |
428 | #define _DECL_TRAIN_FEATURES(type, name, shape, _) \ |
429 | TensorSpec::createSpec<type>(std::string("action_") + #name, shape), |
430 | |
431 | class DevelopmentModeEvictAdvisor : public MLEvictAdvisor { |
432 | public: |
433 | DevelopmentModeEvictAdvisor(const MachineFunction &MF, const RAGreedy &RA, |
434 | MLModelRunner *Runner, |
435 | const MachineBlockFrequencyInfo &MBFI, |
436 | const MachineLoopInfo &Loops, Logger *Log) |
437 | : MLEvictAdvisor(MF, RA, Runner, MBFI, Loops), Log(Log) {} |
438 | |
439 | private: |
440 | int64_t tryFindEvictionCandidatePosition( |
441 | const LiveInterval &VirtReg, const AllocationOrder &Order, |
442 | unsigned OrderLimit, uint8_t CostPerUseLimit, |
443 | const SmallVirtRegSet &FixedRegisters) const override; |
444 | |
445 | Logger *const Log; |
446 | }; |
447 | |
448 | class DevelopmentModeEvictionAdvisorAnalysis final |
449 | : public RegAllocEvictionAdvisorAnalysis { |
450 | public: |
451 | DevelopmentModeEvictionAdvisorAnalysis() |
452 | : RegAllocEvictionAdvisorAnalysis(AdvisorMode::Development) { |
453 | if (EnableDevelopmentFeatures) { |
454 | InputFeatures = {RA_EVICT_FEATURES_LIST( |
455 | _DECL_FEATURES) RA_EVICT_FIRST_DEVELOPMENT_FEATURE(_DECL_FEATURES) |
456 | RA_EVICT_REST_DEVELOPMENT_FEATURES(_DECL_FEATURES)}; |
457 | TrainingInputFeatures = { |
458 | RA_EVICT_FEATURES_LIST(_DECL_TRAIN_FEATURES) |
459 | RA_EVICT_FIRST_DEVELOPMENT_FEATURE(_DECL_TRAIN_FEATURES) |
460 | RA_EVICT_REST_DEVELOPMENT_FEATURES(_DECL_TRAIN_FEATURES) |
461 | TensorSpec::createSpec<float>("action_discount" , {1}), |
462 | TensorSpec::createSpec<int32_t>("action_step_type" , {1}), |
463 | TensorSpec::createSpec<float>("action_reward" , {1})}; |
464 | } else { |
465 | InputFeatures = {RA_EVICT_FEATURES_LIST(_DECL_FEATURES)}; |
466 | TrainingInputFeatures = { |
467 | RA_EVICT_FEATURES_LIST(_DECL_TRAIN_FEATURES) |
468 | TensorSpec::createSpec<float>("action_discount" , {1}), |
469 | TensorSpec::createSpec<int32_t>("action_step_type" , {1}), |
470 | TensorSpec::createSpec<float>("action_reward" , {1})}; |
471 | } |
472 | } |
473 | // support for isa<> and dyn_cast. |
474 | static bool classof(const RegAllocEvictionAdvisorAnalysis *R) { |
475 | return R->getAdvisorMode() == AdvisorMode::Development; |
476 | } |
477 | |
478 | void logRewardIfNeeded(const MachineFunction &MF, |
479 | llvm::function_ref<float()> GetReward) override { |
480 | if (!Log || !Log->hasAnyObservationForContext(MF.getName())) |
481 | return; |
482 | // The function pass manager would run all the function passes for a |
483 | // function, so we assume the last context belongs to this function. If |
484 | // this invariant ever changes, we can implement at that time switching |
485 | // contexts. At this point, it'd be an error |
486 | if (Log->currentContext() != MF.getName()) { |
487 | MF.getFunction().getContext().emitError( |
488 | "The training log context shouldn't have had changed." ); |
489 | } |
490 | if (Log->hasObservationInProgress()) |
491 | Log->logReward<float>(GetReward()); |
492 | } |
493 | |
494 | private: |
495 | std::vector<TensorSpec> InputFeatures; |
496 | std::vector<TensorSpec> TrainingInputFeatures; |
497 | |
498 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
499 | AU.addRequired<MachineBlockFrequencyInfoWrapperPass>(); |
500 | AU.addRequired<MachineLoopInfoWrapperPass>(); |
501 | RegAllocEvictionAdvisorAnalysis::getAnalysisUsage(AU); |
502 | } |
503 | |
504 | bool doInitialization(Module &M) override { |
505 | LLVMContext &Ctx = M.getContext(); |
506 | if (ModelUnderTraining.empty() && TrainingLog.empty()) { |
507 | Ctx.emitError("Regalloc development mode should be requested with at " |
508 | "least logging enabled and/or a training model" ); |
509 | return false; |
510 | } |
511 | if (ModelUnderTraining.empty()) |
512 | Runner = std::make_unique<NoInferenceModelRunner>(Ctx, InputFeatures); |
513 | else |
514 | Runner = ModelUnderTrainingRunner::createAndEnsureValid( |
515 | Ctx, ModelUnderTraining, DecisionName, TrainingInputFeatures); |
516 | if (!Runner) { |
517 | Ctx.emitError("Regalloc: could not set up the model runner" ); |
518 | return false; |
519 | } |
520 | if (TrainingLog.empty()) |
521 | return false; |
522 | std::error_code EC; |
523 | auto OS = std::make_unique<raw_fd_ostream>(TrainingLog, EC); |
524 | if (EC) { |
525 | M.getContext().emitError(EC.message() + ":" + TrainingLog); |
526 | return false; |
527 | } |
528 | std::vector<TensorSpec> LFS = InputFeatures; |
529 | if (auto *MUTR = dyn_cast<ModelUnderTrainingRunner>(Runner.get())) |
530 | append_range(LFS, MUTR->extraOutputsForLoggingSpecs()); |
531 | // We always log the output; in particular, if we're not evaluating, we |
532 | // don't have an output spec json file. That's why we handle the |
533 | // 'normal' output separately. |
534 | LFS.push_back(DecisionSpec); |
535 | |
536 | Log = std::make_unique<Logger>(std::move(OS), LFS, Reward, |
537 | /*IncludeReward*/ true); |
538 | return false; |
539 | } |
540 | |
541 | std::unique_ptr<RegAllocEvictionAdvisor> |
542 | getAdvisor(const MachineFunction &MF, const RAGreedy &RA) override { |
543 | if (!Runner) |
544 | return nullptr; |
545 | if (Log) |
546 | Log->switchContext(MF.getName()); |
547 | return std::make_unique<DevelopmentModeEvictAdvisor>( |
548 | MF, RA, Runner.get(), |
549 | getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI(), |
550 | getAnalysis<MachineLoopInfoWrapperPass>().getLI(), Log.get()); |
551 | } |
552 | |
553 | std::unique_ptr<MLModelRunner> Runner; |
554 | std::unique_ptr<Logger> Log; |
555 | }; |
556 | |
557 | #endif //#ifdef LLVM_HAVE_TFLITE |
558 | } // namespace |
559 | |
560 | float MLEvictAdvisor::getInitialQueueSize(const MachineFunction &MF) { |
561 | auto &MRI = MF.getRegInfo(); |
562 | float Ret = 0.0; |
563 | for (unsigned I = 0, E = MRI.getNumVirtRegs(); I != E; ++I) { |
564 | Register Reg = Register::index2VirtReg(Index: I); |
565 | if (MRI.reg_nodbg_empty(RegNo: Reg)) |
566 | continue; |
567 | ++Ret; |
568 | } |
569 | return Ret; |
570 | } |
571 | |
572 | MLEvictAdvisor::MLEvictAdvisor(const MachineFunction &MF, const RAGreedy &RA, |
573 | MLModelRunner *Runner, |
574 | const MachineBlockFrequencyInfo &MBFI, |
575 | const MachineLoopInfo &Loops) |
576 | : RegAllocEvictionAdvisor(MF, RA), DefaultAdvisor(MF, RA), |
577 | Runner(std::move(Runner)), MBFI(MBFI), Loops(Loops), |
578 | InitialQSize(MLEvictAdvisor::getInitialQueueSize(MF)) { |
579 | assert(this->Runner); |
580 | Runner->switchContext(Name: MF.getName()); |
581 | DoNotNormalize.set(position: FeatureIDs::mask); |
582 | DoNotNormalize.set(position: FeatureIDs::is_free); |
583 | DoNotNormalize.set(position: FeatureIDs::is_hint); |
584 | DoNotNormalize.set(position: FeatureIDs::is_local); |
585 | DoNotNormalize.set(position: FeatureIDs::min_stage); |
586 | DoNotNormalize.set(position: FeatureIDs::max_stage); |
587 | DoNotNormalize.set(position: FeatureIDs::progress); |
588 | } |
589 | |
590 | int64_t MLEvictAdvisor::tryFindEvictionCandidatePosition( |
591 | const LiveInterval &, const AllocationOrder &, unsigned, uint8_t, |
592 | const SmallVirtRegSet &) const { |
593 | int64_t Ret = Runner->evaluate<int64_t>(); |
594 | assert(Ret >= 0); |
595 | assert(Ret <= CandidateVirtRegPos); |
596 | return Ret; |
597 | } |
598 | |
599 | bool MLEvictAdvisor::loadInterferenceFeatures( |
600 | const LiveInterval &VirtReg, MCRegister PhysReg, bool IsHint, |
601 | const SmallVirtRegSet &FixedRegisters, |
602 | llvm::SmallVectorImpl<float> &Largest, size_t Pos, |
603 | llvm::SmallVectorImpl<LRStartEndInfo> &LRPosInfo) const { |
604 | // It is only possible to evict virtual register interference. |
605 | if (Matrix->checkInterference(VirtReg, PhysReg) > LiveRegMatrix::IK_VirtReg) { |
606 | // leave unavailable |
607 | return false; |
608 | } |
609 | |
610 | const bool IsLocal = LIS->intervalIsInOneMBB(LI: VirtReg); |
611 | int64_t LocalIntfs = 0; |
612 | float NrUrgent = 0.0f; |
613 | |
614 | // The cascade tracking is the same as in the default advisor |
615 | unsigned Cascade = RA.getExtraInfo().getCascadeOrCurrentNext(Reg: VirtReg.reg()); |
616 | |
617 | SmallVector<const LiveInterval *, MaxInterferences> InterferingIntervals; |
618 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) { |
619 | LiveIntervalUnion::Query &Q = Matrix->query(LR: VirtReg, RegUnit: Unit); |
620 | // Different from the default heuristic, we don't make any assumptions |
621 | // about what having more than 10 results in the query may mean. |
622 | const auto &IFIntervals = Q.interferingVRegs(MaxInterferingRegs: EvictInterferenceCutoff); |
623 | if (IFIntervals.empty() && InterferingIntervals.empty()) |
624 | continue; |
625 | if (IFIntervals.size() >= EvictInterferenceCutoff) |
626 | return false; |
627 | InterferingIntervals.append(in_start: IFIntervals.begin(), in_end: IFIntervals.end()); |
628 | for (const LiveInterval *Intf : reverse(C: IFIntervals)) { |
629 | assert(Intf->reg().isVirtual() && |
630 | "Only expecting virtual register interference from query" ); |
631 | // This is the same set of legality checks as in the default case: don't |
632 | // try to evict fixed regs or 'done' ones. Also don't break cascades, |
633 | // except in the urgent case, with the same nuances used in the default |
634 | // heuristic. |
635 | // We could try sharing this between the advisors, but it may end up |
636 | // more complex than it is right now. |
637 | if (FixedRegisters.count(V: Intf->reg())) |
638 | return false; |
639 | if (RA.getExtraInfo().getStage(VirtReg: *Intf) == RS_Done) |
640 | return false; |
641 | bool Urgent = |
642 | !VirtReg.isSpillable() && |
643 | (Intf->isSpillable() || |
644 | RegClassInfo.getNumAllocatableRegs(RC: MRI->getRegClass(Reg: VirtReg.reg())) < |
645 | RegClassInfo.getNumAllocatableRegs( |
646 | RC: MRI->getRegClass(Reg: Intf->reg()))); |
647 | // Only evict older cascades or live ranges without a cascade. |
648 | unsigned IntfCascade = RA.getExtraInfo().getCascade(Reg: Intf->reg()); |
649 | if (Cascade <= IntfCascade) { |
650 | if (!Urgent) |
651 | return false; |
652 | ++NrUrgent; |
653 | } |
654 | |
655 | LocalIntfs += (IsLocal && LIS->intervalIsInOneMBB(LI: *Intf) && |
656 | (!EnableLocalReassign || !canReassign(VirtReg: *Intf, FromReg: PhysReg))); |
657 | } |
658 | } |
659 | // OK, so if we made it this far, this LR is an eviction candidate, load its |
660 | // features. |
661 | extractFeatures(Intervals: InterferingIntervals, Largest, Pos, IsHint, LocalIntfsCount: LocalIntfs, |
662 | NrUrgent, LRPosInfo); |
663 | return true; |
664 | } |
665 | |
666 | MCRegister MLEvictAdvisor::tryFindEvictionCandidate( |
667 | const LiveInterval &VirtReg, const AllocationOrder &Order, |
668 | uint8_t CostPerUseLimit, const SmallVirtRegSet &FixedRegisters) const { |
669 | auto MaybeOrderLimit = getOrderLimit(VirtReg, Order, CostPerUseLimit); |
670 | if (!MaybeOrderLimit) |
671 | return MCRegister::NoRegister; |
672 | unsigned OrderLimit = *MaybeOrderLimit; |
673 | |
674 | // The heuristic sets initial costs such as, if CostPerUseLimit is |
675 | // max<uint8_t>, then any of the costs of the legally-evictable intervals |
676 | // would be lower. When that happens, one of those will be selected. |
677 | // Therefore, we allow the candidate be selected, unless the candidate is |
678 | // unspillable, in which case it would be incorrect to not find a register |
679 | // for it. |
680 | const bool MustFindEviction = |
681 | (!VirtReg.isSpillable() && CostPerUseLimit == static_cast<uint8_t>(~0u)); |
682 | // Number of available candidates - if 0, no need to continue. |
683 | size_t Available = 0; |
684 | // Make sure we don't have leftover partial state from an attempt where we |
685 | // had no available candidates and bailed out early. |
686 | resetInputs(Runner&: *Runner); |
687 | |
688 | // Track the index->register mapping because AllocationOrder doesn't do that |
689 | // and we'd have to scan it. |
690 | // Also track their mask, to write asserts/debug. |
691 | CandidateRegList Regs; |
692 | Regs.fill(u: {0, false}); |
693 | |
694 | // Track the largest value of features seen during this eviction session. We |
695 | // only normalize (some of) the float features, but it's just simpler to |
696 | // dimension 'Largest' to all the features, especially since we have the |
697 | // 'DoNotNormalize' list. |
698 | FeaturesListNormalizer Largest(FeatureIDs::FeatureCount, 0.0); |
699 | |
700 | // Same overal idea as in the default eviction policy - we visit the values |
701 | // of AllocationOrder one at a time. If it's not legally available, we mask |
702 | // off the corresponding feature column (==do nothing because we already |
703 | // reset all the features to 0) Use Pos to capture the column we load |
704 | // features at - in AllocationOrder order. |
705 | size_t Pos = 0; |
706 | SmallVector<LRStartEndInfo, NumberOfInterferences> LRPosInfo; |
707 | for (auto I = Order.begin(), E = Order.getOrderLimitEnd(OrderLimit); I != E; |
708 | ++I, ++Pos) { |
709 | MCRegister PhysReg = *I; |
710 | assert(!Regs[Pos].second); |
711 | assert(PhysReg); |
712 | if (!canAllocatePhysReg(CostPerUseLimit, PhysReg)) { |
713 | continue; |
714 | } |
715 | if (loadInterferenceFeatures(VirtReg, PhysReg, IsHint: I.isHint(), FixedRegisters, |
716 | Largest, Pos, LRPosInfo)) { |
717 | ++Available; |
718 | Regs[Pos] = std::make_pair(x&: PhysReg, y: true); |
719 | } |
720 | } |
721 | if (Available == 0) { |
722 | // Nothing to decide, nothing to learn. |
723 | assert(!MustFindEviction); |
724 | return MCRegister::NoRegister; |
725 | } |
726 | const size_t ValidPosLimit = Pos; |
727 | // If we must find eviction, the candidate should be masked out of the |
728 | // decision making process. |
729 | Regs[CandidateVirtRegPos].second = !MustFindEviction; |
730 | if (!MustFindEviction) |
731 | extractFeatures(Intervals: SmallVector<const LiveInterval *, 1>(1, &VirtReg), Largest, |
732 | Pos: CandidateVirtRegPos, /*IsHint*/ 0, |
733 | /*LocalIntfsCount*/ 0, |
734 | /*NrUrgent*/ 0.0, LRPosInfo); |
735 | assert(InitialQSize > 0.0 && "We couldn't have gotten here if we had " |
736 | "nothing to allocate initially." ); |
737 | #ifdef LLVM_HAVE_TFLITE |
738 | if (EnableDevelopmentFeatures) { |
739 | extractInstructionFeatures( |
740 | LRPosInfo, Runner, |
741 | [this](SlotIndex InputIndex) -> int { |
742 | auto *CurrentMachineInstruction = |
743 | LIS->getInstructionFromIndex(InputIndex); |
744 | if (!CurrentMachineInstruction) { |
745 | return -1; |
746 | } |
747 | return CurrentMachineInstruction->getOpcode(); |
748 | }, |
749 | [this](SlotIndex InputIndex) -> float { |
750 | auto *CurrentMachineInstruction = |
751 | LIS->getInstructionFromIndex(InputIndex); |
752 | return MBFI.getBlockFreqRelativeToEntryBlock( |
753 | CurrentMachineInstruction->getParent()); |
754 | }, |
755 | [this](SlotIndex InputIndex) -> MachineBasicBlock * { |
756 | auto *CurrentMachineInstruction = |
757 | LIS->getInstructionFromIndex(InputIndex); |
758 | return CurrentMachineInstruction->getParent(); |
759 | }, |
760 | FeatureIDs::instructions, FeatureIDs::instructions_mapping, |
761 | FeatureIDs::mbb_frequencies, FeatureIDs::mbb_mapping, |
762 | LIS->getSlotIndexes()->getLastIndex()); |
763 | } |
764 | #endif // #ifdef LLVM_HAVE_TFLITE |
765 | // Normalize the features. |
766 | for (auto &V : Largest) |
767 | V = V ? V : 1.0; |
768 | for (size_t FeatureIndex = 0; FeatureIndex < FeatureIDs::FeatureCount; |
769 | ++FeatureIndex) { |
770 | if (DoNotNormalize.test(position: FeatureIndex)) |
771 | continue; |
772 | for (size_t Pos = 0; Pos < NumberOfInterferences; ++Pos) { |
773 | Runner->getTensor<float>(FeatureID: FeatureIndex)[Pos] /= Largest[FeatureIndex]; |
774 | } |
775 | } |
776 | *Runner->getTensor<float>(FeatureID: FeatureIDs::progress) = |
777 | static_cast<float>(RA.getQueueSize()) / InitialQSize; |
778 | |
779 | // Get a decision. |
780 | size_t CandidatePos = tryFindEvictionCandidatePosition( |
781 | VirtReg, Order, OrderLimit, CostPerUseLimit, FixedRegisters); |
782 | // The contract with the ML side is that CandidatePos is mask == 1 (i.e. |
783 | // Regs[CandidatePos].second) |
784 | assert(Regs[CandidatePos].second); |
785 | if (CandidatePos == CandidateVirtRegPos) { |
786 | assert(!MustFindEviction); |
787 | return MCRegister::NoRegister; |
788 | } |
789 | assert(CandidatePos < ValidPosLimit); |
790 | (void)ValidPosLimit; |
791 | return Regs[CandidatePos].first; |
792 | } |
793 | |
794 | const LIFeatureComponents & |
795 | MLEvictAdvisor::getLIFeatureComponents(const LiveInterval &LI) const { |
796 | RegID ID = LI.reg().id(); |
797 | LIFeatureComponents Empty; |
798 | auto I = CachedFeatures.insert(KV: std::make_pair(x&: ID, y&: Empty)); |
799 | LIFeatureComponents &Ret = I.first->getSecond(); |
800 | if (!I.second) |
801 | return Ret; |
802 | |
803 | SmallPtrSet<MachineInstr *, 8> Visited; |
804 | const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo(); |
805 | |
806 | for (MachineRegisterInfo::reg_instr_nodbg_iterator |
807 | I = MRI->reg_instr_nodbg_begin(RegNo: LI.reg()), |
808 | E = MRI->reg_instr_nodbg_end(); |
809 | I != E;) { |
810 | MachineInstr *MI = &*(I++); |
811 | |
812 | ++Ret.NrDefsAndUses; |
813 | if (!Visited.insert(Ptr: MI).second) |
814 | continue; |
815 | |
816 | if (MI->isIdentityCopy() || MI->isImplicitDef()) |
817 | continue; |
818 | |
819 | bool Reads, Writes; |
820 | std::tie(args&: Reads, args&: Writes) = MI->readsWritesVirtualRegister(Reg: LI.reg()); |
821 | |
822 | float Freq = MBFI.getBlockFreqRelativeToEntryBlock(MBB: MI->getParent()); |
823 | Ret.HottestBlockFreq = std::max(a: Freq, b: Ret.HottestBlockFreq); |
824 | |
825 | Ret.R += (Reads && !Writes) * Freq; |
826 | Ret.W += (!Reads && Writes) * Freq; |
827 | Ret.RW += (Reads && Writes) * Freq; |
828 | |
829 | auto *MBB = MI->getParent(); |
830 | auto *Loop = Loops.getLoopFor(BB: MBB); |
831 | bool IsExiting = Loop ? Loop->isLoopExiting(BB: MBB) : false; |
832 | |
833 | if (Writes && IsExiting && LIS->isLiveOutOfMBB(LR: LI, mbb: MBB)) |
834 | Ret.IndVarUpdates += Freq; |
835 | |
836 | if (MI->isCopy() && VirtRegAuxInfo::copyHint(MI, Reg: LI.reg(), TRI, MRI: *MRI)) |
837 | Ret.HintWeights += Freq; |
838 | } |
839 | Ret.IsRemat = VirtRegAuxInfo::isRematerializable( |
840 | LI, LIS: *LIS, VRM: *VRM, TII: *MF.getSubtarget().getInstrInfo()); |
841 | return Ret; |
842 | } |
843 | |
844 | // Overall, this currently mimics what we do for weight calculation, but instead |
845 | // of accummulating the various features, we keep them separate. |
846 | void MLEvictAdvisor::( |
847 | const SmallVectorImpl<const LiveInterval *> &Intervals, |
848 | llvm::SmallVectorImpl<float> &Largest, size_t Pos, int64_t IsHint, |
849 | int64_t LocalIntfsCount, float NrUrgent, |
850 | SmallVectorImpl<LRStartEndInfo> &LRPosInfo) const { |
851 | int64_t NrDefsAndUses = 0; |
852 | int64_t NrBrokenHints = 0; |
853 | double R = 0.0; |
854 | double W = 0.0; |
855 | double RW = 0.0; |
856 | double IndVarUpdates = 0.0; |
857 | double HintWeights = 0.0; |
858 | float StartBBFreq = 0.0; |
859 | float EndBBFreq = 0.0; |
860 | float HottestBlockFreq = 0.0; |
861 | int32_t NrRematerializable = 0; |
862 | float TotalWeight = 0.0; |
863 | |
864 | SlotIndex EndSI = LIS->getSlotIndexes()->getZeroIndex(); |
865 | SlotIndex StartSI = LIS->getSlotIndexes()->getLastIndex(); |
866 | int64_t MaxStage = 0; |
867 | int64_t MinStage = |
868 | Intervals.empty() ? 0 : std::numeric_limits<int64_t>::max(); |
869 | |
870 | for (const auto *L : Intervals) { |
871 | const LiveInterval &LI = *L; |
872 | MaxStage = std::max<int64_t>( |
873 | a: MaxStage, b: static_cast<int64_t>(RA.getExtraInfo().getStage(VirtReg: LI))); |
874 | MinStage = std::min<int64_t>( |
875 | a: MinStage, b: static_cast<int64_t>(RA.getExtraInfo().getStage(VirtReg: LI))); |
876 | |
877 | TotalWeight = std::max(a: TotalWeight, b: LI.weight()); |
878 | |
879 | if (LI.beginIndex() < StartSI) |
880 | StartSI = LI.beginIndex(); |
881 | |
882 | if (LI.endIndex() > EndSI) |
883 | EndSI = LI.endIndex(); |
884 | const LIFeatureComponents &LIFC = getLIFeatureComponents(LI); |
885 | NrBrokenHints += VRM->hasPreferredPhys(VirtReg: LI.reg()); |
886 | |
887 | NrDefsAndUses += LIFC.NrDefsAndUses; |
888 | HottestBlockFreq = std::max(a: HottestBlockFreq, b: LIFC.HottestBlockFreq); |
889 | R += LIFC.R; |
890 | W += LIFC.W; |
891 | RW += LIFC.RW; |
892 | |
893 | IndVarUpdates += LIFC.IndVarUpdates; |
894 | |
895 | HintWeights += LIFC.HintWeights; |
896 | NrRematerializable += LIFC.IsRemat; |
897 | |
898 | if (EnableDevelopmentFeatures) { |
899 | for (auto CurrentSegment : LI) { |
900 | LRPosInfo.push_back( |
901 | Elt: LRStartEndInfo{.Begin: CurrentSegment.start, .End: CurrentSegment.end, .Pos: Pos}); |
902 | } |
903 | } |
904 | } |
905 | size_t Size = 0; |
906 | if (!Intervals.empty()) { |
907 | StartBBFreq = |
908 | MBFI.getBlockFreqRelativeToEntryBlock(MBB: LIS->getMBBFromIndex(index: StartSI)); |
909 | if (EndSI >= LIS->getSlotIndexes()->getLastIndex()) |
910 | EndSI = LIS->getSlotIndexes()->getLastIndex().getPrevIndex(); |
911 | EndBBFreq = |
912 | MBFI.getBlockFreqRelativeToEntryBlock(MBB: LIS->getMBBFromIndex(index: EndSI)); |
913 | Size = StartSI.distance(other: EndSI); |
914 | } |
915 | // Set the features at the column 'Pos'. |
916 | #define SET(ID, TYPE, VAL) \ |
917 | do { \ |
918 | Runner->getTensor<TYPE>(FeatureIDs::ID)[Pos] = static_cast<TYPE>(VAL); \ |
919 | if (!DoNotNormalize.test(FeatureIDs::ID)) \ |
920 | Largest[FeatureIDs::ID] = \ |
921 | std::max(Largest[FeatureIDs::ID], static_cast<float>(VAL)); \ |
922 | } while (false) |
923 | SET(mask, int64_t, 1); |
924 | SET(is_free, int64_t, Intervals.empty()); |
925 | SET(nr_urgent, float, NrUrgent); |
926 | SET(nr_broken_hints, float, NrBrokenHints); |
927 | SET(is_hint, int64_t, IsHint); |
928 | SET(is_local, int64_t, LocalIntfsCount); |
929 | SET(nr_rematerializable, float, NrRematerializable); |
930 | SET(nr_defs_and_uses, float, NrDefsAndUses); |
931 | SET(weighed_reads_by_max, float, R); |
932 | SET(weighed_writes_by_max, float, W); |
933 | SET(weighed_read_writes_by_max, float, RW); |
934 | SET(weighed_indvars_by_max, float, IndVarUpdates); |
935 | SET(hint_weights_by_max, float, HintWeights); |
936 | SET(start_bb_freq_by_max, float, StartBBFreq); |
937 | SET(end_bb_freq_by_max, float, EndBBFreq); |
938 | SET(hottest_bb_freq_by_max, float, HottestBlockFreq); |
939 | SET(liverange_size, float, Size); |
940 | SET(use_def_density, float, TotalWeight); |
941 | SET(max_stage, int64_t, MaxStage); |
942 | SET(min_stage, int64_t, MinStage); |
943 | #undef SET |
944 | } |
945 | |
946 | void ( |
947 | SmallVectorImpl<LRStartEndInfo> &LRPosInfo, MLModelRunner *RegallocRunner, |
948 | function_ref<int(SlotIndex)> GetOpcode, |
949 | function_ref<float(SlotIndex)> GetMBBFreq, |
950 | function_ref<MachineBasicBlock *(SlotIndex)> GetMBBReference, |
951 | const int InstructionsIndex, const int InstructionsMappingIndex, |
952 | const int MBBFreqIndex, const int MBBMappingIndex, |
953 | const SlotIndex LastIndex) { |
954 | // This function extracts instruction based features relevant to the eviction |
955 | // problem currently being solved. This function ends up extracting two |
956 | // tensors. |
957 | // 1 - A vector of size max instruction count. It contains the opcodes of the |
958 | // instructions spanned by all the intervals in the current instance of the |
959 | // eviction problem. |
960 | // 2 - A binary mapping matrix of size (LR count * max |
961 | // instruction count) which maps where the LRs are live to the actual opcodes |
962 | // for which they are live. |
963 | // 3 - A vector of size max supported MBB count storing MBB frequencies, |
964 | // encompassing all of the MBBs covered by the eviction problem. |
965 | // 4 - A vector of size max instruction count of indices to members of the MBB |
966 | // frequency vector, mapping each instruction to its associated MBB. |
967 | |
968 | // Start off by sorting the segments based on the beginning slot index. |
969 | std::sort( |
970 | first: LRPosInfo.begin(), last: LRPosInfo.end(), |
971 | comp: [](LRStartEndInfo A, LRStartEndInfo B) { return A.Begin < B.Begin; }); |
972 | size_t InstructionIndex = 0; |
973 | size_t CurrentSegmentIndex = 0; |
974 | SlotIndex CurrentIndex = LRPosInfo[0].Begin; |
975 | std::map<MachineBasicBlock *, size_t> VisitedMBBs; |
976 | size_t CurrentMBBIndex = 0; |
977 | // This loop processes all the segments sequentially by starting at the |
978 | // beginning slot index of the first segment, iterating through all the slot |
979 | // indices before the end slot index of that segment (while checking for |
980 | // overlaps with segments that start at greater slot indices). After hitting |
981 | // that end index, the current segment being processed gets bumped until they |
982 | // are all processed or the max instruction count is hit, where everything is |
983 | // just truncated. |
984 | while (true) { |
985 | // If the index that we are currently at is within the current segment and |
986 | // we haven't hit the max instruction count, continue processing the current |
987 | // segment. |
988 | while (CurrentIndex <= LRPosInfo[CurrentSegmentIndex].End && |
989 | InstructionIndex < ModelMaxSupportedInstructionCount) { |
990 | int CurrentOpcode = GetOpcode(CurrentIndex); |
991 | // If the current machine instruction is null, skip it |
992 | if (CurrentOpcode == -1) { |
993 | // If we're currently at the last index in the SlotIndex analysis, |
994 | // we can't go any further, so return from the function |
995 | if (CurrentIndex >= LastIndex) { |
996 | return; |
997 | } |
998 | CurrentIndex = CurrentIndex.getNextIndex(); |
999 | continue; |
1000 | } |
1001 | MachineBasicBlock *CurrentMBBReference = GetMBBReference(CurrentIndex); |
1002 | if (VisitedMBBs.count(x: CurrentMBBReference) == 0) { |
1003 | VisitedMBBs[CurrentMBBReference] = CurrentMBBIndex; |
1004 | ++CurrentMBBIndex; |
1005 | } |
1006 | extractMBBFrequency(CurrentIndex, CurrentInstructionIndex: InstructionIndex, VisitedMBBs, |
1007 | GetMBBFreq, CurrentMBBReference, RegallocRunner, |
1008 | MBBFreqIndex, MBBMappingIndex); |
1009 | // Current code assumes we're not going to get any disjointed segments |
1010 | assert(LRPosInfo[CurrentSegmentIndex].Begin <= CurrentIndex); |
1011 | RegallocRunner->getTensor<int64_t>(FeatureID: InstructionsIndex)[InstructionIndex] = |
1012 | CurrentOpcode < OpcodeValueCutoff ? CurrentOpcode : 0; |
1013 | // set value in the binary mapping matrix for the current instruction |
1014 | auto CurrentSegmentPosition = LRPosInfo[CurrentSegmentIndex].Pos; |
1015 | RegallocRunner->getTensor<int64_t>( |
1016 | FeatureID: InstructionsMappingIndex)[CurrentSegmentPosition * |
1017 | ModelMaxSupportedInstructionCount + |
1018 | InstructionIndex] = 1; |
1019 | // All of the segments are sorted based on the beginning slot index, but |
1020 | // this doesn't mean that the beginning slot index of the next segment is |
1021 | // after the end segment of the one being currently processed. This while |
1022 | // loop checks for overlapping segments and modifies the portion of the |
1023 | // column in the mapping matrix for the currently processed instruction |
1024 | // for the LR it is checking. Also make sure that the beginning of the |
1025 | // current segment we're checking for overlap in is less than the current |
1026 | // index, otherwise we're done checking overlaps. |
1027 | size_t OverlapCheckCurrentSegment = CurrentSegmentIndex + 1; |
1028 | while (OverlapCheckCurrentSegment < LRPosInfo.size() && |
1029 | LRPosInfo[OverlapCheckCurrentSegment].Begin <= CurrentIndex) { |
1030 | auto OverlapCurrentSegmentPosition = |
1031 | LRPosInfo[OverlapCheckCurrentSegment].Pos; |
1032 | if (LRPosInfo[OverlapCheckCurrentSegment].End >= CurrentIndex) { |
1033 | RegallocRunner->getTensor<int64_t>( |
1034 | FeatureID: InstructionsMappingIndex)[OverlapCurrentSegmentPosition * |
1035 | ModelMaxSupportedInstructionCount + |
1036 | InstructionIndex] = 1; |
1037 | } |
1038 | ++OverlapCheckCurrentSegment; |
1039 | } |
1040 | ++InstructionIndex; |
1041 | if (CurrentIndex >= LastIndex) { |
1042 | return; |
1043 | } |
1044 | CurrentIndex = CurrentIndex.getNextIndex(); |
1045 | } |
1046 | // if we've just finished processing through the last segment or if we've |
1047 | // hit the maximum number of instructions, break out of the loop. |
1048 | if (CurrentSegmentIndex == LRPosInfo.size() - 1 || |
1049 | InstructionIndex >= ModelMaxSupportedInstructionCount) { |
1050 | break; |
1051 | } |
1052 | // If the segments are not overlapping, we need to move to the beginning |
1053 | // index of the next segment to avoid having instructions not attached to |
1054 | // any register. |
1055 | if (LRPosInfo[CurrentSegmentIndex + 1].Begin > |
1056 | LRPosInfo[CurrentSegmentIndex].End) { |
1057 | CurrentIndex = LRPosInfo[CurrentSegmentIndex + 1].Begin; |
1058 | } |
1059 | ++CurrentSegmentIndex; |
1060 | } |
1061 | } |
1062 | |
1063 | void (const SlotIndex CurrentIndex, |
1064 | const size_t CurrentInstructionIndex, |
1065 | std::map<MachineBasicBlock *, size_t> &VisitedMBBs, |
1066 | function_ref<float(SlotIndex)> GetMBBFreq, |
1067 | MachineBasicBlock *CurrentMBBReference, |
1068 | MLModelRunner *RegallocRunner, const int MBBFreqIndex, |
1069 | const int MBBMappingIndex) { |
1070 | size_t CurrentMBBIndex = VisitedMBBs[CurrentMBBReference]; |
1071 | float CurrentMBBFreq = GetMBBFreq(CurrentIndex); |
1072 | if (CurrentMBBIndex < ModelMaxSupportedMBBCount) { |
1073 | RegallocRunner->getTensor<float>(FeatureID: MBBFreqIndex)[CurrentMBBIndex] = |
1074 | CurrentMBBFreq; |
1075 | RegallocRunner->getTensor<int64_t>( |
1076 | FeatureID: MBBMappingIndex)[CurrentInstructionIndex] = CurrentMBBIndex; |
1077 | } |
1078 | } |
1079 | |
1080 | // Development mode-specific implementations |
1081 | #ifdef LLVM_HAVE_TFLITE |
1082 | |
1083 | RegAllocEvictionAdvisorAnalysis *llvm::createDevelopmentModeAdvisor() { |
1084 | return new DevelopmentModeEvictionAdvisorAnalysis(); |
1085 | } |
1086 | |
1087 | int64_t DevelopmentModeEvictAdvisor::tryFindEvictionCandidatePosition( |
1088 | const LiveInterval &VirtReg, const AllocationOrder &Order, |
1089 | unsigned OrderLimit, uint8_t CostPerUseLimit, |
1090 | const SmallVirtRegSet &FixedRegisters) const { |
1091 | int64_t Ret = 0; |
1092 | if (isa<ModelUnderTrainingRunner>(getRunner())) { |
1093 | Ret = MLEvictAdvisor::tryFindEvictionCandidatePosition( |
1094 | VirtReg, Order, OrderLimit, CostPerUseLimit, FixedRegisters); |
1095 | } else { |
1096 | MCRegister PhysReg = getDefaultAdvisor().tryFindEvictionCandidate( |
1097 | VirtReg, Order, CostPerUseLimit, FixedRegisters); |
1098 | // Find the index of the selected PhysReg. We need it for logging, |
1099 | // otherwise this is wasted cycles (but so would starting development mode |
1100 | // without a model nor logging) |
1101 | if (!PhysReg) |
1102 | Ret = CandidateVirtRegPos; |
1103 | else |
1104 | for (auto I = Order.begin(), E = Order.getOrderLimitEnd(OrderLimit); |
1105 | I != E; ++I, ++Ret) |
1106 | if (*I == PhysReg) |
1107 | break; |
1108 | } |
1109 | if (TrainingLog.empty()) |
1110 | return Ret; |
1111 | // TODO(mtrofin): when we support optional rewards, this can go away. In the |
1112 | // meantime, we log the "pretend" reward (0) for the previous observation |
1113 | // before starting a new one. |
1114 | if (Log->hasObservationInProgress()) |
1115 | Log->logReward<float>(0.0); |
1116 | |
1117 | Log->startObservation(); |
1118 | size_t CurrentFeature = 0; |
1119 | size_t FeatureCount = EnableDevelopmentFeatures |
1120 | ? FeatureIDs::FeaturesWithDevelopmentCount |
1121 | : FeatureIDs::FeatureCount; |
1122 | for (; CurrentFeature < FeatureCount; ++CurrentFeature) { |
1123 | Log->logTensorValue(CurrentFeature, |
1124 | reinterpret_cast<const char *>( |
1125 | getRunner().getTensorUntyped(CurrentFeature))); |
1126 | } |
1127 | if (auto *MUTR = dyn_cast<ModelUnderTrainingRunner>(&getRunner())) |
1128 | for (size_t I = 0; I < MUTR->extraOutputsForLoggingSpecs().size(); |
1129 | ++I, ++CurrentFeature) |
1130 | Log->logTensorValue( |
1131 | CurrentFeature, |
1132 | reinterpret_cast<const char *>(MUTR->getUntypedExtraOutputValue(I))); |
1133 | // The output is right after the features and the extra outputs |
1134 | Log->logTensorValue(CurrentFeature, reinterpret_cast<const char *>(&Ret)); |
1135 | Log->endObservation(); |
1136 | return Ret; |
1137 | } |
1138 | |
1139 | bool RegAllocScoring::runOnMachineFunction(MachineFunction &MF) { |
1140 | std::optional<float> CachedReward; |
1141 | auto GetReward = [&]() { |
1142 | if (!CachedReward) |
1143 | CachedReward = static_cast<float>( |
1144 | calculateRegAllocScore( |
1145 | MF, getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI()) |
1146 | .getScore()); |
1147 | return *CachedReward; |
1148 | }; |
1149 | |
1150 | getAnalysis<RegAllocEvictionAdvisorAnalysis>().logRewardIfNeeded(MF, |
1151 | GetReward); |
1152 | getAnalysis<RegAllocPriorityAdvisorAnalysis>().logRewardIfNeeded(MF, |
1153 | GetReward); |
1154 | return false; |
1155 | } |
1156 | #endif // #ifdef LLVM_HAVE_TFLITE |
1157 | |
1158 | RegAllocEvictionAdvisorAnalysis *llvm::createReleaseModeAdvisor() { |
1159 | return llvm::isEmbeddedModelEvaluatorValid<CompiledModelType>() || |
1160 | !InteractiveChannelBaseName.empty() |
1161 | ? new ReleaseModeEvictionAdvisorAnalysis() |
1162 | : nullptr; |
1163 | } |
1164 | |
1165 | // In all cases except development mode, we don't need scoring. |
1166 | #if !defined(LLVM_HAVE_TFLITE) |
1167 | bool RegAllocScoring::runOnMachineFunction(MachineFunction &) { return false; } |
1168 | #endif |
1169 | |