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