1 //===- MLInlineAdvisor.cpp - machine learned InlineAdvisor ----------------===// 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 // This file implements the interface between the inliner and a learned model. 10 // It delegates model evaluation to either the AOT compiled model (the 11 // 'release' mode) or a runtime-loaded model (the 'development' case). 12 // 13 //===----------------------------------------------------------------------===// 14 #include "llvm/Analysis/MLInlineAdvisor.h" 15 #include "llvm/ADT/SCCIterator.h" 16 #include "llvm/Analysis/AssumptionCache.h" 17 #include "llvm/Analysis/CallGraph.h" 18 #include "llvm/Analysis/FunctionPropertiesAnalysis.h" 19 #include "llvm/Analysis/InlineCost.h" 20 #include "llvm/Analysis/InlineModelFeatureMaps.h" 21 #include "llvm/Analysis/LazyCallGraph.h" 22 #include "llvm/Analysis/LoopInfo.h" 23 #include "llvm/Analysis/MLModelRunner.h" 24 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 25 #include "llvm/Analysis/TargetTransformInfo.h" 26 #include "llvm/IR/InstIterator.h" 27 #include "llvm/IR/PassManager.h" 28 #include "llvm/Support/CommandLine.h" 29 30 using namespace llvm; 31 32 #if defined(LLVM_HAVE_TF_AOT_INLINERSIZEMODEL) 33 #include "llvm/Analysis/ReleaseModeModelRunner.h" 34 // codegen-ed file 35 #include "InlinerSizeModel.h" // NOLINT 36 37 std::unique_ptr<InlineAdvisor> 38 llvm::getReleaseModeAdvisor(Module &M, ModuleAnalysisManager &MAM) { 39 auto AOTRunner = 40 std::make_unique<ReleaseModeModelRunner<llvm::InlinerSizeModel>>( 41 M.getContext(), FeatureMap, DecisionName); 42 return std::make_unique<MLInlineAdvisor>(M, MAM, std::move(AOTRunner)); 43 } 44 #endif 45 46 #define DEBUG_TYPE "inline-ml" 47 48 static cl::opt<float> SizeIncreaseThreshold( 49 "ml-advisor-size-increase-threshold", cl::Hidden, 50 cl::desc("Maximum factor by which expected native size may increase before " 51 "blocking any further inlining."), 52 cl::init(2.0)); 53 54 static cl::opt<bool> KeepFPICache( 55 "ml-advisor-keep-fpi-cache", cl::Hidden, 56 cl::desc( 57 "For test - keep the ML Inline advisor's FunctionPropertiesInfo cache"), 58 cl::init(false)); 59 60 // clang-format off 61 const std::array<TensorSpec, NumberOfFeatures> llvm::FeatureMap{ 62 #define POPULATE_NAMES(_, NAME) TensorSpec::createSpec<int64_t>(NAME, {1} ), 63 // InlineCost features - these must come first 64 INLINE_COST_FEATURE_ITERATOR(POPULATE_NAMES) 65 #undef POPULATE_NAMES 66 67 // Non-cost features 68 #define POPULATE_NAMES(_, NAME, __) TensorSpec::createSpec<int64_t>(NAME, {1} ), 69 INLINE_FEATURE_ITERATOR(POPULATE_NAMES) 70 #undef POPULATE_NAMES 71 }; 72 // clang-format on 73 74 const char *const llvm::DecisionName = "inlining_decision"; 75 const char *const llvm::DefaultDecisionName = "inlining_default"; 76 const char *const llvm::RewardName = "delta_size"; 77 78 CallBase *getInlinableCS(Instruction &I) { 79 if (auto *CS = dyn_cast<CallBase>(&I)) 80 if (Function *Callee = CS->getCalledFunction()) { 81 if (!Callee->isDeclaration()) { 82 return CS; 83 } 84 } 85 return nullptr; 86 } 87 88 MLInlineAdvisor::MLInlineAdvisor(Module &M, ModuleAnalysisManager &MAM, 89 std::unique_ptr<MLModelRunner> Runner) 90 : InlineAdvisor( 91 M, MAM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager()), 92 ModelRunner(std::move(Runner)), 93 CG(MAM.getResult<LazyCallGraphAnalysis>(M)), 94 InitialIRSize(getModuleIRSize()), CurrentIRSize(InitialIRSize) { 95 assert(ModelRunner); 96 97 // Extract the 'call site height' feature - the position of a call site 98 // relative to the farthest statically reachable SCC node. We don't mutate 99 // this value while inlining happens. Empirically, this feature proved 100 // critical in behavioral cloning - i.e. training a model to mimic the manual 101 // heuristic's decisions - and, thus, equally important for training for 102 // improvement. 103 CallGraph CGraph(M); 104 for (auto I = scc_begin(&CGraph); !I.isAtEnd(); ++I) { 105 const std::vector<CallGraphNode *> &CGNodes = *I; 106 unsigned Level = 0; 107 for (auto *CGNode : CGNodes) { 108 Function *F = CGNode->getFunction(); 109 if (!F || F->isDeclaration()) 110 continue; 111 for (auto &I : instructions(F)) { 112 if (auto *CS = getInlinableCS(I)) { 113 auto *Called = CS->getCalledFunction(); 114 auto Pos = FunctionLevels.find(&CG.get(*Called)); 115 // In bottom up traversal, an inlinable callee is either in the 116 // same SCC, or to a function in a visited SCC. So not finding its 117 // level means we haven't visited it yet, meaning it's in this SCC. 118 if (Pos == FunctionLevels.end()) 119 continue; 120 Level = std::max(Level, Pos->second + 1); 121 } 122 } 123 } 124 for (auto *CGNode : CGNodes) { 125 Function *F = CGNode->getFunction(); 126 if (F && !F->isDeclaration()) 127 FunctionLevels[&CG.get(*F)] = Level; 128 } 129 } 130 for (auto KVP : FunctionLevels) { 131 AllNodes.insert(KVP.first); 132 EdgeCount += getLocalCalls(KVP.first->getFunction()); 133 } 134 NodeCount = AllNodes.size(); 135 } 136 137 unsigned MLInlineAdvisor::getInitialFunctionLevel(const Function &F) const { 138 return CG.lookup(F) ? FunctionLevels.at(CG.lookup(F)) : 0; 139 } 140 141 void MLInlineAdvisor::onPassEntry() { 142 if (ForceStop) 143 return; 144 FPICache.clear(); 145 // Function passes executed between InlinerPass runs may have changed the 146 // module-wide features. 147 // The cgscc pass manager rules are such that: 148 // - if a pass leads to merging SCCs, then the pipeline is restarted on the 149 // merged SCC 150 // - if a pass leads to splitting the SCC, then we continue with one of the 151 // splits 152 // This means that the NodesInLastSCC is a superset (not strict) of the nodes 153 // that subsequent passes would have processed 154 // - in addition, if new Nodes were created by a pass (e.g. CoroSplit), 155 // they'd be adjacent to Nodes in the last SCC. So we just need to check the 156 // boundary of Nodes in NodesInLastSCC for Nodes we haven't seen. We don't 157 // care about the nature of the Edge (call or ref). 158 NodeCount -= static_cast<int64_t>(NodesInLastSCC.size()); 159 while (!NodesInLastSCC.empty()) { 160 const auto *N = NodesInLastSCC.front(); 161 NodesInLastSCC.pop_front(); 162 // The Function wrapped by N could have been deleted since we last saw it. 163 if (N->isDead()) { 164 assert(!N->getFunction().isDeclaration()); 165 continue; 166 } 167 ++NodeCount; 168 EdgeCount += getLocalCalls(N->getFunction()); 169 for (const auto &E : *(*N)) { 170 const auto *AdjNode = &E.getNode(); 171 assert(!AdjNode->isDead() && !AdjNode->getFunction().isDeclaration()); 172 auto I = AllNodes.insert(AdjNode); 173 if (I.second) 174 NodesInLastSCC.push_back(AdjNode); 175 } 176 } 177 178 EdgeCount -= EdgesOfLastSeenNodes; 179 EdgesOfLastSeenNodes = 0; 180 } 181 182 void MLInlineAdvisor::onPassExit(LazyCallGraph::SCC *LastSCC) { 183 // No need to keep this around - function passes will invalidate it. 184 if (!KeepFPICache) 185 FPICache.clear(); 186 if (!LastSCC || ForceStop) 187 return; 188 // Keep track of the nodes and edges we last saw. Then, in onPassEntry, 189 // we update the node count and edge count from the subset of these nodes that 190 // survived. 191 assert(NodesInLastSCC.empty()); 192 assert(NodeCount >= LastSCC->size()); 193 EdgesOfLastSeenNodes = 0; 194 for (const auto &N : *LastSCC) { 195 assert(!N.isDead()); 196 EdgesOfLastSeenNodes += getLocalCalls(N.getFunction()); 197 NodesInLastSCC.push_back(&N); 198 } 199 assert(EdgeCount >= EdgesOfLastSeenNodes); 200 } 201 202 int64_t MLInlineAdvisor::getLocalCalls(Function &F) { 203 return getCachedFPI(F).DirectCallsToDefinedFunctions; 204 } 205 206 // Update the internal state of the advisor, and force invalidate feature 207 // analysis. Currently, we maintain minimal (and very simple) global state - the 208 // number of functions and the number of static calls. We also keep track of the 209 // total IR size in this module, to stop misbehaving policies at a certain bloat 210 // factor (SizeIncreaseThreshold) 211 void MLInlineAdvisor::onSuccessfulInlining(const MLInlineAdvice &Advice, 212 bool CalleeWasDeleted) { 213 assert(!ForceStop); 214 Function *Caller = Advice.getCaller(); 215 Function *Callee = Advice.getCallee(); 216 // The caller features aren't valid anymore. 217 { 218 PreservedAnalyses PA = PreservedAnalyses::none(); 219 FAM.invalidate(*Caller, PA); 220 } 221 Advice.updateCachedCallerFPI(FAM.getResult<LoopAnalysis>(*Caller)); 222 int64_t IRSizeAfter = 223 getIRSize(*Caller) + (CalleeWasDeleted ? 0 : Advice.CalleeIRSize); 224 CurrentIRSize += IRSizeAfter - (Advice.CallerIRSize + Advice.CalleeIRSize); 225 if (CurrentIRSize > SizeIncreaseThreshold * InitialIRSize) 226 ForceStop = true; 227 228 // We can delta-update module-wide features. We know the inlining only changed 229 // the caller, and maybe the callee (by deleting the latter). 230 // Nodes are simple to update. 231 // For edges, we 'forget' the edges that the caller and callee used to have 232 // before inlining, and add back what they currently have together. 233 int64_t NewCallerAndCalleeEdges = 234 getCachedFPI(*Caller).DirectCallsToDefinedFunctions; 235 236 if (CalleeWasDeleted) 237 --NodeCount; 238 else 239 NewCallerAndCalleeEdges += 240 getCachedFPI(*Callee).DirectCallsToDefinedFunctions; 241 EdgeCount += (NewCallerAndCalleeEdges - Advice.CallerAndCalleeEdges); 242 assert(CurrentIRSize >= 0 && EdgeCount >= 0 && NodeCount >= 0); 243 } 244 245 int64_t MLInlineAdvisor::getModuleIRSize() const { 246 int64_t Ret = 0; 247 for (auto &F : M) 248 if (!F.isDeclaration()) 249 Ret += getIRSize(F); 250 return Ret; 251 } 252 253 FunctionPropertiesInfo &MLInlineAdvisor::getCachedFPI(Function &F) const { 254 auto InsertPair = 255 FPICache.insert(std::make_pair(&F, FunctionPropertiesInfo())); 256 if (!InsertPair.second) 257 return InsertPair.first->second; 258 InsertPair.first->second = FAM.getResult<FunctionPropertiesAnalysis>(F); 259 return InsertPair.first->second; 260 } 261 262 std::unique_ptr<InlineAdvice> MLInlineAdvisor::getAdviceImpl(CallBase &CB) { 263 auto &Caller = *CB.getCaller(); 264 auto &Callee = *CB.getCalledFunction(); 265 266 auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & { 267 return FAM.getResult<AssumptionAnalysis>(F); 268 }; 269 auto &TIR = FAM.getResult<TargetIRAnalysis>(Callee); 270 auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(Caller); 271 272 auto MandatoryKind = InlineAdvisor::getMandatoryKind(CB, FAM, ORE); 273 // If this is a "never inline" case, there won't be any changes to internal 274 // state we need to track, so we can just return the base InlineAdvice, which 275 // will do nothing interesting. 276 // Same thing if this is a recursive case. 277 if (MandatoryKind == InlineAdvisor::MandatoryInliningKind::Never || 278 &Caller == &Callee) 279 return getMandatoryAdvice(CB, false); 280 281 bool Mandatory = 282 MandatoryKind == InlineAdvisor::MandatoryInliningKind::Always; 283 284 // If we need to stop, we won't want to track anymore any state changes, so 285 // we just return the base InlineAdvice, which acts as a noop. 286 if (ForceStop) { 287 ORE.emit([&] { 288 return OptimizationRemarkMissed(DEBUG_TYPE, "ForceStop", &CB) 289 << "Won't attempt inlining because module size grew too much."; 290 }); 291 return std::make_unique<InlineAdvice>(this, CB, ORE, Mandatory); 292 } 293 294 int CostEstimate = 0; 295 if (!Mandatory) { 296 auto IsCallSiteInlinable = 297 llvm::getInliningCostEstimate(CB, TIR, GetAssumptionCache); 298 if (!IsCallSiteInlinable) { 299 // We can't inline this for correctness reasons, so return the base 300 // InlineAdvice, as we don't care about tracking any state changes (which 301 // won't happen). 302 return std::make_unique<InlineAdvice>(this, CB, ORE, false); 303 } 304 CostEstimate = *IsCallSiteInlinable; 305 } 306 307 const auto CostFeatures = 308 llvm::getInliningCostFeatures(CB, TIR, GetAssumptionCache); 309 if (!CostFeatures) { 310 return std::make_unique<InlineAdvice>(this, CB, ORE, false); 311 } 312 313 if (Mandatory) 314 return getMandatoryAdvice(CB, true); 315 316 auto NrCtantParams = 0; 317 for (auto I = CB.arg_begin(), E = CB.arg_end(); I != E; ++I) { 318 NrCtantParams += (isa<Constant>(*I)); 319 } 320 321 auto &CallerBefore = getCachedFPI(Caller); 322 auto &CalleeBefore = getCachedFPI(Callee); 323 324 *ModelRunner->getTensor<int64_t>(FeatureIndex::CalleeBasicBlockCount) = 325 CalleeBefore.BasicBlockCount; 326 *ModelRunner->getTensor<int64_t>(FeatureIndex::CallSiteHeight) = 327 getInitialFunctionLevel(Caller); 328 *ModelRunner->getTensor<int64_t>(FeatureIndex::NodeCount) = NodeCount; 329 *ModelRunner->getTensor<int64_t>(FeatureIndex::NrCtantParams) = NrCtantParams; 330 *ModelRunner->getTensor<int64_t>(FeatureIndex::EdgeCount) = EdgeCount; 331 *ModelRunner->getTensor<int64_t>(FeatureIndex::CallerUsers) = 332 CallerBefore.Uses; 333 *ModelRunner->getTensor<int64_t>( 334 FeatureIndex::CallerConditionallyExecutedBlocks) = 335 CallerBefore.BlocksReachedFromConditionalInstruction; 336 *ModelRunner->getTensor<int64_t>(FeatureIndex::CallerBasicBlockCount) = 337 CallerBefore.BasicBlockCount; 338 *ModelRunner->getTensor<int64_t>( 339 FeatureIndex::CalleeConditionallyExecutedBlocks) = 340 CalleeBefore.BlocksReachedFromConditionalInstruction; 341 *ModelRunner->getTensor<int64_t>(FeatureIndex::CalleeUsers) = 342 CalleeBefore.Uses; 343 *ModelRunner->getTensor<int64_t>(FeatureIndex::CostEstimate) = CostEstimate; 344 345 // Add the cost features 346 for (size_t I = 0; 347 I < static_cast<size_t>(InlineCostFeatureIndex::NumberOfFeatures); ++I) { 348 *ModelRunner->getTensor<int64_t>(inlineCostFeatureToMlFeature( 349 static_cast<InlineCostFeatureIndex>(I))) = CostFeatures->at(I); 350 } 351 352 return getAdviceFromModel(CB, ORE); 353 } 354 355 std::unique_ptr<MLInlineAdvice> 356 MLInlineAdvisor::getAdviceFromModel(CallBase &CB, 357 OptimizationRemarkEmitter &ORE) { 358 return std::make_unique<MLInlineAdvice>( 359 this, CB, ORE, static_cast<bool>(ModelRunner->evaluate<int64_t>())); 360 } 361 362 std::unique_ptr<InlineAdvice> MLInlineAdvisor::getMandatoryAdvice(CallBase &CB, 363 bool Advice) { 364 // Make sure we track inlinings in all cases - mandatory or not. 365 if (Advice && !ForceStop) 366 return getMandatoryAdviceImpl(CB); 367 368 // If this is a "never inline" case, there won't be any changes to internal 369 // state we need to track, so we can just return the base InlineAdvice, which 370 // will do nothing interesting. 371 // Same if we are forced to stop - we don't track anymore. 372 return std::make_unique<InlineAdvice>(this, CB, getCallerORE(CB), Advice); 373 } 374 375 std::unique_ptr<MLInlineAdvice> 376 MLInlineAdvisor::getMandatoryAdviceImpl(CallBase &CB) { 377 return std::make_unique<MLInlineAdvice>(this, CB, getCallerORE(CB), true); 378 } 379 380 void MLInlineAdvisor::print(raw_ostream &OS) const { 381 OS << "[MLInlineAdvisor] Nodes: " << NodeCount << " Edges: " << EdgeCount 382 << "\n"; 383 OS << "[MLInlineAdvisor] FPI:\n"; 384 for (auto I : FPICache) { 385 OS << I.getFirst()->getName() << ":\n"; 386 I.getSecond().print(OS); 387 OS << "\n"; 388 } 389 OS << "\n"; 390 } 391 392 MLInlineAdvice::MLInlineAdvice(MLInlineAdvisor *Advisor, CallBase &CB, 393 OptimizationRemarkEmitter &ORE, 394 bool Recommendation) 395 : InlineAdvice(Advisor, CB, ORE, Recommendation), 396 CallerIRSize(Advisor->isForcedToStop() ? 0 : Advisor->getIRSize(*Caller)), 397 CalleeIRSize(Advisor->isForcedToStop() ? 0 : Advisor->getIRSize(*Callee)), 398 CallerAndCalleeEdges(Advisor->isForcedToStop() 399 ? 0 400 : (Advisor->getLocalCalls(*Caller) + 401 Advisor->getLocalCalls(*Callee))), 402 PreInlineCallerFPI(Advisor->getCachedFPI(*Caller)) { 403 if (Recommendation) 404 FPU.emplace(Advisor->getCachedFPI(*getCaller()), CB); 405 } 406 407 void MLInlineAdvice::reportContextForRemark( 408 DiagnosticInfoOptimizationBase &OR) { 409 using namespace ore; 410 OR << NV("Callee", Callee->getName()); 411 for (size_t I = 0; I < NumberOfFeatures; ++I) 412 OR << NV(FeatureMap[I].name(), 413 *getAdvisor()->getModelRunner().getTensor<int64_t>(I)); 414 OR << NV("ShouldInline", isInliningRecommended()); 415 } 416 417 void MLInlineAdvice::updateCachedCallerFPI(const LoopInfo &LI) const { 418 FPU->finish(LI); 419 } 420 421 void MLInlineAdvice::recordInliningImpl() { 422 ORE.emit([&]() { 423 OptimizationRemark R(DEBUG_TYPE, "InliningSuccess", DLoc, Block); 424 reportContextForRemark(R); 425 return R; 426 }); 427 getAdvisor()->onSuccessfulInlining(*this, /*CalleeWasDeleted*/ false); 428 } 429 430 void MLInlineAdvice::recordInliningWithCalleeDeletedImpl() { 431 ORE.emit([&]() { 432 OptimizationRemark R(DEBUG_TYPE, "InliningSuccessWithCalleeDeleted", DLoc, 433 Block); 434 reportContextForRemark(R); 435 return R; 436 }); 437 getAdvisor()->onSuccessfulInlining(*this, /*CalleeWasDeleted*/ true); 438 } 439 440 void MLInlineAdvice::recordUnsuccessfulInliningImpl( 441 const InlineResult &Result) { 442 getAdvisor()->getCachedFPI(*Caller) = PreInlineCallerFPI; 443 ORE.emit([&]() { 444 OptimizationRemarkMissed R(DEBUG_TYPE, "InliningAttemptedAndUnsuccessful", 445 DLoc, Block); 446 reportContextForRemark(R); 447 return R; 448 }); 449 } 450 void MLInlineAdvice::recordUnattemptedInliningImpl() { 451 assert(!FPU); 452 ORE.emit([&]() { 453 OptimizationRemarkMissed R(DEBUG_TYPE, "IniningNotAttempted", DLoc, Block); 454 reportContextForRemark(R); 455 return R; 456 }); 457 } 458