//===- MLInlineAdvisor.cpp - machine learned InlineAdvisor ----------------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// // // This file implements the interface between the inliner and a learned model. // It delegates model evaluation to either the AOT compiled model (the // 'release' mode) or a runtime-loaded model (the 'development' case). // //===----------------------------------------------------------------------===// #include "llvm/Analysis/MLInlineAdvisor.h" #include "llvm/ADT/SCCIterator.h" #include "llvm/Analysis/AssumptionCache.h" #include "llvm/Analysis/CallGraph.h" #include "llvm/Analysis/FunctionPropertiesAnalysis.h" #include "llvm/Analysis/InlineCost.h" #include "llvm/Analysis/InlineModelFeatureMaps.h" #include "llvm/Analysis/InteractiveModelRunner.h" #include "llvm/Analysis/LazyCallGraph.h" #include "llvm/Analysis/LoopInfo.h" #include "llvm/Analysis/MLModelRunner.h" #include "llvm/Analysis/OptimizationRemarkEmitter.h" #include "llvm/Analysis/ReleaseModeModelRunner.h" #include "llvm/Analysis/TargetTransformInfo.h" #include "llvm/IR/Dominators.h" #include "llvm/IR/InstIterator.h" #include "llvm/IR/PassManager.h" #include "llvm/Support/CommandLine.h" using namespace llvm; static cl::opt InteractiveChannelBaseName( "inliner-interactive-channel-base", cl::Hidden, cl::desc( "Base file path for the interactive mode. The incoming filename should " "have the name .in, while the " "outgoing name should be .out")); static const std::string InclDefaultMsg = (Twine("In interactive mode, also send the default policy decision: ") + DefaultDecisionName + ".") .str(); static cl::opt InteractiveIncludeDefault("inliner-interactive-include-default", cl::Hidden, cl::desc(InclDefaultMsg)); #if defined(LLVM_HAVE_TF_AOT_INLINERSIZEMODEL) // codegen-ed file #include "InlinerSizeModel.h" // NOLINT using CompiledModelType = llvm::InlinerSizeModel; #else using CompiledModelType = NoopSavedModelImpl; #endif std::unique_ptr llvm::getReleaseModeAdvisor(Module &M, ModuleAnalysisManager &MAM, std::function GetDefaultAdvice) { if (!llvm::isEmbeddedModelEvaluatorValid() && InteractiveChannelBaseName.empty()) return nullptr; std::unique_ptr AOTRunner; if (InteractiveChannelBaseName.empty()) AOTRunner = std::make_unique>( M.getContext(), FeatureMap, DecisionName); else { auto Features = FeatureMap; if (InteractiveIncludeDefault) Features.push_back(DefaultDecisionSpec); AOTRunner = std::make_unique( M.getContext(), Features, InlineDecisionSpec, InteractiveChannelBaseName + ".out", InteractiveChannelBaseName + ".in"); } return std::make_unique(M, MAM, std::move(AOTRunner), GetDefaultAdvice); } #define DEBUG_TYPE "inline-ml" static cl::opt SizeIncreaseThreshold( "ml-advisor-size-increase-threshold", cl::Hidden, cl::desc("Maximum factor by which expected native size may increase before " "blocking any further inlining."), cl::init(2.0)); static cl::opt KeepFPICache( "ml-advisor-keep-fpi-cache", cl::Hidden, cl::desc( "For test - keep the ML Inline advisor's FunctionPropertiesInfo cache"), cl::init(false)); // clang-format off const std::vector llvm::FeatureMap{ #define POPULATE_NAMES(DTYPE, SHAPE, NAME, __) TensorSpec::createSpec(#NAME, SHAPE), // InlineCost features - these must come first INLINE_COST_FEATURE_ITERATOR(POPULATE_NAMES) // Non-cost features INLINE_FEATURE_ITERATOR(POPULATE_NAMES) #undef POPULATE_NAMES }; // clang-format on const char *const llvm::DecisionName = "inlining_decision"; const TensorSpec llvm::InlineDecisionSpec = TensorSpec::createSpec(DecisionName, {1}); const char *const llvm::DefaultDecisionName = "inlining_default"; const TensorSpec llvm::DefaultDecisionSpec = TensorSpec::createSpec(DefaultDecisionName, {1}); const char *const llvm::RewardName = "delta_size"; CallBase *getInlinableCS(Instruction &I) { if (auto *CS = dyn_cast(&I)) if (Function *Callee = CS->getCalledFunction()) { if (!Callee->isDeclaration()) { return CS; } } return nullptr; } MLInlineAdvisor::MLInlineAdvisor( Module &M, ModuleAnalysisManager &MAM, std::unique_ptr Runner, std::function GetDefaultAdvice) : InlineAdvisor( M, MAM.getResult(M).getManager()), ModelRunner(std::move(Runner)), GetDefaultAdvice(GetDefaultAdvice), CG(MAM.getResult(M)), InitialIRSize(getModuleIRSize()), CurrentIRSize(InitialIRSize) { assert(ModelRunner); ModelRunner->switchContext(""); // Extract the 'call site height' feature - the position of a call site // relative to the farthest statically reachable SCC node. We don't mutate // this value while inlining happens. Empirically, this feature proved // critical in behavioral cloning - i.e. training a model to mimic the manual // heuristic's decisions - and, thus, equally important for training for // improvement. CallGraph CGraph(M); for (auto I = scc_begin(&CGraph); !I.isAtEnd(); ++I) { const std::vector &CGNodes = *I; unsigned Level = 0; for (auto *CGNode : CGNodes) { Function *F = CGNode->getFunction(); if (!F || F->isDeclaration()) continue; for (auto &I : instructions(F)) { if (auto *CS = getInlinableCS(I)) { auto *Called = CS->getCalledFunction(); auto Pos = FunctionLevels.find(&CG.get(*Called)); // In bottom up traversal, an inlinable callee is either in the // same SCC, or to a function in a visited SCC. So not finding its // level means we haven't visited it yet, meaning it's in this SCC. if (Pos == FunctionLevels.end()) continue; Level = std::max(Level, Pos->second + 1); } } } for (auto *CGNode : CGNodes) { Function *F = CGNode->getFunction(); if (F && !F->isDeclaration()) FunctionLevels[&CG.get(*F)] = Level; } } for (auto KVP : FunctionLevels) { AllNodes.insert(KVP.first); EdgeCount += getLocalCalls(KVP.first->getFunction()); } NodeCount = AllNodes.size(); } unsigned MLInlineAdvisor::getInitialFunctionLevel(const Function &F) const { return CG.lookup(F) ? FunctionLevels.at(CG.lookup(F)) : 0; } void MLInlineAdvisor::onPassEntry(LazyCallGraph::SCC *LastSCC) { if (!LastSCC || ForceStop) return; FPICache.clear(); // Function passes executed between InlinerPass runs may have changed the // module-wide features. // The cgscc pass manager rules are such that: // - if a pass leads to merging SCCs, then the pipeline is restarted on the // merged SCC // - if a pass leads to splitting the SCC, then we continue with one of the // splits // This means that the NodesInLastSCC is a superset (not strict) of the nodes // that subsequent passes would have processed // - in addition, if new Nodes were created by a pass (e.g. CoroSplit), // they'd be adjacent to Nodes in the last SCC. So we just need to check the // boundary of Nodes in NodesInLastSCC for Nodes we haven't seen. We don't // care about the nature of the Edge (call or ref). NodeCount -= static_cast(NodesInLastSCC.size()); while (!NodesInLastSCC.empty()) { const auto *N = *NodesInLastSCC.begin(); NodesInLastSCC.erase(N); // The Function wrapped by N could have been deleted since we last saw it. if (N->isDead()) { assert(!N->getFunction().isDeclaration()); continue; } ++NodeCount; EdgeCount += getLocalCalls(N->getFunction()); for (const auto &E : *(*N)) { const auto *AdjNode = &E.getNode(); assert(!AdjNode->isDead() && !AdjNode->getFunction().isDeclaration()); auto I = AllNodes.insert(AdjNode); if (I.second) NodesInLastSCC.insert(AdjNode); } } EdgeCount -= EdgesOfLastSeenNodes; EdgesOfLastSeenNodes = 0; // (Re)use NodesInLastSCC to remember the nodes in the SCC right now, // in case the SCC is split before onPassExit and some nodes are split out assert(NodesInLastSCC.empty()); for (const auto &N : *LastSCC) NodesInLastSCC.insert(&N); } void MLInlineAdvisor::onPassExit(LazyCallGraph::SCC *LastSCC) { // No need to keep this around - function passes will invalidate it. if (!KeepFPICache) FPICache.clear(); if (!LastSCC || ForceStop) return; // Keep track of the nodes and edges we last saw. Then, in onPassEntry, // we update the node count and edge count from the subset of these nodes that // survived. EdgesOfLastSeenNodes = 0; // Check on nodes that were in SCC onPassEntry for (auto I = NodesInLastSCC.begin(); I != NodesInLastSCC.end();) { if ((*I)->isDead()) NodesInLastSCC.erase(*I++); else EdgesOfLastSeenNodes += getLocalCalls((*I++)->getFunction()); } // Check on nodes that may have got added to SCC for (const auto &N : *LastSCC) { assert(!N.isDead()); auto I = NodesInLastSCC.insert(&N); if (I.second) EdgesOfLastSeenNodes += getLocalCalls(N.getFunction()); } assert(NodeCount >= NodesInLastSCC.size()); assert(EdgeCount >= EdgesOfLastSeenNodes); } int64_t MLInlineAdvisor::getLocalCalls(Function &F) { return getCachedFPI(F).DirectCallsToDefinedFunctions; } // Update the internal state of the advisor, and force invalidate feature // analysis. Currently, we maintain minimal (and very simple) global state - the // number of functions and the number of static calls. We also keep track of the // total IR size in this module, to stop misbehaving policies at a certain bloat // factor (SizeIncreaseThreshold) void MLInlineAdvisor::onSuccessfulInlining(const MLInlineAdvice &Advice, bool CalleeWasDeleted) { assert(!ForceStop); Function *Caller = Advice.getCaller(); Function *Callee = Advice.getCallee(); // The caller features aren't valid anymore. { PreservedAnalyses PA = PreservedAnalyses::all(); PA.abandon(); PA.abandon(); PA.abandon(); FAM.invalidate(*Caller, PA); } Advice.updateCachedCallerFPI(FAM); int64_t IRSizeAfter = getIRSize(*Caller) + (CalleeWasDeleted ? 0 : Advice.CalleeIRSize); CurrentIRSize += IRSizeAfter - (Advice.CallerIRSize + Advice.CalleeIRSize); if (CurrentIRSize > SizeIncreaseThreshold * InitialIRSize) ForceStop = true; // We can delta-update module-wide features. We know the inlining only changed // the caller, and maybe the callee (by deleting the latter). // Nodes are simple to update. // For edges, we 'forget' the edges that the caller and callee used to have // before inlining, and add back what they currently have together. int64_t NewCallerAndCalleeEdges = getCachedFPI(*Caller).DirectCallsToDefinedFunctions; if (CalleeWasDeleted) --NodeCount; else NewCallerAndCalleeEdges += getCachedFPI(*Callee).DirectCallsToDefinedFunctions; EdgeCount += (NewCallerAndCalleeEdges - Advice.CallerAndCalleeEdges); assert(CurrentIRSize >= 0 && EdgeCount >= 0 && NodeCount >= 0); } int64_t MLInlineAdvisor::getModuleIRSize() const { int64_t Ret = 0; for (auto &F : M) if (!F.isDeclaration()) Ret += getIRSize(F); return Ret; } FunctionPropertiesInfo &MLInlineAdvisor::getCachedFPI(Function &F) const { auto InsertPair = FPICache.insert(std::make_pair(&F, FunctionPropertiesInfo())); if (!InsertPair.second) return InsertPair.first->second; InsertPair.first->second = FAM.getResult(F); return InsertPair.first->second; } std::unique_ptr MLInlineAdvisor::getAdviceImpl(CallBase &CB) { if (auto Skip = getSkipAdviceIfUnreachableCallsite(CB)) return Skip; auto &Caller = *CB.getCaller(); auto &Callee = *CB.getCalledFunction(); auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & { return FAM.getResult(F); }; auto &TIR = FAM.getResult(Callee); auto &ORE = FAM.getResult(Caller); auto MandatoryKind = InlineAdvisor::getMandatoryKind(CB, FAM, ORE); // If this is a "never inline" case, there won't be any changes to internal // state we need to track, so we can just return the base InlineAdvice, which // will do nothing interesting. // Same thing if this is a recursive case. if (MandatoryKind == InlineAdvisor::MandatoryInliningKind::Never || &Caller == &Callee) return getMandatoryAdvice(CB, false); bool Mandatory = MandatoryKind == InlineAdvisor::MandatoryInliningKind::Always; // If we need to stop, we won't want to track anymore any state changes, so // we just return the base InlineAdvice, which acts as a noop. if (ForceStop) { ORE.emit([&] { return OptimizationRemarkMissed(DEBUG_TYPE, "ForceStop", &CB) << "Won't attempt inlining because module size grew too much."; }); return std::make_unique(this, CB, ORE, Mandatory); } int CostEstimate = 0; if (!Mandatory) { auto IsCallSiteInlinable = llvm::getInliningCostEstimate(CB, TIR, GetAssumptionCache); if (!IsCallSiteInlinable) { // We can't inline this for correctness reasons, so return the base // InlineAdvice, as we don't care about tracking any state changes (which // won't happen). return std::make_unique(this, CB, ORE, false); } CostEstimate = *IsCallSiteInlinable; } const auto CostFeatures = llvm::getInliningCostFeatures(CB, TIR, GetAssumptionCache); if (!CostFeatures) { return std::make_unique(this, CB, ORE, false); } if (Mandatory) return getMandatoryAdvice(CB, true); auto NrCtantParams = 0; for (auto I = CB.arg_begin(), E = CB.arg_end(); I != E; ++I) { NrCtantParams += (isa(*I)); } auto &CallerBefore = getCachedFPI(Caller); auto &CalleeBefore = getCachedFPI(Callee); *ModelRunner->getTensor(FeatureIndex::callee_basic_block_count) = CalleeBefore.BasicBlockCount; *ModelRunner->getTensor(FeatureIndex::callsite_height) = getInitialFunctionLevel(Caller); *ModelRunner->getTensor(FeatureIndex::node_count) = NodeCount; *ModelRunner->getTensor(FeatureIndex::nr_ctant_params) = NrCtantParams; *ModelRunner->getTensor(FeatureIndex::edge_count) = EdgeCount; *ModelRunner->getTensor(FeatureIndex::caller_users) = CallerBefore.Uses; *ModelRunner->getTensor( FeatureIndex::caller_conditionally_executed_blocks) = CallerBefore.BlocksReachedFromConditionalInstruction; *ModelRunner->getTensor(FeatureIndex::caller_basic_block_count) = CallerBefore.BasicBlockCount; *ModelRunner->getTensor( FeatureIndex::callee_conditionally_executed_blocks) = CalleeBefore.BlocksReachedFromConditionalInstruction; *ModelRunner->getTensor(FeatureIndex::callee_users) = CalleeBefore.Uses; *ModelRunner->getTensor(FeatureIndex::cost_estimate) = CostEstimate; // Add the cost features for (size_t I = 0; I < static_cast(InlineCostFeatureIndex::NumberOfFeatures); ++I) { *ModelRunner->getTensor(inlineCostFeatureToMlFeature( static_cast(I))) = CostFeatures->at(I); } // This one would have been set up to be right at the end. if (!InteractiveChannelBaseName.empty() && InteractiveIncludeDefault) *ModelRunner->getTensor(InlineCostFeatureIndex::NumberOfFeatures) = GetDefaultAdvice(CB); return getAdviceFromModel(CB, ORE); } std::unique_ptr MLInlineAdvisor::getAdviceFromModel(CallBase &CB, OptimizationRemarkEmitter &ORE) { return std::make_unique( this, CB, ORE, static_cast(ModelRunner->evaluate())); } std::unique_ptr MLInlineAdvisor::getSkipAdviceIfUnreachableCallsite(CallBase &CB) { if (!FAM.getResult(*CB.getCaller()) .isReachableFromEntry(CB.getParent())) return std::make_unique(this, CB, getCallerORE(CB), false); return nullptr; } std::unique_ptr MLInlineAdvisor::getMandatoryAdvice(CallBase &CB, bool Advice) { // Make sure we track inlinings in all cases - mandatory or not. if (auto Skip = getSkipAdviceIfUnreachableCallsite(CB)) return Skip; if (Advice && !ForceStop) return getMandatoryAdviceImpl(CB); // If this is a "never inline" case, there won't be any changes to internal // state we need to track, so we can just return the base InlineAdvice, which // will do nothing interesting. // Same if we are forced to stop - we don't track anymore. return std::make_unique(this, CB, getCallerORE(CB), Advice); } std::unique_ptr MLInlineAdvisor::getMandatoryAdviceImpl(CallBase &CB) { return std::make_unique(this, CB, getCallerORE(CB), true); } void MLInlineAdvisor::print(raw_ostream &OS) const { OS << "[MLInlineAdvisor] Nodes: " << NodeCount << " Edges: " << EdgeCount << " EdgesOfLastSeenNodes: " << EdgesOfLastSeenNodes << "\n"; OS << "[MLInlineAdvisor] FPI:\n"; for (auto I : FPICache) { OS << I.first->getName() << ":\n"; I.second.print(OS); OS << "\n"; } OS << "\n"; } MLInlineAdvice::MLInlineAdvice(MLInlineAdvisor *Advisor, CallBase &CB, OptimizationRemarkEmitter &ORE, bool Recommendation) : InlineAdvice(Advisor, CB, ORE, Recommendation), CallerIRSize(Advisor->isForcedToStop() ? 0 : Advisor->getIRSize(*Caller)), CalleeIRSize(Advisor->isForcedToStop() ? 0 : Advisor->getIRSize(*Callee)), CallerAndCalleeEdges(Advisor->isForcedToStop() ? 0 : (Advisor->getLocalCalls(*Caller) + Advisor->getLocalCalls(*Callee))), PreInlineCallerFPI(Advisor->getCachedFPI(*Caller)) { if (Recommendation) FPU.emplace(Advisor->getCachedFPI(*getCaller()), CB); } void MLInlineAdvice::reportContextForRemark( DiagnosticInfoOptimizationBase &OR) { using namespace ore; OR << NV("Callee", Callee->getName()); for (size_t I = 0; I < NumberOfFeatures; ++I) OR << NV(FeatureMap[I].name(), *getAdvisor()->getModelRunner().getTensor(I)); OR << NV("ShouldInline", isInliningRecommended()); } void MLInlineAdvice::updateCachedCallerFPI(FunctionAnalysisManager &FAM) const { FPU->finish(FAM); } void MLInlineAdvice::recordInliningImpl() { ORE.emit([&]() { OptimizationRemark R(DEBUG_TYPE, "InliningSuccess", DLoc, Block); reportContextForRemark(R); return R; }); getAdvisor()->onSuccessfulInlining(*this, /*CalleeWasDeleted*/ false); } void MLInlineAdvice::recordInliningWithCalleeDeletedImpl() { ORE.emit([&]() { OptimizationRemark R(DEBUG_TYPE, "InliningSuccessWithCalleeDeleted", DLoc, Block); reportContextForRemark(R); return R; }); getAdvisor()->onSuccessfulInlining(*this, /*CalleeWasDeleted*/ true); } void MLInlineAdvice::recordUnsuccessfulInliningImpl( const InlineResult &Result) { getAdvisor()->getCachedFPI(*Caller) = PreInlineCallerFPI; ORE.emit([&]() { OptimizationRemarkMissed R(DEBUG_TYPE, "InliningAttemptedAndUnsuccessful", DLoc, Block); reportContextForRemark(R); return R; }); } void MLInlineAdvice::recordUnattemptedInliningImpl() { assert(!FPU); ORE.emit([&]() { OptimizationRemarkMissed R(DEBUG_TYPE, "IniningNotAttempted", DLoc, Block); reportContextForRemark(R); return R; }); }