1 //===- InlineOrder.cpp - Inlining order abstraction -*- C++ ---*-----------===// 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 #include "llvm/Analysis/InlineOrder.h" 10 #include "llvm/Analysis/AssumptionCache.h" 11 #include "llvm/Analysis/BlockFrequencyInfo.h" 12 #include "llvm/Analysis/GlobalsModRef.h" 13 #include "llvm/Analysis/InlineAdvisor.h" 14 #include "llvm/Analysis/InlineCost.h" 15 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 16 #include "llvm/Analysis/ProfileSummaryInfo.h" 17 #include "llvm/Analysis/TargetLibraryInfo.h" 18 #include "llvm/Analysis/TargetTransformInfo.h" 19 #include "llvm/Support/CommandLine.h" 20 21 using namespace llvm; 22 23 #define DEBUG_TYPE "inline-order" 24 25 enum class InlinePriorityMode : int { Size, Cost, CostBenefit, ML }; 26 27 static cl::opt<InlinePriorityMode> UseInlinePriority( 28 "inline-priority-mode", cl::init(InlinePriorityMode::Size), cl::Hidden, 29 cl::desc("Choose the priority mode to use in module inline"), 30 cl::values(clEnumValN(InlinePriorityMode::Size, "size", 31 "Use callee size priority."), 32 clEnumValN(InlinePriorityMode::Cost, "cost", 33 "Use inline cost priority."), 34 clEnumValN(InlinePriorityMode::CostBenefit, "cost-benefit", 35 "Use cost-benefit ratio."), 36 clEnumValN(InlinePriorityMode::ML, "ml", "Use ML."))); 37 38 static cl::opt<int> ModuleInlinerTopPriorityThreshold( 39 "module-inliner-top-priority-threshold", cl::Hidden, cl::init(0), 40 cl::desc("The cost threshold for call sites that get inlined without the " 41 "cost-benefit analysis")); 42 43 namespace { 44 45 llvm::InlineCost getInlineCostWrapper(CallBase &CB, 46 FunctionAnalysisManager &FAM, 47 const InlineParams &Params) { 48 Function &Caller = *CB.getCaller(); 49 ProfileSummaryInfo *PSI = 50 FAM.getResult<ModuleAnalysisManagerFunctionProxy>(Caller) 51 .getCachedResult<ProfileSummaryAnalysis>( 52 *CB.getParent()->getParent()->getParent()); 53 54 auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(Caller); 55 auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & { 56 return FAM.getResult<AssumptionAnalysis>(F); 57 }; 58 auto GetBFI = [&](Function &F) -> BlockFrequencyInfo & { 59 return FAM.getResult<BlockFrequencyAnalysis>(F); 60 }; 61 auto GetTLI = [&](Function &F) -> const TargetLibraryInfo & { 62 return FAM.getResult<TargetLibraryAnalysis>(F); 63 }; 64 65 Function &Callee = *CB.getCalledFunction(); 66 auto &CalleeTTI = FAM.getResult<TargetIRAnalysis>(Callee); 67 bool RemarksEnabled = 68 Callee.getContext().getDiagHandlerPtr()->isMissedOptRemarkEnabled( 69 DEBUG_TYPE); 70 return getInlineCost(CB, Params, CalleeTTI, GetAssumptionCache, GetTLI, 71 GetBFI, PSI, RemarksEnabled ? &ORE : nullptr); 72 } 73 74 class SizePriority { 75 public: 76 SizePriority() = default; 77 SizePriority(const CallBase *CB, FunctionAnalysisManager &, 78 const InlineParams &) { 79 Function *Callee = CB->getCalledFunction(); 80 Size = Callee->getInstructionCount(); 81 } 82 83 static bool isMoreDesirable(const SizePriority &P1, const SizePriority &P2) { 84 return P1.Size < P2.Size; 85 } 86 87 private: 88 unsigned Size = UINT_MAX; 89 }; 90 91 class CostPriority { 92 public: 93 CostPriority() = default; 94 CostPriority(const CallBase *CB, FunctionAnalysisManager &FAM, 95 const InlineParams &Params) { 96 auto IC = getInlineCostWrapper(const_cast<CallBase &>(*CB), FAM, Params); 97 if (IC.isVariable()) 98 Cost = IC.getCost(); 99 else 100 Cost = IC.isNever() ? INT_MAX : INT_MIN; 101 } 102 103 static bool isMoreDesirable(const CostPriority &P1, const CostPriority &P2) { 104 return P1.Cost < P2.Cost; 105 } 106 107 private: 108 int Cost = INT_MAX; 109 }; 110 111 class CostBenefitPriority { 112 public: 113 CostBenefitPriority() = default; 114 CostBenefitPriority(const CallBase *CB, FunctionAnalysisManager &FAM, 115 const InlineParams &Params) { 116 auto IC = getInlineCostWrapper(const_cast<CallBase &>(*CB), FAM, Params); 117 Cost = IC.getCost(); 118 StaticBonusApplied = IC.getStaticBonusApplied(); 119 CostBenefit = IC.getCostBenefit(); 120 } 121 122 static bool isMoreDesirable(const CostBenefitPriority &P1, 123 const CostBenefitPriority &P2) { 124 // We prioritize call sites in the dictionary order of the following 125 // priorities: 126 // 127 // 1. Those call sites that are expected to reduce the caller size when 128 // inlined. Within them, we prioritize those call sites with bigger 129 // reduction. 130 // 131 // 2. Those call sites that have gone through the cost-benefit analysis. 132 // Currently, they are limited to hot call sites. Within them, we 133 // prioritize those call sites with higher benefit-to-cost ratios. 134 // 135 // 3. Remaining call sites are prioritized according to their costs. 136 137 // We add back StaticBonusApplied to determine whether we expect the caller 138 // to shrink (even if we don't delete the callee). 139 bool P1ReducesCallerSize = 140 P1.Cost + P1.StaticBonusApplied < ModuleInlinerTopPriorityThreshold; 141 bool P2ReducesCallerSize = 142 P2.Cost + P2.StaticBonusApplied < ModuleInlinerTopPriorityThreshold; 143 if (P1ReducesCallerSize || P2ReducesCallerSize) { 144 // If one reduces the caller size while the other doesn't, then return 145 // true iff P1 reduces the caller size. 146 if (P1ReducesCallerSize != P2ReducesCallerSize) 147 return P1ReducesCallerSize; 148 149 // If they both reduce the caller size, pick the one with the smaller 150 // cost. 151 return P1.Cost < P2.Cost; 152 } 153 154 bool P1HasCB = P1.CostBenefit.has_value(); 155 bool P2HasCB = P2.CostBenefit.has_value(); 156 if (P1HasCB || P2HasCB) { 157 // If one has undergone the cost-benefit analysis while the other hasn't, 158 // then return true iff P1 has. 159 if (P1HasCB != P2HasCB) 160 return P1HasCB; 161 162 // If they have undergone the cost-benefit analysis, then pick the one 163 // with a higher benefit-to-cost ratio. 164 APInt LHS = P1.CostBenefit->getBenefit() * P2.CostBenefit->getCost(); 165 APInt RHS = P2.CostBenefit->getBenefit() * P1.CostBenefit->getCost(); 166 return LHS.ugt(RHS); 167 } 168 169 // Remaining call sites are ordered according to their costs. 170 return P1.Cost < P2.Cost; 171 } 172 173 private: 174 int Cost = INT_MAX; 175 int StaticBonusApplied = 0; 176 std::optional<CostBenefitPair> CostBenefit; 177 }; 178 179 class MLPriority { 180 public: 181 MLPriority() = default; 182 MLPriority(const CallBase *CB, FunctionAnalysisManager &FAM, 183 const InlineParams &Params) { 184 auto IC = getInlineCostWrapper(const_cast<CallBase &>(*CB), FAM, Params); 185 if (IC.isVariable()) 186 Cost = IC.getCost(); 187 else 188 Cost = IC.isNever() ? INT_MAX : INT_MIN; 189 } 190 191 static bool isMoreDesirable(const MLPriority &P1, const MLPriority &P2) { 192 return P1.Cost < P2.Cost; 193 } 194 195 private: 196 int Cost = INT_MAX; 197 }; 198 199 template <typename PriorityT> 200 class PriorityInlineOrder : public InlineOrder<std::pair<CallBase *, int>> { 201 using T = std::pair<CallBase *, int>; 202 203 bool hasLowerPriority(const CallBase *L, const CallBase *R) const { 204 const auto I1 = Priorities.find(L); 205 const auto I2 = Priorities.find(R); 206 assert(I1 != Priorities.end() && I2 != Priorities.end()); 207 return PriorityT::isMoreDesirable(I2->second, I1->second); 208 } 209 210 bool updateAndCheckDecreased(const CallBase *CB) { 211 auto It = Priorities.find(CB); 212 const auto OldPriority = It->second; 213 It->second = PriorityT(CB, FAM, Params); 214 const auto NewPriority = It->second; 215 return PriorityT::isMoreDesirable(OldPriority, NewPriority); 216 } 217 218 // A call site could become less desirable for inlining because of the size 219 // growth from prior inlining into the callee. This method is used to lazily 220 // update the desirability of a call site if it's decreasing. It is only 221 // called on pop(), not every time the desirability changes. When the 222 // desirability of the front call site decreases, an updated one would be 223 // pushed right back into the heap. For simplicity, those cases where the 224 // desirability of a call site increases are ignored here. 225 void pop_heap_adjust() { 226 std::pop_heap(Heap.begin(), Heap.end(), isLess); 227 while (updateAndCheckDecreased(Heap.back())) { 228 std::push_heap(Heap.begin(), Heap.end(), isLess); 229 std::pop_heap(Heap.begin(), Heap.end(), isLess); 230 } 231 } 232 233 public: 234 PriorityInlineOrder(FunctionAnalysisManager &FAM, const InlineParams &Params) 235 : FAM(FAM), Params(Params) { 236 isLess = [&](const CallBase *L, const CallBase *R) { 237 return hasLowerPriority(L, R); 238 }; 239 } 240 241 size_t size() override { return Heap.size(); } 242 243 void push(const T &Elt) override { 244 CallBase *CB = Elt.first; 245 const int InlineHistoryID = Elt.second; 246 247 Heap.push_back(CB); 248 Priorities[CB] = PriorityT(CB, FAM, Params); 249 std::push_heap(Heap.begin(), Heap.end(), isLess); 250 InlineHistoryMap[CB] = InlineHistoryID; 251 } 252 253 T pop() override { 254 assert(size() > 0); 255 pop_heap_adjust(); 256 257 CallBase *CB = Heap.pop_back_val(); 258 T Result = std::make_pair(CB, InlineHistoryMap[CB]); 259 InlineHistoryMap.erase(CB); 260 return Result; 261 } 262 263 void erase_if(function_ref<bool(T)> Pred) override { 264 auto PredWrapper = [=](CallBase *CB) -> bool { 265 return Pred(std::make_pair(CB, InlineHistoryMap[CB])); 266 }; 267 llvm::erase_if(Heap, PredWrapper); 268 std::make_heap(Heap.begin(), Heap.end(), isLess); 269 } 270 271 private: 272 SmallVector<CallBase *, 16> Heap; 273 std::function<bool(const CallBase *L, const CallBase *R)> isLess; 274 DenseMap<CallBase *, int> InlineHistoryMap; 275 DenseMap<const CallBase *, PriorityT> Priorities; 276 FunctionAnalysisManager &FAM; 277 const InlineParams &Params; 278 }; 279 280 } // namespace 281 282 AnalysisKey llvm::PluginInlineOrderAnalysis::Key; 283 bool llvm::PluginInlineOrderAnalysis::HasBeenRegistered; 284 285 std::unique_ptr<InlineOrder<std::pair<CallBase *, int>>> 286 llvm::getDefaultInlineOrder(FunctionAnalysisManager &FAM, 287 const InlineParams &Params, 288 ModuleAnalysisManager &MAM, Module &M) { 289 switch (UseInlinePriority) { 290 case InlinePriorityMode::Size: 291 LLVM_DEBUG(dbgs() << " Current used priority: Size priority ---- \n"); 292 return std::make_unique<PriorityInlineOrder<SizePriority>>(FAM, Params); 293 294 case InlinePriorityMode::Cost: 295 LLVM_DEBUG(dbgs() << " Current used priority: Cost priority ---- \n"); 296 return std::make_unique<PriorityInlineOrder<CostPriority>>(FAM, Params); 297 298 case InlinePriorityMode::CostBenefit: 299 LLVM_DEBUG( 300 dbgs() << " Current used priority: cost-benefit priority ---- \n"); 301 return std::make_unique<PriorityInlineOrder<CostBenefitPriority>>(FAM, 302 Params); 303 case InlinePriorityMode::ML: 304 LLVM_DEBUG(dbgs() << " Current used priority: ML priority ---- \n"); 305 return std::make_unique<PriorityInlineOrder<MLPriority>>(FAM, Params); 306 } 307 return nullptr; 308 } 309 310 std::unique_ptr<InlineOrder<std::pair<CallBase *, int>>> 311 llvm::getInlineOrder(FunctionAnalysisManager &FAM, const InlineParams &Params, 312 ModuleAnalysisManager &MAM, Module &M) { 313 if (llvm::PluginInlineOrderAnalysis::isRegistered()) { 314 LLVM_DEBUG(dbgs() << " Current used priority: plugin ---- \n"); 315 return MAM.getResult<PluginInlineOrderAnalysis>(M).Factory(FAM, Params, MAM, 316 M); 317 } 318 return getDefaultInlineOrder(FAM, Params, MAM, M); 319 } 320