Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 14 | pmbaty | 1 | //===- InlineCost.h - Cost analysis for inliner -----------------*- 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 | // This file implements heuristics for inlining decisions. |
||
| 10 | // |
||
| 11 | //===----------------------------------------------------------------------===// |
||
| 12 | |||
| 13 | #ifndef LLVM_ANALYSIS_INLINECOST_H |
||
| 14 | #define LLVM_ANALYSIS_INLINECOST_H |
||
| 15 | |||
| 16 | #include "llvm/ADT/APInt.h" |
||
| 17 | #include "llvm/ADT/STLFunctionalExtras.h" |
||
| 18 | #include "llvm/Analysis/InlineModelFeatureMaps.h" |
||
| 19 | #include "llvm/IR/PassManager.h" |
||
| 20 | #include <cassert> |
||
| 21 | #include <climits> |
||
| 22 | #include <optional> |
||
| 23 | |||
| 24 | namespace llvm { |
||
| 25 | class AssumptionCache; |
||
| 26 | class OptimizationRemarkEmitter; |
||
| 27 | class BlockFrequencyInfo; |
||
| 28 | class CallBase; |
||
| 29 | class DataLayout; |
||
| 30 | class Function; |
||
| 31 | class ProfileSummaryInfo; |
||
| 32 | class TargetTransformInfo; |
||
| 33 | class TargetLibraryInfo; |
||
| 34 | |||
| 35 | namespace InlineConstants { |
||
| 36 | // Various thresholds used by inline cost analysis. |
||
| 37 | /// Use when optsize (-Os) is specified. |
||
| 38 | const int OptSizeThreshold = 50; |
||
| 39 | |||
| 40 | /// Use when minsize (-Oz) is specified. |
||
| 41 | const int OptMinSizeThreshold = 5; |
||
| 42 | |||
| 43 | /// Use when -O3 is specified. |
||
| 44 | const int OptAggressiveThreshold = 250; |
||
| 45 | |||
| 46 | // Various magic constants used to adjust heuristics. |
||
| 47 | int getInstrCost(); |
||
| 48 | const int IndirectCallThreshold = 100; |
||
| 49 | const int LoopPenalty = 25; |
||
| 50 | const int LastCallToStaticBonus = 15000; |
||
| 51 | const int ColdccPenalty = 2000; |
||
| 52 | /// Do not inline functions which allocate this many bytes on the stack |
||
| 53 | /// when the caller is recursive. |
||
| 54 | const unsigned TotalAllocaSizeRecursiveCaller = 1024; |
||
| 55 | /// Do not inline dynamic allocas that have been constant propagated to be |
||
| 56 | /// static allocas above this amount in bytes. |
||
| 57 | const uint64_t MaxSimplifiedDynamicAllocaToInline = 65536; |
||
| 58 | |||
| 59 | const char FunctionInlineCostMultiplierAttributeName[] = |
||
| 60 | "function-inline-cost-multiplier"; |
||
| 61 | |||
| 62 | const char MaxInlineStackSizeAttributeName[] = "inline-max-stacksize"; |
||
| 63 | } // namespace InlineConstants |
||
| 64 | |||
| 65 | // The cost-benefit pair computed by cost-benefit analysis. |
||
| 66 | class CostBenefitPair { |
||
| 67 | public: |
||
| 68 | CostBenefitPair(APInt Cost, APInt Benefit) : Cost(Cost), Benefit(Benefit) {} |
||
| 69 | |||
| 70 | const APInt &getCost() const { return Cost; } |
||
| 71 | |||
| 72 | const APInt &getBenefit() const { return Benefit; } |
||
| 73 | |||
| 74 | private: |
||
| 75 | APInt Cost; |
||
| 76 | APInt Benefit; |
||
| 77 | }; |
||
| 78 | |||
| 79 | /// Represents the cost of inlining a function. |
||
| 80 | /// |
||
| 81 | /// This supports special values for functions which should "always" or |
||
| 82 | /// "never" be inlined. Otherwise, the cost represents a unitless amount; |
||
| 83 | /// smaller values increase the likelihood of the function being inlined. |
||
| 84 | /// |
||
| 85 | /// Objects of this type also provide the adjusted threshold for inlining |
||
| 86 | /// based on the information available for a particular callsite. They can be |
||
| 87 | /// directly tested to determine if inlining should occur given the cost and |
||
| 88 | /// threshold for this cost metric. |
||
| 89 | class InlineCost { |
||
| 90 | enum SentinelValues { AlwaysInlineCost = INT_MIN, NeverInlineCost = INT_MAX }; |
||
| 91 | |||
| 92 | /// The estimated cost of inlining this callsite. |
||
| 93 | int Cost = 0; |
||
| 94 | |||
| 95 | /// The adjusted threshold against which this cost was computed. |
||
| 96 | int Threshold = 0; |
||
| 97 | |||
| 98 | /// The amount of StaticBonus that has been applied. |
||
| 99 | int StaticBonusApplied = 0; |
||
| 100 | |||
| 101 | /// Must be set for Always and Never instances. |
||
| 102 | const char *Reason = nullptr; |
||
| 103 | |||
| 104 | /// The cost-benefit pair computed by cost-benefit analysis. |
||
| 105 | std::optional<CostBenefitPair> CostBenefit; |
||
| 106 | |||
| 107 | // Trivial constructor, interesting logic in the factory functions below. |
||
| 108 | InlineCost(int Cost, int Threshold, int StaticBonusApplied, |
||
| 109 | const char *Reason = nullptr, |
||
| 110 | std::optional<CostBenefitPair> CostBenefit = std::nullopt) |
||
| 111 | : Cost(Cost), Threshold(Threshold), |
||
| 112 | StaticBonusApplied(StaticBonusApplied), Reason(Reason), |
||
| 113 | CostBenefit(CostBenefit) { |
||
| 114 | assert((isVariable() || Reason) && |
||
| 115 | "Reason must be provided for Never or Always"); |
||
| 116 | } |
||
| 117 | |||
| 118 | public: |
||
| 119 | static InlineCost get(int Cost, int Threshold, int StaticBonus = 0) { |
||
| 120 | assert(Cost > AlwaysInlineCost && "Cost crosses sentinel value"); |
||
| 121 | assert(Cost < NeverInlineCost && "Cost crosses sentinel value"); |
||
| 122 | return InlineCost(Cost, Threshold, StaticBonus); |
||
| 123 | } |
||
| 124 | static InlineCost |
||
| 125 | getAlways(const char *Reason, |
||
| 126 | std::optional<CostBenefitPair> CostBenefit = std::nullopt) { |
||
| 127 | return InlineCost(AlwaysInlineCost, 0, 0, Reason, CostBenefit); |
||
| 128 | } |
||
| 129 | static InlineCost |
||
| 130 | getNever(const char *Reason, |
||
| 131 | std::optional<CostBenefitPair> CostBenefit = std::nullopt) { |
||
| 132 | return InlineCost(NeverInlineCost, 0, 0, Reason, CostBenefit); |
||
| 133 | } |
||
| 134 | |||
| 135 | /// Test whether the inline cost is low enough for inlining. |
||
| 136 | explicit operator bool() const { return Cost < Threshold; } |
||
| 137 | |||
| 138 | bool isAlways() const { return Cost == AlwaysInlineCost; } |
||
| 139 | bool isNever() const { return Cost == NeverInlineCost; } |
||
| 140 | bool isVariable() const { return !isAlways() && !isNever(); } |
||
| 141 | |||
| 142 | /// Get the inline cost estimate. |
||
| 143 | /// It is an error to call this on an "always" or "never" InlineCost. |
||
| 144 | int getCost() const { |
||
| 145 | assert(isVariable() && "Invalid access of InlineCost"); |
||
| 146 | return Cost; |
||
| 147 | } |
||
| 148 | |||
| 149 | /// Get the threshold against which the cost was computed |
||
| 150 | int getThreshold() const { |
||
| 151 | assert(isVariable() && "Invalid access of InlineCost"); |
||
| 152 | return Threshold; |
||
| 153 | } |
||
| 154 | |||
| 155 | /// Get the amount of StaticBonus applied. |
||
| 156 | int getStaticBonusApplied() const { |
||
| 157 | assert(isVariable() && "Invalid access of InlineCost"); |
||
| 158 | return StaticBonusApplied; |
||
| 159 | } |
||
| 160 | |||
| 161 | /// Get the cost-benefit pair which was computed by cost-benefit analysis |
||
| 162 | std::optional<CostBenefitPair> getCostBenefit() const { return CostBenefit; } |
||
| 163 | |||
| 164 | /// Get the reason of Always or Never. |
||
| 165 | const char *getReason() const { |
||
| 166 | assert((Reason || isVariable()) && |
||
| 167 | "InlineCost reason must be set for Always or Never"); |
||
| 168 | return Reason; |
||
| 169 | } |
||
| 170 | |||
| 171 | /// Get the cost delta from the threshold for inlining. |
||
| 172 | /// Only valid if the cost is of the variable kind. Returns a negative |
||
| 173 | /// value if the cost is too high to inline. |
||
| 174 | int getCostDelta() const { return Threshold - getCost(); } |
||
| 175 | }; |
||
| 176 | |||
| 177 | /// InlineResult is basically true or false. For false results the message |
||
| 178 | /// describes a reason. |
||
| 179 | class InlineResult { |
||
| 180 | const char *Message = nullptr; |
||
| 181 | InlineResult(const char *Message = nullptr) : Message(Message) {} |
||
| 182 | |||
| 183 | public: |
||
| 184 | static InlineResult success() { return {}; } |
||
| 185 | static InlineResult failure(const char *Reason) { |
||
| 186 | return InlineResult(Reason); |
||
| 187 | } |
||
| 188 | bool isSuccess() const { return Message == nullptr; } |
||
| 189 | const char *getFailureReason() const { |
||
| 190 | assert(!isSuccess() && |
||
| 191 | "getFailureReason should only be called in failure cases"); |
||
| 192 | return Message; |
||
| 193 | } |
||
| 194 | }; |
||
| 195 | |||
| 196 | /// Thresholds to tune inline cost analysis. The inline cost analysis decides |
||
| 197 | /// the condition to apply a threshold and applies it. Otherwise, |
||
| 198 | /// DefaultThreshold is used. If a threshold is Optional, it is applied only |
||
| 199 | /// when it has a valid value. Typically, users of inline cost analysis |
||
| 200 | /// obtain an InlineParams object through one of the \c getInlineParams methods |
||
| 201 | /// and pass it to \c getInlineCost. Some specialized versions of inliner |
||
| 202 | /// (such as the pre-inliner) might have custom logic to compute \c InlineParams |
||
| 203 | /// object. |
||
| 204 | |||
| 205 | struct InlineParams { |
||
| 206 | /// The default threshold to start with for a callee. |
||
| 207 | int DefaultThreshold = -1; |
||
| 208 | |||
| 209 | /// Threshold to use for callees with inline hint. |
||
| 210 | std::optional<int> HintThreshold; |
||
| 211 | |||
| 212 | /// Threshold to use for cold callees. |
||
| 213 | std::optional<int> ColdThreshold; |
||
| 214 | |||
| 215 | /// Threshold to use when the caller is optimized for size. |
||
| 216 | std::optional<int> OptSizeThreshold; |
||
| 217 | |||
| 218 | /// Threshold to use when the caller is optimized for minsize. |
||
| 219 | std::optional<int> OptMinSizeThreshold; |
||
| 220 | |||
| 221 | /// Threshold to use when the callsite is considered hot. |
||
| 222 | std::optional<int> HotCallSiteThreshold; |
||
| 223 | |||
| 224 | /// Threshold to use when the callsite is considered hot relative to function |
||
| 225 | /// entry. |
||
| 226 | std::optional<int> LocallyHotCallSiteThreshold; |
||
| 227 | |||
| 228 | /// Threshold to use when the callsite is considered cold. |
||
| 229 | std::optional<int> ColdCallSiteThreshold; |
||
| 230 | |||
| 231 | /// Compute inline cost even when the cost has exceeded the threshold. |
||
| 232 | std::optional<bool> ComputeFullInlineCost; |
||
| 233 | |||
| 234 | /// Indicate whether we should allow inline deferral. |
||
| 235 | std::optional<bool> EnableDeferral; |
||
| 236 | |||
| 237 | /// Indicate whether we allow inlining for recursive call. |
||
| 238 | std::optional<bool> AllowRecursiveCall = false; |
||
| 239 | }; |
||
| 240 | |||
| 241 | std::optional<int> getStringFnAttrAsInt(CallBase &CB, StringRef AttrKind); |
||
| 242 | |||
| 243 | /// Generate the parameters to tune the inline cost analysis based only on the |
||
| 244 | /// commandline options. |
||
| 245 | InlineParams getInlineParams(); |
||
| 246 | |||
| 247 | /// Generate the parameters to tune the inline cost analysis based on command |
||
| 248 | /// line options. If -inline-threshold option is not explicitly passed, |
||
| 249 | /// \p Threshold is used as the default threshold. |
||
| 250 | InlineParams getInlineParams(int Threshold); |
||
| 251 | |||
| 252 | /// Generate the parameters to tune the inline cost analysis based on command |
||
| 253 | /// line options. If -inline-threshold option is not explicitly passed, |
||
| 254 | /// the default threshold is computed from \p OptLevel and \p SizeOptLevel. |
||
| 255 | /// An \p OptLevel value above 3 is considered an aggressive optimization mode. |
||
| 256 | /// \p SizeOptLevel of 1 corresponds to the -Os flag and 2 corresponds to |
||
| 257 | /// the -Oz flag. |
||
| 258 | InlineParams getInlineParams(unsigned OptLevel, unsigned SizeOptLevel); |
||
| 259 | |||
| 260 | /// Return the cost associated with a callsite, including parameter passing |
||
| 261 | /// and the call/return instruction. |
||
| 262 | int getCallsiteCost(const CallBase &Call, const DataLayout &DL); |
||
| 263 | |||
| 264 | /// Get an InlineCost object representing the cost of inlining this |
||
| 265 | /// callsite. |
||
| 266 | /// |
||
| 267 | /// Note that a default threshold is passed into this function. This threshold |
||
| 268 | /// could be modified based on callsite's properties and only costs below this |
||
| 269 | /// new threshold are computed with any accuracy. The new threshold can be |
||
| 270 | /// used to bound the computation necessary to determine whether the cost is |
||
| 271 | /// sufficiently low to warrant inlining. |
||
| 272 | /// |
||
| 273 | /// Also note that calling this function *dynamically* computes the cost of |
||
| 274 | /// inlining the callsite. It is an expensive, heavyweight call. |
||
| 275 | InlineCost |
||
| 276 | getInlineCost(CallBase &Call, const InlineParams &Params, |
||
| 277 | TargetTransformInfo &CalleeTTI, |
||
| 278 | function_ref<AssumptionCache &(Function &)> GetAssumptionCache, |
||
| 279 | function_ref<const TargetLibraryInfo &(Function &)> GetTLI, |
||
| 280 | function_ref<BlockFrequencyInfo &(Function &)> GetBFI = nullptr, |
||
| 281 | ProfileSummaryInfo *PSI = nullptr, |
||
| 282 | OptimizationRemarkEmitter *ORE = nullptr); |
||
| 283 | |||
| 284 | /// Get an InlineCost with the callee explicitly specified. |
||
| 285 | /// This allows you to calculate the cost of inlining a function via a |
||
| 286 | /// pointer. This behaves exactly as the version with no explicit callee |
||
| 287 | /// parameter in all other respects. |
||
| 288 | // |
||
| 289 | InlineCost |
||
| 290 | getInlineCost(CallBase &Call, Function *Callee, const InlineParams &Params, |
||
| 291 | TargetTransformInfo &CalleeTTI, |
||
| 292 | function_ref<AssumptionCache &(Function &)> GetAssumptionCache, |
||
| 293 | function_ref<const TargetLibraryInfo &(Function &)> GetTLI, |
||
| 294 | function_ref<BlockFrequencyInfo &(Function &)> GetBFI = nullptr, |
||
| 295 | ProfileSummaryInfo *PSI = nullptr, |
||
| 296 | OptimizationRemarkEmitter *ORE = nullptr); |
||
| 297 | |||
| 298 | /// Returns InlineResult::success() if the call site should be always inlined |
||
| 299 | /// because of user directives, and the inlining is viable. Returns |
||
| 300 | /// InlineResult::failure() if the inlining may never happen because of user |
||
| 301 | /// directives or incompatibilities detectable without needing callee traversal. |
||
| 302 | /// Otherwise returns std::nullopt, meaning that inlining should be decided |
||
| 303 | /// based on other criteria (e.g. cost modeling). |
||
| 304 | std::optional<InlineResult> getAttributeBasedInliningDecision( |
||
| 305 | CallBase &Call, Function *Callee, TargetTransformInfo &CalleeTTI, |
||
| 306 | function_ref<const TargetLibraryInfo &(Function &)> GetTLI); |
||
| 307 | |||
| 308 | /// Get the cost estimate ignoring thresholds. This is similar to getInlineCost |
||
| 309 | /// when passed InlineParams::ComputeFullInlineCost, or a non-null ORE. It |
||
| 310 | /// uses default InlineParams otherwise. |
||
| 311 | /// Contrary to getInlineCost, which makes a threshold-based final evaluation of |
||
| 312 | /// should/shouldn't inline, captured in InlineResult, getInliningCostEstimate |
||
| 313 | /// returns: |
||
| 314 | /// - std::nullopt, if the inlining cannot happen (is illegal) |
||
| 315 | /// - an integer, representing the cost. |
||
| 316 | std::optional<int> getInliningCostEstimate( |
||
| 317 | CallBase &Call, TargetTransformInfo &CalleeTTI, |
||
| 318 | function_ref<AssumptionCache &(Function &)> GetAssumptionCache, |
||
| 319 | function_ref<BlockFrequencyInfo &(Function &)> GetBFI = nullptr, |
||
| 320 | ProfileSummaryInfo *PSI = nullptr, |
||
| 321 | OptimizationRemarkEmitter *ORE = nullptr); |
||
| 322 | |||
| 323 | /// Get the expanded cost features. The features are returned unconditionally, |
||
| 324 | /// even if inlining is impossible. |
||
| 325 | std::optional<InlineCostFeatures> getInliningCostFeatures( |
||
| 326 | CallBase &Call, TargetTransformInfo &CalleeTTI, |
||
| 327 | function_ref<AssumptionCache &(Function &)> GetAssumptionCache, |
||
| 328 | function_ref<BlockFrequencyInfo &(Function &)> GetBFI = nullptr, |
||
| 329 | ProfileSummaryInfo *PSI = nullptr, |
||
| 330 | OptimizationRemarkEmitter *ORE = nullptr); |
||
| 331 | |||
| 332 | /// Minimal filter to detect invalid constructs for inlining. |
||
| 333 | InlineResult isInlineViable(Function &Callee); |
||
| 334 | |||
| 335 | // This pass is used to annotate instructions during the inline process for |
||
| 336 | // debugging and analysis. The main purpose of the pass is to see and test |
||
| 337 | // inliner's decisions when creating new optimizations to InlineCost. |
||
| 338 | struct InlineCostAnnotationPrinterPass |
||
| 339 | : PassInfoMixin<InlineCostAnnotationPrinterPass> { |
||
| 340 | raw_ostream &OS; |
||
| 341 | |||
| 342 | public: |
||
| 343 | explicit InlineCostAnnotationPrinterPass(raw_ostream &OS) : OS(OS) {} |
||
| 344 | PreservedAnalyses run(Function &F, FunctionAnalysisManager &FAM); |
||
| 345 | }; |
||
| 346 | } // namespace llvm |
||
| 347 | |||
| 348 | #endif |