Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 1 | //===- InlineCost.h - Cost analysis for inliner -----------------*- C++ -*-===// |
| 2 | // |
Andrew Walbran | 16937d0 | 2019-10-22 13:54:20 +0100 | [diff] [blame] | 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 |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 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/Analysis/AssumptionCache.h" |
| 17 | #include "llvm/Analysis/CallGraphSCCPass.h" |
| 18 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 19 | #include <cassert> |
| 20 | #include <climits> |
| 21 | |
| 22 | namespace llvm { |
| 23 | class AssumptionCacheTracker; |
| 24 | class BlockFrequencyInfo; |
Andrew Walbran | 3d2c197 | 2020-04-07 12:24:26 +0100 | [diff] [blame] | 25 | class CallBase; |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 26 | class DataLayout; |
| 27 | class Function; |
| 28 | class ProfileSummaryInfo; |
| 29 | class TargetTransformInfo; |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame] | 30 | class TargetLibraryInfo; |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 31 | |
| 32 | namespace InlineConstants { |
| 33 | // Various thresholds used by inline cost analysis. |
| 34 | /// Use when optsize (-Os) is specified. |
| 35 | const int OptSizeThreshold = 50; |
| 36 | |
| 37 | /// Use when minsize (-Oz) is specified. |
| 38 | const int OptMinSizeThreshold = 5; |
| 39 | |
| 40 | /// Use when -O3 is specified. |
| 41 | const int OptAggressiveThreshold = 250; |
| 42 | |
| 43 | // Various magic constants used to adjust heuristics. |
| 44 | const int InstrCost = 5; |
| 45 | const int IndirectCallThreshold = 100; |
| 46 | const int CallPenalty = 25; |
| 47 | const int LastCallToStaticBonus = 15000; |
| 48 | const int ColdccPenalty = 2000; |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 49 | /// Do not inline functions which allocate this many bytes on the stack |
| 50 | /// when the caller is recursive. |
| 51 | const unsigned TotalAllocaSizeRecursiveCaller = 1024; |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame] | 52 | /// Do not inline dynamic allocas that have been constant propagated to be |
| 53 | /// static allocas above this amount in bytes. |
| 54 | const uint64_t MaxSimplifiedDynamicAllocaToInline = 65536; |
| 55 | } // namespace InlineConstants |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 56 | |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 57 | /// Represents the cost of inlining a function. |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 58 | /// |
| 59 | /// This supports special values for functions which should "always" or |
| 60 | /// "never" be inlined. Otherwise, the cost represents a unitless amount; |
| 61 | /// smaller values increase the likelihood of the function being inlined. |
| 62 | /// |
| 63 | /// Objects of this type also provide the adjusted threshold for inlining |
| 64 | /// based on the information available for a particular callsite. They can be |
| 65 | /// directly tested to determine if inlining should occur given the cost and |
| 66 | /// threshold for this cost metric. |
| 67 | class InlineCost { |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame] | 68 | enum SentinelValues { AlwaysInlineCost = INT_MIN, NeverInlineCost = INT_MAX }; |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 69 | |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 70 | /// The estimated cost of inlining this callsite. |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame] | 71 | int Cost = 0; |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 72 | |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 73 | /// The adjusted threshold against which this cost was computed. |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame] | 74 | int Threshold = 0; |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 75 | |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 76 | /// Must be set for Always and Never instances. |
| 77 | const char *Reason = nullptr; |
| 78 | |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 79 | // Trivial constructor, interesting logic in the factory functions below. |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 80 | InlineCost(int Cost, int Threshold, const char *Reason = nullptr) |
| 81 | : Cost(Cost), Threshold(Threshold), Reason(Reason) { |
| 82 | assert((isVariable() || Reason) && |
| 83 | "Reason must be provided for Never or Always"); |
| 84 | } |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 85 | |
| 86 | public: |
| 87 | static InlineCost get(int Cost, int Threshold) { |
| 88 | assert(Cost > AlwaysInlineCost && "Cost crosses sentinel value"); |
| 89 | assert(Cost < NeverInlineCost && "Cost crosses sentinel value"); |
| 90 | return InlineCost(Cost, Threshold); |
| 91 | } |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 92 | static InlineCost getAlways(const char *Reason) { |
| 93 | return InlineCost(AlwaysInlineCost, 0, Reason); |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 94 | } |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 95 | static InlineCost getNever(const char *Reason) { |
| 96 | return InlineCost(NeverInlineCost, 0, Reason); |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 97 | } |
| 98 | |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 99 | /// Test whether the inline cost is low enough for inlining. |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame] | 100 | explicit operator bool() const { return Cost < Threshold; } |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 101 | |
| 102 | bool isAlways() const { return Cost == AlwaysInlineCost; } |
| 103 | bool isNever() const { return Cost == NeverInlineCost; } |
| 104 | bool isVariable() const { return !isAlways() && !isNever(); } |
| 105 | |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 106 | /// Get the inline cost estimate. |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 107 | /// It is an error to call this on an "always" or "never" InlineCost. |
| 108 | int getCost() const { |
| 109 | assert(isVariable() && "Invalid access of InlineCost"); |
| 110 | return Cost; |
| 111 | } |
| 112 | |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 113 | /// Get the threshold against which the cost was computed |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 114 | int getThreshold() const { |
| 115 | assert(isVariable() && "Invalid access of InlineCost"); |
| 116 | return Threshold; |
| 117 | } |
| 118 | |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 119 | /// Get the reason of Always or Never. |
| 120 | const char *getReason() const { |
| 121 | assert((Reason || isVariable()) && |
| 122 | "InlineCost reason must be set for Always or Never"); |
| 123 | return Reason; |
| 124 | } |
| 125 | |
| 126 | /// Get the cost delta from the threshold for inlining. |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 127 | /// Only valid if the cost is of the variable kind. Returns a negative |
| 128 | /// value if the cost is too high to inline. |
| 129 | int getCostDelta() const { return Threshold - getCost(); } |
| 130 | }; |
| 131 | |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 132 | /// InlineResult is basically true or false. For false results the message |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame] | 133 | /// describes a reason. |
| 134 | class InlineResult { |
| 135 | const char *Message = nullptr; |
| 136 | InlineResult(const char *Message = nullptr) : Message(Message) {} |
| 137 | |
| 138 | public: |
| 139 | static InlineResult success() { return {}; } |
| 140 | static InlineResult failure(const char *Reason) { |
| 141 | return InlineResult(Reason); |
| 142 | } |
| 143 | bool isSuccess() const { return Message == nullptr; } |
| 144 | const char *getFailureReason() const { |
| 145 | assert(!isSuccess() && |
| 146 | "getFailureReason should only be called in failure cases"); |
| 147 | return Message; |
| 148 | } |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 149 | }; |
| 150 | |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 151 | /// Thresholds to tune inline cost analysis. The inline cost analysis decides |
| 152 | /// the condition to apply a threshold and applies it. Otherwise, |
| 153 | /// DefaultThreshold is used. If a threshold is Optional, it is applied only |
| 154 | /// when it has a valid value. Typically, users of inline cost analysis |
| 155 | /// obtain an InlineParams object through one of the \c getInlineParams methods |
| 156 | /// and pass it to \c getInlineCost. Some specialized versions of inliner |
| 157 | /// (such as the pre-inliner) might have custom logic to compute \c InlineParams |
| 158 | /// object. |
| 159 | |
| 160 | struct InlineParams { |
| 161 | /// The default threshold to start with for a callee. |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame] | 162 | int DefaultThreshold = -1; |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 163 | |
| 164 | /// Threshold to use for callees with inline hint. |
| 165 | Optional<int> HintThreshold; |
| 166 | |
| 167 | /// Threshold to use for cold callees. |
| 168 | Optional<int> ColdThreshold; |
| 169 | |
| 170 | /// Threshold to use when the caller is optimized for size. |
| 171 | Optional<int> OptSizeThreshold; |
| 172 | |
| 173 | /// Threshold to use when the caller is optimized for minsize. |
| 174 | Optional<int> OptMinSizeThreshold; |
| 175 | |
| 176 | /// Threshold to use when the callsite is considered hot. |
| 177 | Optional<int> HotCallSiteThreshold; |
| 178 | |
| 179 | /// Threshold to use when the callsite is considered hot relative to function |
| 180 | /// entry. |
| 181 | Optional<int> LocallyHotCallSiteThreshold; |
| 182 | |
| 183 | /// Threshold to use when the callsite is considered cold. |
| 184 | Optional<int> ColdCallSiteThreshold; |
| 185 | |
| 186 | /// Compute inline cost even when the cost has exceeded the threshold. |
| 187 | Optional<bool> ComputeFullInlineCost; |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame] | 188 | |
| 189 | /// Indicate whether we should allow inline deferral. |
| 190 | Optional<bool> EnableDeferral = true; |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 191 | }; |
| 192 | |
| 193 | /// Generate the parameters to tune the inline cost analysis based only on the |
| 194 | /// commandline options. |
| 195 | InlineParams getInlineParams(); |
| 196 | |
| 197 | /// Generate the parameters to tune the inline cost analysis based on command |
| 198 | /// line options. If -inline-threshold option is not explicitly passed, |
| 199 | /// \p Threshold is used as the default threshold. |
| 200 | InlineParams getInlineParams(int Threshold); |
| 201 | |
| 202 | /// Generate the parameters to tune the inline cost analysis based on command |
| 203 | /// line options. If -inline-threshold option is not explicitly passed, |
| 204 | /// the default threshold is computed from \p OptLevel and \p SizeOptLevel. |
| 205 | /// An \p OptLevel value above 3 is considered an aggressive optimization mode. |
| 206 | /// \p SizeOptLevel of 1 corresponds to the -Os flag and 2 corresponds to |
| 207 | /// the -Oz flag. |
| 208 | InlineParams getInlineParams(unsigned OptLevel, unsigned SizeOptLevel); |
| 209 | |
| 210 | /// Return the cost associated with a callsite, including parameter passing |
| 211 | /// and the call/return instruction. |
Andrew Walbran | 3d2c197 | 2020-04-07 12:24:26 +0100 | [diff] [blame] | 212 | int getCallsiteCost(CallBase &Call, const DataLayout &DL); |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 213 | |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 214 | /// Get an InlineCost object representing the cost of inlining this |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 215 | /// callsite. |
| 216 | /// |
| 217 | /// Note that a default threshold is passed into this function. This threshold |
| 218 | /// could be modified based on callsite's properties and only costs below this |
| 219 | /// new threshold are computed with any accuracy. The new threshold can be |
| 220 | /// used to bound the computation necessary to determine whether the cost is |
| 221 | /// sufficiently low to warrant inlining. |
| 222 | /// |
| 223 | /// Also note that calling this function *dynamically* computes the cost of |
| 224 | /// inlining the callsite. It is an expensive, heavyweight call. |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame] | 225 | InlineCost |
| 226 | getInlineCost(CallBase &Call, const InlineParams &Params, |
| 227 | TargetTransformInfo &CalleeTTI, |
| 228 | function_ref<AssumptionCache &(Function &)> GetAssumptionCache, |
| 229 | function_ref<const TargetLibraryInfo &(Function &)> GetTLI, |
| 230 | function_ref<BlockFrequencyInfo &(Function &)> GetBFI = nullptr, |
| 231 | ProfileSummaryInfo *PSI = nullptr, |
| 232 | OptimizationRemarkEmitter *ORE = nullptr); |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 233 | |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 234 | /// Get an InlineCost with the callee explicitly specified. |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 235 | /// This allows you to calculate the cost of inlining a function via a |
| 236 | /// pointer. This behaves exactly as the version with no explicit callee |
| 237 | /// parameter in all other respects. |
| 238 | // |
| 239 | InlineCost |
Andrew Walbran | 3d2c197 | 2020-04-07 12:24:26 +0100 | [diff] [blame] | 240 | getInlineCost(CallBase &Call, Function *Callee, const InlineParams &Params, |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 241 | TargetTransformInfo &CalleeTTI, |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame] | 242 | function_ref<AssumptionCache &(Function &)> GetAssumptionCache, |
| 243 | function_ref<const TargetLibraryInfo &(Function &)> GetTLI, |
| 244 | function_ref<BlockFrequencyInfo &(Function &)> GetBFI = nullptr, |
| 245 | ProfileSummaryInfo *PSI = nullptr, |
| 246 | OptimizationRemarkEmitter *ORE = nullptr); |
| 247 | |
| 248 | /// Returns InlineResult::success() if the call site should be always inlined |
| 249 | /// because of user directives, and the inlining is viable. Returns |
| 250 | /// InlineResult::failure() if the inlining may never happen because of user |
| 251 | /// directives or incompatibilities detectable without needing callee traversal. |
| 252 | /// Otherwise returns None, meaning that inlining should be decided based on |
| 253 | /// other criteria (e.g. cost modeling). |
| 254 | Optional<InlineResult> getAttributeBasedInliningDecision( |
| 255 | CallBase &Call, Function *Callee, TargetTransformInfo &CalleeTTI, |
| 256 | function_ref<const TargetLibraryInfo &(Function &)> GetTLI); |
| 257 | |
| 258 | /// Get the cost estimate ignoring thresholds. This is similar to getInlineCost |
| 259 | /// when passed InlineParams::ComputeFullInlineCost, or a non-null ORE. It |
| 260 | /// uses default InlineParams otherwise. |
| 261 | /// Contrary to getInlineCost, which makes a threshold-based final evaluation of |
| 262 | /// should/shouldn't inline, captured in InlineResult, getInliningCostEstimate |
| 263 | /// returns: |
| 264 | /// - None, if the inlining cannot happen (is illegal) |
| 265 | /// - an integer, representing the cost. |
| 266 | Optional<int> getInliningCostEstimate( |
| 267 | CallBase &Call, TargetTransformInfo &CalleeTTI, |
| 268 | function_ref<AssumptionCache &(Function &)> GetAssumptionCache, |
| 269 | function_ref<BlockFrequencyInfo &(Function &)> GetBFI = nullptr, |
| 270 | ProfileSummaryInfo *PSI = nullptr, |
| 271 | OptimizationRemarkEmitter *ORE = nullptr); |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 272 | |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 273 | /// Minimal filter to detect invalid constructs for inlining. |
Andrew Walbran | 16937d0 | 2019-10-22 13:54:20 +0100 | [diff] [blame] | 274 | InlineResult isInlineViable(Function &Callee); |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame] | 275 | |
| 276 | // This pass is used to annotate instructions during the inline process for |
| 277 | // debugging and analysis. The main purpose of the pass is to see and test |
| 278 | // inliner's decisions when creating new optimizations to InlineCost. |
| 279 | struct InlineCostAnnotationPrinterPass |
| 280 | : PassInfoMixin<InlineCostAnnotationPrinterPass> { |
| 281 | raw_ostream &OS; |
| 282 | |
| 283 | public: |
| 284 | explicit InlineCostAnnotationPrinterPass(raw_ostream &OS) : OS(OS) {} |
| 285 | PreservedAnalyses run(Function &F, FunctionAnalysisManager &FAM); |
| 286 | }; |
| 287 | } // namespace llvm |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 288 | |
| 289 | #endif |