Back to home page

EIC code displayed by LXR

 
 

    


File indexing completed on 2026-05-10 08:43:11

0001 //===- BranchProbabilityInfo.h - Branch Probability Analysis ----*- C++ -*-===//
0002 //
0003 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
0004 // See https://llvm.org/LICENSE.txt for license information.
0005 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
0006 //
0007 //===----------------------------------------------------------------------===//
0008 //
0009 // This pass is used to evaluate branch probabilties.
0010 //
0011 //===----------------------------------------------------------------------===//
0012 
0013 #ifndef LLVM_ANALYSIS_BRANCHPROBABILITYINFO_H
0014 #define LLVM_ANALYSIS_BRANCHPROBABILITYINFO_H
0015 
0016 #include "llvm/ADT/DenseMap.h"
0017 #include "llvm/ADT/DenseMapInfo.h"
0018 #include "llvm/ADT/DenseSet.h"
0019 #include "llvm/IR/BasicBlock.h"
0020 #include "llvm/IR/CFG.h"
0021 #include "llvm/IR/PassManager.h"
0022 #include "llvm/IR/ValueHandle.h"
0023 #include "llvm/Pass.h"
0024 #include "llvm/Support/BranchProbability.h"
0025 #include <cassert>
0026 #include <cstdint>
0027 #include <memory>
0028 #include <utility>
0029 
0030 namespace llvm {
0031 
0032 class Function;
0033 class Loop;
0034 class LoopInfo;
0035 class raw_ostream;
0036 class DominatorTree;
0037 class PostDominatorTree;
0038 class TargetLibraryInfo;
0039 class Value;
0040 
0041 /// Analysis providing branch probability information.
0042 ///
0043 /// This is a function analysis which provides information on the relative
0044 /// probabilities of each "edge" in the function's CFG where such an edge is
0045 /// defined by a pair (PredBlock and an index in the successors). The
0046 /// probability of an edge from one block is always relative to the
0047 /// probabilities of other edges from the block. The probabilites of all edges
0048 /// from a block sum to exactly one (100%).
0049 /// We use a pair (PredBlock and an index in the successors) to uniquely
0050 /// identify an edge, since we can have multiple edges from Src to Dst.
0051 /// As an example, we can have a switch which jumps to Dst with value 0 and
0052 /// value 10.
0053 ///
0054 /// Process of computing branch probabilities can be logically viewed as three
0055 /// step process:
0056 ///
0057 ///   First, if there is a profile information associated with the branch then
0058 /// it is trivially translated to branch probabilities. There is one exception
0059 /// from this rule though. Probabilities for edges leading to "unreachable"
0060 /// blocks (blocks with the estimated weight not greater than
0061 /// UNREACHABLE_WEIGHT) are evaluated according to static estimation and
0062 /// override profile information. If no branch probabilities were calculated
0063 /// on this step then take the next one.
0064 ///
0065 ///   Second, estimate absolute execution weights for each block based on
0066 /// statically known information. Roots of such information are "cold",
0067 /// "unreachable", "noreturn" and "unwind" blocks. Those blocks get their
0068 /// weights set to BlockExecWeight::COLD, BlockExecWeight::UNREACHABLE,
0069 /// BlockExecWeight::NORETURN and BlockExecWeight::UNWIND respectively. Then the
0070 /// weights are propagated to the other blocks up the domination line. In
0071 /// addition, if all successors have estimated weights set then maximum of these
0072 /// weights assigned to the block itself (while this is not ideal heuristic in
0073 /// theory it's simple and works reasonably well in most cases) and the process
0074 /// repeats. Once the process of weights propagation converges branch
0075 /// probabilities are set for all such branches that have at least one successor
0076 /// with the weight set. Default execution weight (BlockExecWeight::DEFAULT) is
0077 /// used for any successors which doesn't have its weight set. For loop back
0078 /// branches we use their weights scaled by loop trip count equal to
0079 /// 'LBH_TAKEN_WEIGHT/LBH_NOTTAKEN_WEIGHT'.
0080 ///
0081 /// Here is a simple example demonstrating how the described algorithm works.
0082 ///
0083 ///          BB1
0084 ///         /   \
0085 ///        v     v
0086 ///      BB2     BB3
0087 ///     /   \
0088 ///    v     v
0089 ///  ColdBB  UnreachBB
0090 ///
0091 /// Initially, ColdBB is associated with COLD_WEIGHT and UnreachBB with
0092 /// UNREACHABLE_WEIGHT. COLD_WEIGHT is set to BB2 as maximum between its
0093 /// successors. BB1 and BB3 has no explicit estimated weights and assumed to
0094 /// have DEFAULT_WEIGHT. Based on assigned weights branches will have the
0095 /// following probabilities:
0096 /// P(BB1->BB2) = COLD_WEIGHT/(COLD_WEIGHT + DEFAULT_WEIGHT) =
0097 ///   0xffff / (0xffff + 0xfffff) = 0.0588(5.9%)
0098 /// P(BB1->BB3) = DEFAULT_WEIGHT_WEIGHT/(COLD_WEIGHT + DEFAULT_WEIGHT) =
0099 ///          0xfffff / (0xffff + 0xfffff) = 0.941(94.1%)
0100 /// P(BB2->ColdBB) = COLD_WEIGHT/(COLD_WEIGHT + UNREACHABLE_WEIGHT) = 1(100%)
0101 /// P(BB2->UnreachBB) =
0102 ///   UNREACHABLE_WEIGHT/(COLD_WEIGHT+UNREACHABLE_WEIGHT) = 0(0%)
0103 ///
0104 /// If no branch probabilities were calculated on this step then take the next
0105 /// one.
0106 ///
0107 ///   Third, apply different kinds of local heuristics for each individual
0108 /// branch until first match. For example probability of a pointer to be null is
0109 /// estimated as PH_TAKEN_WEIGHT/(PH_TAKEN_WEIGHT + PH_NONTAKEN_WEIGHT). If
0110 /// no local heuristic has been matched then branch is left with no explicit
0111 /// probability set and assumed to have default probability.
0112 class BranchProbabilityInfo {
0113 public:
0114   BranchProbabilityInfo() = default;
0115 
0116   BranchProbabilityInfo(const Function &F, const LoopInfo &LI,
0117                         const TargetLibraryInfo *TLI = nullptr,
0118                         DominatorTree *DT = nullptr,
0119                         PostDominatorTree *PDT = nullptr) {
0120     calculate(F, LI, TLI, DT, PDT);
0121   }
0122 
0123   BranchProbabilityInfo(BranchProbabilityInfo &&Arg)
0124       : Handles(std::move(Arg.Handles)), Probs(std::move(Arg.Probs)),
0125         LastF(Arg.LastF),
0126         EstimatedBlockWeight(std::move(Arg.EstimatedBlockWeight)) {
0127     for (auto &Handle : Handles)
0128       Handle.setBPI(this);
0129   }
0130 
0131   BranchProbabilityInfo(const BranchProbabilityInfo &) = delete;
0132   BranchProbabilityInfo &operator=(const BranchProbabilityInfo &) = delete;
0133 
0134   BranchProbabilityInfo &operator=(BranchProbabilityInfo &&RHS) {
0135     releaseMemory();
0136     Handles = std::move(RHS.Handles);
0137     Probs = std::move(RHS.Probs);
0138     EstimatedBlockWeight = std::move(RHS.EstimatedBlockWeight);
0139     for (auto &Handle : Handles)
0140       Handle.setBPI(this);
0141     return *this;
0142   }
0143 
0144   bool invalidate(Function &, const PreservedAnalyses &PA,
0145                   FunctionAnalysisManager::Invalidator &);
0146 
0147   void releaseMemory();
0148 
0149   void print(raw_ostream &OS) const;
0150 
0151   /// Get an edge's probability, relative to other out-edges of the Src.
0152   ///
0153   /// This routine provides access to the fractional probability between zero
0154   /// (0%) and one (100%) of this edge executing, relative to other edges
0155   /// leaving the 'Src' block. The returned probability is never zero, and can
0156   /// only be one if the source block has only one successor.
0157   BranchProbability getEdgeProbability(const BasicBlock *Src,
0158                                        unsigned IndexInSuccessors) const;
0159 
0160   /// Get the probability of going from Src to Dst.
0161   ///
0162   /// It returns the sum of all probabilities for edges from Src to Dst.
0163   BranchProbability getEdgeProbability(const BasicBlock *Src,
0164                                        const BasicBlock *Dst) const;
0165 
0166   BranchProbability getEdgeProbability(const BasicBlock *Src,
0167                                        const_succ_iterator Dst) const;
0168 
0169   /// Test if an edge is hot relative to other out-edges of the Src.
0170   ///
0171   /// Check whether this edge out of the source block is 'hot'. We define hot
0172   /// as having a relative probability > 80%.
0173   bool isEdgeHot(const BasicBlock *Src, const BasicBlock *Dst) const;
0174 
0175   /// Print an edge's probability.
0176   ///
0177   /// Retrieves an edge's probability similarly to \see getEdgeProbability, but
0178   /// then prints that probability to the provided stream. That stream is then
0179   /// returned.
0180   raw_ostream &printEdgeProbability(raw_ostream &OS, const BasicBlock *Src,
0181                                     const BasicBlock *Dst) const;
0182 
0183 public:
0184   /// Set the raw probabilities for all edges from the given block.
0185   ///
0186   /// This allows a pass to explicitly set edge probabilities for a block. It
0187   /// can be used when updating the CFG to update the branch probability
0188   /// information.
0189   void setEdgeProbability(const BasicBlock *Src,
0190                           const SmallVectorImpl<BranchProbability> &Probs);
0191 
0192   /// Copy outgoing edge probabilities from \p Src to \p Dst.
0193   ///
0194   /// This allows to keep probabilities unset for the destination if they were
0195   /// unset for source.
0196   void copyEdgeProbabilities(BasicBlock *Src, BasicBlock *Dst);
0197 
0198   /// Swap outgoing edges probabilities for \p Src with branch terminator
0199   void swapSuccEdgesProbabilities(const BasicBlock *Src);
0200 
0201   static BranchProbability getBranchProbStackProtector(bool IsLikely) {
0202     static const BranchProbability LikelyProb((1u << 20) - 1, 1u << 20);
0203     return IsLikely ? LikelyProb : LikelyProb.getCompl();
0204   }
0205 
0206   void calculate(const Function &F, const LoopInfo &LI,
0207                  const TargetLibraryInfo *TLI, DominatorTree *DT,
0208                  PostDominatorTree *PDT);
0209 
0210   /// Forget analysis results for the given basic block.
0211   void eraseBlock(const BasicBlock *BB);
0212 
0213   // Data structure to track SCCs for handling irreducible loops.
0214   class SccInfo {
0215     // Enum of types to classify basic blocks in SCC. Basic block belonging to
0216     // SCC is 'Inner' until it is either 'Header' or 'Exiting'. Note that a
0217     // basic block can be 'Header' and 'Exiting' at the same time.
0218     enum SccBlockType {
0219       Inner = 0x0,
0220       Header = 0x1,
0221       Exiting = 0x2,
0222     };
0223     // Map of basic blocks to SCC IDs they belong to. If basic block doesn't
0224     // belong to any SCC it is not in the map.
0225     using SccMap = DenseMap<const BasicBlock *, int>;
0226     // Each basic block in SCC is attributed with one or several types from
0227     // SccBlockType. Map value has uint32_t type (instead of SccBlockType)
0228     // since basic block may be for example "Header" and "Exiting" at the same
0229     // time and we need to be able to keep more than one value from
0230     // SccBlockType.
0231     using SccBlockTypeMap = DenseMap<const BasicBlock *, uint32_t>;
0232     // Vector containing classification of basic blocks for all  SCCs where i'th
0233     // vector element corresponds to SCC with ID equal to i.
0234     using SccBlockTypeMaps = std::vector<SccBlockTypeMap>;
0235 
0236     SccMap SccNums;
0237     SccBlockTypeMaps SccBlocks;
0238 
0239   public:
0240     explicit SccInfo(const Function &F);
0241 
0242     /// If \p BB belongs to some SCC then ID of that SCC is returned, otherwise
0243     /// -1 is returned. If \p BB belongs to more than one SCC at the same time
0244     /// result is undefined.
0245     int getSCCNum(const BasicBlock *BB) const;
0246     /// Returns true if \p BB is a 'header' block in SCC with \p SccNum ID,
0247     /// false otherwise.
0248     bool isSCCHeader(const BasicBlock *BB, int SccNum) const {
0249       return getSccBlockType(BB, SccNum) & Header;
0250     }
0251     /// Returns true if \p BB is an 'exiting' block in SCC with \p SccNum ID,
0252     /// false otherwise.
0253     bool isSCCExitingBlock(const BasicBlock *BB, int SccNum) const {
0254       return getSccBlockType(BB, SccNum) & Exiting;
0255     }
0256     /// Fills in \p Enters vector with all such blocks that don't belong to
0257     /// SCC with \p SccNum ID but there is an edge to a block belonging to the
0258     /// SCC.
0259     void getSccEnterBlocks(int SccNum,
0260                            SmallVectorImpl<BasicBlock *> &Enters) const;
0261     /// Fills in \p Exits vector with all such blocks that don't belong to
0262     /// SCC with \p SccNum ID but there is an edge from a block belonging to the
0263     /// SCC.
0264     void getSccExitBlocks(int SccNum,
0265                           SmallVectorImpl<BasicBlock *> &Exits) const;
0266 
0267   private:
0268     /// Returns \p BB's type according to classification given by SccBlockType
0269     /// enum. Please note that \p BB must belong to SSC with \p SccNum ID.
0270     uint32_t getSccBlockType(const BasicBlock *BB, int SccNum) const;
0271     /// Calculates \p BB's type and stores it in internal data structures for
0272     /// future use. Please note that \p BB must belong to SSC with \p SccNum ID.
0273     void calculateSccBlockType(const BasicBlock *BB, int SccNum);
0274   };
0275 
0276 private:
0277   // We need to store CallbackVH's in order to correctly handle basic block
0278   // removal.
0279   class BasicBlockCallbackVH final : public CallbackVH {
0280     BranchProbabilityInfo *BPI;
0281 
0282     void deleted() override {
0283       assert(BPI != nullptr);
0284       BPI->eraseBlock(cast<BasicBlock>(getValPtr()));
0285     }
0286 
0287   public:
0288     void setBPI(BranchProbabilityInfo *BPI) { this->BPI = BPI; }
0289 
0290     BasicBlockCallbackVH(const Value *V, BranchProbabilityInfo *BPI = nullptr)
0291         : CallbackVH(const_cast<Value *>(V)), BPI(BPI) {}
0292   };
0293 
0294   /// Pair of Loop and SCC ID number. Used to unify handling of normal and
0295   /// SCC based loop representations.
0296   using LoopData = std::pair<Loop *, int>;
0297   /// Helper class to keep basic block along with its loop data information.
0298   class LoopBlock {
0299   public:
0300     explicit LoopBlock(const BasicBlock *BB, const LoopInfo &LI,
0301                        const SccInfo &SccI);
0302 
0303     const BasicBlock *getBlock() const { return BB; }
0304     BasicBlock *getBlock() { return const_cast<BasicBlock *>(BB); }
0305     LoopData getLoopData() const { return LD; }
0306     Loop *getLoop() const { return LD.first; }
0307     int getSccNum() const { return LD.second; }
0308 
0309     bool belongsToLoop() const { return getLoop() || getSccNum() != -1; }
0310     bool belongsToSameLoop(const LoopBlock &LB) const {
0311       return (LB.getLoop() && getLoop() == LB.getLoop()) ||
0312              (LB.getSccNum() != -1 && getSccNum() == LB.getSccNum());
0313     }
0314 
0315   private:
0316     const BasicBlock *const BB = nullptr;
0317     LoopData LD = {nullptr, -1};
0318   };
0319 
0320   // Pair of LoopBlocks representing an edge from first to second block.
0321   using LoopEdge = std::pair<const LoopBlock &, const LoopBlock &>;
0322 
0323   DenseSet<BasicBlockCallbackVH, DenseMapInfo<Value*>> Handles;
0324 
0325   // Since we allow duplicate edges from one basic block to another, we use
0326   // a pair (PredBlock and an index in the successors) to specify an edge.
0327   using Edge = std::pair<const BasicBlock *, unsigned>;
0328 
0329   DenseMap<Edge, BranchProbability> Probs;
0330 
0331   /// Track the last function we run over for printing.
0332   const Function *LastF = nullptr;
0333 
0334   const LoopInfo *LI = nullptr;
0335 
0336   /// Keeps information about all SCCs in a function.
0337   std::unique_ptr<const SccInfo> SccI;
0338 
0339   /// Keeps mapping of a basic block to its estimated weight.
0340   SmallDenseMap<const BasicBlock *, uint32_t> EstimatedBlockWeight;
0341 
0342   /// Keeps mapping of a loop to estimated weight to enter the loop.
0343   SmallDenseMap<LoopData, uint32_t> EstimatedLoopWeight;
0344 
0345   /// Helper to construct LoopBlock for \p BB.
0346   LoopBlock getLoopBlock(const BasicBlock *BB) const {
0347     return LoopBlock(BB, *LI, *SccI);
0348   }
0349 
0350   /// Returns true if destination block belongs to some loop and source block is
0351   /// either doesn't belong to any loop or belongs to a loop which is not inner
0352   /// relative to the destination block.
0353   bool isLoopEnteringEdge(const LoopEdge &Edge) const;
0354   /// Returns true if source block belongs to some loop and destination block is
0355   /// either doesn't belong to any loop or belongs to a loop which is not inner
0356   /// relative to the source block.
0357   bool isLoopExitingEdge(const LoopEdge &Edge) const;
0358   /// Returns true if \p Edge is either enters to or exits from some loop, false
0359   /// in all other cases.
0360   bool isLoopEnteringExitingEdge(const LoopEdge &Edge) const;
0361   /// Returns true if source and destination blocks belongs to the same loop and
0362   /// destination block is loop header.
0363   bool isLoopBackEdge(const LoopEdge &Edge) const;
0364   // Fills in \p Enters vector with all "enter" blocks to a loop \LB belongs to.
0365   void getLoopEnterBlocks(const LoopBlock &LB,
0366                           SmallVectorImpl<BasicBlock *> &Enters) const;
0367   // Fills in \p Exits vector with all "exit" blocks from a loop \LB belongs to.
0368   void getLoopExitBlocks(const LoopBlock &LB,
0369                          SmallVectorImpl<BasicBlock *> &Exits) const;
0370 
0371   /// Returns estimated weight for \p BB. std::nullopt if \p BB has no estimated
0372   /// weight.
0373   std::optional<uint32_t> getEstimatedBlockWeight(const BasicBlock *BB) const;
0374 
0375   /// Returns estimated weight to enter \p L. In other words it is weight of
0376   /// loop's header block not scaled by trip count. Returns std::nullopt if \p L
0377   /// has no no estimated weight.
0378   std::optional<uint32_t> getEstimatedLoopWeight(const LoopData &L) const;
0379 
0380   /// Return estimated weight for \p Edge. Returns std::nullopt if estimated
0381   /// weight is unknown.
0382   std::optional<uint32_t> getEstimatedEdgeWeight(const LoopEdge &Edge) const;
0383 
0384   /// Iterates over all edges leading from \p SrcBB to \p Successors and
0385   /// returns maximum of all estimated weights. If at least one edge has unknown
0386   /// estimated weight std::nullopt is returned.
0387   template <class IterT>
0388   std::optional<uint32_t>
0389   getMaxEstimatedEdgeWeight(const LoopBlock &SrcBB,
0390                             iterator_range<IterT> Successors) const;
0391 
0392   /// If \p LoopBB has no estimated weight then set it to \p BBWeight and
0393   /// return true. Otherwise \p BB's weight remains unchanged and false is
0394   /// returned. In addition all blocks/loops that might need their weight to be
0395   /// re-estimated are put into BlockWorkList/LoopWorkList.
0396   bool updateEstimatedBlockWeight(LoopBlock &LoopBB, uint32_t BBWeight,
0397                                   SmallVectorImpl<BasicBlock *> &BlockWorkList,
0398                                   SmallVectorImpl<LoopBlock> &LoopWorkList);
0399 
0400   /// Starting from \p LoopBB (including \p LoopBB itself) propagate \p BBWeight
0401   /// up the domination tree.
0402   void propagateEstimatedBlockWeight(const LoopBlock &LoopBB, DominatorTree *DT,
0403                                      PostDominatorTree *PDT, uint32_t BBWeight,
0404                                      SmallVectorImpl<BasicBlock *> &WorkList,
0405                                      SmallVectorImpl<LoopBlock> &LoopWorkList);
0406 
0407   /// Returns block's weight encoded in the IR.
0408   std::optional<uint32_t> getInitialEstimatedBlockWeight(const BasicBlock *BB);
0409 
0410   // Computes estimated weights for all blocks in \p F.
0411   void computeEestimateBlockWeight(const Function &F, DominatorTree *DT,
0412                                    PostDominatorTree *PDT);
0413 
0414   /// Based on computed weights by \p computeEstimatedBlockWeight set
0415   /// probabilities on branches.
0416   bool calcEstimatedHeuristics(const BasicBlock *BB);
0417   bool calcMetadataWeights(const BasicBlock *BB);
0418   bool calcPointerHeuristics(const BasicBlock *BB);
0419   bool calcZeroHeuristics(const BasicBlock *BB, const TargetLibraryInfo *TLI);
0420   bool calcFloatingPointHeuristics(const BasicBlock *BB);
0421 };
0422 
0423 /// Analysis pass which computes \c BranchProbabilityInfo.
0424 class BranchProbabilityAnalysis
0425     : public AnalysisInfoMixin<BranchProbabilityAnalysis> {
0426   friend AnalysisInfoMixin<BranchProbabilityAnalysis>;
0427 
0428   static AnalysisKey Key;
0429 
0430 public:
0431   /// Provide the result type for this analysis pass.
0432   using Result = BranchProbabilityInfo;
0433 
0434   /// Run the analysis pass over a function and produce BPI.
0435   BranchProbabilityInfo run(Function &F, FunctionAnalysisManager &AM);
0436 };
0437 
0438 /// Printer pass for the \c BranchProbabilityAnalysis results.
0439 class BranchProbabilityPrinterPass
0440     : public PassInfoMixin<BranchProbabilityPrinterPass> {
0441   raw_ostream &OS;
0442 
0443 public:
0444   explicit BranchProbabilityPrinterPass(raw_ostream &OS) : OS(OS) {}
0445 
0446   PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
0447 
0448   static bool isRequired() { return true; }
0449 };
0450 
0451 /// Legacy analysis pass which computes \c BranchProbabilityInfo.
0452 class BranchProbabilityInfoWrapperPass : public FunctionPass {
0453   BranchProbabilityInfo BPI;
0454 
0455 public:
0456   static char ID;
0457 
0458   BranchProbabilityInfoWrapperPass();
0459 
0460   BranchProbabilityInfo &getBPI() { return BPI; }
0461   const BranchProbabilityInfo &getBPI() const { return BPI; }
0462 
0463   void getAnalysisUsage(AnalysisUsage &AU) const override;
0464   bool runOnFunction(Function &F) override;
0465   void releaseMemory() override;
0466   void print(raw_ostream &OS, const Module *M = nullptr) const override;
0467 };
0468 
0469 } // end namespace llvm
0470 
0471 #endif // LLVM_ANALYSIS_BRANCHPROBABILITYINFO_H