| 1 | //===- BranchProbabilityInfo.cpp - Branch Probability Analysis ------------===// |
| 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 | // Loops should be simplified before this analysis. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "llvm/Analysis/BranchProbabilityInfo.h" |
| 14 | #include "llvm/ADT/PostOrderIterator.h" |
| 15 | #include "llvm/ADT/STLExtras.h" |
| 16 | #include "llvm/ADT/SmallVector.h" |
| 17 | #include "llvm/Analysis/ConstantFolding.h" |
| 18 | #include "llvm/Analysis/CycleAnalysis.h" |
| 19 | #include "llvm/Analysis/PostDominators.h" |
| 20 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 21 | #include "llvm/IR/Attributes.h" |
| 22 | #include "llvm/IR/BasicBlock.h" |
| 23 | #include "llvm/IR/CFG.h" |
| 24 | #include "llvm/IR/Constants.h" |
| 25 | #include "llvm/IR/Dominators.h" |
| 26 | #include "llvm/IR/Function.h" |
| 27 | #include "llvm/IR/InstrTypes.h" |
| 28 | #include "llvm/IR/Instruction.h" |
| 29 | #include "llvm/IR/Instructions.h" |
| 30 | #include "llvm/IR/LLVMContext.h" |
| 31 | #include "llvm/IR/Metadata.h" |
| 32 | #include "llvm/IR/PassManager.h" |
| 33 | #include "llvm/IR/ProfDataUtils.h" |
| 34 | #include "llvm/IR/Type.h" |
| 35 | #include "llvm/IR/Value.h" |
| 36 | #include "llvm/InitializePasses.h" |
| 37 | #include "llvm/Pass.h" |
| 38 | #include "llvm/Support/BranchProbability.h" |
| 39 | #include "llvm/Support/Casting.h" |
| 40 | #include "llvm/Support/CommandLine.h" |
| 41 | #include "llvm/Support/Debug.h" |
| 42 | #include "llvm/Support/raw_ostream.h" |
| 43 | #include <cassert> |
| 44 | #include <cstdint> |
| 45 | #include <utility> |
| 46 | |
| 47 | using namespace llvm; |
| 48 | |
| 49 | #define DEBUG_TYPE "branch-prob" |
| 50 | |
| 51 | static cl::opt<bool> PrintBranchProb( |
| 52 | "print-bpi" , cl::init(Val: false), cl::Hidden, |
| 53 | cl::desc("Print the branch probability info." )); |
| 54 | |
| 55 | static cl::opt<std::string> PrintBranchProbFuncName( |
| 56 | "print-bpi-func-name" , cl::Hidden, |
| 57 | cl::desc("The option to specify the name of the function " |
| 58 | "whose branch probability info is printed." )); |
| 59 | |
| 60 | INITIALIZE_PASS_BEGIN(BranchProbabilityInfoWrapperPass, "branch-prob" , |
| 61 | "Branch Probability Analysis" , false, true) |
| 62 | INITIALIZE_PASS_DEPENDENCY(CycleInfoWrapperPass) |
| 63 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) |
| 64 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
| 65 | INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass) |
| 66 | INITIALIZE_PASS_END(BranchProbabilityInfoWrapperPass, "branch-prob" , |
| 67 | "Branch Probability Analysis" , false, true) |
| 68 | |
| 69 | BranchProbabilityInfoWrapperPass::BranchProbabilityInfoWrapperPass() |
| 70 | : FunctionPass(ID) {} |
| 71 | |
| 72 | char BranchProbabilityInfoWrapperPass::ID = 0; |
| 73 | |
| 74 | // Weights are for internal use only. They are used by heuristics to help to |
| 75 | // estimate edges' probability. Example: |
| 76 | // |
| 77 | // Using "Loop Branch Heuristics" we predict weights of edges for the |
| 78 | // block BB2. |
| 79 | // ... |
| 80 | // | |
| 81 | // V |
| 82 | // BB1<-+ |
| 83 | // | | |
| 84 | // | | (Weight = 124) |
| 85 | // V | |
| 86 | // BB2--+ |
| 87 | // | |
| 88 | // | (Weight = 4) |
| 89 | // V |
| 90 | // BB3 |
| 91 | // |
| 92 | // Probability of the edge BB2->BB1 = 124 / (124 + 4) = 0.96875 |
| 93 | // Probability of the edge BB2->BB3 = 4 / (124 + 4) = 0.03125 |
| 94 | static const uint32_t LBH_TAKEN_WEIGHT = 124; |
| 95 | static const uint32_t LBH_NONTAKEN_WEIGHT = 4; |
| 96 | |
| 97 | /// Unreachable-terminating branch taken probability. |
| 98 | /// |
| 99 | /// This is the probability for a branch being taken to a block that terminates |
| 100 | /// (eventually) in unreachable. These are predicted as unlikely as possible. |
| 101 | /// All reachable probability will proportionally share the remaining part. |
| 102 | static constexpr BranchProbability UR_TAKEN_PROB = BranchProbability::getRaw(N: 1); |
| 103 | |
| 104 | /// Heuristics and lookup tables for non-loop branches: |
| 105 | /// Pointer Heuristics (PH) |
| 106 | static const uint32_t PH_TAKEN_WEIGHT = 20; |
| 107 | static const uint32_t PH_NONTAKEN_WEIGHT = 12; |
| 108 | static constexpr BranchProbability |
| 109 | PtrTakenProb(PH_TAKEN_WEIGHT, PH_TAKEN_WEIGHT + PH_NONTAKEN_WEIGHT); |
| 110 | static constexpr BranchProbability |
| 111 | PtrUntakenProb(PH_NONTAKEN_WEIGHT, PH_TAKEN_WEIGHT + PH_NONTAKEN_WEIGHT); |
| 112 | |
| 113 | /// Zero Heuristics (ZH) |
| 114 | static const uint32_t ZH_TAKEN_WEIGHT = 20; |
| 115 | static const uint32_t ZH_NONTAKEN_WEIGHT = 12; |
| 116 | static constexpr BranchProbability |
| 117 | ZeroTakenProb(ZH_TAKEN_WEIGHT, ZH_TAKEN_WEIGHT + ZH_NONTAKEN_WEIGHT); |
| 118 | static constexpr BranchProbability |
| 119 | ZeroUntakenProb(ZH_NONTAKEN_WEIGHT, ZH_TAKEN_WEIGHT + ZH_NONTAKEN_WEIGHT); |
| 120 | |
| 121 | // Floating-Point Heuristics (FPH) |
| 122 | static const uint32_t FPH_TAKEN_WEIGHT = 20; |
| 123 | static const uint32_t FPH_NONTAKEN_WEIGHT = 12; |
| 124 | |
| 125 | /// This is the probability for an ordered floating point comparison. |
| 126 | static const uint32_t FPH_ORD_WEIGHT = 1024 * 1024 - 1; |
| 127 | /// This is the probability for an unordered floating point comparison, it means |
| 128 | /// one or two of the operands are NaN. Usually it is used to test for an |
| 129 | /// exceptional case, so the result is unlikely. |
| 130 | static const uint32_t FPH_UNO_WEIGHT = 1; |
| 131 | |
| 132 | static constexpr BranchProbability |
| 133 | FPOrdTakenProb(FPH_ORD_WEIGHT, FPH_ORD_WEIGHT + FPH_UNO_WEIGHT); |
| 134 | static constexpr BranchProbability |
| 135 | FPOrdUntakenProb(FPH_UNO_WEIGHT, FPH_ORD_WEIGHT + FPH_UNO_WEIGHT); |
| 136 | static constexpr BranchProbability |
| 137 | FPTakenProb(FPH_TAKEN_WEIGHT, FPH_TAKEN_WEIGHT + FPH_NONTAKEN_WEIGHT); |
| 138 | static constexpr BranchProbability |
| 139 | FPUntakenProb(FPH_NONTAKEN_WEIGHT, FPH_TAKEN_WEIGHT + FPH_NONTAKEN_WEIGHT); |
| 140 | |
| 141 | /// Set of dedicated "absolute" execution weights for a block. These weights are |
| 142 | /// meaningful relative to each other and their derivatives only. |
| 143 | enum class BlockExecWeight : std::uint32_t { |
| 144 | /// Special weight used for cases with exact zero probability. |
| 145 | ZERO = 0x0, |
| 146 | /// Minimal possible non zero weight. |
| 147 | LOWEST_NON_ZERO = 0x1, |
| 148 | /// Weight to an 'unreachable' block. |
| 149 | UNREACHABLE = ZERO, |
| 150 | /// Weight to a block containing non returning call. |
| 151 | NORETURN = LOWEST_NON_ZERO, |
| 152 | /// Weight to 'unwind' block of an invoke instruction. |
| 153 | UNWIND = LOWEST_NON_ZERO, |
| 154 | /// Weight to a 'cold' block. Cold blocks are the ones containing calls marked |
| 155 | /// with attribute 'cold'. |
| 156 | COLD = 0xffff, |
| 157 | /// Default weight is used in cases when there is no dedicated execution |
| 158 | /// weight set. It is not propagated through the domination line either. |
| 159 | DEFAULT = 0xfffff |
| 160 | }; |
| 161 | |
| 162 | namespace { |
| 163 | class BPIConstruction { |
| 164 | public: |
| 165 | BPIConstruction(BranchProbabilityInfo &BPI) : BPI(BPI) {} |
| 166 | void calculate(const Function &F, const CycleInfo &CI, |
| 167 | const TargetLibraryInfo *TLI, DominatorTree *DT, |
| 168 | PostDominatorTree *PDT); |
| 169 | |
| 170 | private: |
| 171 | // Pair representing an edge from first to second block. |
| 172 | using LoopEdge = std::pair<const BasicBlock *, const BasicBlock *>; |
| 173 | |
| 174 | /// Returns true if destination block belongs to some loop and source block is |
| 175 | /// either doesn't belong to any loop or belongs to a loop which is not inner |
| 176 | /// relative to the destination block. |
| 177 | bool isLoopEnteringEdge(const LoopEdge &Edge) const; |
| 178 | /// Returns true if source block belongs to some loop and destination block is |
| 179 | /// either doesn't belong to any loop or belongs to a loop which is not inner |
| 180 | /// relative to the source block. |
| 181 | bool isLoopExitingEdge(const LoopEdge &Edge) const; |
| 182 | /// Returns true if \p Edge is either enters to or exits from some loop, false |
| 183 | /// in all other cases. |
| 184 | bool isLoopEnteringExitingEdge(const LoopEdge &Edge) const; |
| 185 | // Fills in \p Enters vector with all "enter" blocks to a loop \LB belongs to. |
| 186 | void getLoopEnterBlocks(const BasicBlock *LB, |
| 187 | SmallVectorImpl<const BasicBlock *> &Enters) const; |
| 188 | |
| 189 | /// Returns estimated weight for \p BB. std::nullopt if \p BB has no estimated |
| 190 | /// weight. |
| 191 | std::optional<uint32_t> getEstimatedBlockWeight(const BasicBlock *BB) const; |
| 192 | |
| 193 | /// Returns estimated weight to enter \p L. In other words it is weight of |
| 194 | /// loop's header block not scaled by trip count. Returns std::nullopt if \p C |
| 195 | /// has no no estimated weight. |
| 196 | std::optional<uint32_t> getEstimatedLoopWeight(CycleRef C) const; |
| 197 | |
| 198 | /// Return estimated weight for \p Edge. Returns std::nullopt if estimated |
| 199 | /// weight is unknown. |
| 200 | std::optional<uint32_t> getEstimatedEdgeWeight(const LoopEdge &Edge) const; |
| 201 | |
| 202 | /// Iterates over all edges leading from \p SrcBB to \p Successors and |
| 203 | /// returns maximum of all estimated weights. If at least one edge has unknown |
| 204 | /// estimated weight std::nullopt is returned. |
| 205 | template <class IterT> |
| 206 | std::optional<uint32_t> |
| 207 | getMaxEstimatedEdgeWeight(const BasicBlock *SrcBB, |
| 208 | iterator_range<IterT> Successors) const; |
| 209 | |
| 210 | /// If \p LoopBB has no estimated weight then set it to \p BBWeight and |
| 211 | /// return true. Otherwise \p BB's weight remains unchanged and false is |
| 212 | /// returned. In addition all blocks/loops that might need their weight to be |
| 213 | /// re-estimated are put into BlockWorkList/LoopWorkList. |
| 214 | bool |
| 215 | updateEstimatedBlockWeight(const BasicBlock *BB, uint32_t BBWeight, |
| 216 | SmallVectorImpl<const BasicBlock *> &BlockWorkList, |
| 217 | SmallVectorImpl<const BasicBlock *> &LoopWorkList); |
| 218 | |
| 219 | /// Starting from \p LoopBB (including \p LoopBB itself) propagate \p BBWeight |
| 220 | /// up the domination tree. |
| 221 | void propagateEstimatedBlockWeight( |
| 222 | const BasicBlock *BB, DominatorTree *DT, PostDominatorTree *PDT, |
| 223 | uint32_t BBWeight, SmallVectorImpl<const BasicBlock *> &WorkList, |
| 224 | SmallVectorImpl<const BasicBlock *> &LoopWorkList); |
| 225 | |
| 226 | /// Returns block's weight encoded in the IR. |
| 227 | std::optional<uint32_t> getInitialEstimatedBlockWeight(const BasicBlock *BB); |
| 228 | |
| 229 | // Computes estimated weights for all blocks in \p F. |
| 230 | void estimateBlockWeights(const Function &F, DominatorTree *DT, |
| 231 | PostDominatorTree *PDT); |
| 232 | |
| 233 | /// Based on computed weights by \p computeEstimatedBlockWeight set |
| 234 | /// probabilities on branches. |
| 235 | bool calcEstimatedHeuristics(const BasicBlock *BB); |
| 236 | bool calcMetadataWeights(const BasicBlock *BB); |
| 237 | bool calcPointerHeuristics(const BasicBlock *BB); |
| 238 | bool calcZeroHeuristics(const BasicBlock *BB, const TargetLibraryInfo *TLI); |
| 239 | bool calcFloatingPointHeuristics(const BasicBlock *BB); |
| 240 | |
| 241 | BranchProbabilityInfo &BPI; |
| 242 | |
| 243 | const CycleInfo *CI = nullptr; |
| 244 | |
| 245 | /// Keeps mapping of a basic block to its estimated weight. |
| 246 | SmallDenseMap<const BasicBlock *, uint32_t> EstimatedBlockWeight; |
| 247 | |
| 248 | /// Keeps mapping of a loop to estimated weight to enter the loop. |
| 249 | SmallDenseMap<CycleRef, uint32_t> EstimatedLoopWeight; |
| 250 | }; |
| 251 | |
| 252 | bool BPIConstruction::isLoopEnteringEdge(const LoopEdge &Edge) const { |
| 253 | CycleRef SrcCycle = CI->getCycle(Block: Edge.first); |
| 254 | CycleRef DstCycle = CI->getCycle(Block: Edge.second); |
| 255 | if (!DstCycle) // Edge into no-cycle is not entering. |
| 256 | return false; |
| 257 | if (!SrcCycle) // Edge from no-cycle into cycle is entering. |
| 258 | return true; |
| 259 | return !CI->contains(Outer: DstCycle, Inner: SrcCycle); |
| 260 | } |
| 261 | |
| 262 | bool BPIConstruction::isLoopExitingEdge(const LoopEdge &Edge) const { |
| 263 | return isLoopEnteringEdge(Edge: {Edge.second, Edge.first}); |
| 264 | } |
| 265 | |
| 266 | bool BPIConstruction::isLoopEnteringExitingEdge(const LoopEdge &Edge) const { |
| 267 | return isLoopEnteringEdge(Edge) || isLoopExitingEdge(Edge); |
| 268 | } |
| 269 | |
| 270 | void BPIConstruction::getLoopEnterBlocks( |
| 271 | const BasicBlock *BB, SmallVectorImpl<const BasicBlock *> &Enters) const { |
| 272 | CycleRef C = CI->getCycle(Block: BB); |
| 273 | for (BasicBlock *Entry : CI->getEntries(C)) |
| 274 | for (const auto *Pred : predecessors(BB: Entry)) |
| 275 | if (!CI->contains(C, Block: Pred)) |
| 276 | Enters.push_back(Elt: Pred); |
| 277 | } |
| 278 | |
| 279 | // Propagate existing explicit probabilities from either profile data or |
| 280 | // 'expect' intrinsic processing. Examine metadata against unreachable |
| 281 | // heuristic. The probability of the edge coming to unreachable block is |
| 282 | // set to min of metadata and unreachable heuristic. |
| 283 | bool BPIConstruction::calcMetadataWeights(const BasicBlock *BB) { |
| 284 | const Instruction *TI = BB->getTerminator(); |
| 285 | assert(TI->getNumSuccessors() > 1 && "expected more than one successor!" ); |
| 286 | if (!(isa<CondBrInst>(Val: TI) || isa<SwitchInst>(Val: TI) || isa<IndirectBrInst>(Val: TI) || |
| 287 | isa<InvokeInst>(Val: TI) || isa<CallBrInst>(Val: TI))) |
| 288 | return false; |
| 289 | |
| 290 | MDNode *WeightsNode = getValidBranchWeightMDNode(I: *TI); |
| 291 | if (!WeightsNode) |
| 292 | return false; |
| 293 | |
| 294 | // Check that the number of successors is manageable. |
| 295 | assert(TI->getNumSuccessors() < UINT32_MAX && "Too many successors" ); |
| 296 | |
| 297 | // Build up the final weights that will be used in a temporary buffer. |
| 298 | // Compute the sum of all weights to later decide whether they need to |
| 299 | // be scaled to fit in 32 bits. |
| 300 | uint64_t WeightSum = 0; |
| 301 | SmallVector<uint32_t, 2> Weights; |
| 302 | SmallVector<unsigned, 2> UnreachableIdxs; |
| 303 | SmallVector<unsigned, 2> ReachableIdxs; |
| 304 | |
| 305 | extractBranchWeights(ProfileData: WeightsNode, Weights); |
| 306 | auto Succs = succ_begin(I: TI); |
| 307 | for (unsigned I = 0, E = Weights.size(); I != E; ++I) { |
| 308 | WeightSum += Weights[I]; |
| 309 | auto EstimatedWeight = getEstimatedEdgeWeight(Edge: {BB, *Succs++}); |
| 310 | if (EstimatedWeight && |
| 311 | *EstimatedWeight <= static_cast<uint32_t>(BlockExecWeight::UNREACHABLE)) |
| 312 | UnreachableIdxs.push_back(Elt: I); |
| 313 | else |
| 314 | ReachableIdxs.push_back(Elt: I); |
| 315 | } |
| 316 | assert(Weights.size() == TI->getNumSuccessors() && "Checked above" ); |
| 317 | |
| 318 | // If the sum of weights does not fit in 32 bits, scale every weight down |
| 319 | // accordingly. |
| 320 | uint64_t ScalingFactor = |
| 321 | (WeightSum > UINT32_MAX) ? WeightSum / UINT32_MAX + 1 : 1; |
| 322 | |
| 323 | if (ScalingFactor > 1) { |
| 324 | WeightSum = 0; |
| 325 | for (unsigned I = 0, E = TI->getNumSuccessors(); I != E; ++I) { |
| 326 | Weights[I] /= ScalingFactor; |
| 327 | WeightSum += Weights[I]; |
| 328 | } |
| 329 | } |
| 330 | assert(WeightSum <= UINT32_MAX && |
| 331 | "Expected weights to scale down to 32 bits" ); |
| 332 | |
| 333 | if (WeightSum == 0 || ReachableIdxs.size() == 0) { |
| 334 | for (unsigned I = 0, E = TI->getNumSuccessors(); I != E; ++I) |
| 335 | Weights[I] = 1; |
| 336 | WeightSum = TI->getNumSuccessors(); |
| 337 | } |
| 338 | |
| 339 | // Set the probability. |
| 340 | SmallVector<BranchProbability, 2> BP; |
| 341 | for (unsigned I = 0, E = TI->getNumSuccessors(); I != E; ++I) |
| 342 | BP.push_back(Elt: { Weights[I], static_cast<uint32_t>(WeightSum) }); |
| 343 | |
| 344 | // Examine the metadata against unreachable heuristic. |
| 345 | // If the unreachable heuristic is more strong then we use it for this edge. |
| 346 | if (UnreachableIdxs.size() == 0 || ReachableIdxs.size() == 0) { |
| 347 | BPI.setEdgeProbability(Src: BB, Probs: BP); |
| 348 | return true; |
| 349 | } |
| 350 | |
| 351 | auto UnreachableProb = UR_TAKEN_PROB; |
| 352 | for (auto I : UnreachableIdxs) |
| 353 | if (UnreachableProb < BP[I]) { |
| 354 | BP[I] = UnreachableProb; |
| 355 | } |
| 356 | |
| 357 | // Sum of all edge probabilities must be 1.0. If we modified the probability |
| 358 | // of some edges then we must distribute the introduced difference over the |
| 359 | // reachable blocks. |
| 360 | // |
| 361 | // Proportional distribution: the relation between probabilities of the |
| 362 | // reachable edges is kept unchanged. That is for any reachable edges i and j: |
| 363 | // newBP[i] / newBP[j] == oldBP[i] / oldBP[j] => |
| 364 | // newBP[i] / oldBP[i] == newBP[j] / oldBP[j] == K |
| 365 | // Where K is independent of i,j. |
| 366 | // newBP[i] == oldBP[i] * K |
| 367 | // We need to find K. |
| 368 | // Make sum of all reachables of the left and right parts: |
| 369 | // sum_of_reachable(newBP) == K * sum_of_reachable(oldBP) |
| 370 | // Sum of newBP must be equal to 1.0: |
| 371 | // sum_of_reachable(newBP) + sum_of_unreachable(newBP) == 1.0 => |
| 372 | // sum_of_reachable(newBP) = 1.0 - sum_of_unreachable(newBP) |
| 373 | // Where sum_of_unreachable(newBP) is what has been just changed. |
| 374 | // Finally: |
| 375 | // K == sum_of_reachable(newBP) / sum_of_reachable(oldBP) => |
| 376 | // K == (1.0 - sum_of_unreachable(newBP)) / sum_of_reachable(oldBP) |
| 377 | BranchProbability NewUnreachableSum = BranchProbability::getZero(); |
| 378 | for (auto I : UnreachableIdxs) |
| 379 | NewUnreachableSum += BP[I]; |
| 380 | |
| 381 | BranchProbability NewReachableSum = |
| 382 | BranchProbability::getOne() - NewUnreachableSum; |
| 383 | |
| 384 | BranchProbability OldReachableSum = BranchProbability::getZero(); |
| 385 | for (auto I : ReachableIdxs) |
| 386 | OldReachableSum += BP[I]; |
| 387 | |
| 388 | if (OldReachableSum != NewReachableSum) { // Anything to dsitribute? |
| 389 | if (OldReachableSum.isZero()) { |
| 390 | // If all oldBP[i] are zeroes then the proportional distribution results |
| 391 | // in all zero probabilities and the error stays big. In this case we |
| 392 | // evenly spread NewReachableSum over the reachable edges. |
| 393 | BranchProbability PerEdge = NewReachableSum / ReachableIdxs.size(); |
| 394 | for (auto I : ReachableIdxs) |
| 395 | BP[I] = PerEdge; |
| 396 | } else { |
| 397 | for (auto I : ReachableIdxs) { |
| 398 | // We use uint64_t to avoid double rounding error of the following |
| 399 | // calculation: BP[i] = BP[i] * NewReachableSum / OldReachableSum |
| 400 | // The formula is taken from the private constructor |
| 401 | // BranchProbability(uint32_t Numerator, uint32_t Denominator) |
| 402 | uint64_t Mul = static_cast<uint64_t>(NewReachableSum.getNumerator()) * |
| 403 | BP[I].getNumerator(); |
| 404 | uint32_t Div = static_cast<uint32_t>( |
| 405 | divideNearest(Numerator: Mul, Denominator: OldReachableSum.getNumerator())); |
| 406 | BP[I] = BranchProbability::getRaw(N: Div); |
| 407 | } |
| 408 | } |
| 409 | } |
| 410 | |
| 411 | BPI.setEdgeProbability(Src: BB, Probs: BP); |
| 412 | |
| 413 | return true; |
| 414 | } |
| 415 | |
| 416 | // Calculate Edge Weights using "Pointer Heuristics". Predict a comparison |
| 417 | // between two pointer or pointer and NULL will fail. |
| 418 | bool BPIConstruction::calcPointerHeuristics(const BasicBlock *BB) { |
| 419 | const CondBrInst *BI = dyn_cast<CondBrInst>(Val: BB->getTerminator()); |
| 420 | if (!BI) |
| 421 | return false; |
| 422 | |
| 423 | Value *Cond = BI->getCondition(); |
| 424 | ICmpInst *CI = dyn_cast<ICmpInst>(Val: Cond); |
| 425 | if (!CI || !CI->isEquality()) |
| 426 | return false; |
| 427 | |
| 428 | Value *LHS = CI->getOperand(i_nocapture: 0); |
| 429 | |
| 430 | if (!LHS->getType()->isPointerTy()) |
| 431 | return false; |
| 432 | |
| 433 | assert(CI->getOperand(1)->getType()->isPointerTy()); |
| 434 | |
| 435 | switch (CI->getPredicate()) { |
| 436 | case ICmpInst::ICMP_NE: // p != q -> Likely |
| 437 | BPI.setEdgeProbability(Src: BB, Probs: {PtrTakenProb, PtrUntakenProb}); |
| 438 | return true; |
| 439 | case ICmpInst::ICMP_EQ: // p == q -> Unlikely |
| 440 | BPI.setEdgeProbability(Src: BB, Probs: {PtrUntakenProb, PtrTakenProb}); |
| 441 | return true; |
| 442 | default: |
| 443 | return false; |
| 444 | } |
| 445 | } |
| 446 | |
| 447 | // Compute the unlikely successors to the block BB in the cycle C, specifically |
| 448 | // those that are unlikely because this is a loop, and add them to the |
| 449 | // UnlikelyBlocks set. |
| 450 | static void |
| 451 | computeUnlikelySuccessors(const BasicBlock *BB, const CycleInfo &CI, CycleRef C, |
| 452 | SmallPtrSetImpl<const BasicBlock *> &UnlikelyBlocks) { |
| 453 | // Sometimes in a loop we have a branch whose condition is made false by |
| 454 | // taking it. This is typically something like |
| 455 | // int n = 0; |
| 456 | // while (...) { |
| 457 | // if (++n >= MAX) { |
| 458 | // n = 0; |
| 459 | // } |
| 460 | // } |
| 461 | // In this sort of situation taking the branch means that at the very least it |
| 462 | // won't be taken again in the next iteration of the loop, so we should |
| 463 | // consider it less likely than a typical branch. |
| 464 | // |
| 465 | // We detect this by looking back through the graph of PHI nodes that sets the |
| 466 | // value that the condition depends on, and seeing if we can reach a successor |
| 467 | // block which can be determined to make the condition false. |
| 468 | // |
| 469 | // FIXME: We currently consider unlikely blocks to be half as likely as other |
| 470 | // blocks, but if we consider the example above the likelyhood is actually |
| 471 | // 1/MAX. We could therefore be more precise in how unlikely we consider |
| 472 | // blocks to be, but it would require more careful examination of the form |
| 473 | // of the comparison expression. |
| 474 | const CondBrInst *BI = dyn_cast<CondBrInst>(Val: BB->getTerminator()); |
| 475 | if (!BI) |
| 476 | return; |
| 477 | |
| 478 | // Check if the branch is based on an instruction compared with a constant |
| 479 | CmpInst *Cmp = dyn_cast<CmpInst>(Val: BI->getCondition()); |
| 480 | if (!Cmp || !isa<Instruction>(Val: Cmp->getOperand(i_nocapture: 0)) || |
| 481 | !isa<Constant>(Val: Cmp->getOperand(i_nocapture: 1))) |
| 482 | return; |
| 483 | |
| 484 | // Either the instruction must be a PHI, or a chain of operations involving |
| 485 | // constants that ends in a PHI which we can then collapse into a single value |
| 486 | // if the PHI value is known. |
| 487 | Instruction *CmpLHS = dyn_cast<Instruction>(Val: Cmp->getOperand(i_nocapture: 0)); |
| 488 | PHINode *CmpPHI = dyn_cast<PHINode>(Val: CmpLHS); |
| 489 | Constant *CmpConst = dyn_cast<Constant>(Val: Cmp->getOperand(i_nocapture: 1)); |
| 490 | // Collect the instructions until we hit a PHI |
| 491 | SmallVector<BinaryOperator *, 1> InstChain; |
| 492 | while (!CmpPHI && CmpLHS && isa<BinaryOperator>(Val: CmpLHS) && |
| 493 | isa<Constant>(Val: CmpLHS->getOperand(i: 1))) { |
| 494 | // Stop if the chain extends outside of the loop |
| 495 | if (!CI.contains(C, Block: CmpLHS->getParent())) |
| 496 | return; |
| 497 | InstChain.push_back(Elt: cast<BinaryOperator>(Val: CmpLHS)); |
| 498 | CmpLHS = dyn_cast<Instruction>(Val: CmpLHS->getOperand(i: 0)); |
| 499 | if (CmpLHS) |
| 500 | CmpPHI = dyn_cast<PHINode>(Val: CmpLHS); |
| 501 | } |
| 502 | if (!CmpPHI || !CI.contains(C, Block: CmpPHI->getParent())) |
| 503 | return; |
| 504 | |
| 505 | // Trace the phi node to find all values that come from successors of BB |
| 506 | SmallPtrSet<PHINode*, 8> VisitedInsts; |
| 507 | SmallVector<PHINode*, 8> WorkList; |
| 508 | WorkList.push_back(Elt: CmpPHI); |
| 509 | VisitedInsts.insert(Ptr: CmpPHI); |
| 510 | while (!WorkList.empty()) { |
| 511 | PHINode *P = WorkList.pop_back_val(); |
| 512 | for (BasicBlock *B : P->blocks()) { |
| 513 | // Skip blocks that aren't part of the loop |
| 514 | if (!CI.contains(C, Block: B)) |
| 515 | continue; |
| 516 | Value *V = P->getIncomingValueForBlock(BB: B); |
| 517 | // If the source is a PHI add it to the work list if we haven't |
| 518 | // already visited it. |
| 519 | if (PHINode *PN = dyn_cast<PHINode>(Val: V)) { |
| 520 | if (VisitedInsts.insert(Ptr: PN).second) |
| 521 | WorkList.push_back(Elt: PN); |
| 522 | continue; |
| 523 | } |
| 524 | // If this incoming value is a constant and B is a successor of BB, then |
| 525 | // we can constant-evaluate the compare to see if it makes the branch be |
| 526 | // taken or not. |
| 527 | Constant *CmpLHSConst = dyn_cast<Constant>(Val: V); |
| 528 | if (!CmpLHSConst || !llvm::is_contained(Range: successors(BB), Element: B)) |
| 529 | continue; |
| 530 | // First collapse InstChain |
| 531 | const DataLayout &DL = BB->getDataLayout(); |
| 532 | for (Instruction *I : llvm::reverse(C&: InstChain)) { |
| 533 | CmpLHSConst = ConstantFoldBinaryOpOperands( |
| 534 | Opcode: I->getOpcode(), LHS: CmpLHSConst, RHS: cast<Constant>(Val: I->getOperand(i: 1)), DL); |
| 535 | if (!CmpLHSConst) |
| 536 | break; |
| 537 | } |
| 538 | if (!CmpLHSConst) |
| 539 | continue; |
| 540 | // Now constant-evaluate the compare |
| 541 | Constant *Result = ConstantFoldCompareInstOperands( |
| 542 | Predicate: Cmp->getPredicate(), LHS: CmpLHSConst, RHS: CmpConst, DL); |
| 543 | // If the result means we don't branch to the block then that block is |
| 544 | // unlikely. |
| 545 | if (Result && ((Result->isNullValue() && B == BI->getSuccessor(i: 0)) || |
| 546 | (Result->isOneValue() && B == BI->getSuccessor(i: 1)))) |
| 547 | UnlikelyBlocks.insert(Ptr: B); |
| 548 | } |
| 549 | } |
| 550 | } |
| 551 | |
| 552 | std::optional<uint32_t> |
| 553 | BPIConstruction::getEstimatedBlockWeight(const BasicBlock *BB) const { |
| 554 | auto WeightIt = EstimatedBlockWeight.find(Val: BB); |
| 555 | if (WeightIt == EstimatedBlockWeight.end()) |
| 556 | return std::nullopt; |
| 557 | return WeightIt->second; |
| 558 | } |
| 559 | |
| 560 | std::optional<uint32_t> |
| 561 | BPIConstruction::getEstimatedLoopWeight(CycleRef C) const { |
| 562 | auto WeightIt = EstimatedLoopWeight.find(Val: C); |
| 563 | if (WeightIt == EstimatedLoopWeight.end()) |
| 564 | return std::nullopt; |
| 565 | return WeightIt->second; |
| 566 | } |
| 567 | |
| 568 | std::optional<uint32_t> |
| 569 | BPIConstruction::getEstimatedEdgeWeight(const LoopEdge &Edge) const { |
| 570 | // For edges entering a loop take weight of a loop rather than an individual |
| 571 | // block in the loop. |
| 572 | return isLoopEnteringEdge(Edge) |
| 573 | ? getEstimatedLoopWeight(C: CI->getCycle(Block: Edge.second)) |
| 574 | : getEstimatedBlockWeight(BB: Edge.second); |
| 575 | } |
| 576 | |
| 577 | template <class IterT> |
| 578 | std::optional<uint32_t> BPIConstruction::getMaxEstimatedEdgeWeight( |
| 579 | const BasicBlock *SrcBB, iterator_range<IterT> Successors) const { |
| 580 | std::optional<uint32_t> MaxWeight; |
| 581 | for (const BasicBlock *DstBB : Successors) { |
| 582 | auto Weight = getEstimatedEdgeWeight(Edge: {SrcBB, DstBB}); |
| 583 | if (!Weight) |
| 584 | return std::nullopt; |
| 585 | if (!MaxWeight || *MaxWeight < *Weight) |
| 586 | MaxWeight = Weight; |
| 587 | } |
| 588 | |
| 589 | return MaxWeight; |
| 590 | } |
| 591 | |
| 592 | // Updates \p LoopBB's weight and returns true. If \p LoopBB has already |
| 593 | // an associated weight it is unchanged and false is returned. |
| 594 | // |
| 595 | // Please note by the algorithm the weight is not expected to change once set |
| 596 | // thus 'false' status is used to track visited blocks. |
| 597 | bool BPIConstruction::updateEstimatedBlockWeight( |
| 598 | const BasicBlock *BB, uint32_t BBWeight, |
| 599 | SmallVectorImpl<const BasicBlock *> &BlockWorkList, |
| 600 | SmallVectorImpl<const BasicBlock *> &LoopWorkList) { |
| 601 | // In general, weight is assigned to a block when it has final value and |
| 602 | // can't/shouldn't be changed. However, there are cases when a block |
| 603 | // inherently has several (possibly "contradicting") weights. For example, |
| 604 | // "unwind" block may also contain "cold" call. In that case the first |
| 605 | // set weight is favored and all consequent weights are ignored. |
| 606 | if (!EstimatedBlockWeight.insert(KV: {BB, BBWeight}).second) |
| 607 | return false; |
| 608 | |
| 609 | for (const BasicBlock *PredBlock : predecessors(BB)) { |
| 610 | // Add affected block/loop to a working list. |
| 611 | if (isLoopExitingEdge(Edge: {PredBlock, BB})) { |
| 612 | if (!EstimatedLoopWeight.count(Val: CI->getCycle(Block: PredBlock))) |
| 613 | LoopWorkList.push_back(Elt: PredBlock); |
| 614 | } else if (!EstimatedBlockWeight.count(Val: PredBlock)) |
| 615 | BlockWorkList.push_back(Elt: PredBlock); |
| 616 | } |
| 617 | return true; |
| 618 | } |
| 619 | |
| 620 | // Starting from \p BB traverse through dominator blocks and assign \p BBWeight |
| 621 | // to all such blocks that are post dominated by \BB. In other words to all |
| 622 | // blocks that the one is executed if and only if another one is executed. |
| 623 | // Importantly, we skip loops here for two reasons. First weights of blocks in |
| 624 | // a loop should be scaled by trip count (yet possibly unknown). Second there is |
| 625 | // no any value in doing that because that doesn't give any additional |
| 626 | // information regarding distribution of probabilities inside the loop. |
| 627 | // Exception is loop 'enter' and 'exit' edges that are handled in a special way |
| 628 | // at calcEstimatedHeuristics. |
| 629 | // |
| 630 | // In addition, \p WorkList is populated with basic blocks if at leas one |
| 631 | // successor has updated estimated weight. |
| 632 | void BPIConstruction::propagateEstimatedBlockWeight( |
| 633 | const BasicBlock *BB, DominatorTree *DT, PostDominatorTree *PDT, |
| 634 | uint32_t BBWeight, SmallVectorImpl<const BasicBlock *> &BlockWorkList, |
| 635 | SmallVectorImpl<const BasicBlock *> &LoopWorkList) { |
| 636 | const auto *DTStartNode = DT->getNode(BB); |
| 637 | const auto *PDTStartNode = PDT->getNode(BB); |
| 638 | |
| 639 | // TODO: Consider propagating weight down the domination line as well. |
| 640 | for (const auto *DTNode = DTStartNode; DTNode != nullptr; |
| 641 | DTNode = DTNode->getIDom()) { |
| 642 | auto *DomBB = DTNode->getBlock(); |
| 643 | // Consider blocks which lie on one 'line'. |
| 644 | if (!PDT->dominates(A: PDTStartNode, B: PDT->getNode(BB: DomBB))) |
| 645 | // If BB doesn't post dominate DomBB it will not post dominate dominators |
| 646 | // of DomBB as well. |
| 647 | break; |
| 648 | |
| 649 | const LoopEdge Edge{DomBB, BB}; |
| 650 | // Don't propagate weight to blocks belonging to different loops. |
| 651 | if (!isLoopEnteringExitingEdge(Edge)) { |
| 652 | if (!updateEstimatedBlockWeight(BB: DomBB, BBWeight, BlockWorkList, |
| 653 | LoopWorkList)) |
| 654 | // If DomBB has weight set then all it's predecessors are already |
| 655 | // processed (since we propagate weight up to the top of IR each time). |
| 656 | break; |
| 657 | } else if (isLoopExitingEdge(Edge)) { |
| 658 | LoopWorkList.push_back(Elt: DomBB); |
| 659 | } |
| 660 | } |
| 661 | } |
| 662 | |
| 663 | std::optional<uint32_t> |
| 664 | BPIConstruction::getInitialEstimatedBlockWeight(const BasicBlock *BB) { |
| 665 | // Returns true if \p BB has call marked with "NoReturn" attribute. |
| 666 | auto hasNoReturn = [&](const BasicBlock *BB) { |
| 667 | for (const auto &I : reverse(C: *BB)) |
| 668 | if (const CallInst *CI = dyn_cast<CallInst>(Val: &I)) |
| 669 | if (CI->hasFnAttr(Kind: Attribute::NoReturn)) |
| 670 | return true; |
| 671 | |
| 672 | return false; |
| 673 | }; |
| 674 | |
| 675 | // Important note regarding the order of checks. They are ordered by weight |
| 676 | // from lowest to highest. Doing that allows to avoid "unstable" results |
| 677 | // when several conditions heuristics can be applied simultaneously. |
| 678 | if (isa<UnreachableInst>(Val: BB->getTerminator()) || |
| 679 | // If this block is terminated by a call to |
| 680 | // @llvm.experimental.deoptimize then treat it like an unreachable |
| 681 | // since it is expected to practically never execute. |
| 682 | // TODO: Should we actually treat as never returning call? |
| 683 | BB->getTerminatingDeoptimizeCall()) |
| 684 | return hasNoReturn(BB) |
| 685 | ? static_cast<uint32_t>(BlockExecWeight::NORETURN) |
| 686 | : static_cast<uint32_t>(BlockExecWeight::UNREACHABLE); |
| 687 | |
| 688 | // Check if the block is an exception handling block. |
| 689 | if (BB->isEHPad()) |
| 690 | return static_cast<uint32_t>(BlockExecWeight::UNWIND); |
| 691 | |
| 692 | // Check if the block contains 'cold' call. |
| 693 | for (const auto &I : *BB) |
| 694 | if (const CallInst *CI = dyn_cast<CallInst>(Val: &I)) |
| 695 | if (CI->hasFnAttr(Kind: Attribute::Cold)) |
| 696 | return static_cast<uint32_t>(BlockExecWeight::COLD); |
| 697 | |
| 698 | return std::nullopt; |
| 699 | } |
| 700 | |
| 701 | // Does RPO traversal over all blocks in \p F and assigns weights to |
| 702 | // 'unreachable', 'noreturn', 'cold', 'unwind' blocks. In addition it does its |
| 703 | // best to propagate the weight to up/down the IR. |
| 704 | void BPIConstruction::estimateBlockWeights(const Function &F, DominatorTree *DT, |
| 705 | PostDominatorTree *PDT) { |
| 706 | SmallVector<const BasicBlock *, 8> BlockWorkList; |
| 707 | SmallVector<const BasicBlock *, 8> LoopWorkList; |
| 708 | SmallDenseMap<CycleRef, SmallVector<BasicBlock *, 4>> LoopExitBlocks; |
| 709 | |
| 710 | // By doing RPO we make sure that all predecessors already have weights |
| 711 | // calculated before visiting theirs successors. |
| 712 | ReversePostOrderTraversal<const Function *> RPOT(&F); |
| 713 | for (const auto *BB : RPOT) |
| 714 | if (auto BBWeight = getInitialEstimatedBlockWeight(BB)) |
| 715 | // If we were able to find estimated weight for the block set it to this |
| 716 | // block and propagate up the IR. |
| 717 | propagateEstimatedBlockWeight(BB, DT, PDT, BBWeight: *BBWeight, BlockWorkList, |
| 718 | LoopWorkList); |
| 719 | |
| 720 | // BlockWorklist/LoopWorkList contains blocks/loops with at least one |
| 721 | // successor/exit having estimated weight. Try to propagate weight to such |
| 722 | // blocks/loops from successors/exits. |
| 723 | // Process loops and blocks. Order is not important. |
| 724 | do { |
| 725 | while (!LoopWorkList.empty()) { |
| 726 | const BasicBlock *LoopBB = LoopWorkList.pop_back_val(); |
| 727 | CycleRef C = CI->getCycle(Block: LoopBB); |
| 728 | if (EstimatedLoopWeight.count(Val: C)) |
| 729 | continue; |
| 730 | |
| 731 | auto Res = LoopExitBlocks.try_emplace(Key: C); |
| 732 | SmallVectorImpl<BasicBlock *> &Exits = Res.first->second; |
| 733 | if (Res.second) |
| 734 | CI->getExitBlocks(C, TmpStorage&: Exits); |
| 735 | auto LoopWeight = getMaxEstimatedEdgeWeight( |
| 736 | SrcBB: LoopBB, Successors: make_range(x: Exits.begin(), y: Exits.end())); |
| 737 | |
| 738 | if (LoopWeight) { |
| 739 | // If we never exit the loop then we can enter it once at maximum. |
| 740 | if (LoopWeight <= static_cast<uint32_t>(BlockExecWeight::UNREACHABLE)) |
| 741 | LoopWeight = static_cast<uint32_t>(BlockExecWeight::LOWEST_NON_ZERO); |
| 742 | |
| 743 | EstimatedLoopWeight.insert(KV: {C, *LoopWeight}); |
| 744 | // Add all blocks entering the loop into working list. |
| 745 | getLoopEnterBlocks(BB: LoopBB, Enters&: BlockWorkList); |
| 746 | } |
| 747 | } |
| 748 | |
| 749 | while (!BlockWorkList.empty()) { |
| 750 | // We can reach here only if BlockWorkList is not empty. |
| 751 | const BasicBlock *BB = BlockWorkList.pop_back_val(); |
| 752 | if (EstimatedBlockWeight.count(Val: BB)) |
| 753 | continue; |
| 754 | |
| 755 | // We take maximum over all weights of successors. In other words we take |
| 756 | // weight of "hot" path. In theory we can probably find a better function |
| 757 | // which gives higher accuracy results (comparing to "maximum") but I |
| 758 | // can't |
| 759 | // think of any right now. And I doubt it will make any difference in |
| 760 | // practice. |
| 761 | auto MaxWeight = getMaxEstimatedEdgeWeight(SrcBB: BB, Successors: successors(BB)); |
| 762 | |
| 763 | if (MaxWeight) |
| 764 | propagateEstimatedBlockWeight(BB, DT, PDT, BBWeight: *MaxWeight, BlockWorkList, |
| 765 | LoopWorkList); |
| 766 | } |
| 767 | } while (!BlockWorkList.empty() || !LoopWorkList.empty()); |
| 768 | } |
| 769 | |
| 770 | // Calculate edge probabilities based on block's estimated weight. |
| 771 | // Note that gathered weights were not scaled for loops. Thus edges entering |
| 772 | // and exiting loops requires special processing. |
| 773 | bool BPIConstruction::calcEstimatedHeuristics(const BasicBlock *BB) { |
| 774 | assert(BB->getTerminator()->getNumSuccessors() > 1 && |
| 775 | "expected more than one successor!" ); |
| 776 | |
| 777 | CycleRef BBCycle = CI->getCycle(Block: BB); |
| 778 | |
| 779 | SmallPtrSet<const BasicBlock *, 8> UnlikelyBlocks; |
| 780 | uint32_t TC = LBH_TAKEN_WEIGHT / LBH_NONTAKEN_WEIGHT; |
| 781 | if (BBCycle) |
| 782 | computeUnlikelySuccessors(BB, CI: *CI, C: BBCycle, UnlikelyBlocks); |
| 783 | |
| 784 | // Changed to 'true' if at least one successor has estimated weight. |
| 785 | bool FoundEstimatedWeight = false; |
| 786 | SmallVector<uint32_t, 4> SuccWeights; |
| 787 | uint64_t TotalWeight = 0; |
| 788 | // Go over all successors of BB and put their weights into SuccWeights. |
| 789 | for (const BasicBlock *SuccBB : successors(BB)) { |
| 790 | std::optional<uint32_t> Weight; |
| 791 | const LoopEdge Edge{BB, SuccBB}; |
| 792 | |
| 793 | Weight = getEstimatedEdgeWeight(Edge); |
| 794 | |
| 795 | if (isLoopExitingEdge(Edge) && |
| 796 | // Avoid adjustment of ZERO weight since it should remain unchanged. |
| 797 | Weight != static_cast<uint32_t>(BlockExecWeight::ZERO)) { |
| 798 | // Scale down loop exiting weight by trip count. |
| 799 | Weight = std::max( |
| 800 | a: static_cast<uint32_t>(BlockExecWeight::LOWEST_NON_ZERO), |
| 801 | b: Weight.value_or(u: static_cast<uint32_t>(BlockExecWeight::DEFAULT)) / |
| 802 | TC); |
| 803 | } |
| 804 | bool IsUnlikelyEdge = BBCycle && UnlikelyBlocks.contains(Ptr: SuccBB); |
| 805 | if (IsUnlikelyEdge && |
| 806 | // Avoid adjustment of ZERO weight since it should remain unchanged. |
| 807 | Weight != static_cast<uint32_t>(BlockExecWeight::ZERO)) { |
| 808 | // 'Unlikely' blocks have twice lower weight. |
| 809 | Weight = std::max( |
| 810 | a: static_cast<uint32_t>(BlockExecWeight::LOWEST_NON_ZERO), |
| 811 | b: Weight.value_or(u: static_cast<uint32_t>(BlockExecWeight::DEFAULT)) / 2); |
| 812 | } |
| 813 | |
| 814 | if (Weight) |
| 815 | FoundEstimatedWeight = true; |
| 816 | |
| 817 | auto WeightVal = |
| 818 | Weight.value_or(u: static_cast<uint32_t>(BlockExecWeight::DEFAULT)); |
| 819 | TotalWeight += WeightVal; |
| 820 | SuccWeights.push_back(Elt: WeightVal); |
| 821 | } |
| 822 | |
| 823 | // If non of blocks have estimated weight bail out. |
| 824 | // If TotalWeight is 0 that means weight of each successor is 0 as well and |
| 825 | // equally likely. Bail out early to not deal with devision by zero. |
| 826 | if (!FoundEstimatedWeight || TotalWeight == 0) |
| 827 | return false; |
| 828 | |
| 829 | assert(SuccWeights.size() == succ_size(BB) && "Missed successor?" ); |
| 830 | const unsigned SuccCount = SuccWeights.size(); |
| 831 | |
| 832 | // If the sum of weights does not fit in 32 bits, scale every weight down |
| 833 | // accordingly. |
| 834 | if (TotalWeight > UINT32_MAX) { |
| 835 | uint64_t ScalingFactor = TotalWeight / UINT32_MAX + 1; |
| 836 | TotalWeight = 0; |
| 837 | for (unsigned Idx = 0; Idx < SuccCount; ++Idx) { |
| 838 | SuccWeights[Idx] /= ScalingFactor; |
| 839 | if (SuccWeights[Idx] == static_cast<uint32_t>(BlockExecWeight::ZERO)) |
| 840 | SuccWeights[Idx] = |
| 841 | static_cast<uint32_t>(BlockExecWeight::LOWEST_NON_ZERO); |
| 842 | TotalWeight += SuccWeights[Idx]; |
| 843 | } |
| 844 | assert(TotalWeight <= UINT32_MAX && "Total weight overflows" ); |
| 845 | } |
| 846 | |
| 847 | // Finally set probabilities to edges according to estimated block weights. |
| 848 | SmallVector<BranchProbability, 4> EdgeProbabilities( |
| 849 | SuccCount, BranchProbability::getUnknown()); |
| 850 | |
| 851 | for (unsigned Idx = 0; Idx < SuccCount; ++Idx) { |
| 852 | EdgeProbabilities[Idx] = |
| 853 | BranchProbability(SuccWeights[Idx], (uint32_t)TotalWeight); |
| 854 | } |
| 855 | BPI.setEdgeProbability(Src: BB, Probs: EdgeProbabilities); |
| 856 | return true; |
| 857 | } |
| 858 | |
| 859 | bool BPIConstruction::calcZeroHeuristics(const BasicBlock *BB, |
| 860 | const TargetLibraryInfo *TLI) { |
| 861 | const CondBrInst *BI = dyn_cast<CondBrInst>(Val: BB->getTerminator()); |
| 862 | if (!BI) |
| 863 | return false; |
| 864 | |
| 865 | Value *Cond = BI->getCondition(); |
| 866 | ICmpInst *CI = dyn_cast<ICmpInst>(Val: Cond); |
| 867 | if (!CI) |
| 868 | return false; |
| 869 | |
| 870 | auto GetConstantInt = [](Value *V) { |
| 871 | if (auto *I = dyn_cast<BitCastInst>(Val: V)) |
| 872 | return dyn_cast<ConstantInt>(Val: I->getOperand(i_nocapture: 0)); |
| 873 | return dyn_cast<ConstantInt>(Val: V); |
| 874 | }; |
| 875 | |
| 876 | Value *RHS = CI->getOperand(i_nocapture: 1); |
| 877 | ConstantInt *CV = GetConstantInt(RHS); |
| 878 | if (!CV) |
| 879 | return false; |
| 880 | |
| 881 | // If the LHS is the result of AND'ing a value with a single bit bitmask, |
| 882 | // we don't have information about probabilities. |
| 883 | if (Instruction *LHS = dyn_cast<Instruction>(Val: CI->getOperand(i_nocapture: 0))) |
| 884 | if (LHS->getOpcode() == Instruction::And) |
| 885 | if (ConstantInt *AndRHS = GetConstantInt(LHS->getOperand(i: 1))) |
| 886 | if (AndRHS->getValue().isPowerOf2()) |
| 887 | return false; |
| 888 | |
| 889 | // Check if the LHS is the return value of a library function |
| 890 | LibFunc Func = LibFunc::NotLibFunc; |
| 891 | if (TLI) |
| 892 | if (CallInst *Call = dyn_cast<CallInst>(Val: CI->getOperand(i_nocapture: 0))) |
| 893 | if (Function *CalledFn = Call->getCalledFunction()) |
| 894 | Func = TLI->getLibFunc(FDecl: *CalledFn); |
| 895 | |
| 896 | bool Likely; |
| 897 | if (Func == LibFunc_strcasecmp || |
| 898 | Func == LibFunc_strcmp || |
| 899 | Func == LibFunc_strncasecmp || |
| 900 | Func == LibFunc_strncmp || |
| 901 | Func == LibFunc_memcmp || |
| 902 | Func == LibFunc_bcmp) { |
| 903 | /// strcmp and similar functions return zero, negative, or positive, if the |
| 904 | /// first string is equal, less, or greater than the second. We consider it |
| 905 | /// likely that the strings are not equal, so a comparison with zero is |
| 906 | /// probably false, but also a comparison with any other number is also |
| 907 | /// probably false given that what exactly is returned for nonzero values is |
| 908 | /// not specified. Any kind of comparison other than equality we know |
| 909 | /// nothing about. |
| 910 | // clang-format off |
| 911 | switch (CI->getPredicate()) { |
| 912 | case CmpInst::ICMP_EQ: Likely = false; break; |
| 913 | case CmpInst::ICMP_NE: Likely = true; break; |
| 914 | default: return false; |
| 915 | } |
| 916 | // clang-format on |
| 917 | } else if (CV->isZero()) { |
| 918 | // clang-format off |
| 919 | switch (CI->getPredicate()) { |
| 920 | case CmpInst::ICMP_EQ: Likely = false; break; |
| 921 | case CmpInst::ICMP_NE: Likely = true; break; |
| 922 | case CmpInst::ICMP_SLT: Likely = false; break; |
| 923 | case CmpInst::ICMP_SGT: Likely = true; break; |
| 924 | default: return false; |
| 925 | } |
| 926 | // clang-format on |
| 927 | } else if (CV->isOne()) { |
| 928 | // clang-format off |
| 929 | switch (CI->getPredicate()) { |
| 930 | case CmpInst::ICMP_SLT: Likely = false; break; |
| 931 | default: return false; |
| 932 | } |
| 933 | // clang-format on |
| 934 | } else if (CV->isMinusOne()) { |
| 935 | // clang-format off |
| 936 | switch (CI->getPredicate()) { |
| 937 | case CmpInst::ICMP_EQ: Likely = false; break; |
| 938 | case CmpInst::ICMP_NE: Likely = true; break; |
| 939 | // InstCombine canonicalizes X >= 0 into X > -1 |
| 940 | case CmpInst::ICMP_SGT: Likely = true; break; |
| 941 | default: return false; |
| 942 | } |
| 943 | // clang-format on |
| 944 | } else { |
| 945 | return false; |
| 946 | } |
| 947 | |
| 948 | if (Likely) |
| 949 | BPI.setEdgeProbability(Src: BB, Probs: {ZeroTakenProb, ZeroUntakenProb}); |
| 950 | else |
| 951 | BPI.setEdgeProbability(Src: BB, Probs: {ZeroUntakenProb, ZeroTakenProb}); |
| 952 | return true; |
| 953 | } |
| 954 | |
| 955 | bool BPIConstruction::calcFloatingPointHeuristics(const BasicBlock *BB) { |
| 956 | const CondBrInst *BI = dyn_cast<CondBrInst>(Val: BB->getTerminator()); |
| 957 | if (!BI) |
| 958 | return false; |
| 959 | |
| 960 | Value *Cond = BI->getCondition(); |
| 961 | FCmpInst *FCmp = dyn_cast<FCmpInst>(Val: Cond); |
| 962 | if (!FCmp) |
| 963 | return false; |
| 964 | |
| 965 | if (FCmp->isEquality()) { |
| 966 | if (!FCmp->isTrueWhenEqual()) // f1 == f2 -> Unlikely |
| 967 | BPI.setEdgeProbability(Src: BB, Probs: {FPTakenProb, FPUntakenProb}); |
| 968 | else // f1 != f2 -> Likely |
| 969 | BPI.setEdgeProbability(Src: BB, Probs: {FPUntakenProb, FPTakenProb}); |
| 970 | } else if (FCmp->getPredicate() == FCmpInst::FCMP_ORD) { |
| 971 | BPI.setEdgeProbability( |
| 972 | Src: BB, Probs: {FPOrdTakenProb, FPOrdUntakenProb}); // !isnan -> Likely |
| 973 | } else if (FCmp->getPredicate() == FCmpInst::FCMP_UNO) { |
| 974 | BPI.setEdgeProbability( |
| 975 | Src: BB, Probs: {FPOrdUntakenProb, FPOrdTakenProb}); // isnan -> Unlikely |
| 976 | } else { |
| 977 | return false; |
| 978 | } |
| 979 | return true; |
| 980 | } |
| 981 | void BPIConstruction::calculate(const Function &F, const CycleInfo &CycleI, |
| 982 | const TargetLibraryInfo *TLI, DominatorTree *DT, |
| 983 | PostDominatorTree *PDT) { |
| 984 | CI = &CycleI; |
| 985 | |
| 986 | std::unique_ptr<DominatorTree> DTPtr; |
| 987 | std::unique_ptr<PostDominatorTree> PDTPtr; |
| 988 | |
| 989 | if (!DT) { |
| 990 | DTPtr = std::make_unique<DominatorTree>(args&: const_cast<Function &>(F)); |
| 991 | DT = DTPtr.get(); |
| 992 | } |
| 993 | |
| 994 | if (!PDT) { |
| 995 | PDTPtr = std::make_unique<PostDominatorTree>(args&: const_cast<Function &>(F)); |
| 996 | PDT = PDTPtr.get(); |
| 997 | } |
| 998 | |
| 999 | estimateBlockWeights(F, DT, PDT); |
| 1000 | |
| 1001 | // Walk the basic blocks in post-order so that we can build up state about |
| 1002 | // the successors of a block iteratively. |
| 1003 | for (const auto *BB : post_order(G: &F.getEntryBlock())) { |
| 1004 | LLVM_DEBUG(dbgs() << "Computing probabilities for " << BB->getName() |
| 1005 | << "\n" ); |
| 1006 | // If there is no at least two successors, no sense to set probability. |
| 1007 | if (BB->getTerminator()->getNumSuccessors() < 2) |
| 1008 | continue; |
| 1009 | if (calcMetadataWeights(BB)) |
| 1010 | continue; |
| 1011 | if (calcEstimatedHeuristics(BB)) |
| 1012 | continue; |
| 1013 | if (calcPointerHeuristics(BB)) |
| 1014 | continue; |
| 1015 | if (calcZeroHeuristics(BB, TLI)) |
| 1016 | continue; |
| 1017 | if (calcFloatingPointHeuristics(BB)) |
| 1018 | continue; |
| 1019 | } |
| 1020 | } |
| 1021 | |
| 1022 | } // end anonymous namespace |
| 1023 | |
| 1024 | MutableArrayRef<BranchProbability> |
| 1025 | BranchProbabilityInfo::allocEdges(const BasicBlock *BB) { |
| 1026 | assert(BB->getParent() == LastF); |
| 1027 | assert(BlockNumberEpoch == LastF->getBlockNumberEpoch()); |
| 1028 | unsigned NumSuccs = succ_size(BB); |
| 1029 | if (NumSuccs == 0) { |
| 1030 | eraseBlock(BB); |
| 1031 | return {}; |
| 1032 | } |
| 1033 | if (EdgeStarts.size() <= BB->getNumber()) |
| 1034 | EdgeStarts.resize(N: LastF->getMaxBlockNumber(), NV: 0); |
| 1035 | unsigned EdgeStart = Probs.size(); |
| 1036 | EdgeStarts[BB->getNumber()] = EdgeStart + 1; // 0 = no edges. |
| 1037 | Probs.append(NumInputs: NumSuccs, Elt: {}); |
| 1038 | return MutableArrayRef(&Probs[EdgeStart], NumSuccs); |
| 1039 | } |
| 1040 | |
| 1041 | ArrayRef<BranchProbability> |
| 1042 | BranchProbabilityInfo::getEdges(const BasicBlock *BB) const { |
| 1043 | assert(BB->getParent() == LastF); |
| 1044 | assert(BlockNumberEpoch == LastF->getBlockNumberEpoch()); |
| 1045 | if (EdgeStarts.size() <= BB->getNumber()) |
| 1046 | return {}; |
| 1047 | if (unsigned EdgeStart = EdgeStarts[BB->getNumber()]) { |
| 1048 | const BranchProbability *Start = &Probs[EdgeStart - 1]; // 0 = no edges. |
| 1049 | size_t Count = SIZE_MAX; // Avoid querying num successors in release builds. |
| 1050 | #ifndef NDEBUG |
| 1051 | Count = succ_size(BB); |
| 1052 | #endif |
| 1053 | return ArrayRef(Start, Count); |
| 1054 | } |
| 1055 | return {}; |
| 1056 | } |
| 1057 | |
| 1058 | bool BranchProbabilityInfo::invalidate(Function &, const PreservedAnalyses &PA, |
| 1059 | FunctionAnalysisManager::Invalidator &) { |
| 1060 | // Check whether the analysis, all analyses on functions, or the function's |
| 1061 | // CFG have been preserved. |
| 1062 | auto PAC = PA.getChecker<BranchProbabilityAnalysis>(); |
| 1063 | return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>() || |
| 1064 | PAC.preservedSet<CFGAnalyses>()); |
| 1065 | } |
| 1066 | |
| 1067 | void BranchProbabilityInfo::print(raw_ostream &OS) const { |
| 1068 | OS << "---- Branch Probabilities ----\n" ; |
| 1069 | // We print the probabilities from the last function the analysis ran over, |
| 1070 | // or the function it is currently running over. |
| 1071 | assert(LastF && "Cannot print prior to running over a function" ); |
| 1072 | for (const auto &BI : *LastF) { |
| 1073 | for (const BasicBlock *Succ : successors(BB: &BI)) |
| 1074 | printEdgeProbability(OS&: OS << " " , Src: &BI, Dst: Succ); |
| 1075 | } |
| 1076 | } |
| 1077 | |
| 1078 | bool BranchProbabilityInfo:: |
| 1079 | isEdgeHot(const BasicBlock *Src, const BasicBlock *Dst) const { |
| 1080 | // Hot probability is at least 4/5 = 80% |
| 1081 | // FIXME: Compare against a static "hot" BranchProbability. |
| 1082 | return getEdgeProbability(Src, Dst) > BranchProbability(4, 5); |
| 1083 | } |
| 1084 | |
| 1085 | /// Get the raw edge probability for the edge. If can't find it, return a |
| 1086 | /// default probability 1/N where N is the number of successors. Here an edge is |
| 1087 | /// specified using PredBlock and an |
| 1088 | /// index to the successors. |
| 1089 | BranchProbability |
| 1090 | BranchProbabilityInfo::getEdgeProbability(const BasicBlock *Src, |
| 1091 | unsigned IndexInSuccessors) const { |
| 1092 | if (ArrayRef<BranchProbability> P = getEdges(BB: Src); !P.empty()) |
| 1093 | return P[IndexInSuccessors]; |
| 1094 | return {1, static_cast<uint32_t>(succ_size(BB: Src))}; |
| 1095 | } |
| 1096 | |
| 1097 | /// Get the raw edge probability calculated for the block pair. This returns the |
| 1098 | /// sum of all raw edge probabilities from Src to Dst. |
| 1099 | BranchProbability |
| 1100 | BranchProbabilityInfo::getEdgeProbability(const BasicBlock *Src, |
| 1101 | const BasicBlock *Dst) const { |
| 1102 | ArrayRef<BranchProbability> P = getEdges(BB: Src); |
| 1103 | if (P.empty()) |
| 1104 | return BranchProbability(llvm::count(Range: successors(BB: Src), Element: Dst), succ_size(BB: Src)); |
| 1105 | |
| 1106 | auto Prob = BranchProbability::getZero(); |
| 1107 | for (auto It : enumerate(First: successors(BB: Src))) |
| 1108 | if (It.value() == Dst) |
| 1109 | Prob += P[It.index()]; |
| 1110 | |
| 1111 | return Prob; |
| 1112 | } |
| 1113 | |
| 1114 | /// Set the edge probability for all edges at once. |
| 1115 | void BranchProbabilityInfo::setEdgeProbability( |
| 1116 | const BasicBlock *Src, ArrayRef<BranchProbability> Probs) { |
| 1117 | assert(Src->getTerminator()->getNumSuccessors() == Probs.size()); |
| 1118 | MutableArrayRef<BranchProbability> P = allocEdges(BB: Src); |
| 1119 | uint64_t TotalNumerator = 0; |
| 1120 | for (unsigned SuccIdx = 0; SuccIdx < Probs.size(); ++SuccIdx) { |
| 1121 | P[SuccIdx] = Probs[SuccIdx]; |
| 1122 | LLVM_DEBUG(dbgs() << "set edge " << Src->getName() << " -> " << SuccIdx |
| 1123 | << " successor probability to " << Probs[SuccIdx] |
| 1124 | << "\n" ); |
| 1125 | TotalNumerator += Probs[SuccIdx].getNumerator(); |
| 1126 | } |
| 1127 | |
| 1128 | // Because of rounding errors the total probability cannot be checked to be |
| 1129 | // 1.0 exactly. That is TotalNumerator == BranchProbability::getDenominator. |
| 1130 | // Instead, every single probability in Probs must be as accurate as possible. |
| 1131 | // This results in error 1/denominator at most, thus the total absolute error |
| 1132 | // should be within Probs.size / BranchProbability::getDenominator. |
| 1133 | if (P.empty()) |
| 1134 | return; // If we store no probabilities, TotalNumerator is zero. |
| 1135 | assert(TotalNumerator <= BranchProbability::getDenominator() + Probs.size()); |
| 1136 | assert(TotalNumerator >= BranchProbability::getDenominator() - Probs.size()); |
| 1137 | (void)TotalNumerator; |
| 1138 | } |
| 1139 | |
| 1140 | void BranchProbabilityInfo::copyEdgeProbabilities(BasicBlock *Src, |
| 1141 | BasicBlock *Dst) { |
| 1142 | assert(succ_size(Src) == succ_size(Dst)); |
| 1143 | // allocEdges can reallocate and must be called first. |
| 1144 | MutableArrayRef<BranchProbability> DstP = allocEdges(BB: Dst); |
| 1145 | ArrayRef<BranchProbability> SrcP = getEdges(BB: Src); |
| 1146 | if (SrcP.empty()) { |
| 1147 | // Nothing to copy from, erase again. |
| 1148 | eraseBlock(BB: Dst); |
| 1149 | return; |
| 1150 | } |
| 1151 | for (unsigned i = 0; i != DstP.size(); ++i) { |
| 1152 | DstP[i] = SrcP[i]; |
| 1153 | LLVM_DEBUG(dbgs() << "set edge " << Dst->getName() << " -> " << i |
| 1154 | << " successor probability to " << SrcP[i] << "\n" ); |
| 1155 | } |
| 1156 | } |
| 1157 | |
| 1158 | void BranchProbabilityInfo::swapSuccEdgesProbabilities(const BasicBlock *Src) { |
| 1159 | assert(Src->getTerminator()->getNumSuccessors() == 2); |
| 1160 | ArrayRef<BranchProbability> P = getEdges(BB: Src); |
| 1161 | if (P.empty()) |
| 1162 | return; |
| 1163 | MutableArrayRef<BranchProbability> MP( |
| 1164 | const_cast<BranchProbability *>(P.data()), P.size()); |
| 1165 | std::swap(a&: MP[0], b&: MP[1]); |
| 1166 | } |
| 1167 | |
| 1168 | raw_ostream & |
| 1169 | BranchProbabilityInfo::printEdgeProbability(raw_ostream &OS, |
| 1170 | const BasicBlock *Src, |
| 1171 | const BasicBlock *Dst) const { |
| 1172 | const BranchProbability Prob = getEdgeProbability(Src, Dst); |
| 1173 | OS << "edge " ; |
| 1174 | Src->printAsOperand(O&: OS, PrintType: false, M: Src->getModule()); |
| 1175 | OS << " -> " ; |
| 1176 | Dst->printAsOperand(O&: OS, PrintType: false, M: Dst->getModule()); |
| 1177 | OS << " probability is " << Prob |
| 1178 | << (isEdgeHot(Src, Dst) ? " [HOT edge]\n" : "\n" ); |
| 1179 | |
| 1180 | return OS; |
| 1181 | } |
| 1182 | |
| 1183 | void BranchProbabilityInfo::eraseBlock(const BasicBlock *BB) { |
| 1184 | LLVM_DEBUG(dbgs() << "eraseBlock " << BB->getName() << "\n" ); |
| 1185 | assert(BB->getParent() == LastF); |
| 1186 | assert(BlockNumberEpoch == LastF->getBlockNumberEpoch()); |
| 1187 | if (EdgeStarts.size() > BB->getNumber()) |
| 1188 | EdgeStarts[BB->getNumber()] = 0; |
| 1189 | } |
| 1190 | |
| 1191 | void BranchProbabilityInfo::calculate(const Function &F, |
| 1192 | const CycleInfo &CycleI, |
| 1193 | const TargetLibraryInfo *TLI, |
| 1194 | DominatorTree *DT, |
| 1195 | PostDominatorTree *PDT) { |
| 1196 | LLVM_DEBUG(dbgs() << "---- Branch Probability Info : " << F.getName() |
| 1197 | << " ----\n\n" ); |
| 1198 | LastF = &F; // Store the last function we ran on for printing. |
| 1199 | BlockNumberEpoch = F.getBlockNumberEpoch(); |
| 1200 | Probs.clear(); |
| 1201 | EdgeStarts.clear(); |
| 1202 | BPIConstruction(*this).calculate(F, CycleI, TLI, DT, PDT); |
| 1203 | |
| 1204 | if (PrintBranchProb && (PrintBranchProbFuncName.empty() || |
| 1205 | F.getName() == PrintBranchProbFuncName)) { |
| 1206 | print(OS&: dbgs()); |
| 1207 | } |
| 1208 | } |
| 1209 | |
| 1210 | void BranchProbabilityInfoWrapperPass::getAnalysisUsage( |
| 1211 | AnalysisUsage &AU) const { |
| 1212 | // We require DT so it's available when LI is available. The LI updating code |
| 1213 | // asserts that DT is also present so if we don't make sure that we have DT |
| 1214 | // here, that assert will trigger. |
| 1215 | AU.addRequired<DominatorTreeWrapperPass>(); |
| 1216 | AU.addRequired<CycleInfoWrapperPass>(); |
| 1217 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
| 1218 | AU.addRequired<DominatorTreeWrapperPass>(); |
| 1219 | AU.addRequired<PostDominatorTreeWrapperPass>(); |
| 1220 | AU.setPreservesAll(); |
| 1221 | } |
| 1222 | |
| 1223 | bool BranchProbabilityInfoWrapperPass::runOnFunction(Function &F) { |
| 1224 | const CycleInfo &CI = getAnalysis<CycleInfoWrapperPass>().getResult(); |
| 1225 | const TargetLibraryInfo &TLI = |
| 1226 | getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); |
| 1227 | DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
| 1228 | PostDominatorTree &PDT = |
| 1229 | getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree(); |
| 1230 | BPI.calculate(F, CycleI: CI, TLI: &TLI, DT: &DT, PDT: &PDT); |
| 1231 | return false; |
| 1232 | } |
| 1233 | |
| 1234 | void BranchProbabilityInfoWrapperPass::print(raw_ostream &OS, |
| 1235 | const Module *) const { |
| 1236 | BPI.print(OS); |
| 1237 | } |
| 1238 | |
| 1239 | AnalysisKey BranchProbabilityAnalysis::Key; |
| 1240 | BranchProbabilityInfo |
| 1241 | BranchProbabilityAnalysis::run(Function &F, FunctionAnalysisManager &AM) { |
| 1242 | auto &CI = AM.getResult<CycleAnalysis>(IR&: F); |
| 1243 | auto &TLI = AM.getResult<TargetLibraryAnalysis>(IR&: F); |
| 1244 | auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
| 1245 | auto &PDT = AM.getResult<PostDominatorTreeAnalysis>(IR&: F); |
| 1246 | BranchProbabilityInfo BPI; |
| 1247 | BPI.calculate(F, CycleI: CI, TLI: &TLI, DT: &DT, PDT: &PDT); |
| 1248 | return BPI; |
| 1249 | } |
| 1250 | |
| 1251 | PreservedAnalyses |
| 1252 | BranchProbabilityPrinterPass::run(Function &F, FunctionAnalysisManager &AM) { |
| 1253 | OS << "Printing analysis 'Branch Probability Analysis' for function '" |
| 1254 | << F.getName() << "':\n" ; |
| 1255 | AM.getResult<BranchProbabilityAnalysis>(IR&: F).print(OS); |
| 1256 | return PreservedAnalyses::all(); |
| 1257 | } |
| 1258 | |