| File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Support/GenericDomTree.h |
| Warning: | line 494, column 12 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- LoopPeel.cpp -------------------------------------------------------===// | |||
| 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 | // Loop Peeling Utilities. | |||
| 10 | //===----------------------------------------------------------------------===// | |||
| 11 | ||||
| 12 | #include "llvm/Transforms/Utils/LoopPeel.h" | |||
| 13 | #include "llvm/ADT/DenseMap.h" | |||
| 14 | #include "llvm/ADT/Optional.h" | |||
| 15 | #include "llvm/ADT/SmallVector.h" | |||
| 16 | #include "llvm/ADT/Statistic.h" | |||
| 17 | #include "llvm/Analysis/LoopInfo.h" | |||
| 18 | #include "llvm/Analysis/LoopIterator.h" | |||
| 19 | #include "llvm/Analysis/ScalarEvolution.h" | |||
| 20 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" | |||
| 21 | #include "llvm/Analysis/TargetTransformInfo.h" | |||
| 22 | #include "llvm/IR/BasicBlock.h" | |||
| 23 | #include "llvm/IR/Dominators.h" | |||
| 24 | #include "llvm/IR/Function.h" | |||
| 25 | #include "llvm/IR/InstrTypes.h" | |||
| 26 | #include "llvm/IR/Instruction.h" | |||
| 27 | #include "llvm/IR/Instructions.h" | |||
| 28 | #include "llvm/IR/LLVMContext.h" | |||
| 29 | #include "llvm/IR/MDBuilder.h" | |||
| 30 | #include "llvm/IR/Metadata.h" | |||
| 31 | #include "llvm/IR/PatternMatch.h" | |||
| 32 | #include "llvm/Support/Casting.h" | |||
| 33 | #include "llvm/Support/CommandLine.h" | |||
| 34 | #include "llvm/Support/Debug.h" | |||
| 35 | #include "llvm/Support/raw_ostream.h" | |||
| 36 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | |||
| 37 | #include "llvm/Transforms/Utils/Cloning.h" | |||
| 38 | #include "llvm/Transforms/Utils/LoopSimplify.h" | |||
| 39 | #include "llvm/Transforms/Utils/LoopUtils.h" | |||
| 40 | #include "llvm/Transforms/Utils/UnrollLoop.h" | |||
| 41 | #include "llvm/Transforms/Utils/ValueMapper.h" | |||
| 42 | #include <algorithm> | |||
| 43 | #include <cassert> | |||
| 44 | #include <cstdint> | |||
| 45 | #include <limits> | |||
| 46 | ||||
| 47 | using namespace llvm; | |||
| 48 | using namespace llvm::PatternMatch; | |||
| 49 | ||||
| 50 | #define DEBUG_TYPE"loop-peel" "loop-peel" | |||
| 51 | ||||
| 52 | STATISTIC(NumPeeled, "Number of loops peeled")static llvm::Statistic NumPeeled = {"loop-peel", "NumPeeled", "Number of loops peeled"}; | |||
| 53 | ||||
| 54 | static cl::opt<unsigned> UnrollPeelCount( | |||
| 55 | "unroll-peel-count", cl::Hidden, | |||
| 56 | cl::desc("Set the unroll peeling count, for testing purposes")); | |||
| 57 | ||||
| 58 | static cl::opt<bool> | |||
| 59 | UnrollAllowPeeling("unroll-allow-peeling", cl::init(true), cl::Hidden, | |||
| 60 | cl::desc("Allows loops to be peeled when the dynamic " | |||
| 61 | "trip count is known to be low.")); | |||
| 62 | ||||
| 63 | static cl::opt<bool> | |||
| 64 | UnrollAllowLoopNestsPeeling("unroll-allow-loop-nests-peeling", | |||
| 65 | cl::init(false), cl::Hidden, | |||
| 66 | cl::desc("Allows loop nests to be peeled.")); | |||
| 67 | ||||
| 68 | static cl::opt<unsigned> UnrollPeelMaxCount( | |||
| 69 | "unroll-peel-max-count", cl::init(7), cl::Hidden, | |||
| 70 | cl::desc("Max average trip count which will cause loop peeling.")); | |||
| 71 | ||||
| 72 | static cl::opt<unsigned> UnrollForcePeelCount( | |||
| 73 | "unroll-force-peel-count", cl::init(0), cl::Hidden, | |||
| 74 | cl::desc("Force a peel count regardless of profiling information.")); | |||
| 75 | ||||
| 76 | static cl::opt<bool> UnrollPeelMultiDeoptExit( | |||
| 77 | "unroll-peel-multi-deopt-exit", cl::init(true), cl::Hidden, | |||
| 78 | cl::desc("Allow peeling of loops with multiple deopt exits.")); | |||
| 79 | ||||
| 80 | static const char *PeeledCountMetaData = "llvm.loop.peeled.count"; | |||
| 81 | ||||
| 82 | // Designates that a Phi is estimated to become invariant after an "infinite" | |||
| 83 | // number of loop iterations (i.e. only may become an invariant if the loop is | |||
| 84 | // fully unrolled). | |||
| 85 | static const unsigned InfiniteIterationsToInvariance = | |||
| 86 | std::numeric_limits<unsigned>::max(); | |||
| 87 | ||||
| 88 | // Check whether we are capable of peeling this loop. | |||
| 89 | bool llvm::canPeel(Loop *L) { | |||
| 90 | // Make sure the loop is in simplified form | |||
| 91 | if (!L->isLoopSimplifyForm()) | |||
| 92 | return false; | |||
| 93 | ||||
| 94 | if (UnrollPeelMultiDeoptExit) { | |||
| 95 | SmallVector<BasicBlock *, 4> Exits; | |||
| 96 | L->getUniqueNonLatchExitBlocks(Exits); | |||
| 97 | ||||
| 98 | if (!Exits.empty()) { | |||
| 99 | // Latch's terminator is a conditional branch, Latch is exiting and | |||
| 100 | // all non Latch exits ends up with deoptimize. | |||
| 101 | const BasicBlock *Latch = L->getLoopLatch(); | |||
| 102 | const BranchInst *T = dyn_cast<BranchInst>(Latch->getTerminator()); | |||
| 103 | return T && T->isConditional() && L->isLoopExiting(Latch) && | |||
| 104 | all_of(Exits, [](const BasicBlock *BB) { | |||
| 105 | return BB->getTerminatingDeoptimizeCall(); | |||
| 106 | }); | |||
| 107 | } | |||
| 108 | } | |||
| 109 | ||||
| 110 | // Only peel loops that contain a single exit | |||
| 111 | if (!L->getExitingBlock() || !L->getUniqueExitBlock()) | |||
| 112 | return false; | |||
| 113 | ||||
| 114 | // Don't try to peel loops where the latch is not the exiting block. | |||
| 115 | // This can be an indication of two different things: | |||
| 116 | // 1) The loop is not rotated. | |||
| 117 | // 2) The loop contains irreducible control flow that involves the latch. | |||
| 118 | const BasicBlock *Latch = L->getLoopLatch(); | |||
| 119 | if (Latch != L->getExitingBlock()) | |||
| 120 | return false; | |||
| 121 | ||||
| 122 | // Peeling is only supported if the latch is a branch. | |||
| 123 | if (!isa<BranchInst>(Latch->getTerminator())) | |||
| 124 | return false; | |||
| 125 | ||||
| 126 | return true; | |||
| 127 | } | |||
| 128 | ||||
| 129 | // This function calculates the number of iterations after which the given Phi | |||
| 130 | // becomes an invariant. The pre-calculated values are memorized in the map. The | |||
| 131 | // function (shortcut is I) is calculated according to the following definition: | |||
| 132 | // Given %x = phi <Inputs from above the loop>, ..., [%y, %back.edge]. | |||
| 133 | // If %y is a loop invariant, then I(%x) = 1. | |||
| 134 | // If %y is a Phi from the loop header, I(%x) = I(%y) + 1. | |||
| 135 | // Otherwise, I(%x) is infinite. | |||
| 136 | // TODO: Actually if %y is an expression that depends only on Phi %z and some | |||
| 137 | // loop invariants, we can estimate I(%x) = I(%z) + 1. The example | |||
| 138 | // looks like: | |||
| 139 | // %x = phi(0, %a), <-- becomes invariant starting from 3rd iteration. | |||
| 140 | // %y = phi(0, 5), | |||
| 141 | // %a = %y + 1. | |||
| 142 | static unsigned calculateIterationsToInvariance( | |||
| 143 | PHINode *Phi, Loop *L, BasicBlock *BackEdge, | |||
| 144 | SmallDenseMap<PHINode *, unsigned> &IterationsToInvariance) { | |||
| 145 | assert(Phi->getParent() == L->getHeader() &&((void)0) | |||
| 146 | "Non-loop Phi should not be checked for turning into invariant.")((void)0); | |||
| 147 | assert(BackEdge == L->getLoopLatch() && "Wrong latch?")((void)0); | |||
| 148 | // If we already know the answer, take it from the map. | |||
| 149 | auto I = IterationsToInvariance.find(Phi); | |||
| 150 | if (I != IterationsToInvariance.end()) | |||
| 151 | return I->second; | |||
| 152 | ||||
| 153 | // Otherwise we need to analyze the input from the back edge. | |||
| 154 | Value *Input = Phi->getIncomingValueForBlock(BackEdge); | |||
| 155 | // Place infinity to map to avoid infinite recursion for cycled Phis. Such | |||
| 156 | // cycles can never stop on an invariant. | |||
| 157 | IterationsToInvariance[Phi] = InfiniteIterationsToInvariance; | |||
| 158 | unsigned ToInvariance = InfiniteIterationsToInvariance; | |||
| 159 | ||||
| 160 | if (L->isLoopInvariant(Input)) | |||
| 161 | ToInvariance = 1u; | |||
| 162 | else if (PHINode *IncPhi = dyn_cast<PHINode>(Input)) { | |||
| 163 | // Only consider Phis in header block. | |||
| 164 | if (IncPhi->getParent() != L->getHeader()) | |||
| 165 | return InfiniteIterationsToInvariance; | |||
| 166 | // If the input becomes an invariant after X iterations, then our Phi | |||
| 167 | // becomes an invariant after X + 1 iterations. | |||
| 168 | unsigned InputToInvariance = calculateIterationsToInvariance( | |||
| 169 | IncPhi, L, BackEdge, IterationsToInvariance); | |||
| 170 | if (InputToInvariance != InfiniteIterationsToInvariance) | |||
| 171 | ToInvariance = InputToInvariance + 1u; | |||
| 172 | } | |||
| 173 | ||||
| 174 | // If we found that this Phi lies in an invariant chain, update the map. | |||
| 175 | if (ToInvariance != InfiniteIterationsToInvariance) | |||
| 176 | IterationsToInvariance[Phi] = ToInvariance; | |||
| 177 | return ToInvariance; | |||
| 178 | } | |||
| 179 | ||||
| 180 | // Return the number of iterations to peel off that make conditions in the | |||
| 181 | // body true/false. For example, if we peel 2 iterations off the loop below, | |||
| 182 | // the condition i < 2 can be evaluated at compile time. | |||
| 183 | // for (i = 0; i < n; i++) | |||
| 184 | // if (i < 2) | |||
| 185 | // .. | |||
| 186 | // else | |||
| 187 | // .. | |||
| 188 | // } | |||
| 189 | static unsigned countToEliminateCompares(Loop &L, unsigned MaxPeelCount, | |||
| 190 | ScalarEvolution &SE) { | |||
| 191 | assert(L.isLoopSimplifyForm() && "Loop needs to be in loop simplify form")((void)0); | |||
| 192 | unsigned DesiredPeelCount = 0; | |||
| 193 | ||||
| 194 | for (auto *BB : L.blocks()) { | |||
| 195 | auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); | |||
| 196 | if (!BI || BI->isUnconditional()) | |||
| 197 | continue; | |||
| 198 | ||||
| 199 | // Ignore loop exit condition. | |||
| 200 | if (L.getLoopLatch() == BB) | |||
| 201 | continue; | |||
| 202 | ||||
| 203 | Value *Condition = BI->getCondition(); | |||
| 204 | Value *LeftVal, *RightVal; | |||
| 205 | CmpInst::Predicate Pred; | |||
| 206 | if (!match(Condition, m_ICmp(Pred, m_Value(LeftVal), m_Value(RightVal)))) | |||
| 207 | continue; | |||
| 208 | ||||
| 209 | const SCEV *LeftSCEV = SE.getSCEV(LeftVal); | |||
| 210 | const SCEV *RightSCEV = SE.getSCEV(RightVal); | |||
| 211 | ||||
| 212 | // Do not consider predicates that are known to be true or false | |||
| 213 | // independently of the loop iteration. | |||
| 214 | if (SE.evaluatePredicate(Pred, LeftSCEV, RightSCEV)) | |||
| 215 | continue; | |||
| 216 | ||||
| 217 | // Check if we have a condition with one AddRec and one non AddRec | |||
| 218 | // expression. Normalize LeftSCEV to be the AddRec. | |||
| 219 | if (!isa<SCEVAddRecExpr>(LeftSCEV)) { | |||
| 220 | if (isa<SCEVAddRecExpr>(RightSCEV)) { | |||
| 221 | std::swap(LeftSCEV, RightSCEV); | |||
| 222 | Pred = ICmpInst::getSwappedPredicate(Pred); | |||
| 223 | } else | |||
| 224 | continue; | |||
| 225 | } | |||
| 226 | ||||
| 227 | const SCEVAddRecExpr *LeftAR = cast<SCEVAddRecExpr>(LeftSCEV); | |||
| 228 | ||||
| 229 | // Avoid huge SCEV computations in the loop below, make sure we only | |||
| 230 | // consider AddRecs of the loop we are trying to peel. | |||
| 231 | if (!LeftAR->isAffine() || LeftAR->getLoop() != &L) | |||
| 232 | continue; | |||
| 233 | if (!(ICmpInst::isEquality(Pred) && LeftAR->hasNoSelfWrap()) && | |||
| 234 | !SE.getMonotonicPredicateType(LeftAR, Pred)) | |||
| 235 | continue; | |||
| 236 | ||||
| 237 | // Check if extending the current DesiredPeelCount lets us evaluate Pred | |||
| 238 | // or !Pred in the loop body statically. | |||
| 239 | unsigned NewPeelCount = DesiredPeelCount; | |||
| 240 | ||||
| 241 | const SCEV *IterVal = LeftAR->evaluateAtIteration( | |||
| 242 | SE.getConstant(LeftSCEV->getType(), NewPeelCount), SE); | |||
| 243 | ||||
| 244 | // If the original condition is not known, get the negated predicate | |||
| 245 | // (which holds on the else branch) and check if it is known. This allows | |||
| 246 | // us to peel of iterations that make the original condition false. | |||
| 247 | if (!SE.isKnownPredicate(Pred, IterVal, RightSCEV)) | |||
| 248 | Pred = ICmpInst::getInversePredicate(Pred); | |||
| 249 | ||||
| 250 | const SCEV *Step = LeftAR->getStepRecurrence(SE); | |||
| 251 | const SCEV *NextIterVal = SE.getAddExpr(IterVal, Step); | |||
| 252 | auto PeelOneMoreIteration = [&IterVal, &NextIterVal, &SE, Step, | |||
| 253 | &NewPeelCount]() { | |||
| 254 | IterVal = NextIterVal; | |||
| 255 | NextIterVal = SE.getAddExpr(IterVal, Step); | |||
| 256 | NewPeelCount++; | |||
| 257 | }; | |||
| 258 | ||||
| 259 | auto CanPeelOneMoreIteration = [&NewPeelCount, &MaxPeelCount]() { | |||
| 260 | return NewPeelCount < MaxPeelCount; | |||
| 261 | }; | |||
| 262 | ||||
| 263 | while (CanPeelOneMoreIteration() && | |||
| 264 | SE.isKnownPredicate(Pred, IterVal, RightSCEV)) | |||
| 265 | PeelOneMoreIteration(); | |||
| 266 | ||||
| 267 | // With *that* peel count, does the predicate !Pred become known in the | |||
| 268 | // first iteration of the loop body after peeling? | |||
| 269 | if (!SE.isKnownPredicate(ICmpInst::getInversePredicate(Pred), IterVal, | |||
| 270 | RightSCEV)) | |||
| 271 | continue; // If not, give up. | |||
| 272 | ||||
| 273 | // However, for equality comparisons, that isn't always sufficient to | |||
| 274 | // eliminate the comparsion in loop body, we may need to peel one more | |||
| 275 | // iteration. See if that makes !Pred become unknown again. | |||
| 276 | if (ICmpInst::isEquality(Pred) && | |||
| 277 | !SE.isKnownPredicate(ICmpInst::getInversePredicate(Pred), NextIterVal, | |||
| 278 | RightSCEV) && | |||
| 279 | !SE.isKnownPredicate(Pred, IterVal, RightSCEV) && | |||
| 280 | SE.isKnownPredicate(Pred, NextIterVal, RightSCEV)) { | |||
| 281 | if (!CanPeelOneMoreIteration()) | |||
| 282 | continue; // Need to peel one more iteration, but can't. Give up. | |||
| 283 | PeelOneMoreIteration(); // Great! | |||
| 284 | } | |||
| 285 | ||||
| 286 | DesiredPeelCount = std::max(DesiredPeelCount, NewPeelCount); | |||
| 287 | } | |||
| 288 | ||||
| 289 | return DesiredPeelCount; | |||
| 290 | } | |||
| 291 | ||||
| 292 | // Return the number of iterations we want to peel off. | |||
| 293 | void llvm::computePeelCount(Loop *L, unsigned LoopSize, | |||
| 294 | TargetTransformInfo::PeelingPreferences &PP, | |||
| 295 | unsigned &TripCount, ScalarEvolution &SE, | |||
| 296 | unsigned Threshold) { | |||
| 297 | assert(LoopSize > 0 && "Zero loop size is not allowed!")((void)0); | |||
| 298 | // Save the PP.PeelCount value set by the target in | |||
| 299 | // TTI.getPeelingPreferences or by the flag -unroll-peel-count. | |||
| 300 | unsigned TargetPeelCount = PP.PeelCount; | |||
| 301 | PP.PeelCount = 0; | |||
| 302 | if (!canPeel(L)) | |||
| 303 | return; | |||
| 304 | ||||
| 305 | // Only try to peel innermost loops by default. | |||
| 306 | // The constraint can be relaxed by the target in TTI.getUnrollingPreferences | |||
| 307 | // or by the flag -unroll-allow-loop-nests-peeling. | |||
| 308 | if (!PP.AllowLoopNestsPeeling && !L->isInnermost()) | |||
| 309 | return; | |||
| 310 | ||||
| 311 | // If the user provided a peel count, use that. | |||
| 312 | bool UserPeelCount = UnrollForcePeelCount.getNumOccurrences() > 0; | |||
| 313 | if (UserPeelCount) { | |||
| 314 | LLVM_DEBUG(dbgs() << "Force-peeling first " << UnrollForcePeelCountdo { } while (false) | |||
| 315 | << " iterations.\n")do { } while (false); | |||
| 316 | PP.PeelCount = UnrollForcePeelCount; | |||
| 317 | PP.PeelProfiledIterations = true; | |||
| 318 | return; | |||
| 319 | } | |||
| 320 | ||||
| 321 | // Skip peeling if it's disabled. | |||
| 322 | if (!PP.AllowPeeling) | |||
| 323 | return; | |||
| 324 | ||||
| 325 | unsigned AlreadyPeeled = 0; | |||
| 326 | if (auto Peeled = getOptionalIntLoopAttribute(L, PeeledCountMetaData)) | |||
| 327 | AlreadyPeeled = *Peeled; | |||
| 328 | // Stop if we already peeled off the maximum number of iterations. | |||
| 329 | if (AlreadyPeeled >= UnrollPeelMaxCount) | |||
| 330 | return; | |||
| 331 | ||||
| 332 | // Here we try to get rid of Phis which become invariants after 1, 2, ..., N | |||
| 333 | // iterations of the loop. For this we compute the number for iterations after | |||
| 334 | // which every Phi is guaranteed to become an invariant, and try to peel the | |||
| 335 | // maximum number of iterations among these values, thus turning all those | |||
| 336 | // Phis into invariants. | |||
| 337 | // First, check that we can peel at least one iteration. | |||
| 338 | if (2 * LoopSize <= Threshold && UnrollPeelMaxCount > 0) { | |||
| 339 | // Store the pre-calculated values here. | |||
| 340 | SmallDenseMap<PHINode *, unsigned> IterationsToInvariance; | |||
| 341 | // Now go through all Phis to calculate their the number of iterations they | |||
| 342 | // need to become invariants. | |||
| 343 | // Start the max computation with the UP.PeelCount value set by the target | |||
| 344 | // in TTI.getUnrollingPreferences or by the flag -unroll-peel-count. | |||
| 345 | unsigned DesiredPeelCount = TargetPeelCount; | |||
| 346 | BasicBlock *BackEdge = L->getLoopLatch(); | |||
| 347 | assert(BackEdge && "Loop is not in simplified form?")((void)0); | |||
| 348 | for (auto BI = L->getHeader()->begin(); isa<PHINode>(&*BI); ++BI) { | |||
| 349 | PHINode *Phi = cast<PHINode>(&*BI); | |||
| 350 | unsigned ToInvariance = calculateIterationsToInvariance( | |||
| 351 | Phi, L, BackEdge, IterationsToInvariance); | |||
| 352 | if (ToInvariance != InfiniteIterationsToInvariance) | |||
| 353 | DesiredPeelCount = std::max(DesiredPeelCount, ToInvariance); | |||
| 354 | } | |||
| 355 | ||||
| 356 | // Pay respect to limitations implied by loop size and the max peel count. | |||
| 357 | unsigned MaxPeelCount = UnrollPeelMaxCount; | |||
| 358 | MaxPeelCount = std::min(MaxPeelCount, Threshold / LoopSize - 1); | |||
| 359 | ||||
| 360 | DesiredPeelCount = std::max(DesiredPeelCount, | |||
| 361 | countToEliminateCompares(*L, MaxPeelCount, SE)); | |||
| 362 | ||||
| 363 | if (DesiredPeelCount > 0) { | |||
| 364 | DesiredPeelCount = std::min(DesiredPeelCount, MaxPeelCount); | |||
| 365 | // Consider max peel count limitation. | |||
| 366 | assert(DesiredPeelCount > 0 && "Wrong loop size estimation?")((void)0); | |||
| 367 | if (DesiredPeelCount + AlreadyPeeled <= UnrollPeelMaxCount) { | |||
| 368 | LLVM_DEBUG(dbgs() << "Peel " << DesiredPeelCountdo { } while (false) | |||
| 369 | << " iteration(s) to turn"do { } while (false) | |||
| 370 | << " some Phis into invariants.\n")do { } while (false); | |||
| 371 | PP.PeelCount = DesiredPeelCount; | |||
| 372 | PP.PeelProfiledIterations = false; | |||
| 373 | return; | |||
| 374 | } | |||
| 375 | } | |||
| 376 | } | |||
| 377 | ||||
| 378 | // Bail if we know the statically calculated trip count. | |||
| 379 | // In this case we rather prefer partial unrolling. | |||
| 380 | if (TripCount) | |||
| 381 | return; | |||
| 382 | ||||
| 383 | // Do not apply profile base peeling if it is disabled. | |||
| 384 | if (!PP.PeelProfiledIterations) | |||
| 385 | return; | |||
| 386 | // If we don't know the trip count, but have reason to believe the average | |||
| 387 | // trip count is low, peeling should be beneficial, since we will usually | |||
| 388 | // hit the peeled section. | |||
| 389 | // We only do this in the presence of profile information, since otherwise | |||
| 390 | // our estimates of the trip count are not reliable enough. | |||
| 391 | if (L->getHeader()->getParent()->hasProfileData()) { | |||
| 392 | Optional<unsigned> PeelCount = getLoopEstimatedTripCount(L); | |||
| 393 | if (!PeelCount) | |||
| 394 | return; | |||
| 395 | ||||
| 396 | LLVM_DEBUG(dbgs() << "Profile-based estimated trip count is " << *PeelCountdo { } while (false) | |||
| 397 | << "\n")do { } while (false); | |||
| 398 | ||||
| 399 | if (*PeelCount) { | |||
| 400 | if ((*PeelCount + AlreadyPeeled <= UnrollPeelMaxCount) && | |||
| 401 | (LoopSize * (*PeelCount + 1) <= Threshold)) { | |||
| 402 | LLVM_DEBUG(dbgs() << "Peeling first " << *PeelCountdo { } while (false) | |||
| 403 | << " iterations.\n")do { } while (false); | |||
| 404 | PP.PeelCount = *PeelCount; | |||
| 405 | return; | |||
| 406 | } | |||
| 407 | LLVM_DEBUG(dbgs() << "Requested peel count: " << *PeelCount << "\n")do { } while (false); | |||
| 408 | LLVM_DEBUG(dbgs() << "Already peel count: " << AlreadyPeeled << "\n")do { } while (false); | |||
| 409 | LLVM_DEBUG(dbgs() << "Max peel count: " << UnrollPeelMaxCount << "\n")do { } while (false); | |||
| 410 | LLVM_DEBUG(dbgs() << "Peel cost: " << LoopSize * (*PeelCount + 1)do { } while (false) | |||
| 411 | << "\n")do { } while (false); | |||
| 412 | LLVM_DEBUG(dbgs() << "Max peel cost: " << Threshold << "\n")do { } while (false); | |||
| 413 | } | |||
| 414 | } | |||
| 415 | } | |||
| 416 | ||||
| 417 | /// Update the branch weights of the latch of a peeled-off loop | |||
| 418 | /// iteration. | |||
| 419 | /// This sets the branch weights for the latch of the recently peeled off loop | |||
| 420 | /// iteration correctly. | |||
| 421 | /// Let F is a weight of the edge from latch to header. | |||
| 422 | /// Let E is a weight of the edge from latch to exit. | |||
| 423 | /// F/(F+E) is a probability to go to loop and E/(F+E) is a probability to | |||
| 424 | /// go to exit. | |||
| 425 | /// Then, Estimated TripCount = F / E. | |||
| 426 | /// For I-th (counting from 0) peeled off iteration we set the the weights for | |||
| 427 | /// the peeled latch as (TC - I, 1). It gives us reasonable distribution, | |||
| 428 | /// The probability to go to exit 1/(TC-I) increases. At the same time | |||
| 429 | /// the estimated trip count of remaining loop reduces by I. | |||
| 430 | /// To avoid dealing with division rounding we can just multiple both part | |||
| 431 | /// of weights to E and use weight as (F - I * E, E). | |||
| 432 | /// | |||
| 433 | /// \param Header The copy of the header block that belongs to next iteration. | |||
| 434 | /// \param LatchBR The copy of the latch branch that belongs to this iteration. | |||
| 435 | /// \param[in,out] FallThroughWeight The weight of the edge from latch to | |||
| 436 | /// header before peeling (in) and after peeled off one iteration (out). | |||
| 437 | static void updateBranchWeights(BasicBlock *Header, BranchInst *LatchBR, | |||
| 438 | uint64_t ExitWeight, | |||
| 439 | uint64_t &FallThroughWeight) { | |||
| 440 | // FallThroughWeight is 0 means that there is no branch weights on original | |||
| 441 | // latch block or estimated trip count is zero. | |||
| 442 | if (!FallThroughWeight) | |||
| 443 | return; | |||
| 444 | ||||
| 445 | unsigned HeaderIdx = (LatchBR->getSuccessor(0) == Header ? 0 : 1); | |||
| 446 | MDBuilder MDB(LatchBR->getContext()); | |||
| 447 | MDNode *WeightNode = | |||
| 448 | HeaderIdx ? MDB.createBranchWeights(ExitWeight, FallThroughWeight) | |||
| 449 | : MDB.createBranchWeights(FallThroughWeight, ExitWeight); | |||
| 450 | LatchBR->setMetadata(LLVMContext::MD_prof, WeightNode); | |||
| 451 | FallThroughWeight = | |||
| 452 | FallThroughWeight > ExitWeight ? FallThroughWeight - ExitWeight : 1; | |||
| 453 | } | |||
| 454 | ||||
| 455 | /// Initialize the weights. | |||
| 456 | /// | |||
| 457 | /// \param Header The header block. | |||
| 458 | /// \param LatchBR The latch branch. | |||
| 459 | /// \param[out] ExitWeight The weight of the edge from Latch to Exit. | |||
| 460 | /// \param[out] FallThroughWeight The weight of the edge from Latch to Header. | |||
| 461 | static void initBranchWeights(BasicBlock *Header, BranchInst *LatchBR, | |||
| 462 | uint64_t &ExitWeight, | |||
| 463 | uint64_t &FallThroughWeight) { | |||
| 464 | uint64_t TrueWeight, FalseWeight; | |||
| 465 | if (!LatchBR->extractProfMetadata(TrueWeight, FalseWeight)) | |||
| 466 | return; | |||
| 467 | unsigned HeaderIdx = LatchBR->getSuccessor(0) == Header ? 0 : 1; | |||
| 468 | ExitWeight = HeaderIdx ? TrueWeight : FalseWeight; | |||
| 469 | FallThroughWeight = HeaderIdx ? FalseWeight : TrueWeight; | |||
| 470 | } | |||
| 471 | ||||
| 472 | /// Update the weights of original Latch block after peeling off all iterations. | |||
| 473 | /// | |||
| 474 | /// \param Header The header block. | |||
| 475 | /// \param LatchBR The latch branch. | |||
| 476 | /// \param ExitWeight The weight of the edge from Latch to Exit. | |||
| 477 | /// \param FallThroughWeight The weight of the edge from Latch to Header. | |||
| 478 | static void fixupBranchWeights(BasicBlock *Header, BranchInst *LatchBR, | |||
| 479 | uint64_t ExitWeight, | |||
| 480 | uint64_t FallThroughWeight) { | |||
| 481 | // FallThroughWeight is 0 means that there is no branch weights on original | |||
| 482 | // latch block or estimated trip count is zero. | |||
| 483 | if (!FallThroughWeight) | |||
| 484 | return; | |||
| 485 | ||||
| 486 | // Sets the branch weights on the loop exit. | |||
| 487 | MDBuilder MDB(LatchBR->getContext()); | |||
| 488 | unsigned HeaderIdx = LatchBR->getSuccessor(0) == Header ? 0 : 1; | |||
| 489 | MDNode *WeightNode = | |||
| 490 | HeaderIdx ? MDB.createBranchWeights(ExitWeight, FallThroughWeight) | |||
| 491 | : MDB.createBranchWeights(FallThroughWeight, ExitWeight); | |||
| 492 | LatchBR->setMetadata(LLVMContext::MD_prof, WeightNode); | |||
| 493 | } | |||
| 494 | ||||
| 495 | /// Clones the body of the loop L, putting it between \p InsertTop and \p | |||
| 496 | /// InsertBot. | |||
| 497 | /// \param IterNumber The serial number of the iteration currently being | |||
| 498 | /// peeled off. | |||
| 499 | /// \param ExitEdges The exit edges of the original loop. | |||
| 500 | /// \param[out] NewBlocks A list of the blocks in the newly created clone | |||
| 501 | /// \param[out] VMap The value map between the loop and the new clone. | |||
| 502 | /// \param LoopBlocks A helper for DFS-traversal of the loop. | |||
| 503 | /// \param LVMap A value-map that maps instructions from the original loop to | |||
| 504 | /// instructions in the last peeled-off iteration. | |||
| 505 | static void cloneLoopBlocks( | |||
| 506 | Loop *L, unsigned IterNumber, BasicBlock *InsertTop, BasicBlock *InsertBot, | |||
| 507 | SmallVectorImpl<std::pair<BasicBlock *, BasicBlock *>> &ExitEdges, | |||
| 508 | SmallVectorImpl<BasicBlock *> &NewBlocks, LoopBlocksDFS &LoopBlocks, | |||
| 509 | ValueToValueMapTy &VMap, ValueToValueMapTy &LVMap, DominatorTree *DT, | |||
| 510 | LoopInfo *LI, ArrayRef<MDNode *> LoopLocalNoAliasDeclScopes) { | |||
| 511 | BasicBlock *Header = L->getHeader(); | |||
| 512 | BasicBlock *Latch = L->getLoopLatch(); | |||
| 513 | BasicBlock *PreHeader = L->getLoopPreheader(); | |||
| 514 | ||||
| 515 | Function *F = Header->getParent(); | |||
| 516 | LoopBlocksDFS::RPOIterator BlockBegin = LoopBlocks.beginRPO(); | |||
| 517 | LoopBlocksDFS::RPOIterator BlockEnd = LoopBlocks.endRPO(); | |||
| 518 | Loop *ParentLoop = L->getParentLoop(); | |||
| 519 | ||||
| 520 | // For each block in the original loop, create a new copy, | |||
| 521 | // and update the value map with the newly created values. | |||
| 522 | for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) { | |||
| 523 | BasicBlock *NewBB = CloneBasicBlock(*BB, VMap, ".peel", F); | |||
| 524 | NewBlocks.push_back(NewBB); | |||
| 525 | ||||
| 526 | // If an original block is an immediate child of the loop L, its copy | |||
| 527 | // is a child of a ParentLoop after peeling. If a block is a child of | |||
| 528 | // a nested loop, it is handled in the cloneLoop() call below. | |||
| 529 | if (ParentLoop && LI->getLoopFor(*BB) == L) | |||
| 530 | ParentLoop->addBasicBlockToLoop(NewBB, *LI); | |||
| 531 | ||||
| 532 | VMap[*BB] = NewBB; | |||
| 533 | ||||
| 534 | // If dominator tree is available, insert nodes to represent cloned blocks. | |||
| 535 | if (DT) { | |||
| 536 | if (Header == *BB) | |||
| 537 | DT->addNewBlock(NewBB, InsertTop); | |||
| 538 | else { | |||
| 539 | DomTreeNode *IDom = DT->getNode(*BB)->getIDom(); | |||
| 540 | // VMap must contain entry for IDom, as the iteration order is RPO. | |||
| 541 | DT->addNewBlock(NewBB, cast<BasicBlock>(VMap[IDom->getBlock()])); | |||
| 542 | } | |||
| 543 | } | |||
| 544 | } | |||
| 545 | ||||
| 546 | { | |||
| 547 | // Identify what other metadata depends on the cloned version. After | |||
| 548 | // cloning, replace the metadata with the corrected version for both | |||
| 549 | // memory instructions and noalias intrinsics. | |||
| 550 | std::string Ext = (Twine("Peel") + Twine(IterNumber)).str(); | |||
| 551 | cloneAndAdaptNoAliasScopes(LoopLocalNoAliasDeclScopes, NewBlocks, | |||
| 552 | Header->getContext(), Ext); | |||
| 553 | } | |||
| 554 | ||||
| 555 | // Recursively create the new Loop objects for nested loops, if any, | |||
| 556 | // to preserve LoopInfo. | |||
| 557 | for (Loop *ChildLoop : *L) { | |||
| 558 | cloneLoop(ChildLoop, ParentLoop, VMap, LI, nullptr); | |||
| 559 | } | |||
| 560 | ||||
| 561 | // Hook-up the control flow for the newly inserted blocks. | |||
| 562 | // The new header is hooked up directly to the "top", which is either | |||
| 563 | // the original loop preheader (for the first iteration) or the previous | |||
| 564 | // iteration's exiting block (for every other iteration) | |||
| 565 | InsertTop->getTerminator()->setSuccessor(0, cast<BasicBlock>(VMap[Header])); | |||
| 566 | ||||
| 567 | // Similarly, for the latch: | |||
| 568 | // The original exiting edge is still hooked up to the loop exit. | |||
| 569 | // The backedge now goes to the "bottom", which is either the loop's real | |||
| 570 | // header (for the last peeled iteration) or the copied header of the next | |||
| 571 | // iteration (for every other iteration) | |||
| 572 | BasicBlock *NewLatch = cast<BasicBlock>(VMap[Latch]); | |||
| 573 | BranchInst *LatchBR = cast<BranchInst>(NewLatch->getTerminator()); | |||
| 574 | for (unsigned idx = 0, e = LatchBR->getNumSuccessors(); idx < e; ++idx) | |||
| 575 | if (LatchBR->getSuccessor(idx) == Header) { | |||
| 576 | LatchBR->setSuccessor(idx, InsertBot); | |||
| 577 | break; | |||
| 578 | } | |||
| 579 | if (DT) | |||
| 580 | DT->changeImmediateDominator(InsertBot, NewLatch); | |||
| 581 | ||||
| 582 | // The new copy of the loop body starts with a bunch of PHI nodes | |||
| 583 | // that pick an incoming value from either the preheader, or the previous | |||
| 584 | // loop iteration. Since this copy is no longer part of the loop, we | |||
| 585 | // resolve this statically: | |||
| 586 | // For the first iteration, we use the value from the preheader directly. | |||
| 587 | // For any other iteration, we replace the phi with the value generated by | |||
| 588 | // the immediately preceding clone of the loop body (which represents | |||
| 589 | // the previous iteration). | |||
| 590 | for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { | |||
| 591 | PHINode *NewPHI = cast<PHINode>(VMap[&*I]); | |||
| 592 | if (IterNumber == 0) { | |||
| 593 | VMap[&*I] = NewPHI->getIncomingValueForBlock(PreHeader); | |||
| 594 | } else { | |||
| 595 | Value *LatchVal = NewPHI->getIncomingValueForBlock(Latch); | |||
| 596 | Instruction *LatchInst = dyn_cast<Instruction>(LatchVal); | |||
| 597 | if (LatchInst && L->contains(LatchInst)) | |||
| 598 | VMap[&*I] = LVMap[LatchInst]; | |||
| 599 | else | |||
| 600 | VMap[&*I] = LatchVal; | |||
| 601 | } | |||
| 602 | cast<BasicBlock>(VMap[Header])->getInstList().erase(NewPHI); | |||
| 603 | } | |||
| 604 | ||||
| 605 | // Fix up the outgoing values - we need to add a value for the iteration | |||
| 606 | // we've just created. Note that this must happen *after* the incoming | |||
| 607 | // values are adjusted, since the value going out of the latch may also be | |||
| 608 | // a value coming into the header. | |||
| 609 | for (auto Edge : ExitEdges) | |||
| 610 | for (PHINode &PHI : Edge.second->phis()) { | |||
| 611 | Value *LatchVal = PHI.getIncomingValueForBlock(Edge.first); | |||
| 612 | Instruction *LatchInst = dyn_cast<Instruction>(LatchVal); | |||
| 613 | if (LatchInst && L->contains(LatchInst)) | |||
| 614 | LatchVal = VMap[LatchVal]; | |||
| 615 | PHI.addIncoming(LatchVal, cast<BasicBlock>(VMap[Edge.first])); | |||
| 616 | } | |||
| 617 | ||||
| 618 | // LastValueMap is updated with the values for the current loop | |||
| 619 | // which are used the next time this function is called. | |||
| 620 | for (auto KV : VMap) | |||
| 621 | LVMap[KV.first] = KV.second; | |||
| 622 | } | |||
| 623 | ||||
| 624 | TargetTransformInfo::PeelingPreferences llvm::gatherPeelingPreferences( | |||
| 625 | Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI, | |||
| 626 | Optional<bool> UserAllowPeeling, | |||
| 627 | Optional<bool> UserAllowProfileBasedPeeling, bool UnrollingSpecficValues) { | |||
| 628 | TargetTransformInfo::PeelingPreferences PP; | |||
| 629 | ||||
| 630 | // Set the default values. | |||
| 631 | PP.PeelCount = 0; | |||
| 632 | PP.AllowPeeling = true; | |||
| 633 | PP.AllowLoopNestsPeeling = false; | |||
| 634 | PP.PeelProfiledIterations = true; | |||
| 635 | ||||
| 636 | // Get the target specifc values. | |||
| 637 | TTI.getPeelingPreferences(L, SE, PP); | |||
| 638 | ||||
| 639 | // User specified values using cl::opt. | |||
| 640 | if (UnrollingSpecficValues) { | |||
| 641 | if (UnrollPeelCount.getNumOccurrences() > 0) | |||
| 642 | PP.PeelCount = UnrollPeelCount; | |||
| 643 | if (UnrollAllowPeeling.getNumOccurrences() > 0) | |||
| 644 | PP.AllowPeeling = UnrollAllowPeeling; | |||
| 645 | if (UnrollAllowLoopNestsPeeling.getNumOccurrences() > 0) | |||
| 646 | PP.AllowLoopNestsPeeling = UnrollAllowLoopNestsPeeling; | |||
| 647 | } | |||
| 648 | ||||
| 649 | // User specifed values provided by argument. | |||
| 650 | if (UserAllowPeeling.hasValue()) | |||
| 651 | PP.AllowPeeling = *UserAllowPeeling; | |||
| 652 | if (UserAllowProfileBasedPeeling.hasValue()) | |||
| 653 | PP.PeelProfiledIterations = *UserAllowProfileBasedPeeling; | |||
| 654 | ||||
| 655 | return PP; | |||
| 656 | } | |||
| 657 | ||||
| 658 | /// Peel off the first \p PeelCount iterations of loop \p L. | |||
| 659 | /// | |||
| 660 | /// Note that this does not peel them off as a single straight-line block. | |||
| 661 | /// Rather, each iteration is peeled off separately, and needs to check the | |||
| 662 | /// exit condition. | |||
| 663 | /// For loops that dynamically execute \p PeelCount iterations or less | |||
| 664 | /// this provides a benefit, since the peeled off iterations, which account | |||
| 665 | /// for the bulk of dynamic execution, can be further simplified by scalar | |||
| 666 | /// optimizations. | |||
| 667 | bool llvm::peelLoop(Loop *L, unsigned PeelCount, LoopInfo *LI, | |||
| 668 | ScalarEvolution *SE, DominatorTree *DT, AssumptionCache *AC, | |||
| 669 | bool PreserveLCSSA) { | |||
| 670 | assert(PeelCount > 0 && "Attempt to peel out zero iterations?")((void)0); | |||
| 671 | assert(canPeel(L) && "Attempt to peel a loop which is not peelable?")((void)0); | |||
| 672 | ||||
| 673 | LoopBlocksDFS LoopBlocks(L); | |||
| 674 | LoopBlocks.perform(LI); | |||
| 675 | ||||
| 676 | BasicBlock *Header = L->getHeader(); | |||
| 677 | BasicBlock *PreHeader = L->getLoopPreheader(); | |||
| 678 | BasicBlock *Latch = L->getLoopLatch(); | |||
| 679 | SmallVector<std::pair<BasicBlock *, BasicBlock *>, 4> ExitEdges; | |||
| 680 | L->getExitEdges(ExitEdges); | |||
| 681 | ||||
| 682 | DenseMap<BasicBlock *, BasicBlock *> ExitIDom; | |||
| 683 | if (DT) { | |||
| ||||
| 684 | // We'd like to determine the idom of exit block after peeling one | |||
| 685 | // iteration. | |||
| 686 | // Let Exit is exit block. | |||
| 687 | // Let ExitingSet - is a set of predecessors of Exit block. They are exiting | |||
| 688 | // blocks. | |||
| 689 | // Let Latch' and ExitingSet' are copies after a peeling. | |||
| 690 | // We'd like to find an idom'(Exit) - idom of Exit after peeling. | |||
| 691 | // It is an evident that idom'(Exit) will be the nearest common dominator | |||
| 692 | // of ExitingSet and ExitingSet'. | |||
| 693 | // idom(Exit) is a nearest common dominator of ExitingSet. | |||
| 694 | // idom(Exit)' is a nearest common dominator of ExitingSet'. | |||
| 695 | // Taking into account that we have a single Latch, Latch' will dominate | |||
| 696 | // Header and idom(Exit). | |||
| 697 | // So the idom'(Exit) is nearest common dominator of idom(Exit)' and Latch'. | |||
| 698 | // All these basic blocks are in the same loop, so what we find is | |||
| 699 | // (nearest common dominator of idom(Exit) and Latch)'. | |||
| 700 | // In the loop below we remember nearest common dominator of idom(Exit) and | |||
| 701 | // Latch to update idom of Exit later. | |||
| 702 | assert(L->hasDedicatedExits() && "No dedicated exits?")((void)0); | |||
| 703 | for (auto Edge : ExitEdges) { | |||
| 704 | if (ExitIDom.count(Edge.second)) | |||
| 705 | continue; | |||
| 706 | BasicBlock *BB = DT->findNearestCommonDominator( | |||
| 707 | DT->getNode(Edge.second)->getIDom()->getBlock(), Latch); | |||
| 708 | assert(L->contains(BB) && "IDom is not in a loop")((void)0); | |||
| 709 | ExitIDom[Edge.second] = BB; | |||
| 710 | } | |||
| 711 | } | |||
| 712 | ||||
| 713 | Function *F = Header->getParent(); | |||
| 714 | ||||
| 715 | // Set up all the necessary basic blocks. It is convenient to split the | |||
| 716 | // preheader into 3 parts - two blocks to anchor the peeled copy of the loop | |||
| 717 | // body, and a new preheader for the "real" loop. | |||
| 718 | ||||
| 719 | // Peeling the first iteration transforms. | |||
| 720 | // | |||
| 721 | // PreHeader: | |||
| 722 | // ... | |||
| 723 | // Header: | |||
| 724 | // LoopBody | |||
| 725 | // If (cond) goto Header | |||
| 726 | // Exit: | |||
| 727 | // | |||
| 728 | // into | |||
| 729 | // | |||
| 730 | // InsertTop: | |||
| 731 | // LoopBody | |||
| 732 | // If (!cond) goto Exit | |||
| 733 | // InsertBot: | |||
| 734 | // NewPreHeader: | |||
| 735 | // ... | |||
| 736 | // Header: | |||
| 737 | // LoopBody | |||
| 738 | // If (cond) goto Header | |||
| 739 | // Exit: | |||
| 740 | // | |||
| 741 | // Each following iteration will split the current bottom anchor in two, | |||
| 742 | // and put the new copy of the loop body between these two blocks. That is, | |||
| 743 | // after peeling another iteration from the example above, we'll split | |||
| 744 | // InsertBot, and get: | |||
| 745 | // | |||
| 746 | // InsertTop: | |||
| 747 | // LoopBody | |||
| 748 | // If (!cond) goto Exit | |||
| 749 | // InsertBot: | |||
| 750 | // LoopBody | |||
| 751 | // If (!cond) goto Exit | |||
| 752 | // InsertBot.next: | |||
| 753 | // NewPreHeader: | |||
| 754 | // ... | |||
| 755 | // Header: | |||
| 756 | // LoopBody | |||
| 757 | // If (cond) goto Header | |||
| 758 | // Exit: | |||
| 759 | ||||
| 760 | BasicBlock *InsertTop = SplitEdge(PreHeader, Header, DT, LI); | |||
| 761 | BasicBlock *InsertBot = | |||
| 762 | SplitBlock(InsertTop, InsertTop->getTerminator(), DT, LI); | |||
| 763 | BasicBlock *NewPreHeader = | |||
| 764 | SplitBlock(InsertBot, InsertBot->getTerminator(), DT, LI); | |||
| 765 | ||||
| 766 | InsertTop->setName(Header->getName() + ".peel.begin"); | |||
| 767 | InsertBot->setName(Header->getName() + ".peel.next"); | |||
| 768 | NewPreHeader->setName(PreHeader->getName() + ".peel.newph"); | |||
| 769 | ||||
| 770 | ValueToValueMapTy LVMap; | |||
| 771 | ||||
| 772 | // If we have branch weight information, we'll want to update it for the | |||
| 773 | // newly created branches. | |||
| 774 | BranchInst *LatchBR = | |||
| 775 | cast<BranchInst>(cast<BasicBlock>(Latch)->getTerminator()); | |||
| 776 | uint64_t ExitWeight = 0, FallThroughWeight = 0; | |||
| 777 | initBranchWeights(Header, LatchBR, ExitWeight, FallThroughWeight); | |||
| 778 | ||||
| 779 | // Identify what noalias metadata is inside the loop: if it is inside the | |||
| 780 | // loop, the associated metadata must be cloned for each iteration. | |||
| 781 | SmallVector<MDNode *, 6> LoopLocalNoAliasDeclScopes; | |||
| 782 | identifyNoAliasScopesToClone(L->getBlocks(), LoopLocalNoAliasDeclScopes); | |||
| 783 | ||||
| 784 | // For each peeled-off iteration, make a copy of the loop. | |||
| 785 | for (unsigned Iter = 0; Iter < PeelCount; ++Iter) { | |||
| 786 | SmallVector<BasicBlock *, 8> NewBlocks; | |||
| 787 | ValueToValueMapTy VMap; | |||
| 788 | ||||
| 789 | cloneLoopBlocks(L, Iter, InsertTop, InsertBot, ExitEdges, NewBlocks, | |||
| 790 | LoopBlocks, VMap, LVMap, DT, LI, | |||
| 791 | LoopLocalNoAliasDeclScopes); | |||
| 792 | ||||
| 793 | // Remap to use values from the current iteration instead of the | |||
| 794 | // previous one. | |||
| 795 | remapInstructionsInBlocks(NewBlocks, VMap); | |||
| 796 | ||||
| 797 | if (DT) { | |||
| 798 | // Latches of the cloned loops dominate over the loop exit, so idom of the | |||
| 799 | // latter is the first cloned loop body, as original PreHeader dominates | |||
| 800 | // the original loop body. | |||
| 801 | if (Iter == 0) | |||
| 802 | for (auto Exit : ExitIDom) | |||
| 803 | DT->changeImmediateDominator(Exit.first, | |||
| 804 | cast<BasicBlock>(LVMap[Exit.second])); | |||
| 805 | #ifdef EXPENSIVE_CHECKS | |||
| 806 | assert(DT->verify(DominatorTree::VerificationLevel::Fast))((void)0); | |||
| 807 | #endif | |||
| 808 | } | |||
| 809 | ||||
| 810 | auto *LatchBRCopy = cast<BranchInst>(VMap[LatchBR]); | |||
| 811 | updateBranchWeights(InsertBot, LatchBRCopy, ExitWeight, FallThroughWeight); | |||
| 812 | // Remove Loop metadata from the latch branch instruction | |||
| 813 | // because it is not the Loop's latch branch anymore. | |||
| 814 | LatchBRCopy->setMetadata(LLVMContext::MD_loop, nullptr); | |||
| 815 | ||||
| 816 | InsertTop = InsertBot; | |||
| 817 | InsertBot = SplitBlock(InsertBot, InsertBot->getTerminator(), DT, LI); | |||
| 818 | InsertBot->setName(Header->getName() + ".peel.next"); | |||
| 819 | ||||
| 820 | F->getBasicBlockList().splice(InsertTop->getIterator(), | |||
| 821 | F->getBasicBlockList(), | |||
| 822 | NewBlocks[0]->getIterator(), F->end()); | |||
| 823 | } | |||
| 824 | ||||
| 825 | // Now adjust the phi nodes in the loop header to get their initial values | |||
| 826 | // from the last peeled-off iteration instead of the preheader. | |||
| 827 | for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { | |||
| 828 | PHINode *PHI = cast<PHINode>(I); | |||
| 829 | Value *NewVal = PHI->getIncomingValueForBlock(Latch); | |||
| 830 | Instruction *LatchInst = dyn_cast<Instruction>(NewVal); | |||
| 831 | if (LatchInst && L->contains(LatchInst)) | |||
| 832 | NewVal = LVMap[LatchInst]; | |||
| 833 | ||||
| 834 | PHI->setIncomingValueForBlock(NewPreHeader, NewVal); | |||
| 835 | } | |||
| 836 | ||||
| 837 | fixupBranchWeights(Header, LatchBR, ExitWeight, FallThroughWeight); | |||
| 838 | ||||
| 839 | // Update Metadata for count of peeled off iterations. | |||
| 840 | unsigned AlreadyPeeled = 0; | |||
| 841 | if (auto Peeled = getOptionalIntLoopAttribute(L, PeeledCountMetaData)) | |||
| 842 | AlreadyPeeled = *Peeled; | |||
| 843 | addStringMetadataToLoop(L, PeeledCountMetaData, AlreadyPeeled + PeelCount); | |||
| 844 | ||||
| 845 | if (Loop *ParentLoop = L->getParentLoop()) | |||
| 846 | L = ParentLoop; | |||
| 847 | ||||
| 848 | // We modified the loop, update SE. | |||
| 849 | SE->forgetTopmostLoop(L); | |||
| 850 | ||||
| 851 | // Finally DomtTree must be correct. | |||
| 852 | assert(DT->verify(DominatorTree::VerificationLevel::Fast))((void)0); | |||
| 853 | ||||
| 854 | // FIXME: Incrementally update loop-simplify | |||
| 855 | simplifyLoop(L, DT, LI, SE, AC, nullptr, PreserveLCSSA); | |||
| 856 | ||||
| 857 | NumPeeled++; | |||
| 858 | ||||
| 859 | return true; | |||
| 860 | } |
| 1 | //===- GenericDomTree.h - Generic dominator trees for graphs ----*- C++ -*-===// | |||
| 2 | // | |||
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
| 4 | // See https://llvm.org/LICENSE.txt for license information. | |||
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
| 6 | // | |||
| 7 | //===----------------------------------------------------------------------===// | |||
| 8 | /// \file | |||
| 9 | /// | |||
| 10 | /// This file defines a set of templates that efficiently compute a dominator | |||
| 11 | /// tree over a generic graph. This is used typically in LLVM for fast | |||
| 12 | /// dominance queries on the CFG, but is fully generic w.r.t. the underlying | |||
| 13 | /// graph types. | |||
| 14 | /// | |||
| 15 | /// Unlike ADT/* graph algorithms, generic dominator tree has more requirements | |||
| 16 | /// on the graph's NodeRef. The NodeRef should be a pointer and, | |||
| 17 | /// NodeRef->getParent() must return the parent node that is also a pointer. | |||
| 18 | /// | |||
| 19 | /// FIXME: Maybe GenericDomTree needs a TreeTraits, instead of GraphTraits. | |||
| 20 | /// | |||
| 21 | //===----------------------------------------------------------------------===// | |||
| 22 | ||||
| 23 | #ifndef LLVM_SUPPORT_GENERICDOMTREE_H | |||
| 24 | #define LLVM_SUPPORT_GENERICDOMTREE_H | |||
| 25 | ||||
| 26 | #include "llvm/ADT/DenseMap.h" | |||
| 27 | #include "llvm/ADT/GraphTraits.h" | |||
| 28 | #include "llvm/ADT/STLExtras.h" | |||
| 29 | #include "llvm/ADT/SmallPtrSet.h" | |||
| 30 | #include "llvm/ADT/SmallVector.h" | |||
| 31 | #include "llvm/Support/CFGDiff.h" | |||
| 32 | #include "llvm/Support/CFGUpdate.h" | |||
| 33 | #include "llvm/Support/raw_ostream.h" | |||
| 34 | #include <algorithm> | |||
| 35 | #include <cassert> | |||
| 36 | #include <cstddef> | |||
| 37 | #include <iterator> | |||
| 38 | #include <memory> | |||
| 39 | #include <type_traits> | |||
| 40 | #include <utility> | |||
| 41 | ||||
| 42 | namespace llvm { | |||
| 43 | ||||
| 44 | template <typename NodeT, bool IsPostDom> | |||
| 45 | class DominatorTreeBase; | |||
| 46 | ||||
| 47 | namespace DomTreeBuilder { | |||
| 48 | template <typename DomTreeT> | |||
| 49 | struct SemiNCAInfo; | |||
| 50 | } // namespace DomTreeBuilder | |||
| 51 | ||||
| 52 | /// Base class for the actual dominator tree node. | |||
| 53 | template <class NodeT> class DomTreeNodeBase { | |||
| 54 | friend class PostDominatorTree; | |||
| 55 | friend class DominatorTreeBase<NodeT, false>; | |||
| 56 | friend class DominatorTreeBase<NodeT, true>; | |||
| 57 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase<NodeT, false>>; | |||
| 58 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase<NodeT, true>>; | |||
| 59 | ||||
| 60 | NodeT *TheBB; | |||
| 61 | DomTreeNodeBase *IDom; | |||
| 62 | unsigned Level; | |||
| 63 | SmallVector<DomTreeNodeBase *, 4> Children; | |||
| 64 | mutable unsigned DFSNumIn = ~0; | |||
| 65 | mutable unsigned DFSNumOut = ~0; | |||
| 66 | ||||
| 67 | public: | |||
| 68 | DomTreeNodeBase(NodeT *BB, DomTreeNodeBase *iDom) | |||
| 69 | : TheBB(BB), IDom(iDom), Level(IDom ? IDom->Level + 1 : 0) {} | |||
| 70 | ||||
| 71 | using iterator = typename SmallVector<DomTreeNodeBase *, 4>::iterator; | |||
| 72 | using const_iterator = | |||
| 73 | typename SmallVector<DomTreeNodeBase *, 4>::const_iterator; | |||
| 74 | ||||
| 75 | iterator begin() { return Children.begin(); } | |||
| 76 | iterator end() { return Children.end(); } | |||
| 77 | const_iterator begin() const { return Children.begin(); } | |||
| 78 | const_iterator end() const { return Children.end(); } | |||
| 79 | ||||
| 80 | DomTreeNodeBase *const &back() const { return Children.back(); } | |||
| 81 | DomTreeNodeBase *&back() { return Children.back(); } | |||
| 82 | ||||
| 83 | iterator_range<iterator> children() { return make_range(begin(), end()); } | |||
| 84 | iterator_range<const_iterator> children() const { | |||
| 85 | return make_range(begin(), end()); | |||
| 86 | } | |||
| 87 | ||||
| 88 | NodeT *getBlock() const { return TheBB; } | |||
| 89 | DomTreeNodeBase *getIDom() const { return IDom; } | |||
| 90 | unsigned getLevel() const { return Level; } | |||
| 91 | ||||
| 92 | std::unique_ptr<DomTreeNodeBase> addChild( | |||
| 93 | std::unique_ptr<DomTreeNodeBase> C) { | |||
| 94 | Children.push_back(C.get()); | |||
| 95 | return C; | |||
| 96 | } | |||
| 97 | ||||
| 98 | bool isLeaf() const { return Children.empty(); } | |||
| 99 | size_t getNumChildren() const { return Children.size(); } | |||
| 100 | ||||
| 101 | void clearAllChildren() { Children.clear(); } | |||
| 102 | ||||
| 103 | bool compare(const DomTreeNodeBase *Other) const { | |||
| 104 | if (getNumChildren() != Other->getNumChildren()) | |||
| 105 | return true; | |||
| 106 | ||||
| 107 | if (Level != Other->Level) return true; | |||
| 108 | ||||
| 109 | SmallPtrSet<const NodeT *, 4> OtherChildren; | |||
| 110 | for (const DomTreeNodeBase *I : *Other) { | |||
| 111 | const NodeT *Nd = I->getBlock(); | |||
| 112 | OtherChildren.insert(Nd); | |||
| 113 | } | |||
| 114 | ||||
| 115 | for (const DomTreeNodeBase *I : *this) { | |||
| 116 | const NodeT *N = I->getBlock(); | |||
| 117 | if (OtherChildren.count(N) == 0) | |||
| 118 | return true; | |||
| 119 | } | |||
| 120 | return false; | |||
| 121 | } | |||
| 122 | ||||
| 123 | void setIDom(DomTreeNodeBase *NewIDom) { | |||
| 124 | assert(IDom && "No immediate dominator?")((void)0); | |||
| 125 | if (IDom == NewIDom) return; | |||
| 126 | ||||
| 127 | auto I = find(IDom->Children, this); | |||
| 128 | assert(I != IDom->Children.end() &&((void)0) | |||
| 129 | "Not in immediate dominator children set!")((void)0); | |||
| 130 | // I am no longer your child... | |||
| 131 | IDom->Children.erase(I); | |||
| 132 | ||||
| 133 | // Switch to new dominator | |||
| 134 | IDom = NewIDom; | |||
| 135 | IDom->Children.push_back(this); | |||
| 136 | ||||
| 137 | UpdateLevel(); | |||
| 138 | } | |||
| 139 | ||||
| 140 | /// getDFSNumIn/getDFSNumOut - These return the DFS visitation order for nodes | |||
| 141 | /// in the dominator tree. They are only guaranteed valid if | |||
| 142 | /// updateDFSNumbers() has been called. | |||
| 143 | unsigned getDFSNumIn() const { return DFSNumIn; } | |||
| 144 | unsigned getDFSNumOut() const { return DFSNumOut; } | |||
| 145 | ||||
| 146 | private: | |||
| 147 | // Return true if this node is dominated by other. Use this only if DFS info | |||
| 148 | // is valid. | |||
| 149 | bool DominatedBy(const DomTreeNodeBase *other) const { | |||
| 150 | return this->DFSNumIn >= other->DFSNumIn && | |||
| 151 | this->DFSNumOut <= other->DFSNumOut; | |||
| 152 | } | |||
| 153 | ||||
| 154 | void UpdateLevel() { | |||
| 155 | assert(IDom)((void)0); | |||
| 156 | if (Level == IDom->Level + 1) return; | |||
| 157 | ||||
| 158 | SmallVector<DomTreeNodeBase *, 64> WorkStack = {this}; | |||
| 159 | ||||
| 160 | while (!WorkStack.empty()) { | |||
| 161 | DomTreeNodeBase *Current = WorkStack.pop_back_val(); | |||
| 162 | Current->Level = Current->IDom->Level + 1; | |||
| 163 | ||||
| 164 | for (DomTreeNodeBase *C : *Current) { | |||
| 165 | assert(C->IDom)((void)0); | |||
| 166 | if (C->Level != C->IDom->Level + 1) WorkStack.push_back(C); | |||
| 167 | } | |||
| 168 | } | |||
| 169 | } | |||
| 170 | }; | |||
| 171 | ||||
| 172 | template <class NodeT> | |||
| 173 | raw_ostream &operator<<(raw_ostream &O, const DomTreeNodeBase<NodeT> *Node) { | |||
| 174 | if (Node->getBlock()) | |||
| 175 | Node->getBlock()->printAsOperand(O, false); | |||
| 176 | else | |||
| 177 | O << " <<exit node>>"; | |||
| 178 | ||||
| 179 | O << " {" << Node->getDFSNumIn() << "," << Node->getDFSNumOut() << "} [" | |||
| 180 | << Node->getLevel() << "]\n"; | |||
| 181 | ||||
| 182 | return O; | |||
| 183 | } | |||
| 184 | ||||
| 185 | template <class NodeT> | |||
| 186 | void PrintDomTree(const DomTreeNodeBase<NodeT> *N, raw_ostream &O, | |||
| 187 | unsigned Lev) { | |||
| 188 | O.indent(2 * Lev) << "[" << Lev << "] " << N; | |||
| 189 | for (typename DomTreeNodeBase<NodeT>::const_iterator I = N->begin(), | |||
| 190 | E = N->end(); | |||
| 191 | I != E; ++I) | |||
| 192 | PrintDomTree<NodeT>(*I, O, Lev + 1); | |||
| 193 | } | |||
| 194 | ||||
| 195 | namespace DomTreeBuilder { | |||
| 196 | // The routines below are provided in a separate header but referenced here. | |||
| 197 | template <typename DomTreeT> | |||
| 198 | void Calculate(DomTreeT &DT); | |||
| 199 | ||||
| 200 | template <typename DomTreeT> | |||
| 201 | void CalculateWithUpdates(DomTreeT &DT, | |||
| 202 | ArrayRef<typename DomTreeT::UpdateType> Updates); | |||
| 203 | ||||
| 204 | template <typename DomTreeT> | |||
| 205 | void InsertEdge(DomTreeT &DT, typename DomTreeT::NodePtr From, | |||
| 206 | typename DomTreeT::NodePtr To); | |||
| 207 | ||||
| 208 | template <typename DomTreeT> | |||
| 209 | void DeleteEdge(DomTreeT &DT, typename DomTreeT::NodePtr From, | |||
| 210 | typename DomTreeT::NodePtr To); | |||
| 211 | ||||
| 212 | template <typename DomTreeT> | |||
| 213 | void ApplyUpdates(DomTreeT &DT, | |||
| 214 | GraphDiff<typename DomTreeT::NodePtr, | |||
| 215 | DomTreeT::IsPostDominator> &PreViewCFG, | |||
| 216 | GraphDiff<typename DomTreeT::NodePtr, | |||
| 217 | DomTreeT::IsPostDominator> *PostViewCFG); | |||
| 218 | ||||
| 219 | template <typename DomTreeT> | |||
| 220 | bool Verify(const DomTreeT &DT, typename DomTreeT::VerificationLevel VL); | |||
| 221 | } // namespace DomTreeBuilder | |||
| 222 | ||||
| 223 | /// Core dominator tree base class. | |||
| 224 | /// | |||
| 225 | /// This class is a generic template over graph nodes. It is instantiated for | |||
| 226 | /// various graphs in the LLVM IR or in the code generator. | |||
| 227 | template <typename NodeT, bool IsPostDom> | |||
| 228 | class DominatorTreeBase { | |||
| 229 | public: | |||
| 230 | static_assert(std::is_pointer<typename GraphTraits<NodeT *>::NodeRef>::value, | |||
| 231 | "Currently DominatorTreeBase supports only pointer nodes"); | |||
| 232 | using NodeType = NodeT; | |||
| 233 | using NodePtr = NodeT *; | |||
| 234 | using ParentPtr = decltype(std::declval<NodeT *>()->getParent()); | |||
| 235 | static_assert(std::is_pointer<ParentPtr>::value, | |||
| 236 | "Currently NodeT's parent must be a pointer type"); | |||
| 237 | using ParentType = std::remove_pointer_t<ParentPtr>; | |||
| 238 | static constexpr bool IsPostDominator = IsPostDom; | |||
| 239 | ||||
| 240 | using UpdateType = cfg::Update<NodePtr>; | |||
| 241 | using UpdateKind = cfg::UpdateKind; | |||
| 242 | static constexpr UpdateKind Insert = UpdateKind::Insert; | |||
| 243 | static constexpr UpdateKind Delete = UpdateKind::Delete; | |||
| 244 | ||||
| 245 | enum class VerificationLevel { Fast, Basic, Full }; | |||
| 246 | ||||
| 247 | protected: | |||
| 248 | // Dominators always have a single root, postdominators can have more. | |||
| 249 | SmallVector<NodeT *, IsPostDom ? 4 : 1> Roots; | |||
| 250 | ||||
| 251 | using DomTreeNodeMapType = | |||
| 252 | DenseMap<NodeT *, std::unique_ptr<DomTreeNodeBase<NodeT>>>; | |||
| 253 | DomTreeNodeMapType DomTreeNodes; | |||
| 254 | DomTreeNodeBase<NodeT> *RootNode = nullptr; | |||
| 255 | ParentPtr Parent = nullptr; | |||
| 256 | ||||
| 257 | mutable bool DFSInfoValid = false; | |||
| 258 | mutable unsigned int SlowQueries = 0; | |||
| 259 | ||||
| 260 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase>; | |||
| 261 | ||||
| 262 | public: | |||
| 263 | DominatorTreeBase() {} | |||
| 264 | ||||
| 265 | DominatorTreeBase(DominatorTreeBase &&Arg) | |||
| 266 | : Roots(std::move(Arg.Roots)), | |||
| 267 | DomTreeNodes(std::move(Arg.DomTreeNodes)), | |||
| 268 | RootNode(Arg.RootNode), | |||
| 269 | Parent(Arg.Parent), | |||
| 270 | DFSInfoValid(Arg.DFSInfoValid), | |||
| 271 | SlowQueries(Arg.SlowQueries) { | |||
| 272 | Arg.wipe(); | |||
| 273 | } | |||
| 274 | ||||
| 275 | DominatorTreeBase &operator=(DominatorTreeBase &&RHS) { | |||
| 276 | Roots = std::move(RHS.Roots); | |||
| 277 | DomTreeNodes = std::move(RHS.DomTreeNodes); | |||
| 278 | RootNode = RHS.RootNode; | |||
| 279 | Parent = RHS.Parent; | |||
| 280 | DFSInfoValid = RHS.DFSInfoValid; | |||
| 281 | SlowQueries = RHS.SlowQueries; | |||
| 282 | RHS.wipe(); | |||
| 283 | return *this; | |||
| 284 | } | |||
| 285 | ||||
| 286 | DominatorTreeBase(const DominatorTreeBase &) = delete; | |||
| 287 | DominatorTreeBase &operator=(const DominatorTreeBase &) = delete; | |||
| 288 | ||||
| 289 | /// Iteration over roots. | |||
| 290 | /// | |||
| 291 | /// This may include multiple blocks if we are computing post dominators. | |||
| 292 | /// For forward dominators, this will always be a single block (the entry | |||
| 293 | /// block). | |||
| 294 | using root_iterator = typename SmallVectorImpl<NodeT *>::iterator; | |||
| 295 | using const_root_iterator = typename SmallVectorImpl<NodeT *>::const_iterator; | |||
| 296 | ||||
| 297 | root_iterator root_begin() { return Roots.begin(); } | |||
| 298 | const_root_iterator root_begin() const { return Roots.begin(); } | |||
| 299 | root_iterator root_end() { return Roots.end(); } | |||
| 300 | const_root_iterator root_end() const { return Roots.end(); } | |||
| 301 | ||||
| 302 | size_t root_size() const { return Roots.size(); } | |||
| 303 | ||||
| 304 | iterator_range<root_iterator> roots() { | |||
| 305 | return make_range(root_begin(), root_end()); | |||
| 306 | } | |||
| 307 | iterator_range<const_root_iterator> roots() const { | |||
| 308 | return make_range(root_begin(), root_end()); | |||
| 309 | } | |||
| 310 | ||||
| 311 | /// isPostDominator - Returns true if analysis based of postdoms | |||
| 312 | /// | |||
| 313 | bool isPostDominator() const { return IsPostDominator; } | |||
| 314 | ||||
| 315 | /// compare - Return false if the other dominator tree base matches this | |||
| 316 | /// dominator tree base. Otherwise return true. | |||
| 317 | bool compare(const DominatorTreeBase &Other) const { | |||
| 318 | if (Parent != Other.Parent) return true; | |||
| 319 | ||||
| 320 | if (Roots.size() != Other.Roots.size()) | |||
| 321 | return true; | |||
| 322 | ||||
| 323 | if (!std::is_permutation(Roots.begin(), Roots.end(), Other.Roots.begin())) | |||
| 324 | return true; | |||
| 325 | ||||
| 326 | const DomTreeNodeMapType &OtherDomTreeNodes = Other.DomTreeNodes; | |||
| 327 | if (DomTreeNodes.size() != OtherDomTreeNodes.size()) | |||
| 328 | return true; | |||
| 329 | ||||
| 330 | for (const auto &DomTreeNode : DomTreeNodes) { | |||
| 331 | NodeT *BB = DomTreeNode.first; | |||
| 332 | typename DomTreeNodeMapType::const_iterator OI = | |||
| 333 | OtherDomTreeNodes.find(BB); | |||
| 334 | if (OI == OtherDomTreeNodes.end()) | |||
| 335 | return true; | |||
| 336 | ||||
| 337 | DomTreeNodeBase<NodeT> &MyNd = *DomTreeNode.second; | |||
| 338 | DomTreeNodeBase<NodeT> &OtherNd = *OI->second; | |||
| 339 | ||||
| 340 | if (MyNd.compare(&OtherNd)) | |||
| 341 | return true; | |||
| 342 | } | |||
| 343 | ||||
| 344 | return false; | |||
| 345 | } | |||
| 346 | ||||
| 347 | /// getNode - return the (Post)DominatorTree node for the specified basic | |||
| 348 | /// block. This is the same as using operator[] on this class. The result | |||
| 349 | /// may (but is not required to) be null for a forward (backwards) | |||
| 350 | /// statically unreachable block. | |||
| 351 | DomTreeNodeBase<NodeT> *getNode(const NodeT *BB) const { | |||
| 352 | auto I = DomTreeNodes.find(BB); | |||
| 353 | if (I != DomTreeNodes.end()) | |||
| 354 | return I->second.get(); | |||
| 355 | return nullptr; | |||
| 356 | } | |||
| 357 | ||||
| 358 | /// See getNode. | |||
| 359 | DomTreeNodeBase<NodeT> *operator[](const NodeT *BB) const { | |||
| 360 | return getNode(BB); | |||
| 361 | } | |||
| 362 | ||||
| 363 | /// getRootNode - This returns the entry node for the CFG of the function. If | |||
| 364 | /// this tree represents the post-dominance relations for a function, however, | |||
| 365 | /// this root may be a node with the block == NULL. This is the case when | |||
| 366 | /// there are multiple exit nodes from a particular function. Consumers of | |||
| 367 | /// post-dominance information must be capable of dealing with this | |||
| 368 | /// possibility. | |||
| 369 | /// | |||
| 370 | DomTreeNodeBase<NodeT> *getRootNode() { return RootNode; } | |||
| 371 | const DomTreeNodeBase<NodeT> *getRootNode() const { return RootNode; } | |||
| 372 | ||||
| 373 | /// Get all nodes dominated by R, including R itself. | |||
| 374 | void getDescendants(NodeT *R, SmallVectorImpl<NodeT *> &Result) const { | |||
| 375 | Result.clear(); | |||
| 376 | const DomTreeNodeBase<NodeT> *RN = getNode(R); | |||
| 377 | if (!RN) | |||
| 378 | return; // If R is unreachable, it will not be present in the DOM tree. | |||
| 379 | SmallVector<const DomTreeNodeBase<NodeT> *, 8> WL; | |||
| 380 | WL.push_back(RN); | |||
| 381 | ||||
| 382 | while (!WL.empty()) { | |||
| 383 | const DomTreeNodeBase<NodeT> *N = WL.pop_back_val(); | |||
| 384 | Result.push_back(N->getBlock()); | |||
| 385 | WL.append(N->begin(), N->end()); | |||
| 386 | } | |||
| 387 | } | |||
| 388 | ||||
| 389 | /// properlyDominates - Returns true iff A dominates B and A != B. | |||
| 390 | /// Note that this is not a constant time operation! | |||
| 391 | /// | |||
| 392 | bool properlyDominates(const DomTreeNodeBase<NodeT> *A, | |||
| 393 | const DomTreeNodeBase<NodeT> *B) const { | |||
| 394 | if (!A || !B) | |||
| 395 | return false; | |||
| 396 | if (A == B) | |||
| 397 | return false; | |||
| 398 | return dominates(A, B); | |||
| 399 | } | |||
| 400 | ||||
| 401 | bool properlyDominates(const NodeT *A, const NodeT *B) const; | |||
| 402 | ||||
| 403 | /// isReachableFromEntry - Return true if A is dominated by the entry | |||
| 404 | /// block of the function containing it. | |||
| 405 | bool isReachableFromEntry(const NodeT *A) const { | |||
| 406 | assert(!this->isPostDominator() &&((void)0) | |||
| 407 | "This is not implemented for post dominators")((void)0); | |||
| 408 | return isReachableFromEntry(getNode(const_cast<NodeT *>(A))); | |||
| 409 | } | |||
| 410 | ||||
| 411 | bool isReachableFromEntry(const DomTreeNodeBase<NodeT> *A) const { return A; } | |||
| 412 | ||||
| 413 | /// dominates - Returns true iff A dominates B. Note that this is not a | |||
| 414 | /// constant time operation! | |||
| 415 | /// | |||
| 416 | bool dominates(const DomTreeNodeBase<NodeT> *A, | |||
| 417 | const DomTreeNodeBase<NodeT> *B) const { | |||
| 418 | // A node trivially dominates itself. | |||
| 419 | if (B == A) | |||
| 420 | return true; | |||
| 421 | ||||
| 422 | // An unreachable node is dominated by anything. | |||
| 423 | if (!isReachableFromEntry(B)) | |||
| 424 | return true; | |||
| 425 | ||||
| 426 | // And dominates nothing. | |||
| 427 | if (!isReachableFromEntry(A)) | |||
| 428 | return false; | |||
| 429 | ||||
| 430 | if (B->getIDom() == A) return true; | |||
| 431 | ||||
| 432 | if (A->getIDom() == B) return false; | |||
| 433 | ||||
| 434 | // A can only dominate B if it is higher in the tree. | |||
| 435 | if (A->getLevel() >= B->getLevel()) return false; | |||
| 436 | ||||
| 437 | // Compare the result of the tree walk and the dfs numbers, if expensive | |||
| 438 | // checks are enabled. | |||
| 439 | #ifdef EXPENSIVE_CHECKS | |||
| 440 | assert((!DFSInfoValid ||((void)0) | |||
| 441 | (dominatedBySlowTreeWalk(A, B) == B->DominatedBy(A))) &&((void)0) | |||
| 442 | "Tree walk disagrees with dfs numbers!")((void)0); | |||
| 443 | #endif | |||
| 444 | ||||
| 445 | if (DFSInfoValid) | |||
| 446 | return B->DominatedBy(A); | |||
| 447 | ||||
| 448 | // If we end up with too many slow queries, just update the | |||
| 449 | // DFS numbers on the theory that we are going to keep querying. | |||
| 450 | SlowQueries++; | |||
| 451 | if (SlowQueries > 32) { | |||
| 452 | updateDFSNumbers(); | |||
| 453 | return B->DominatedBy(A); | |||
| 454 | } | |||
| 455 | ||||
| 456 | return dominatedBySlowTreeWalk(A, B); | |||
| 457 | } | |||
| 458 | ||||
| 459 | bool dominates(const NodeT *A, const NodeT *B) const; | |||
| 460 | ||||
| 461 | NodeT *getRoot() const { | |||
| 462 | assert(this->Roots.size() == 1 && "Should always have entry node!")((void)0); | |||
| 463 | return this->Roots[0]; | |||
| 464 | } | |||
| 465 | ||||
| 466 | /// Find nearest common dominator basic block for basic block A and B. A and B | |||
| 467 | /// must have tree nodes. | |||
| 468 | NodeT *findNearestCommonDominator(NodeT *A, NodeT *B) const { | |||
| 469 | assert(A && B && "Pointers are not valid")((void)0); | |||
| 470 | assert(A->getParent() == B->getParent() &&((void)0) | |||
| 471 | "Two blocks are not in same function")((void)0); | |||
| 472 | ||||
| 473 | // If either A or B is a entry block then it is nearest common dominator | |||
| 474 | // (for forward-dominators). | |||
| 475 | if (!isPostDominator()) { | |||
| 476 | NodeT &Entry = A->getParent()->front(); | |||
| 477 | if (A == &Entry || B == &Entry) | |||
| 478 | return &Entry; | |||
| 479 | } | |||
| 480 | ||||
| 481 | DomTreeNodeBase<NodeT> *NodeA = getNode(A); | |||
| 482 | DomTreeNodeBase<NodeT> *NodeB = getNode(B); | |||
| 483 | assert(NodeA && "A must be in the tree")((void)0); | |||
| 484 | assert(NodeB && "B must be in the tree")((void)0); | |||
| 485 | ||||
| 486 | // Use level information to go up the tree until the levels match. Then | |||
| 487 | // continue going up til we arrive at the same node. | |||
| 488 | while (NodeA != NodeB) { | |||
| 489 | if (NodeA->getLevel() < NodeB->getLevel()) std::swap(NodeA, NodeB); | |||
| 490 | ||||
| 491 | NodeA = NodeA->IDom; | |||
| 492 | } | |||
| 493 | ||||
| 494 | return NodeA->getBlock(); | |||
| ||||
| 495 | } | |||
| 496 | ||||
| 497 | const NodeT *findNearestCommonDominator(const NodeT *A, | |||
| 498 | const NodeT *B) const { | |||
| 499 | // Cast away the const qualifiers here. This is ok since | |||
| 500 | // const is re-introduced on the return type. | |||
| 501 | return findNearestCommonDominator(const_cast<NodeT *>(A), | |||
| 502 | const_cast<NodeT *>(B)); | |||
| 503 | } | |||
| 504 | ||||
| 505 | bool isVirtualRoot(const DomTreeNodeBase<NodeT> *A) const { | |||
| 506 | return isPostDominator() && !A->getBlock(); | |||
| 507 | } | |||
| 508 | ||||
| 509 | //===--------------------------------------------------------------------===// | |||
| 510 | // API to update (Post)DominatorTree information based on modifications to | |||
| 511 | // the CFG... | |||
| 512 | ||||
| 513 | /// Inform the dominator tree about a sequence of CFG edge insertions and | |||
| 514 | /// deletions and perform a batch update on the tree. | |||
| 515 | /// | |||
| 516 | /// This function should be used when there were multiple CFG updates after | |||
| 517 | /// the last dominator tree update. It takes care of performing the updates | |||
| 518 | /// in sync with the CFG and optimizes away the redundant operations that | |||
| 519 | /// cancel each other. | |||
| 520 | /// The functions expects the sequence of updates to be balanced. Eg.: | |||
| 521 | /// - {{Insert, A, B}, {Delete, A, B}, {Insert, A, B}} is fine, because | |||
| 522 | /// logically it results in a single insertions. | |||
| 523 | /// - {{Insert, A, B}, {Insert, A, B}} is invalid, because it doesn't make | |||
| 524 | /// sense to insert the same edge twice. | |||
| 525 | /// | |||
| 526 | /// What's more, the functions assumes that it's safe to ask every node in the | |||
| 527 | /// CFG about its children and inverse children. This implies that deletions | |||
| 528 | /// of CFG edges must not delete the CFG nodes before calling this function. | |||
| 529 | /// | |||
| 530 | /// The applyUpdates function can reorder the updates and remove redundant | |||
| 531 | /// ones internally. The batch updater is also able to detect sequences of | |||
| 532 | /// zero and exactly one update -- it's optimized to do less work in these | |||
| 533 | /// cases. | |||
| 534 | /// | |||
| 535 | /// Note that for postdominators it automatically takes care of applying | |||
| 536 | /// updates on reverse edges internally (so there's no need to swap the | |||
| 537 | /// From and To pointers when constructing DominatorTree::UpdateType). | |||
| 538 | /// The type of updates is the same for DomTreeBase<T> and PostDomTreeBase<T> | |||
| 539 | /// with the same template parameter T. | |||
| 540 | /// | |||
| 541 | /// \param Updates An unordered sequence of updates to perform. The current | |||
| 542 | /// CFG and the reverse of these updates provides the pre-view of the CFG. | |||
| 543 | /// | |||
| 544 | void applyUpdates(ArrayRef<UpdateType> Updates) { | |||
| 545 | GraphDiff<NodePtr, IsPostDominator> PreViewCFG( | |||
| 546 | Updates, /*ReverseApplyUpdates=*/true); | |||
| 547 | DomTreeBuilder::ApplyUpdates(*this, PreViewCFG, nullptr); | |||
| 548 | } | |||
| 549 | ||||
| 550 | /// \param Updates An unordered sequence of updates to perform. The current | |||
| 551 | /// CFG and the reverse of these updates provides the pre-view of the CFG. | |||
| 552 | /// \param PostViewUpdates An unordered sequence of update to perform in order | |||
| 553 | /// to obtain a post-view of the CFG. The DT will be updated assuming the | |||
| 554 | /// obtained PostViewCFG is the desired end state. | |||
| 555 | void applyUpdates(ArrayRef<UpdateType> Updates, | |||
| 556 | ArrayRef<UpdateType> PostViewUpdates) { | |||
| 557 | if (Updates.empty()) { | |||
| 558 | GraphDiff<NodePtr, IsPostDom> PostViewCFG(PostViewUpdates); | |||
| 559 | DomTreeBuilder::ApplyUpdates(*this, PostViewCFG, &PostViewCFG); | |||
| 560 | } else { | |||
| 561 | // PreViewCFG needs to merge Updates and PostViewCFG. The updates in | |||
| 562 | // Updates need to be reversed, and match the direction in PostViewCFG. | |||
| 563 | // The PostViewCFG is created with updates reversed (equivalent to changes | |||
| 564 | // made to the CFG), so the PreViewCFG needs all the updates reverse | |||
| 565 | // applied. | |||
| 566 | SmallVector<UpdateType> AllUpdates(Updates.begin(), Updates.end()); | |||
| 567 | append_range(AllUpdates, PostViewUpdates); | |||
| 568 | GraphDiff<NodePtr, IsPostDom> PreViewCFG(AllUpdates, | |||
| 569 | /*ReverseApplyUpdates=*/true); | |||
| 570 | GraphDiff<NodePtr, IsPostDom> PostViewCFG(PostViewUpdates); | |||
| 571 | DomTreeBuilder::ApplyUpdates(*this, PreViewCFG, &PostViewCFG); | |||
| 572 | } | |||
| 573 | } | |||
| 574 | ||||
| 575 | /// Inform the dominator tree about a CFG edge insertion and update the tree. | |||
| 576 | /// | |||
| 577 | /// This function has to be called just before or just after making the update | |||
| 578 | /// on the actual CFG. There cannot be any other updates that the dominator | |||
| 579 | /// tree doesn't know about. | |||
| 580 | /// | |||
| 581 | /// Note that for postdominators it automatically takes care of inserting | |||
| 582 | /// a reverse edge internally (so there's no need to swap the parameters). | |||
| 583 | /// | |||
| 584 | void insertEdge(NodeT *From, NodeT *To) { | |||
| 585 | assert(From)((void)0); | |||
| 586 | assert(To)((void)0); | |||
| 587 | assert(From->getParent() == Parent)((void)0); | |||
| 588 | assert(To->getParent() == Parent)((void)0); | |||
| 589 | DomTreeBuilder::InsertEdge(*this, From, To); | |||
| 590 | } | |||
| 591 | ||||
| 592 | /// Inform the dominator tree about a CFG edge deletion and update the tree. | |||
| 593 | /// | |||
| 594 | /// This function has to be called just after making the update on the actual | |||
| 595 | /// CFG. An internal functions checks if the edge doesn't exist in the CFG in | |||
| 596 | /// DEBUG mode. There cannot be any other updates that the | |||
| 597 | /// dominator tree doesn't know about. | |||
| 598 | /// | |||
| 599 | /// Note that for postdominators it automatically takes care of deleting | |||
| 600 | /// a reverse edge internally (so there's no need to swap the parameters). | |||
| 601 | /// | |||
| 602 | void deleteEdge(NodeT *From, NodeT *To) { | |||
| 603 | assert(From)((void)0); | |||
| 604 | assert(To)((void)0); | |||
| 605 | assert(From->getParent() == Parent)((void)0); | |||
| 606 | assert(To->getParent() == Parent)((void)0); | |||
| 607 | DomTreeBuilder::DeleteEdge(*this, From, To); | |||
| 608 | } | |||
| 609 | ||||
| 610 | /// Add a new node to the dominator tree information. | |||
| 611 | /// | |||
| 612 | /// This creates a new node as a child of DomBB dominator node, linking it | |||
| 613 | /// into the children list of the immediate dominator. | |||
| 614 | /// | |||
| 615 | /// \param BB New node in CFG. | |||
| 616 | /// \param DomBB CFG node that is dominator for BB. | |||
| 617 | /// \returns New dominator tree node that represents new CFG node. | |||
| 618 | /// | |||
| 619 | DomTreeNodeBase<NodeT> *addNewBlock(NodeT *BB, NodeT *DomBB) { | |||
| 620 | assert(getNode(BB) == nullptr && "Block already in dominator tree!")((void)0); | |||
| 621 | DomTreeNodeBase<NodeT> *IDomNode = getNode(DomBB); | |||
| 622 | assert(IDomNode && "Not immediate dominator specified for block!")((void)0); | |||
| 623 | DFSInfoValid = false; | |||
| 624 | return createChild(BB, IDomNode); | |||
| 625 | } | |||
| 626 | ||||
| 627 | /// Add a new node to the forward dominator tree and make it a new root. | |||
| 628 | /// | |||
| 629 | /// \param BB New node in CFG. | |||
| 630 | /// \returns New dominator tree node that represents new CFG node. | |||
| 631 | /// | |||
| 632 | DomTreeNodeBase<NodeT> *setNewRoot(NodeT *BB) { | |||
| 633 | assert(getNode(BB) == nullptr && "Block already in dominator tree!")((void)0); | |||
| 634 | assert(!this->isPostDominator() &&((void)0) | |||
| 635 | "Cannot change root of post-dominator tree")((void)0); | |||
| 636 | DFSInfoValid = false; | |||
| 637 | DomTreeNodeBase<NodeT> *NewNode = createNode(BB); | |||
| 638 | if (Roots.empty()) { | |||
| 639 | addRoot(BB); | |||
| 640 | } else { | |||
| 641 | assert(Roots.size() == 1)((void)0); | |||
| 642 | NodeT *OldRoot = Roots.front(); | |||
| 643 | auto &OldNode = DomTreeNodes[OldRoot]; | |||
| 644 | OldNode = NewNode->addChild(std::move(DomTreeNodes[OldRoot])); | |||
| 645 | OldNode->IDom = NewNode; | |||
| 646 | OldNode->UpdateLevel(); | |||
| 647 | Roots[0] = BB; | |||
| 648 | } | |||
| 649 | return RootNode = NewNode; | |||
| 650 | } | |||
| 651 | ||||
| 652 | /// changeImmediateDominator - This method is used to update the dominator | |||
| 653 | /// tree information when a node's immediate dominator changes. | |||
| 654 | /// | |||
| 655 | void changeImmediateDominator(DomTreeNodeBase<NodeT> *N, | |||
| 656 | DomTreeNodeBase<NodeT> *NewIDom) { | |||
| 657 | assert(N && NewIDom && "Cannot change null node pointers!")((void)0); | |||
| 658 | DFSInfoValid = false; | |||
| 659 | N->setIDom(NewIDom); | |||
| 660 | } | |||
| 661 | ||||
| 662 | void changeImmediateDominator(NodeT *BB, NodeT *NewBB) { | |||
| 663 | changeImmediateDominator(getNode(BB), getNode(NewBB)); | |||
| 664 | } | |||
| 665 | ||||
| 666 | /// eraseNode - Removes a node from the dominator tree. Block must not | |||
| 667 | /// dominate any other blocks. Removes node from its immediate dominator's | |||
| 668 | /// children list. Deletes dominator node associated with basic block BB. | |||
| 669 | void eraseNode(NodeT *BB) { | |||
| 670 | DomTreeNodeBase<NodeT> *Node = getNode(BB); | |||
| 671 | assert(Node && "Removing node that isn't in dominator tree.")((void)0); | |||
| 672 | assert(Node->isLeaf() && "Node is not a leaf node.")((void)0); | |||
| 673 | ||||
| 674 | DFSInfoValid = false; | |||
| 675 | ||||
| 676 | // Remove node from immediate dominator's children list. | |||
| 677 | DomTreeNodeBase<NodeT> *IDom = Node->getIDom(); | |||
| 678 | if (IDom) { | |||
| 679 | const auto I = find(IDom->Children, Node); | |||
| 680 | assert(I != IDom->Children.end() &&((void)0) | |||
| 681 | "Not in immediate dominator children set!")((void)0); | |||
| 682 | // I am no longer your child... | |||
| 683 | IDom->Children.erase(I); | |||
| 684 | } | |||
| 685 | ||||
| 686 | DomTreeNodes.erase(BB); | |||
| 687 | ||||
| 688 | if (!IsPostDom) return; | |||
| 689 | ||||
| 690 | // Remember to update PostDominatorTree roots. | |||
| 691 | auto RIt = llvm::find(Roots, BB); | |||
| 692 | if (RIt != Roots.end()) { | |||
| 693 | std::swap(*RIt, Roots.back()); | |||
| 694 | Roots.pop_back(); | |||
| 695 | } | |||
| 696 | } | |||
| 697 | ||||
| 698 | /// splitBlock - BB is split and now it has one successor. Update dominator | |||
| 699 | /// tree to reflect this change. | |||
| 700 | void splitBlock(NodeT *NewBB) { | |||
| 701 | if (IsPostDominator) | |||
| 702 | Split<Inverse<NodeT *>>(NewBB); | |||
| 703 | else | |||
| 704 | Split<NodeT *>(NewBB); | |||
| 705 | } | |||
| 706 | ||||
| 707 | /// print - Convert to human readable form | |||
| 708 | /// | |||
| 709 | void print(raw_ostream &O) const { | |||
| 710 | O << "=============================--------------------------------\n"; | |||
| 711 | if (IsPostDominator) | |||
| 712 | O << "Inorder PostDominator Tree: "; | |||
| 713 | else | |||
| 714 | O << "Inorder Dominator Tree: "; | |||
| 715 | if (!DFSInfoValid) | |||
| 716 | O << "DFSNumbers invalid: " << SlowQueries << " slow queries."; | |||
| 717 | O << "\n"; | |||
| 718 | ||||
| 719 | // The postdom tree can have a null root if there are no returns. | |||
| 720 | if (getRootNode()) PrintDomTree<NodeT>(getRootNode(), O, 1); | |||
| 721 | O << "Roots: "; | |||
| 722 | for (const NodePtr Block : Roots) { | |||
| 723 | Block->printAsOperand(O, false); | |||
| 724 | O << " "; | |||
| 725 | } | |||
| 726 | O << "\n"; | |||
| 727 | } | |||
| 728 | ||||
| 729 | public: | |||
| 730 | /// updateDFSNumbers - Assign In and Out numbers to the nodes while walking | |||
| 731 | /// dominator tree in dfs order. | |||
| 732 | void updateDFSNumbers() const { | |||
| 733 | if (DFSInfoValid) { | |||
| 734 | SlowQueries = 0; | |||
| 735 | return; | |||
| 736 | } | |||
| 737 | ||||
| 738 | SmallVector<std::pair<const DomTreeNodeBase<NodeT> *, | |||
| 739 | typename DomTreeNodeBase<NodeT>::const_iterator>, | |||
| 740 | 32> WorkStack; | |||
| 741 | ||||
| 742 | const DomTreeNodeBase<NodeT> *ThisRoot = getRootNode(); | |||
| 743 | assert((!Parent || ThisRoot) && "Empty constructed DomTree")((void)0); | |||
| 744 | if (!ThisRoot) | |||
| 745 | return; | |||
| 746 | ||||
| 747 | // Both dominators and postdominators have a single root node. In the case | |||
| 748 | // case of PostDominatorTree, this node is a virtual root. | |||
| 749 | WorkStack.push_back({ThisRoot, ThisRoot->begin()}); | |||
| 750 | ||||
| 751 | unsigned DFSNum = 0; | |||
| 752 | ThisRoot->DFSNumIn = DFSNum++; | |||
| 753 | ||||
| 754 | while (!WorkStack.empty()) { | |||
| 755 | const DomTreeNodeBase<NodeT> *Node = WorkStack.back().first; | |||
| 756 | const auto ChildIt = WorkStack.back().second; | |||
| 757 | ||||
| 758 | // If we visited all of the children of this node, "recurse" back up the | |||
| 759 | // stack setting the DFOutNum. | |||
| 760 | if (ChildIt == Node->end()) { | |||
| 761 | Node->DFSNumOut = DFSNum++; | |||
| 762 | WorkStack.pop_back(); | |||
| 763 | } else { | |||
| 764 | // Otherwise, recursively visit this child. | |||
| 765 | const DomTreeNodeBase<NodeT> *Child = *ChildIt; | |||
| 766 | ++WorkStack.back().second; | |||
| 767 | ||||
| 768 | WorkStack.push_back({Child, Child->begin()}); | |||
| 769 | Child->DFSNumIn = DFSNum++; | |||
| 770 | } | |||
| 771 | } | |||
| 772 | ||||
| 773 | SlowQueries = 0; | |||
| 774 | DFSInfoValid = true; | |||
| 775 | } | |||
| 776 | ||||
| 777 | /// recalculate - compute a dominator tree for the given function | |||
| 778 | void recalculate(ParentType &Func) { | |||
| 779 | Parent = &Func; | |||
| 780 | DomTreeBuilder::Calculate(*this); | |||
| 781 | } | |||
| 782 | ||||
| 783 | void recalculate(ParentType &Func, ArrayRef<UpdateType> Updates) { | |||
| 784 | Parent = &Func; | |||
| 785 | DomTreeBuilder::CalculateWithUpdates(*this, Updates); | |||
| 786 | } | |||
| 787 | ||||
| 788 | /// verify - checks if the tree is correct. There are 3 level of verification: | |||
| 789 | /// - Full -- verifies if the tree is correct by making sure all the | |||
| 790 | /// properties (including the parent and the sibling property) | |||
| 791 | /// hold. | |||
| 792 | /// Takes O(N^3) time. | |||
| 793 | /// | |||
| 794 | /// - Basic -- checks if the tree is correct, but compares it to a freshly | |||
| 795 | /// constructed tree instead of checking the sibling property. | |||
| 796 | /// Takes O(N^2) time. | |||
| 797 | /// | |||
| 798 | /// - Fast -- checks basic tree structure and compares it with a freshly | |||
| 799 | /// constructed tree. | |||
| 800 | /// Takes O(N^2) time worst case, but is faster in practise (same | |||
| 801 | /// as tree construction). | |||
| 802 | bool verify(VerificationLevel VL = VerificationLevel::Full) const { | |||
| 803 | return DomTreeBuilder::Verify(*this, VL); | |||
| 804 | } | |||
| 805 | ||||
| 806 | void reset() { | |||
| 807 | DomTreeNodes.clear(); | |||
| 808 | Roots.clear(); | |||
| 809 | RootNode = nullptr; | |||
| 810 | Parent = nullptr; | |||
| 811 | DFSInfoValid = false; | |||
| 812 | SlowQueries = 0; | |||
| 813 | } | |||
| 814 | ||||
| 815 | protected: | |||
| 816 | void addRoot(NodeT *BB) { this->Roots.push_back(BB); } | |||
| 817 | ||||
| 818 | DomTreeNodeBase<NodeT> *createChild(NodeT *BB, DomTreeNodeBase<NodeT> *IDom) { | |||
| 819 | return (DomTreeNodes[BB] = IDom->addChild( | |||
| 820 | std::make_unique<DomTreeNodeBase<NodeT>>(BB, IDom))) | |||
| 821 | .get(); | |||
| 822 | } | |||
| 823 | ||||
| 824 | DomTreeNodeBase<NodeT> *createNode(NodeT *BB) { | |||
| 825 | return (DomTreeNodes[BB] = | |||
| 826 | std::make_unique<DomTreeNodeBase<NodeT>>(BB, nullptr)) | |||
| 827 | .get(); | |||
| 828 | } | |||
| 829 | ||||
| 830 | // NewBB is split and now it has one successor. Update dominator tree to | |||
| 831 | // reflect this change. | |||
| 832 | template <class N> | |||
| 833 | void Split(typename GraphTraits<N>::NodeRef NewBB) { | |||
| 834 | using GraphT = GraphTraits<N>; | |||
| 835 | using NodeRef = typename GraphT::NodeRef; | |||
| 836 | assert(std::distance(GraphT::child_begin(NewBB),((void)0) | |||
| 837 | GraphT::child_end(NewBB)) == 1 &&((void)0) | |||
| 838 | "NewBB should have a single successor!")((void)0); | |||
| 839 | NodeRef NewBBSucc = *GraphT::child_begin(NewBB); | |||
| 840 | ||||
| 841 | SmallVector<NodeRef, 4> PredBlocks(children<Inverse<N>>(NewBB)); | |||
| 842 | ||||
| 843 | assert(!PredBlocks.empty() && "No predblocks?")((void)0); | |||
| 844 | ||||
| 845 | bool NewBBDominatesNewBBSucc = true; | |||
| 846 | for (auto Pred : children<Inverse<N>>(NewBBSucc)) { | |||
| 847 | if (Pred != NewBB && !dominates(NewBBSucc, Pred) && | |||
| 848 | isReachableFromEntry(Pred)) { | |||
| 849 | NewBBDominatesNewBBSucc = false; | |||
| 850 | break; | |||
| 851 | } | |||
| 852 | } | |||
| 853 | ||||
| 854 | // Find NewBB's immediate dominator and create new dominator tree node for | |||
| 855 | // NewBB. | |||
| 856 | NodeT *NewBBIDom = nullptr; | |||
| 857 | unsigned i = 0; | |||
| 858 | for (i = 0; i < PredBlocks.size(); ++i) | |||
| 859 | if (isReachableFromEntry(PredBlocks[i])) { | |||
| 860 | NewBBIDom = PredBlocks[i]; | |||
| 861 | break; | |||
| 862 | } | |||
| 863 | ||||
| 864 | // It's possible that none of the predecessors of NewBB are reachable; | |||
| 865 | // in that case, NewBB itself is unreachable, so nothing needs to be | |||
| 866 | // changed. | |||
| 867 | if (!NewBBIDom) return; | |||
| 868 | ||||
| 869 | for (i = i + 1; i < PredBlocks.size(); ++i) { | |||
| 870 | if (isReachableFromEntry(PredBlocks[i])) | |||
| 871 | NewBBIDom = findNearestCommonDominator(NewBBIDom, PredBlocks[i]); | |||
| 872 | } | |||
| 873 | ||||
| 874 | // Create the new dominator tree node... and set the idom of NewBB. | |||
| 875 | DomTreeNodeBase<NodeT> *NewBBNode = addNewBlock(NewBB, NewBBIDom); | |||
| 876 | ||||
| 877 | // If NewBB strictly dominates other blocks, then it is now the immediate | |||
| 878 | // dominator of NewBBSucc. Update the dominator tree as appropriate. | |||
| 879 | if (NewBBDominatesNewBBSucc) { | |||
| 880 | DomTreeNodeBase<NodeT> *NewBBSuccNode = getNode(NewBBSucc); | |||
| 881 | changeImmediateDominator(NewBBSuccNode, NewBBNode); | |||
| 882 | } | |||
| 883 | } | |||
| 884 | ||||
| 885 | private: | |||
| 886 | bool dominatedBySlowTreeWalk(const DomTreeNodeBase<NodeT> *A, | |||
| 887 | const DomTreeNodeBase<NodeT> *B) const { | |||
| 888 | assert(A != B)((void)0); | |||
| 889 | assert(isReachableFromEntry(B))((void)0); | |||
| 890 | assert(isReachableFromEntry(A))((void)0); | |||
| 891 | ||||
| 892 | const unsigned ALevel = A->getLevel(); | |||
| 893 | const DomTreeNodeBase<NodeT> *IDom; | |||
| 894 | ||||
| 895 | // Don't walk nodes above A's subtree. When we reach A's level, we must | |||
| 896 | // either find A or be in some other subtree not dominated by A. | |||
| 897 | while ((IDom = B->getIDom()) != nullptr && IDom->getLevel() >= ALevel) | |||
| 898 | B = IDom; // Walk up the tree | |||
| 899 | ||||
| 900 | return B == A; | |||
| 901 | } | |||
| 902 | ||||
| 903 | /// Wipe this tree's state without releasing any resources. | |||
| 904 | /// | |||
| 905 | /// This is essentially a post-move helper only. It leaves the object in an | |||
| 906 | /// assignable and destroyable state, but otherwise invalid. | |||
| 907 | void wipe() { | |||
| 908 | DomTreeNodes.clear(); | |||
| 909 | RootNode = nullptr; | |||
| 910 | Parent = nullptr; | |||
| 911 | } | |||
| 912 | }; | |||
| 913 | ||||
| 914 | template <typename T> | |||
| 915 | using DomTreeBase = DominatorTreeBase<T, false>; | |||
| 916 | ||||
| 917 | template <typename T> | |||
| 918 | using PostDomTreeBase = DominatorTreeBase<T, true>; | |||
| 919 | ||||
| 920 | // These two functions are declared out of line as a workaround for building | |||
| 921 | // with old (< r147295) versions of clang because of pr11642. | |||
| 922 | template <typename NodeT, bool IsPostDom> | |||
| 923 | bool DominatorTreeBase<NodeT, IsPostDom>::dominates(const NodeT *A, | |||
| 924 | const NodeT *B) const { | |||
| 925 | if (A == B) | |||
| 926 | return true; | |||
| 927 | ||||
| 928 | // Cast away the const qualifiers here. This is ok since | |||
| 929 | // this function doesn't actually return the values returned | |||
| 930 | // from getNode. | |||
| 931 | return dominates(getNode(const_cast<NodeT *>(A)), | |||
| 932 | getNode(const_cast<NodeT *>(B))); | |||
| 933 | } | |||
| 934 | template <typename NodeT, bool IsPostDom> | |||
| 935 | bool DominatorTreeBase<NodeT, IsPostDom>::properlyDominates( | |||
| 936 | const NodeT *A, const NodeT *B) const { | |||
| 937 | if (A == B) | |||
| 938 | return false; | |||
| 939 | ||||
| 940 | // Cast away the const qualifiers here. This is ok since | |||
| 941 | // this function doesn't actually return the values returned | |||
| 942 | // from getNode. | |||
| 943 | return dominates(getNode(const_cast<NodeT *>(A)), | |||
| 944 | getNode(const_cast<NodeT *>(B))); | |||
| 945 | } | |||
| 946 | ||||
| 947 | } // end namespace llvm | |||
| 948 | ||||
| 949 | #endif // LLVM_SUPPORT_GENERICDOMTREE_H |
| 1 | //===- llvm/ADT/DenseMap.h - Dense probed hash table ------------*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file defines the DenseMap class. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #ifndef LLVM_ADT_DENSEMAP_H |
| 14 | #define LLVM_ADT_DENSEMAP_H |
| 15 | |
| 16 | #include "llvm/ADT/DenseMapInfo.h" |
| 17 | #include "llvm/ADT/EpochTracker.h" |
| 18 | #include "llvm/Support/AlignOf.h" |
| 19 | #include "llvm/Support/Compiler.h" |
| 20 | #include "llvm/Support/MathExtras.h" |
| 21 | #include "llvm/Support/MemAlloc.h" |
| 22 | #include "llvm/Support/ReverseIteration.h" |
| 23 | #include "llvm/Support/type_traits.h" |
| 24 | #include <algorithm> |
| 25 | #include <cassert> |
| 26 | #include <cstddef> |
| 27 | #include <cstring> |
| 28 | #include <initializer_list> |
| 29 | #include <iterator> |
| 30 | #include <new> |
| 31 | #include <type_traits> |
| 32 | #include <utility> |
| 33 | |
| 34 | namespace llvm { |
| 35 | |
| 36 | namespace detail { |
| 37 | |
| 38 | // We extend a pair to allow users to override the bucket type with their own |
| 39 | // implementation without requiring two members. |
| 40 | template <typename KeyT, typename ValueT> |
| 41 | struct DenseMapPair : public std::pair<KeyT, ValueT> { |
| 42 | using std::pair<KeyT, ValueT>::pair; |
| 43 | |
| 44 | KeyT &getFirst() { return std::pair<KeyT, ValueT>::first; } |
| 45 | const KeyT &getFirst() const { return std::pair<KeyT, ValueT>::first; } |
| 46 | ValueT &getSecond() { return std::pair<KeyT, ValueT>::second; } |
| 47 | const ValueT &getSecond() const { return std::pair<KeyT, ValueT>::second; } |
| 48 | }; |
| 49 | |
| 50 | } // end namespace detail |
| 51 | |
| 52 | template <typename KeyT, typename ValueT, |
| 53 | typename KeyInfoT = DenseMapInfo<KeyT>, |
| 54 | typename Bucket = llvm::detail::DenseMapPair<KeyT, ValueT>, |
| 55 | bool IsConst = false> |
| 56 | class DenseMapIterator; |
| 57 | |
| 58 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
| 59 | typename BucketT> |
| 60 | class DenseMapBase : public DebugEpochBase { |
| 61 | template <typename T> |
| 62 | using const_arg_type_t = typename const_pointer_or_const_ref<T>::type; |
| 63 | |
| 64 | public: |
| 65 | using size_type = unsigned; |
| 66 | using key_type = KeyT; |
| 67 | using mapped_type = ValueT; |
| 68 | using value_type = BucketT; |
| 69 | |
| 70 | using iterator = DenseMapIterator<KeyT, ValueT, KeyInfoT, BucketT>; |
| 71 | using const_iterator = |
| 72 | DenseMapIterator<KeyT, ValueT, KeyInfoT, BucketT, true>; |
| 73 | |
| 74 | inline iterator begin() { |
| 75 | // When the map is empty, avoid the overhead of advancing/retreating past |
| 76 | // empty buckets. |
| 77 | if (empty()) |
| 78 | return end(); |
| 79 | if (shouldReverseIterate<KeyT>()) |
| 80 | return makeIterator(getBucketsEnd() - 1, getBuckets(), *this); |
| 81 | return makeIterator(getBuckets(), getBucketsEnd(), *this); |
| 82 | } |
| 83 | inline iterator end() { |
| 84 | return makeIterator(getBucketsEnd(), getBucketsEnd(), *this, true); |
| 85 | } |
| 86 | inline const_iterator begin() const { |
| 87 | if (empty()) |
| 88 | return end(); |
| 89 | if (shouldReverseIterate<KeyT>()) |
| 90 | return makeConstIterator(getBucketsEnd() - 1, getBuckets(), *this); |
| 91 | return makeConstIterator(getBuckets(), getBucketsEnd(), *this); |
| 92 | } |
| 93 | inline const_iterator end() const { |
| 94 | return makeConstIterator(getBucketsEnd(), getBucketsEnd(), *this, true); |
| 95 | } |
| 96 | |
| 97 | LLVM_NODISCARD[[clang::warn_unused_result]] bool empty() const { |
| 98 | return getNumEntries() == 0; |
| 99 | } |
| 100 | unsigned size() const { return getNumEntries(); } |
| 101 | |
| 102 | /// Grow the densemap so that it can contain at least \p NumEntries items |
| 103 | /// before resizing again. |
| 104 | void reserve(size_type NumEntries) { |
| 105 | auto NumBuckets = getMinBucketToReserveForEntries(NumEntries); |
| 106 | incrementEpoch(); |
| 107 | if (NumBuckets > getNumBuckets()) |
| 108 | grow(NumBuckets); |
| 109 | } |
| 110 | |
| 111 | void clear() { |
| 112 | incrementEpoch(); |
| 113 | if (getNumEntries() == 0 && getNumTombstones() == 0) return; |
| 114 | |
| 115 | // If the capacity of the array is huge, and the # elements used is small, |
| 116 | // shrink the array. |
| 117 | if (getNumEntries() * 4 < getNumBuckets() && getNumBuckets() > 64) { |
| 118 | shrink_and_clear(); |
| 119 | return; |
| 120 | } |
| 121 | |
| 122 | const KeyT EmptyKey = getEmptyKey(), TombstoneKey = getTombstoneKey(); |
| 123 | if (std::is_trivially_destructible<ValueT>::value) { |
| 124 | // Use a simpler loop when values don't need destruction. |
| 125 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) |
| 126 | P->getFirst() = EmptyKey; |
| 127 | } else { |
| 128 | unsigned NumEntries = getNumEntries(); |
| 129 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) { |
| 130 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey)) { |
| 131 | if (!KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) { |
| 132 | P->getSecond().~ValueT(); |
| 133 | --NumEntries; |
| 134 | } |
| 135 | P->getFirst() = EmptyKey; |
| 136 | } |
| 137 | } |
| 138 | assert(NumEntries == 0 && "Node count imbalance!")((void)0); |
| 139 | } |
| 140 | setNumEntries(0); |
| 141 | setNumTombstones(0); |
| 142 | } |
| 143 | |
| 144 | /// Return 1 if the specified key is in the map, 0 otherwise. |
| 145 | size_type count(const_arg_type_t<KeyT> Val) const { |
| 146 | const BucketT *TheBucket; |
| 147 | return LookupBucketFor(Val, TheBucket) ? 1 : 0; |
| 148 | } |
| 149 | |
| 150 | iterator find(const_arg_type_t<KeyT> Val) { |
| 151 | BucketT *TheBucket; |
| 152 | if (LookupBucketFor(Val, TheBucket)) |
| 153 | return makeIterator(TheBucket, |
| 154 | shouldReverseIterate<KeyT>() ? getBuckets() |
| 155 | : getBucketsEnd(), |
| 156 | *this, true); |
| 157 | return end(); |
| 158 | } |
| 159 | const_iterator find(const_arg_type_t<KeyT> Val) const { |
| 160 | const BucketT *TheBucket; |
| 161 | if (LookupBucketFor(Val, TheBucket)) |
| 162 | return makeConstIterator(TheBucket, |
| 163 | shouldReverseIterate<KeyT>() ? getBuckets() |
| 164 | : getBucketsEnd(), |
| 165 | *this, true); |
| 166 | return end(); |
| 167 | } |
| 168 | |
| 169 | /// Alternate version of find() which allows a different, and possibly |
| 170 | /// less expensive, key type. |
| 171 | /// The DenseMapInfo is responsible for supplying methods |
| 172 | /// getHashValue(LookupKeyT) and isEqual(LookupKeyT, KeyT) for each key |
| 173 | /// type used. |
| 174 | template<class LookupKeyT> |
| 175 | iterator find_as(const LookupKeyT &Val) { |
| 176 | BucketT *TheBucket; |
| 177 | if (LookupBucketFor(Val, TheBucket)) |
| 178 | return makeIterator(TheBucket, |
| 179 | shouldReverseIterate<KeyT>() ? getBuckets() |
| 180 | : getBucketsEnd(), |
| 181 | *this, true); |
| 182 | return end(); |
| 183 | } |
| 184 | template<class LookupKeyT> |
| 185 | const_iterator find_as(const LookupKeyT &Val) const { |
| 186 | const BucketT *TheBucket; |
| 187 | if (LookupBucketFor(Val, TheBucket)) |
| 188 | return makeConstIterator(TheBucket, |
| 189 | shouldReverseIterate<KeyT>() ? getBuckets() |
| 190 | : getBucketsEnd(), |
| 191 | *this, true); |
| 192 | return end(); |
| 193 | } |
| 194 | |
| 195 | /// lookup - Return the entry for the specified key, or a default |
| 196 | /// constructed value if no such entry exists. |
| 197 | ValueT lookup(const_arg_type_t<KeyT> Val) const { |
| 198 | const BucketT *TheBucket; |
| 199 | if (LookupBucketFor(Val, TheBucket)) |
| 200 | return TheBucket->getSecond(); |
| 201 | return ValueT(); |
| 202 | } |
| 203 | |
| 204 | // Inserts key,value pair into the map if the key isn't already in the map. |
| 205 | // If the key is already in the map, it returns false and doesn't update the |
| 206 | // value. |
| 207 | std::pair<iterator, bool> insert(const std::pair<KeyT, ValueT> &KV) { |
| 208 | return try_emplace(KV.first, KV.second); |
| 209 | } |
| 210 | |
| 211 | // Inserts key,value pair into the map if the key isn't already in the map. |
| 212 | // If the key is already in the map, it returns false and doesn't update the |
| 213 | // value. |
| 214 | std::pair<iterator, bool> insert(std::pair<KeyT, ValueT> &&KV) { |
| 215 | return try_emplace(std::move(KV.first), std::move(KV.second)); |
| 216 | } |
| 217 | |
| 218 | // Inserts key,value pair into the map if the key isn't already in the map. |
| 219 | // The value is constructed in-place if the key is not in the map, otherwise |
| 220 | // it is not moved. |
| 221 | template <typename... Ts> |
| 222 | std::pair<iterator, bool> try_emplace(KeyT &&Key, Ts &&... Args) { |
| 223 | BucketT *TheBucket; |
| 224 | if (LookupBucketFor(Key, TheBucket)) |
| 225 | return std::make_pair(makeIterator(TheBucket, |
| 226 | shouldReverseIterate<KeyT>() |
| 227 | ? getBuckets() |
| 228 | : getBucketsEnd(), |
| 229 | *this, true), |
| 230 | false); // Already in map. |
| 231 | |
| 232 | // Otherwise, insert the new element. |
| 233 | TheBucket = |
| 234 | InsertIntoBucket(TheBucket, std::move(Key), std::forward<Ts>(Args)...); |
| 235 | return std::make_pair(makeIterator(TheBucket, |
| 236 | shouldReverseIterate<KeyT>() |
| 237 | ? getBuckets() |
| 238 | : getBucketsEnd(), |
| 239 | *this, true), |
| 240 | true); |
| 241 | } |
| 242 | |
| 243 | // Inserts key,value pair into the map if the key isn't already in the map. |
| 244 | // The value is constructed in-place if the key is not in the map, otherwise |
| 245 | // it is not moved. |
| 246 | template <typename... Ts> |
| 247 | std::pair<iterator, bool> try_emplace(const KeyT &Key, Ts &&... Args) { |
| 248 | BucketT *TheBucket; |
| 249 | if (LookupBucketFor(Key, TheBucket)) |
| 250 | return std::make_pair(makeIterator(TheBucket, |
| 251 | shouldReverseIterate<KeyT>() |
| 252 | ? getBuckets() |
| 253 | : getBucketsEnd(), |
| 254 | *this, true), |
| 255 | false); // Already in map. |
| 256 | |
| 257 | // Otherwise, insert the new element. |
| 258 | TheBucket = InsertIntoBucket(TheBucket, Key, std::forward<Ts>(Args)...); |
| 259 | return std::make_pair(makeIterator(TheBucket, |
| 260 | shouldReverseIterate<KeyT>() |
| 261 | ? getBuckets() |
| 262 | : getBucketsEnd(), |
| 263 | *this, true), |
| 264 | true); |
| 265 | } |
| 266 | |
| 267 | /// Alternate version of insert() which allows a different, and possibly |
| 268 | /// less expensive, key type. |
| 269 | /// The DenseMapInfo is responsible for supplying methods |
| 270 | /// getHashValue(LookupKeyT) and isEqual(LookupKeyT, KeyT) for each key |
| 271 | /// type used. |
| 272 | template <typename LookupKeyT> |
| 273 | std::pair<iterator, bool> insert_as(std::pair<KeyT, ValueT> &&KV, |
| 274 | const LookupKeyT &Val) { |
| 275 | BucketT *TheBucket; |
| 276 | if (LookupBucketFor(Val, TheBucket)) |
| 277 | return std::make_pair(makeIterator(TheBucket, |
| 278 | shouldReverseIterate<KeyT>() |
| 279 | ? getBuckets() |
| 280 | : getBucketsEnd(), |
| 281 | *this, true), |
| 282 | false); // Already in map. |
| 283 | |
| 284 | // Otherwise, insert the new element. |
| 285 | TheBucket = InsertIntoBucketWithLookup(TheBucket, std::move(KV.first), |
| 286 | std::move(KV.second), Val); |
| 287 | return std::make_pair(makeIterator(TheBucket, |
| 288 | shouldReverseIterate<KeyT>() |
| 289 | ? getBuckets() |
| 290 | : getBucketsEnd(), |
| 291 | *this, true), |
| 292 | true); |
| 293 | } |
| 294 | |
| 295 | /// insert - Range insertion of pairs. |
| 296 | template<typename InputIt> |
| 297 | void insert(InputIt I, InputIt E) { |
| 298 | for (; I != E; ++I) |
| 299 | insert(*I); |
| 300 | } |
| 301 | |
| 302 | bool erase(const KeyT &Val) { |
| 303 | BucketT *TheBucket; |
| 304 | if (!LookupBucketFor(Val, TheBucket)) |
| 305 | return false; // not in map. |
| 306 | |
| 307 | TheBucket->getSecond().~ValueT(); |
| 308 | TheBucket->getFirst() = getTombstoneKey(); |
| 309 | decrementNumEntries(); |
| 310 | incrementNumTombstones(); |
| 311 | return true; |
| 312 | } |
| 313 | void erase(iterator I) { |
| 314 | BucketT *TheBucket = &*I; |
| 315 | TheBucket->getSecond().~ValueT(); |
| 316 | TheBucket->getFirst() = getTombstoneKey(); |
| 317 | decrementNumEntries(); |
| 318 | incrementNumTombstones(); |
| 319 | } |
| 320 | |
| 321 | value_type& FindAndConstruct(const KeyT &Key) { |
| 322 | BucketT *TheBucket; |
| 323 | if (LookupBucketFor(Key, TheBucket)) |
| 324 | return *TheBucket; |
| 325 | |
| 326 | return *InsertIntoBucket(TheBucket, Key); |
| 327 | } |
| 328 | |
| 329 | ValueT &operator[](const KeyT &Key) { |
| 330 | return FindAndConstruct(Key).second; |
| 331 | } |
| 332 | |
| 333 | value_type& FindAndConstruct(KeyT &&Key) { |
| 334 | BucketT *TheBucket; |
| 335 | if (LookupBucketFor(Key, TheBucket)) |
| 336 | return *TheBucket; |
| 337 | |
| 338 | return *InsertIntoBucket(TheBucket, std::move(Key)); |
| 339 | } |
| 340 | |
| 341 | ValueT &operator[](KeyT &&Key) { |
| 342 | return FindAndConstruct(std::move(Key)).second; |
| 343 | } |
| 344 | |
| 345 | /// isPointerIntoBucketsArray - Return true if the specified pointer points |
| 346 | /// somewhere into the DenseMap's array of buckets (i.e. either to a key or |
| 347 | /// value in the DenseMap). |
| 348 | bool isPointerIntoBucketsArray(const void *Ptr) const { |
| 349 | return Ptr >= getBuckets() && Ptr < getBucketsEnd(); |
| 350 | } |
| 351 | |
| 352 | /// getPointerIntoBucketsArray() - Return an opaque pointer into the buckets |
| 353 | /// array. In conjunction with the previous method, this can be used to |
| 354 | /// determine whether an insertion caused the DenseMap to reallocate. |
| 355 | const void *getPointerIntoBucketsArray() const { return getBuckets(); } |
| 356 | |
| 357 | protected: |
| 358 | DenseMapBase() = default; |
| 359 | |
| 360 | void destroyAll() { |
| 361 | if (getNumBuckets() == 0) // Nothing to do. |
| 362 | return; |
| 363 | |
| 364 | const KeyT EmptyKey = getEmptyKey(), TombstoneKey = getTombstoneKey(); |
| 365 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) { |
| 366 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey) && |
| 367 | !KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) |
| 368 | P->getSecond().~ValueT(); |
| 369 | P->getFirst().~KeyT(); |
| 370 | } |
| 371 | } |
| 372 | |
| 373 | void initEmpty() { |
| 374 | setNumEntries(0); |
| 375 | setNumTombstones(0); |
| 376 | |
| 377 | assert((getNumBuckets() & (getNumBuckets()-1)) == 0 &&((void)0) |
| 378 | "# initial buckets must be a power of two!")((void)0); |
| 379 | const KeyT EmptyKey = getEmptyKey(); |
| 380 | for (BucketT *B = getBuckets(), *E = getBucketsEnd(); B != E; ++B) |
| 381 | ::new (&B->getFirst()) KeyT(EmptyKey); |
| 382 | } |
| 383 | |
| 384 | /// Returns the number of buckets to allocate to ensure that the DenseMap can |
| 385 | /// accommodate \p NumEntries without need to grow(). |
| 386 | unsigned getMinBucketToReserveForEntries(unsigned NumEntries) { |
| 387 | // Ensure that "NumEntries * 4 < NumBuckets * 3" |
| 388 | if (NumEntries == 0) |
| 389 | return 0; |
| 390 | // +1 is required because of the strict equality. |
| 391 | // For example if NumEntries is 48, we need to return 401. |
| 392 | return NextPowerOf2(NumEntries * 4 / 3 + 1); |
| 393 | } |
| 394 | |
| 395 | void moveFromOldBuckets(BucketT *OldBucketsBegin, BucketT *OldBucketsEnd) { |
| 396 | initEmpty(); |
| 397 | |
| 398 | // Insert all the old elements. |
| 399 | const KeyT EmptyKey = getEmptyKey(); |
| 400 | const KeyT TombstoneKey = getTombstoneKey(); |
| 401 | for (BucketT *B = OldBucketsBegin, *E = OldBucketsEnd; B != E; ++B) { |
| 402 | if (!KeyInfoT::isEqual(B->getFirst(), EmptyKey) && |
| 403 | !KeyInfoT::isEqual(B->getFirst(), TombstoneKey)) { |
| 404 | // Insert the key/value into the new table. |
| 405 | BucketT *DestBucket; |
| 406 | bool FoundVal = LookupBucketFor(B->getFirst(), DestBucket); |
| 407 | (void)FoundVal; // silence warning. |
| 408 | assert(!FoundVal && "Key already in new map?")((void)0); |
| 409 | DestBucket->getFirst() = std::move(B->getFirst()); |
| 410 | ::new (&DestBucket->getSecond()) ValueT(std::move(B->getSecond())); |
| 411 | incrementNumEntries(); |
| 412 | |
| 413 | // Free the value. |
| 414 | B->getSecond().~ValueT(); |
| 415 | } |
| 416 | B->getFirst().~KeyT(); |
| 417 | } |
| 418 | } |
| 419 | |
| 420 | template <typename OtherBaseT> |
| 421 | void copyFrom( |
| 422 | const DenseMapBase<OtherBaseT, KeyT, ValueT, KeyInfoT, BucketT> &other) { |
| 423 | assert(&other != this)((void)0); |
| 424 | assert(getNumBuckets() == other.getNumBuckets())((void)0); |
| 425 | |
| 426 | setNumEntries(other.getNumEntries()); |
| 427 | setNumTombstones(other.getNumTombstones()); |
| 428 | |
| 429 | if (std::is_trivially_copyable<KeyT>::value && |
| 430 | std::is_trivially_copyable<ValueT>::value) |
| 431 | memcpy(reinterpret_cast<void *>(getBuckets()), other.getBuckets(), |
| 432 | getNumBuckets() * sizeof(BucketT)); |
| 433 | else |
| 434 | for (size_t i = 0; i < getNumBuckets(); ++i) { |
| 435 | ::new (&getBuckets()[i].getFirst()) |
| 436 | KeyT(other.getBuckets()[i].getFirst()); |
| 437 | if (!KeyInfoT::isEqual(getBuckets()[i].getFirst(), getEmptyKey()) && |
| 438 | !KeyInfoT::isEqual(getBuckets()[i].getFirst(), getTombstoneKey())) |
| 439 | ::new (&getBuckets()[i].getSecond()) |
| 440 | ValueT(other.getBuckets()[i].getSecond()); |
| 441 | } |
| 442 | } |
| 443 | |
| 444 | static unsigned getHashValue(const KeyT &Val) { |
| 445 | return KeyInfoT::getHashValue(Val); |
| 446 | } |
| 447 | |
| 448 | template<typename LookupKeyT> |
| 449 | static unsigned getHashValue(const LookupKeyT &Val) { |
| 450 | return KeyInfoT::getHashValue(Val); |
| 451 | } |
| 452 | |
| 453 | static const KeyT getEmptyKey() { |
| 454 | static_assert(std::is_base_of<DenseMapBase, DerivedT>::value, |
| 455 | "Must pass the derived type to this template!"); |
| 456 | return KeyInfoT::getEmptyKey(); |
| 457 | } |
| 458 | |
| 459 | static const KeyT getTombstoneKey() { |
| 460 | return KeyInfoT::getTombstoneKey(); |
| 461 | } |
| 462 | |
| 463 | private: |
| 464 | iterator makeIterator(BucketT *P, BucketT *E, |
| 465 | DebugEpochBase &Epoch, |
| 466 | bool NoAdvance=false) { |
| 467 | if (shouldReverseIterate<KeyT>()) { |
| 468 | BucketT *B = P == getBucketsEnd() ? getBuckets() : P + 1; |
| 469 | return iterator(B, E, Epoch, NoAdvance); |
| 470 | } |
| 471 | return iterator(P, E, Epoch, NoAdvance); |
| 472 | } |
| 473 | |
| 474 | const_iterator makeConstIterator(const BucketT *P, const BucketT *E, |
| 475 | const DebugEpochBase &Epoch, |
| 476 | const bool NoAdvance=false) const { |
| 477 | if (shouldReverseIterate<KeyT>()) { |
| 478 | const BucketT *B = P == getBucketsEnd() ? getBuckets() : P + 1; |
| 479 | return const_iterator(B, E, Epoch, NoAdvance); |
| 480 | } |
| 481 | return const_iterator(P, E, Epoch, NoAdvance); |
| 482 | } |
| 483 | |
| 484 | unsigned getNumEntries() const { |
| 485 | return static_cast<const DerivedT *>(this)->getNumEntries(); |
| 486 | } |
| 487 | |
| 488 | void setNumEntries(unsigned Num) { |
| 489 | static_cast<DerivedT *>(this)->setNumEntries(Num); |
| 490 | } |
| 491 | |
| 492 | void incrementNumEntries() { |
| 493 | setNumEntries(getNumEntries() + 1); |
| 494 | } |
| 495 | |
| 496 | void decrementNumEntries() { |
| 497 | setNumEntries(getNumEntries() - 1); |
| 498 | } |
| 499 | |
| 500 | unsigned getNumTombstones() const { |
| 501 | return static_cast<const DerivedT *>(this)->getNumTombstones(); |
| 502 | } |
| 503 | |
| 504 | void setNumTombstones(unsigned Num) { |
| 505 | static_cast<DerivedT *>(this)->setNumTombstones(Num); |
| 506 | } |
| 507 | |
| 508 | void incrementNumTombstones() { |
| 509 | setNumTombstones(getNumTombstones() + 1); |
| 510 | } |
| 511 | |
| 512 | void decrementNumTombstones() { |
| 513 | setNumTombstones(getNumTombstones() - 1); |
| 514 | } |
| 515 | |
| 516 | const BucketT *getBuckets() const { |
| 517 | return static_cast<const DerivedT *>(this)->getBuckets(); |
| 518 | } |
| 519 | |
| 520 | BucketT *getBuckets() { |
| 521 | return static_cast<DerivedT *>(this)->getBuckets(); |
| 522 | } |
| 523 | |
| 524 | unsigned getNumBuckets() const { |
| 525 | return static_cast<const DerivedT *>(this)->getNumBuckets(); |
| 526 | } |
| 527 | |
| 528 | BucketT *getBucketsEnd() { |
| 529 | return getBuckets() + getNumBuckets(); |
| 530 | } |
| 531 | |
| 532 | const BucketT *getBucketsEnd() const { |
| 533 | return getBuckets() + getNumBuckets(); |
| 534 | } |
| 535 | |
| 536 | void grow(unsigned AtLeast) { |
| 537 | static_cast<DerivedT *>(this)->grow(AtLeast); |
| 538 | } |
| 539 | |
| 540 | void shrink_and_clear() { |
| 541 | static_cast<DerivedT *>(this)->shrink_and_clear(); |
| 542 | } |
| 543 | |
| 544 | template <typename KeyArg, typename... ValueArgs> |
| 545 | BucketT *InsertIntoBucket(BucketT *TheBucket, KeyArg &&Key, |
| 546 | ValueArgs &&... Values) { |
| 547 | TheBucket = InsertIntoBucketImpl(Key, Key, TheBucket); |
| 548 | |
| 549 | TheBucket->getFirst() = std::forward<KeyArg>(Key); |
| 550 | ::new (&TheBucket->getSecond()) ValueT(std::forward<ValueArgs>(Values)...); |
| 551 | return TheBucket; |
| 552 | } |
| 553 | |
| 554 | template <typename LookupKeyT> |
| 555 | BucketT *InsertIntoBucketWithLookup(BucketT *TheBucket, KeyT &&Key, |
| 556 | ValueT &&Value, LookupKeyT &Lookup) { |
| 557 | TheBucket = InsertIntoBucketImpl(Key, Lookup, TheBucket); |
| 558 | |
| 559 | TheBucket->getFirst() = std::move(Key); |
| 560 | ::new (&TheBucket->getSecond()) ValueT(std::move(Value)); |
| 561 | return TheBucket; |
| 562 | } |
| 563 | |
| 564 | template <typename LookupKeyT> |
| 565 | BucketT *InsertIntoBucketImpl(const KeyT &Key, const LookupKeyT &Lookup, |
| 566 | BucketT *TheBucket) { |
| 567 | incrementEpoch(); |
| 568 | |
| 569 | // If the load of the hash table is more than 3/4, or if fewer than 1/8 of |
| 570 | // the buckets are empty (meaning that many are filled with tombstones), |
| 571 | // grow the table. |
| 572 | // |
| 573 | // The later case is tricky. For example, if we had one empty bucket with |
| 574 | // tons of tombstones, failing lookups (e.g. for insertion) would have to |
| 575 | // probe almost the entire table until it found the empty bucket. If the |
| 576 | // table completely filled with tombstones, no lookup would ever succeed, |
| 577 | // causing infinite loops in lookup. |
| 578 | unsigned NewNumEntries = getNumEntries() + 1; |
| 579 | unsigned NumBuckets = getNumBuckets(); |
| 580 | if (LLVM_UNLIKELY(NewNumEntries * 4 >= NumBuckets * 3)__builtin_expect((bool)(NewNumEntries * 4 >= NumBuckets * 3 ), false)) { |
| 581 | this->grow(NumBuckets * 2); |
| 582 | LookupBucketFor(Lookup, TheBucket); |
| 583 | NumBuckets = getNumBuckets(); |
| 584 | } else if (LLVM_UNLIKELY(NumBuckets-(NewNumEntries+getNumTombstones()) <=__builtin_expect((bool)(NumBuckets-(NewNumEntries+getNumTombstones ()) <= NumBuckets/8), false) |
| 585 | NumBuckets/8)__builtin_expect((bool)(NumBuckets-(NewNumEntries+getNumTombstones ()) <= NumBuckets/8), false)) { |
| 586 | this->grow(NumBuckets); |
| 587 | LookupBucketFor(Lookup, TheBucket); |
| 588 | } |
| 589 | assert(TheBucket)((void)0); |
| 590 | |
| 591 | // Only update the state after we've grown our bucket space appropriately |
| 592 | // so that when growing buckets we have self-consistent entry count. |
| 593 | incrementNumEntries(); |
| 594 | |
| 595 | // If we are writing over a tombstone, remember this. |
| 596 | const KeyT EmptyKey = getEmptyKey(); |
| 597 | if (!KeyInfoT::isEqual(TheBucket->getFirst(), EmptyKey)) |
| 598 | decrementNumTombstones(); |
| 599 | |
| 600 | return TheBucket; |
| 601 | } |
| 602 | |
| 603 | /// LookupBucketFor - Lookup the appropriate bucket for Val, returning it in |
| 604 | /// FoundBucket. If the bucket contains the key and a value, this returns |
| 605 | /// true, otherwise it returns a bucket with an empty marker or tombstone and |
| 606 | /// returns false. |
| 607 | template<typename LookupKeyT> |
| 608 | bool LookupBucketFor(const LookupKeyT &Val, |
| 609 | const BucketT *&FoundBucket) const { |
| 610 | const BucketT *BucketsPtr = getBuckets(); |
| 611 | const unsigned NumBuckets = getNumBuckets(); |
| 612 | |
| 613 | if (NumBuckets == 0) { |
| 614 | FoundBucket = nullptr; |
| 615 | return false; |
| 616 | } |
| 617 | |
| 618 | // FoundTombstone - Keep track of whether we find a tombstone while probing. |
| 619 | const BucketT *FoundTombstone = nullptr; |
| 620 | const KeyT EmptyKey = getEmptyKey(); |
| 621 | const KeyT TombstoneKey = getTombstoneKey(); |
| 622 | assert(!KeyInfoT::isEqual(Val, EmptyKey) &&((void)0) |
| 623 | !KeyInfoT::isEqual(Val, TombstoneKey) &&((void)0) |
| 624 | "Empty/Tombstone value shouldn't be inserted into map!")((void)0); |
| 625 | |
| 626 | unsigned BucketNo = getHashValue(Val) & (NumBuckets-1); |
| 627 | unsigned ProbeAmt = 1; |
| 628 | while (true) { |
| 629 | const BucketT *ThisBucket = BucketsPtr + BucketNo; |
| 630 | // Found Val's bucket? If so, return it. |
| 631 | if (LLVM_LIKELY(KeyInfoT::isEqual(Val, ThisBucket->getFirst()))__builtin_expect((bool)(KeyInfoT::isEqual(Val, ThisBucket-> getFirst())), true)) { |
| 632 | FoundBucket = ThisBucket; |
| 633 | return true; |
| 634 | } |
| 635 | |
| 636 | // If we found an empty bucket, the key doesn't exist in the set. |
| 637 | // Insert it and return the default value. |
| 638 | if (LLVM_LIKELY(KeyInfoT::isEqual(ThisBucket->getFirst(), EmptyKey))__builtin_expect((bool)(KeyInfoT::isEqual(ThisBucket->getFirst (), EmptyKey)), true)) { |
| 639 | // If we've already seen a tombstone while probing, fill it in instead |
| 640 | // of the empty bucket we eventually probed to. |
| 641 | FoundBucket = FoundTombstone ? FoundTombstone : ThisBucket; |
| 642 | return false; |
| 643 | } |
| 644 | |
| 645 | // If this is a tombstone, remember it. If Val ends up not in the map, we |
| 646 | // prefer to return it than something that would require more probing. |
| 647 | if (KeyInfoT::isEqual(ThisBucket->getFirst(), TombstoneKey) && |
| 648 | !FoundTombstone) |
| 649 | FoundTombstone = ThisBucket; // Remember the first tombstone found. |
| 650 | |
| 651 | // Otherwise, it's a hash collision or a tombstone, continue quadratic |
| 652 | // probing. |
| 653 | BucketNo += ProbeAmt++; |
| 654 | BucketNo &= (NumBuckets-1); |
| 655 | } |
| 656 | } |
| 657 | |
| 658 | template <typename LookupKeyT> |
| 659 | bool LookupBucketFor(const LookupKeyT &Val, BucketT *&FoundBucket) { |
| 660 | const BucketT *ConstFoundBucket; |
| 661 | bool Result = const_cast<const DenseMapBase *>(this) |
| 662 | ->LookupBucketFor(Val, ConstFoundBucket); |
| 663 | FoundBucket = const_cast<BucketT *>(ConstFoundBucket); |
| 664 | return Result; |
| 665 | } |
| 666 | |
| 667 | public: |
| 668 | /// Return the approximate size (in bytes) of the actual map. |
| 669 | /// This is just the raw memory used by DenseMap. |
| 670 | /// If entries are pointers to objects, the size of the referenced objects |
| 671 | /// are not included. |
| 672 | size_t getMemorySize() const { |
| 673 | return getNumBuckets() * sizeof(BucketT); |
| 674 | } |
| 675 | }; |
| 676 | |
| 677 | /// Equality comparison for DenseMap. |
| 678 | /// |
| 679 | /// Iterates over elements of LHS confirming that each (key, value) pair in LHS |
| 680 | /// is also in RHS, and that no additional pairs are in RHS. |
| 681 | /// Equivalent to N calls to RHS.find and N value comparisons. Amortized |
| 682 | /// complexity is linear, worst case is O(N^2) (if every hash collides). |
| 683 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
| 684 | typename BucketT> |
| 685 | bool operator==( |
| 686 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &LHS, |
| 687 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &RHS) { |
| 688 | if (LHS.size() != RHS.size()) |
| 689 | return false; |
| 690 | |
| 691 | for (auto &KV : LHS) { |
| 692 | auto I = RHS.find(KV.first); |
| 693 | if (I == RHS.end() || I->second != KV.second) |
| 694 | return false; |
| 695 | } |
| 696 | |
| 697 | return true; |
| 698 | } |
| 699 | |
| 700 | /// Inequality comparison for DenseMap. |
| 701 | /// |
| 702 | /// Equivalent to !(LHS == RHS). See operator== for performance notes. |
| 703 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
| 704 | typename BucketT> |
| 705 | bool operator!=( |
| 706 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &LHS, |
| 707 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &RHS) { |
| 708 | return !(LHS == RHS); |
| 709 | } |
| 710 | |
| 711 | template <typename KeyT, typename ValueT, |
| 712 | typename KeyInfoT = DenseMapInfo<KeyT>, |
| 713 | typename BucketT = llvm::detail::DenseMapPair<KeyT, ValueT>> |
| 714 | class DenseMap : public DenseMapBase<DenseMap<KeyT, ValueT, KeyInfoT, BucketT>, |
| 715 | KeyT, ValueT, KeyInfoT, BucketT> { |
| 716 | friend class DenseMapBase<DenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
| 717 | |
| 718 | // Lift some types from the dependent base class into this class for |
| 719 | // simplicity of referring to them. |
| 720 | using BaseT = DenseMapBase<DenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
| 721 | |
| 722 | BucketT *Buckets; |
| 723 | unsigned NumEntries; |
| 724 | unsigned NumTombstones; |
| 725 | unsigned NumBuckets; |
| 726 | |
| 727 | public: |
| 728 | /// Create a DenseMap with an optional \p InitialReserve that guarantee that |
| 729 | /// this number of elements can be inserted in the map without grow() |
| 730 | explicit DenseMap(unsigned InitialReserve = 0) { init(InitialReserve); } |
| 731 | |
| 732 | DenseMap(const DenseMap &other) : BaseT() { |
| 733 | init(0); |
| 734 | copyFrom(other); |
| 735 | } |
| 736 | |
| 737 | DenseMap(DenseMap &&other) : BaseT() { |
| 738 | init(0); |
| 739 | swap(other); |
| 740 | } |
| 741 | |
| 742 | template<typename InputIt> |
| 743 | DenseMap(const InputIt &I, const InputIt &E) { |
| 744 | init(std::distance(I, E)); |
| 745 | this->insert(I, E); |
| 746 | } |
| 747 | |
| 748 | DenseMap(std::initializer_list<typename BaseT::value_type> Vals) { |
| 749 | init(Vals.size()); |
| 750 | this->insert(Vals.begin(), Vals.end()); |
| 751 | } |
| 752 | |
| 753 | ~DenseMap() { |
| 754 | this->destroyAll(); |
| 755 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
| 756 | } |
| 757 | |
| 758 | void swap(DenseMap& RHS) { |
| 759 | this->incrementEpoch(); |
| 760 | RHS.incrementEpoch(); |
| 761 | std::swap(Buckets, RHS.Buckets); |
| 762 | std::swap(NumEntries, RHS.NumEntries); |
| 763 | std::swap(NumTombstones, RHS.NumTombstones); |
| 764 | std::swap(NumBuckets, RHS.NumBuckets); |
| 765 | } |
| 766 | |
| 767 | DenseMap& operator=(const DenseMap& other) { |
| 768 | if (&other != this) |
| 769 | copyFrom(other); |
| 770 | return *this; |
| 771 | } |
| 772 | |
| 773 | DenseMap& operator=(DenseMap &&other) { |
| 774 | this->destroyAll(); |
| 775 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
| 776 | init(0); |
| 777 | swap(other); |
| 778 | return *this; |
| 779 | } |
| 780 | |
| 781 | void copyFrom(const DenseMap& other) { |
| 782 | this->destroyAll(); |
| 783 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
| 784 | if (allocateBuckets(other.NumBuckets)) { |
| 785 | this->BaseT::copyFrom(other); |
| 786 | } else { |
| 787 | NumEntries = 0; |
| 788 | NumTombstones = 0; |
| 789 | } |
| 790 | } |
| 791 | |
| 792 | void init(unsigned InitNumEntries) { |
| 793 | auto InitBuckets = BaseT::getMinBucketToReserveForEntries(InitNumEntries); |
| 794 | if (allocateBuckets(InitBuckets)) { |
| 795 | this->BaseT::initEmpty(); |
| 796 | } else { |
| 797 | NumEntries = 0; |
| 798 | NumTombstones = 0; |
| 799 | } |
| 800 | } |
| 801 | |
| 802 | void grow(unsigned AtLeast) { |
| 803 | unsigned OldNumBuckets = NumBuckets; |
| 804 | BucketT *OldBuckets = Buckets; |
| 805 | |
| 806 | allocateBuckets(std::max<unsigned>(64, static_cast<unsigned>(NextPowerOf2(AtLeast-1)))); |
| 807 | assert(Buckets)((void)0); |
| 808 | if (!OldBuckets) { |
| 809 | this->BaseT::initEmpty(); |
| 810 | return; |
| 811 | } |
| 812 | |
| 813 | this->moveFromOldBuckets(OldBuckets, OldBuckets+OldNumBuckets); |
| 814 | |
| 815 | // Free the old table. |
| 816 | deallocate_buffer(OldBuckets, sizeof(BucketT) * OldNumBuckets, |
| 817 | alignof(BucketT)); |
| 818 | } |
| 819 | |
| 820 | void shrink_and_clear() { |
| 821 | unsigned OldNumBuckets = NumBuckets; |
| 822 | unsigned OldNumEntries = NumEntries; |
| 823 | this->destroyAll(); |
| 824 | |
| 825 | // Reduce the number of buckets. |
| 826 | unsigned NewNumBuckets = 0; |
| 827 | if (OldNumEntries) |
| 828 | NewNumBuckets = std::max(64, 1 << (Log2_32_Ceil(OldNumEntries) + 1)); |
| 829 | if (NewNumBuckets == NumBuckets) { |
| 830 | this->BaseT::initEmpty(); |
| 831 | return; |
| 832 | } |
| 833 | |
| 834 | deallocate_buffer(Buckets, sizeof(BucketT) * OldNumBuckets, |
| 835 | alignof(BucketT)); |
| 836 | init(NewNumBuckets); |
| 837 | } |
| 838 | |
| 839 | private: |
| 840 | unsigned getNumEntries() const { |
| 841 | return NumEntries; |
| 842 | } |
| 843 | |
| 844 | void setNumEntries(unsigned Num) { |
| 845 | NumEntries = Num; |
| 846 | } |
| 847 | |
| 848 | unsigned getNumTombstones() const { |
| 849 | return NumTombstones; |
| 850 | } |
| 851 | |
| 852 | void setNumTombstones(unsigned Num) { |
| 853 | NumTombstones = Num; |
| 854 | } |
| 855 | |
| 856 | BucketT *getBuckets() const { |
| 857 | return Buckets; |
| 858 | } |
| 859 | |
| 860 | unsigned getNumBuckets() const { |
| 861 | return NumBuckets; |
| 862 | } |
| 863 | |
| 864 | bool allocateBuckets(unsigned Num) { |
| 865 | NumBuckets = Num; |
| 866 | if (NumBuckets == 0) { |
| 867 | Buckets = nullptr; |
| 868 | return false; |
| 869 | } |
| 870 | |
| 871 | Buckets = static_cast<BucketT *>( |
| 872 | allocate_buffer(sizeof(BucketT) * NumBuckets, alignof(BucketT))); |
| 873 | return true; |
| 874 | } |
| 875 | }; |
| 876 | |
| 877 | template <typename KeyT, typename ValueT, unsigned InlineBuckets = 4, |
| 878 | typename KeyInfoT = DenseMapInfo<KeyT>, |
| 879 | typename BucketT = llvm::detail::DenseMapPair<KeyT, ValueT>> |
| 880 | class SmallDenseMap |
| 881 | : public DenseMapBase< |
| 882 | SmallDenseMap<KeyT, ValueT, InlineBuckets, KeyInfoT, BucketT>, KeyT, |
| 883 | ValueT, KeyInfoT, BucketT> { |
| 884 | friend class DenseMapBase<SmallDenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
| 885 | |
| 886 | // Lift some types from the dependent base class into this class for |
| 887 | // simplicity of referring to them. |
| 888 | using BaseT = DenseMapBase<SmallDenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
| 889 | |
| 890 | static_assert(isPowerOf2_64(InlineBuckets), |
| 891 | "InlineBuckets must be a power of 2."); |
| 892 | |
| 893 | unsigned Small : 1; |
| 894 | unsigned NumEntries : 31; |
| 895 | unsigned NumTombstones; |
| 896 | |
| 897 | struct LargeRep { |
| 898 | BucketT *Buckets; |
| 899 | unsigned NumBuckets; |
| 900 | }; |
| 901 | |
| 902 | /// A "union" of an inline bucket array and the struct representing |
| 903 | /// a large bucket. This union will be discriminated by the 'Small' bit. |
| 904 | AlignedCharArrayUnion<BucketT[InlineBuckets], LargeRep> storage; |
| 905 | |
| 906 | public: |
| 907 | explicit SmallDenseMap(unsigned NumInitBuckets = 0) { |
| 908 | init(NumInitBuckets); |
| 909 | } |
| 910 | |
| 911 | SmallDenseMap(const SmallDenseMap &other) : BaseT() { |
| 912 | init(0); |
| 913 | copyFrom(other); |
| 914 | } |
| 915 | |
| 916 | SmallDenseMap(SmallDenseMap &&other) : BaseT() { |
| 917 | init(0); |
| 918 | swap(other); |
| 919 | } |
| 920 | |
| 921 | template<typename InputIt> |
| 922 | SmallDenseMap(const InputIt &I, const InputIt &E) { |
| 923 | init(NextPowerOf2(std::distance(I, E))); |
| 924 | this->insert(I, E); |
| 925 | } |
| 926 | |
| 927 | SmallDenseMap(std::initializer_list<typename BaseT::value_type> Vals) |
| 928 | : SmallDenseMap(Vals.begin(), Vals.end()) {} |
| 929 | |
| 930 | ~SmallDenseMap() { |
| 931 | this->destroyAll(); |
| 932 | deallocateBuckets(); |
| 933 | } |
| 934 | |
| 935 | void swap(SmallDenseMap& RHS) { |
| 936 | unsigned TmpNumEntries = RHS.NumEntries; |
| 937 | RHS.NumEntries = NumEntries; |
| 938 | NumEntries = TmpNumEntries; |
| 939 | std::swap(NumTombstones, RHS.NumTombstones); |
| 940 | |
| 941 | const KeyT EmptyKey = this->getEmptyKey(); |
| 942 | const KeyT TombstoneKey = this->getTombstoneKey(); |
| 943 | if (Small && RHS.Small) { |
| 944 | // If we're swapping inline bucket arrays, we have to cope with some of |
| 945 | // the tricky bits of DenseMap's storage system: the buckets are not |
| 946 | // fully initialized. Thus we swap every key, but we may have |
| 947 | // a one-directional move of the value. |
| 948 | for (unsigned i = 0, e = InlineBuckets; i != e; ++i) { |
| 949 | BucketT *LHSB = &getInlineBuckets()[i], |
| 950 | *RHSB = &RHS.getInlineBuckets()[i]; |
| 951 | bool hasLHSValue = (!KeyInfoT::isEqual(LHSB->getFirst(), EmptyKey) && |
| 952 | !KeyInfoT::isEqual(LHSB->getFirst(), TombstoneKey)); |
| 953 | bool hasRHSValue = (!KeyInfoT::isEqual(RHSB->getFirst(), EmptyKey) && |
| 954 | !KeyInfoT::isEqual(RHSB->getFirst(), TombstoneKey)); |
| 955 | if (hasLHSValue && hasRHSValue) { |
| 956 | // Swap together if we can... |
| 957 | std::swap(*LHSB, *RHSB); |
| 958 | continue; |
| 959 | } |
| 960 | // Swap separately and handle any asymmetry. |
| 961 | std::swap(LHSB->getFirst(), RHSB->getFirst()); |
| 962 | if (hasLHSValue) { |
| 963 | ::new (&RHSB->getSecond()) ValueT(std::move(LHSB->getSecond())); |
| 964 | LHSB->getSecond().~ValueT(); |
| 965 | } else if (hasRHSValue) { |
| 966 | ::new (&LHSB->getSecond()) ValueT(std::move(RHSB->getSecond())); |
| 967 | RHSB->getSecond().~ValueT(); |
| 968 | } |
| 969 | } |
| 970 | return; |
| 971 | } |
| 972 | if (!Small && !RHS.Small) { |
| 973 | std::swap(getLargeRep()->Buckets, RHS.getLargeRep()->Buckets); |
| 974 | std::swap(getLargeRep()->NumBuckets, RHS.getLargeRep()->NumBuckets); |
| 975 | return; |
| 976 | } |
| 977 | |
| 978 | SmallDenseMap &SmallSide = Small ? *this : RHS; |
| 979 | SmallDenseMap &LargeSide = Small ? RHS : *this; |
| 980 | |
| 981 | // First stash the large side's rep and move the small side across. |
| 982 | LargeRep TmpRep = std::move(*LargeSide.getLargeRep()); |
| 983 | LargeSide.getLargeRep()->~LargeRep(); |
| 984 | LargeSide.Small = true; |
| 985 | // This is similar to the standard move-from-old-buckets, but the bucket |
| 986 | // count hasn't actually rotated in this case. So we have to carefully |
| 987 | // move construct the keys and values into their new locations, but there |
| 988 | // is no need to re-hash things. |
| 989 | for (unsigned i = 0, e = InlineBuckets; i != e; ++i) { |
| 990 | BucketT *NewB = &LargeSide.getInlineBuckets()[i], |
| 991 | *OldB = &SmallSide.getInlineBuckets()[i]; |
| 992 | ::new (&NewB->getFirst()) KeyT(std::move(OldB->getFirst())); |
| 993 | OldB->getFirst().~KeyT(); |
| 994 | if (!KeyInfoT::isEqual(NewB->getFirst(), EmptyKey) && |
| 995 | !KeyInfoT::isEqual(NewB->getFirst(), TombstoneKey)) { |
| 996 | ::new (&NewB->getSecond()) ValueT(std::move(OldB->getSecond())); |
| 997 | OldB->getSecond().~ValueT(); |
| 998 | } |
| 999 | } |
| 1000 | |
| 1001 | // The hard part of moving the small buckets across is done, just move |
| 1002 | // the TmpRep into its new home. |
| 1003 | SmallSide.Small = false; |
| 1004 | new (SmallSide.getLargeRep()) LargeRep(std::move(TmpRep)); |
| 1005 | } |
| 1006 | |
| 1007 | SmallDenseMap& operator=(const SmallDenseMap& other) { |
| 1008 | if (&other != this) |
| 1009 | copyFrom(other); |
| 1010 | return *this; |
| 1011 | } |
| 1012 | |
| 1013 | SmallDenseMap& operator=(SmallDenseMap &&other) { |
| 1014 | this->destroyAll(); |
| 1015 | deallocateBuckets(); |
| 1016 | init(0); |
| 1017 | swap(other); |
| 1018 | return *this; |
| 1019 | } |
| 1020 | |
| 1021 | void copyFrom(const SmallDenseMap& other) { |
| 1022 | this->destroyAll(); |
| 1023 | deallocateBuckets(); |
| 1024 | Small = true; |
| 1025 | if (other.getNumBuckets() > InlineBuckets) { |
| 1026 | Small = false; |
| 1027 | new (getLargeRep()) LargeRep(allocateBuckets(other.getNumBuckets())); |
| 1028 | } |
| 1029 | this->BaseT::copyFrom(other); |
| 1030 | } |
| 1031 | |
| 1032 | void init(unsigned InitBuckets) { |
| 1033 | Small = true; |
| 1034 | if (InitBuckets > InlineBuckets) { |
| 1035 | Small = false; |
| 1036 | new (getLargeRep()) LargeRep(allocateBuckets(InitBuckets)); |
| 1037 | } |
| 1038 | this->BaseT::initEmpty(); |
| 1039 | } |
| 1040 | |
| 1041 | void grow(unsigned AtLeast) { |
| 1042 | if (AtLeast > InlineBuckets) |
| 1043 | AtLeast = std::max<unsigned>(64, NextPowerOf2(AtLeast-1)); |
| 1044 | |
| 1045 | if (Small) { |
| 1046 | // First move the inline buckets into a temporary storage. |
| 1047 | AlignedCharArrayUnion<BucketT[InlineBuckets]> TmpStorage; |
| 1048 | BucketT *TmpBegin = reinterpret_cast<BucketT *>(&TmpStorage); |
| 1049 | BucketT *TmpEnd = TmpBegin; |
| 1050 | |
| 1051 | // Loop over the buckets, moving non-empty, non-tombstones into the |
| 1052 | // temporary storage. Have the loop move the TmpEnd forward as it goes. |
| 1053 | const KeyT EmptyKey = this->getEmptyKey(); |
| 1054 | const KeyT TombstoneKey = this->getTombstoneKey(); |
| 1055 | for (BucketT *P = getBuckets(), *E = P + InlineBuckets; P != E; ++P) { |
| 1056 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey) && |
| 1057 | !KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) { |
| 1058 | assert(size_t(TmpEnd - TmpBegin) < InlineBuckets &&((void)0) |
| 1059 | "Too many inline buckets!")((void)0); |
| 1060 | ::new (&TmpEnd->getFirst()) KeyT(std::move(P->getFirst())); |
| 1061 | ::new (&TmpEnd->getSecond()) ValueT(std::move(P->getSecond())); |
| 1062 | ++TmpEnd; |
| 1063 | P->getSecond().~ValueT(); |
| 1064 | } |
| 1065 | P->getFirst().~KeyT(); |
| 1066 | } |
| 1067 | |
| 1068 | // AtLeast == InlineBuckets can happen if there are many tombstones, |
| 1069 | // and grow() is used to remove them. Usually we always switch to the |
| 1070 | // large rep here. |
| 1071 | if (AtLeast > InlineBuckets) { |
| 1072 | Small = false; |
| 1073 | new (getLargeRep()) LargeRep(allocateBuckets(AtLeast)); |
| 1074 | } |
| 1075 | this->moveFromOldBuckets(TmpBegin, TmpEnd); |
| 1076 | return; |
| 1077 | } |
| 1078 | |
| 1079 | LargeRep OldRep = std::move(*getLargeRep()); |
| 1080 | getLargeRep()->~LargeRep(); |
| 1081 | if (AtLeast <= InlineBuckets) { |
| 1082 | Small = true; |
| 1083 | } else { |
| 1084 | new (getLargeRep()) LargeRep(allocateBuckets(AtLeast)); |
| 1085 | } |
| 1086 | |
| 1087 | this->moveFromOldBuckets(OldRep.Buckets, OldRep.Buckets+OldRep.NumBuckets); |
| 1088 | |
| 1089 | // Free the old table. |
| 1090 | deallocate_buffer(OldRep.Buckets, sizeof(BucketT) * OldRep.NumBuckets, |
| 1091 | alignof(BucketT)); |
| 1092 | } |
| 1093 | |
| 1094 | void shrink_and_clear() { |
| 1095 | unsigned OldSize = this->size(); |
| 1096 | this->destroyAll(); |
| 1097 | |
| 1098 | // Reduce the number of buckets. |
| 1099 | unsigned NewNumBuckets = 0; |
| 1100 | if (OldSize) { |
| 1101 | NewNumBuckets = 1 << (Log2_32_Ceil(OldSize) + 1); |
| 1102 | if (NewNumBuckets > InlineBuckets && NewNumBuckets < 64u) |
| 1103 | NewNumBuckets = 64; |
| 1104 | } |
| 1105 | if ((Small && NewNumBuckets <= InlineBuckets) || |
| 1106 | (!Small && NewNumBuckets == getLargeRep()->NumBuckets)) { |
| 1107 | this->BaseT::initEmpty(); |
| 1108 | return; |
| 1109 | } |
| 1110 | |
| 1111 | deallocateBuckets(); |
| 1112 | init(NewNumBuckets); |
| 1113 | } |
| 1114 | |
| 1115 | private: |
| 1116 | unsigned getNumEntries() const { |
| 1117 | return NumEntries; |
| 1118 | } |
| 1119 | |
| 1120 | void setNumEntries(unsigned Num) { |
| 1121 | // NumEntries is hardcoded to be 31 bits wide. |
| 1122 | assert(Num < (1U << 31) && "Cannot support more than 1<<31 entries")((void)0); |
| 1123 | NumEntries = Num; |
| 1124 | } |
| 1125 | |
| 1126 | unsigned getNumTombstones() const { |
| 1127 | return NumTombstones; |
| 1128 | } |
| 1129 | |
| 1130 | void setNumTombstones(unsigned Num) { |
| 1131 | NumTombstones = Num; |
| 1132 | } |
| 1133 | |
| 1134 | const BucketT *getInlineBuckets() const { |
| 1135 | assert(Small)((void)0); |
| 1136 | // Note that this cast does not violate aliasing rules as we assert that |
| 1137 | // the memory's dynamic type is the small, inline bucket buffer, and the |
| 1138 | // 'storage' is a POD containing a char buffer. |
| 1139 | return reinterpret_cast<const BucketT *>(&storage); |
| 1140 | } |
| 1141 | |
| 1142 | BucketT *getInlineBuckets() { |
| 1143 | return const_cast<BucketT *>( |
| 1144 | const_cast<const SmallDenseMap *>(this)->getInlineBuckets()); |
| 1145 | } |
| 1146 | |
| 1147 | const LargeRep *getLargeRep() const { |
| 1148 | assert(!Small)((void)0); |
| 1149 | // Note, same rule about aliasing as with getInlineBuckets. |
| 1150 | return reinterpret_cast<const LargeRep *>(&storage); |
| 1151 | } |
| 1152 | |
| 1153 | LargeRep *getLargeRep() { |
| 1154 | return const_cast<LargeRep *>( |
| 1155 | const_cast<const SmallDenseMap *>(this)->getLargeRep()); |
| 1156 | } |
| 1157 | |
| 1158 | const BucketT *getBuckets() const { |
| 1159 | return Small ? getInlineBuckets() : getLargeRep()->Buckets; |
| 1160 | } |
| 1161 | |
| 1162 | BucketT *getBuckets() { |
| 1163 | return const_cast<BucketT *>( |
| 1164 | const_cast<const SmallDenseMap *>(this)->getBuckets()); |
| 1165 | } |
| 1166 | |
| 1167 | unsigned getNumBuckets() const { |
| 1168 | return Small ? InlineBuckets : getLargeRep()->NumBuckets; |
| 1169 | } |
| 1170 | |
| 1171 | void deallocateBuckets() { |
| 1172 | if (Small) |
| 1173 | return; |
| 1174 | |
| 1175 | deallocate_buffer(getLargeRep()->Buckets, |
| 1176 | sizeof(BucketT) * getLargeRep()->NumBuckets, |
| 1177 | alignof(BucketT)); |
| 1178 | getLargeRep()->~LargeRep(); |
| 1179 | } |
| 1180 | |
| 1181 | LargeRep allocateBuckets(unsigned Num) { |
| 1182 | assert(Num > InlineBuckets && "Must allocate more buckets than are inline")((void)0); |
| 1183 | LargeRep Rep = {static_cast<BucketT *>(allocate_buffer( |
| 1184 | sizeof(BucketT) * Num, alignof(BucketT))), |
| 1185 | Num}; |
| 1186 | return Rep; |
| 1187 | } |
| 1188 | }; |
| 1189 | |
| 1190 | template <typename KeyT, typename ValueT, typename KeyInfoT, typename Bucket, |
| 1191 | bool IsConst> |
| 1192 | class DenseMapIterator : DebugEpochBase::HandleBase { |
| 1193 | friend class DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, true>; |
| 1194 | friend class DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, false>; |
| 1195 | |
| 1196 | public: |
| 1197 | using difference_type = ptrdiff_t; |
| 1198 | using value_type = |
| 1199 | typename std::conditional<IsConst, const Bucket, Bucket>::type; |
| 1200 | using pointer = value_type *; |
| 1201 | using reference = value_type &; |
| 1202 | using iterator_category = std::forward_iterator_tag; |
| 1203 | |
| 1204 | private: |
| 1205 | pointer Ptr = nullptr; |
| 1206 | pointer End = nullptr; |
| 1207 | |
| 1208 | public: |
| 1209 | DenseMapIterator() = default; |
| 1210 | |
| 1211 | DenseMapIterator(pointer Pos, pointer E, const DebugEpochBase &Epoch, |
| 1212 | bool NoAdvance = false) |
| 1213 | : DebugEpochBase::HandleBase(&Epoch), Ptr(Pos), End(E) { |
| 1214 | assert(isHandleInSync() && "invalid construction!")((void)0); |
| 1215 | |
| 1216 | if (NoAdvance) return; |
| 1217 | if (shouldReverseIterate<KeyT>()) { |
| 1218 | RetreatPastEmptyBuckets(); |
| 1219 | return; |
| 1220 | } |
| 1221 | AdvancePastEmptyBuckets(); |
| 1222 | } |
| 1223 | |
| 1224 | // Converting ctor from non-const iterators to const iterators. SFINAE'd out |
| 1225 | // for const iterator destinations so it doesn't end up as a user defined copy |
| 1226 | // constructor. |
| 1227 | template <bool IsConstSrc, |
| 1228 | typename = std::enable_if_t<!IsConstSrc && IsConst>> |
| 1229 | DenseMapIterator( |
| 1230 | const DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, IsConstSrc> &I) |
| 1231 | : DebugEpochBase::HandleBase(I), Ptr(I.Ptr), End(I.End) {} |
| 1232 | |
| 1233 | reference operator*() const { |
| 1234 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
| 1235 | assert(Ptr != End && "dereferencing end() iterator")((void)0); |
| 1236 | if (shouldReverseIterate<KeyT>()) |
| 1237 | return Ptr[-1]; |
| 1238 | return *Ptr; |
| 1239 | } |
| 1240 | pointer operator->() const { |
| 1241 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
| 1242 | assert(Ptr != End && "dereferencing end() iterator")((void)0); |
| 1243 | if (shouldReverseIterate<KeyT>()) |
| 1244 | return &(Ptr[-1]); |
| 1245 | return Ptr; |
| 1246 | } |
| 1247 | |
| 1248 | friend bool operator==(const DenseMapIterator &LHS, |
| 1249 | const DenseMapIterator &RHS) { |
| 1250 | assert((!LHS.Ptr || LHS.isHandleInSync()) && "handle not in sync!")((void)0); |
| 1251 | assert((!RHS.Ptr || RHS.isHandleInSync()) && "handle not in sync!")((void)0); |
| 1252 | assert(LHS.getEpochAddress() == RHS.getEpochAddress() &&((void)0) |
| 1253 | "comparing incomparable iterators!")((void)0); |
| 1254 | return LHS.Ptr == RHS.Ptr; |
| 1255 | } |
| 1256 | |
| 1257 | friend bool operator!=(const DenseMapIterator &LHS, |
| 1258 | const DenseMapIterator &RHS) { |
| 1259 | return !(LHS == RHS); |
| 1260 | } |
| 1261 | |
| 1262 | inline DenseMapIterator& operator++() { // Preincrement |
| 1263 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
| 1264 | assert(Ptr != End && "incrementing end() iterator")((void)0); |
| 1265 | if (shouldReverseIterate<KeyT>()) { |
| 1266 | --Ptr; |
| 1267 | RetreatPastEmptyBuckets(); |
| 1268 | return *this; |
| 1269 | } |
| 1270 | ++Ptr; |
| 1271 | AdvancePastEmptyBuckets(); |
| 1272 | return *this; |
| 1273 | } |
| 1274 | DenseMapIterator operator++(int) { // Postincrement |
| 1275 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
| 1276 | DenseMapIterator tmp = *this; ++*this; return tmp; |
| 1277 | } |
| 1278 | |
| 1279 | private: |
| 1280 | void AdvancePastEmptyBuckets() { |
| 1281 | assert(Ptr <= End)((void)0); |
| 1282 | const KeyT Empty = KeyInfoT::getEmptyKey(); |
| 1283 | const KeyT Tombstone = KeyInfoT::getTombstoneKey(); |
| 1284 | |
| 1285 | while (Ptr != End && (KeyInfoT::isEqual(Ptr->getFirst(), Empty) || |
| 1286 | KeyInfoT::isEqual(Ptr->getFirst(), Tombstone))) |
| 1287 | ++Ptr; |
| 1288 | } |
| 1289 | |
| 1290 | void RetreatPastEmptyBuckets() { |
| 1291 | assert(Ptr >= End)((void)0); |
| 1292 | const KeyT Empty = KeyInfoT::getEmptyKey(); |
| 1293 | const KeyT Tombstone = KeyInfoT::getTombstoneKey(); |
| 1294 | |
| 1295 | while (Ptr != End && (KeyInfoT::isEqual(Ptr[-1].getFirst(), Empty) || |
| 1296 | KeyInfoT::isEqual(Ptr[-1].getFirst(), Tombstone))) |
| 1297 | --Ptr; |
| 1298 | } |
| 1299 | }; |
| 1300 | |
| 1301 | template <typename KeyT, typename ValueT, typename KeyInfoT> |
| 1302 | inline size_t capacity_in_bytes(const DenseMap<KeyT, ValueT, KeyInfoT> &X) { |
| 1303 | return X.getMemorySize(); |
| 1304 | } |
| 1305 | |
| 1306 | } // end namespace llvm |
| 1307 | |
| 1308 | #endif // LLVM_ADT_DENSEMAP_H |