3228 lines · cpp
1//===- JumpThreading.cpp - Thread control through conditional blocks ------===//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-exception6//7//===----------------------------------------------------------------------===//8//9// This file implements the Jump Threading pass.10//11//===----------------------------------------------------------------------===//12 13#include "llvm/Transforms/Scalar/JumpThreading.h"14#include "llvm/ADT/DenseMap.h"15#include "llvm/ADT/MapVector.h"16#include "llvm/ADT/STLExtras.h"17#include "llvm/ADT/ScopeExit.h"18#include "llvm/ADT/SmallPtrSet.h"19#include "llvm/ADT/SmallVector.h"20#include "llvm/ADT/Statistic.h"21#include "llvm/Analysis/AliasAnalysis.h"22#include "llvm/Analysis/BlockFrequencyInfo.h"23#include "llvm/Analysis/BranchProbabilityInfo.h"24#include "llvm/Analysis/CFG.h"25#include "llvm/Analysis/ConstantFolding.h"26#include "llvm/Analysis/GlobalsModRef.h"27#include "llvm/Analysis/GuardUtils.h"28#include "llvm/Analysis/InstructionSimplify.h"29#include "llvm/Analysis/LazyValueInfo.h"30#include "llvm/Analysis/Loads.h"31#include "llvm/Analysis/LoopInfo.h"32#include "llvm/Analysis/MemoryLocation.h"33#include "llvm/Analysis/PostDominators.h"34#include "llvm/Analysis/TargetLibraryInfo.h"35#include "llvm/Analysis/TargetTransformInfo.h"36#include "llvm/Analysis/ValueTracking.h"37#include "llvm/IR/BasicBlock.h"38#include "llvm/IR/CFG.h"39#include "llvm/IR/Constant.h"40#include "llvm/IR/ConstantRange.h"41#include "llvm/IR/Constants.h"42#include "llvm/IR/DataLayout.h"43#include "llvm/IR/DebugInfo.h"44#include "llvm/IR/Dominators.h"45#include "llvm/IR/Function.h"46#include "llvm/IR/InstrTypes.h"47#include "llvm/IR/Instruction.h"48#include "llvm/IR/Instructions.h"49#include "llvm/IR/IntrinsicInst.h"50#include "llvm/IR/Intrinsics.h"51#include "llvm/IR/LLVMContext.h"52#include "llvm/IR/MDBuilder.h"53#include "llvm/IR/Metadata.h"54#include "llvm/IR/Module.h"55#include "llvm/IR/PassManager.h"56#include "llvm/IR/PatternMatch.h"57#include "llvm/IR/ProfDataUtils.h"58#include "llvm/IR/Type.h"59#include "llvm/IR/Use.h"60#include "llvm/IR/Value.h"61#include "llvm/Support/BlockFrequency.h"62#include "llvm/Support/BranchProbability.h"63#include "llvm/Support/Casting.h"64#include "llvm/Support/CommandLine.h"65#include "llvm/Support/Debug.h"66#include "llvm/Support/raw_ostream.h"67#include "llvm/Transforms/Utils/BasicBlockUtils.h"68#include "llvm/Transforms/Utils/Cloning.h"69#include "llvm/Transforms/Utils/Local.h"70#include "llvm/Transforms/Utils/SSAUpdater.h"71#include "llvm/Transforms/Utils/ValueMapper.h"72#include <cassert>73#include <cstdint>74#include <iterator>75#include <memory>76#include <utility>77 78using namespace llvm;79using namespace jumpthreading;80 81#define DEBUG_TYPE "jump-threading"82 83STATISTIC(NumThreads, "Number of jumps threaded");84STATISTIC(NumFolds, "Number of terminators folded");85STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");86 87static cl::opt<unsigned>88BBDuplicateThreshold("jump-threading-threshold",89 cl::desc("Max block size to duplicate for jump threading"),90 cl::init(6), cl::Hidden);91 92static cl::opt<unsigned>93ImplicationSearchThreshold(94 "jump-threading-implication-search-threshold",95 cl::desc("The number of predecessors to search for a stronger "96 "condition to use to thread over a weaker condition"),97 cl::init(3), cl::Hidden);98 99static cl::opt<unsigned> PhiDuplicateThreshold(100 "jump-threading-phi-threshold",101 cl::desc("Max PHIs in BB to duplicate for jump threading"), cl::init(76),102 cl::Hidden);103 104static cl::opt<bool> ThreadAcrossLoopHeaders(105 "jump-threading-across-loop-headers",106 cl::desc("Allow JumpThreading to thread across loop headers, for testing"),107 cl::init(false), cl::Hidden);108 109JumpThreadingPass::JumpThreadingPass(int T) {110 DefaultBBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);111}112 113// Update branch probability information according to conditional114// branch probability. This is usually made possible for cloned branches115// in inline instances by the context specific profile in the caller.116// For instance,117//118// [Block PredBB]119// [Branch PredBr]120// if (t) {121// Block A;122// } else {123// Block B;124// }125//126// [Block BB]127// cond = PN([true, %A], [..., %B]); // PHI node128// [Branch CondBr]129// if (cond) {130// ... // P(cond == true) = 1%131// }132//133// Here we know that when block A is taken, cond must be true, which means134// P(cond == true | A) = 1135//136// Given that P(cond == true) = P(cond == true | A) * P(A) +137// P(cond == true | B) * P(B)138// we get:139// P(cond == true ) = P(A) + P(cond == true | B) * P(B)140//141// which gives us:142// P(A) is less than P(cond == true), i.e.143// P(t == true) <= P(cond == true)144//145// In other words, if we know P(cond == true) is unlikely, we know146// that P(t == true) is also unlikely.147//148static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB) {149 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());150 if (!CondBr)151 return;152 153 uint64_t TrueWeight, FalseWeight;154 if (!extractBranchWeights(*CondBr, TrueWeight, FalseWeight))155 return;156 157 if (TrueWeight + FalseWeight == 0)158 // Zero branch_weights do not give a hint for getting branch probabilities.159 // Technically it would result in division by zero denominator, which is160 // TrueWeight + FalseWeight.161 return;162 163 // Returns the outgoing edge of the dominating predecessor block164 // that leads to the PhiNode's incoming block:165 auto GetPredOutEdge =166 [](BasicBlock *IncomingBB,167 BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> {168 auto *PredBB = IncomingBB;169 auto *SuccBB = PhiBB;170 SmallPtrSet<BasicBlock *, 16> Visited;171 while (true) {172 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());173 if (PredBr && PredBr->isConditional())174 return {PredBB, SuccBB};175 Visited.insert(PredBB);176 auto *SinglePredBB = PredBB->getSinglePredecessor();177 if (!SinglePredBB)178 return {nullptr, nullptr};179 180 // Stop searching when SinglePredBB has been visited. It means we see181 // an unreachable loop.182 if (Visited.count(SinglePredBB))183 return {nullptr, nullptr};184 185 SuccBB = PredBB;186 PredBB = SinglePredBB;187 }188 };189 190 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {191 Value *PhiOpnd = PN->getIncomingValue(i);192 ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd);193 194 if (!CI || !CI->getType()->isIntegerTy(1))195 continue;196 197 BranchProbability BP =198 (CI->isOne() ? BranchProbability::getBranchProbability(199 TrueWeight, TrueWeight + FalseWeight)200 : BranchProbability::getBranchProbability(201 FalseWeight, TrueWeight + FalseWeight));202 203 auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB);204 if (!PredOutEdge.first)205 return;206 207 BasicBlock *PredBB = PredOutEdge.first;208 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());209 if (!PredBr)210 return;211 212 uint64_t PredTrueWeight, PredFalseWeight;213 // FIXME: We currently only set the profile data when it is missing.214 // With PGO, this can be used to refine even existing profile data with215 // context information. This needs to be done after more performance216 // testing.217 if (extractBranchWeights(*PredBr, PredTrueWeight, PredFalseWeight))218 continue;219 220 // We can not infer anything useful when BP >= 50%, because BP is the221 // upper bound probability value.222 if (BP >= BranchProbability(50, 100))223 continue;224 225 uint32_t Weights[2];226 if (PredBr->getSuccessor(0) == PredOutEdge.second) {227 Weights[0] = BP.getNumerator();228 Weights[1] = BP.getCompl().getNumerator();229 } else {230 Weights[0] = BP.getCompl().getNumerator();231 Weights[1] = BP.getNumerator();232 }233 setBranchWeights(*PredBr, Weights, hasBranchWeightOrigin(*PredBr));234 }235}236 237PreservedAnalyses JumpThreadingPass::run(Function &F,238 FunctionAnalysisManager &AM) {239 auto &TTI = AM.getResult<TargetIRAnalysis>(F);240 // Jump Threading has no sense for the targets with divergent CF241 if (TTI.hasBranchDivergence(&F))242 return PreservedAnalyses::all();243 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);244 auto &LVI = AM.getResult<LazyValueAnalysis>(F);245 auto &AA = AM.getResult<AAManager>(F);246 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);247 248 bool Changed =249 runImpl(F, &AM, &TLI, &TTI, &LVI, &AA,250 std::make_unique<DomTreeUpdater>(251 &DT, nullptr, DomTreeUpdater::UpdateStrategy::Lazy),252 nullptr, nullptr);253 254 if (!Changed)255 return PreservedAnalyses::all();256 257 258 getDomTreeUpdater()->flush();259 260#if defined(EXPENSIVE_CHECKS)261 assert(getDomTreeUpdater()->getDomTree().verify(262 DominatorTree::VerificationLevel::Full) &&263 "DT broken after JumpThreading");264 assert((!getDomTreeUpdater()->hasPostDomTree() ||265 getDomTreeUpdater()->getPostDomTree().verify(266 PostDominatorTree::VerificationLevel::Full)) &&267 "PDT broken after JumpThreading");268#else269 assert(getDomTreeUpdater()->getDomTree().verify(270 DominatorTree::VerificationLevel::Fast) &&271 "DT broken after JumpThreading");272 assert((!getDomTreeUpdater()->hasPostDomTree() ||273 getDomTreeUpdater()->getPostDomTree().verify(274 PostDominatorTree::VerificationLevel::Fast)) &&275 "PDT broken after JumpThreading");276#endif277 278 return getPreservedAnalysis();279}280 281bool JumpThreadingPass::runImpl(Function &F_, FunctionAnalysisManager *FAM_,282 TargetLibraryInfo *TLI_,283 TargetTransformInfo *TTI_, LazyValueInfo *LVI_,284 AliasAnalysis *AA_,285 std::unique_ptr<DomTreeUpdater> DTU_,286 BlockFrequencyInfo *BFI_,287 BranchProbabilityInfo *BPI_) {288 LLVM_DEBUG(dbgs() << "Jump threading on function '" << F_.getName() << "'\n");289 F = &F_;290 FAM = FAM_;291 TLI = TLI_;292 TTI = TTI_;293 LVI = LVI_;294 AA = AA_;295 DTU = std::move(DTU_);296 BFI = BFI_;297 BPI = BPI_;298 auto *GuardDecl = Intrinsic::getDeclarationIfExists(299 F->getParent(), Intrinsic::experimental_guard);300 HasGuards = GuardDecl && !GuardDecl->use_empty();301 302 // Reduce the number of instructions duplicated when optimizing strictly for303 // size.304 if (BBDuplicateThreshold.getNumOccurrences())305 BBDupThreshold = BBDuplicateThreshold;306 else if (F->hasMinSize())307 BBDupThreshold = 3;308 else309 BBDupThreshold = DefaultBBDupThreshold;310 311 assert(DTU && "DTU isn't passed into JumpThreading before using it.");312 assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed.");313 DominatorTree &DT = DTU->getDomTree();314 315 Unreachable.clear();316 for (auto &BB : *F)317 if (!DT.isReachableFromEntry(&BB))318 Unreachable.insert(&BB);319 320 if (!ThreadAcrossLoopHeaders)321 findLoopHeaders(*F);322 323 bool EverChanged = false;324 bool Changed;325 do {326 Changed = false;327 for (auto &BB : *F) {328 if (Unreachable.count(&BB))329 continue;330 while (processBlock(&BB)) // Thread all of the branches we can over BB.331 Changed = ChangedSinceLastAnalysisUpdate = true;332 333 // Stop processing BB if it's the entry or is now deleted. The following334 // routines attempt to eliminate BB and locating a suitable replacement335 // for the entry is non-trivial.336 if (&BB == &F->getEntryBlock() || DTU->isBBPendingDeletion(&BB))337 continue;338 339 if (pred_empty(&BB)) {340 // When processBlock makes BB unreachable it doesn't bother to fix up341 // the instructions in it. We must remove BB to prevent invalid IR.342 LLVM_DEBUG(dbgs() << " JT: Deleting dead block '" << BB.getName()343 << "' with terminator: " << *BB.getTerminator()344 << '\n');345 LoopHeaders.erase(&BB);346 LVI->eraseBlock(&BB);347 DeleteDeadBlock(&BB, DTU.get());348 Changed = ChangedSinceLastAnalysisUpdate = true;349 continue;350 }351 352 // processBlock doesn't thread BBs with unconditional TIs. However, if BB353 // is "almost empty", we attempt to merge BB with its sole successor.354 auto *BI = dyn_cast<BranchInst>(BB.getTerminator());355 if (BI && BI->isUnconditional()) {356 BasicBlock *Succ = BI->getSuccessor(0);357 if (358 // The terminator must be the only non-phi instruction in BB.359 BB.getFirstNonPHIOrDbg(true)->isTerminator() &&360 // Don't alter Loop headers and latches to ensure another pass can361 // detect and transform nested loops later.362 !LoopHeaders.count(&BB) && !LoopHeaders.count(Succ) &&363 TryToSimplifyUncondBranchFromEmptyBlock(&BB, DTU.get())) {364 // BB is valid for cleanup here because we passed in DTU. F remains365 // BB's parent until a DTU->getDomTree() event.366 LVI->eraseBlock(&BB);367 Changed = ChangedSinceLastAnalysisUpdate = true;368 }369 }370 }371 EverChanged |= Changed;372 } while (Changed);373 374 // Jump threading may have introduced redundant debug values into F which375 // should be removed.376 if (EverChanged)377 for (auto &BB : *F) {378 RemoveRedundantDbgInstrs(&BB);379 }380 381 LoopHeaders.clear();382 return EverChanged;383}384 385// Replace uses of Cond with ToVal when safe to do so. If all uses are386// replaced, we can remove Cond. We cannot blindly replace all uses of Cond387// because we may incorrectly replace uses when guards/assumes are uses of388// of `Cond` and we used the guards/assume to reason about the `Cond` value389// at the end of block. RAUW unconditionally replaces all uses390// including the guards/assumes themselves and the uses before the391// guard/assume.392static bool replaceFoldableUses(Instruction *Cond, Value *ToVal,393 BasicBlock *KnownAtEndOfBB) {394 bool Changed = false;395 assert(Cond->getType() == ToVal->getType());396 // We can unconditionally replace all uses in non-local blocks (i.e. uses397 // strictly dominated by BB), since LVI information is true from the398 // terminator of BB.399 if (Cond->getParent() == KnownAtEndOfBB)400 Changed |= replaceNonLocalUsesWith(Cond, ToVal);401 for (Instruction &I : reverse(*KnownAtEndOfBB)) {402 // Replace any debug-info record users of Cond with ToVal.403 for (DbgVariableRecord &DVR : filterDbgVars(I.getDbgRecordRange()))404 DVR.replaceVariableLocationOp(Cond, ToVal, true);405 406 // Reached the Cond whose uses we are trying to replace, so there are no407 // more uses.408 if (&I == Cond)409 break;410 // We only replace uses in instructions that are guaranteed to reach the end411 // of BB, where we know Cond is ToVal.412 if (!isGuaranteedToTransferExecutionToSuccessor(&I))413 break;414 Changed |= I.replaceUsesOfWith(Cond, ToVal);415 }416 if (Cond->use_empty() && !Cond->mayHaveSideEffects()) {417 Cond->eraseFromParent();418 Changed = true;419 }420 return Changed;421}422 423/// Return the cost of duplicating a piece of this block from first non-phi424/// and before StopAt instruction to thread across it. Stop scanning the block425/// when exceeding the threshold. If duplication is impossible, returns ~0U.426static unsigned getJumpThreadDuplicationCost(const TargetTransformInfo *TTI,427 BasicBlock *BB,428 Instruction *StopAt,429 unsigned Threshold) {430 assert(StopAt->getParent() == BB && "Not an instruction from proper BB?");431 432 // Do not duplicate the BB if it has a lot of PHI nodes.433 // If a threadable chain is too long then the number of PHI nodes can add up,434 // leading to a substantial increase in compile time when rewriting the SSA.435 unsigned PhiCount = 0;436 Instruction *FirstNonPHI = nullptr;437 for (Instruction &I : *BB) {438 if (!isa<PHINode>(&I)) {439 FirstNonPHI = &I;440 break;441 }442 if (++PhiCount > PhiDuplicateThreshold)443 return ~0U;444 }445 446 /// Ignore PHI nodes, these will be flattened when duplication happens.447 BasicBlock::const_iterator I(FirstNonPHI);448 449 // FIXME: THREADING will delete values that are just used to compute the450 // branch, so they shouldn't count against the duplication cost.451 452 unsigned Bonus = 0;453 if (BB->getTerminator() == StopAt) {454 // Threading through a switch statement is particularly profitable. If this455 // block ends in a switch, decrease its cost to make it more likely to456 // happen.457 if (isa<SwitchInst>(StopAt))458 Bonus = 6;459 460 // The same holds for indirect branches, but slightly more so.461 if (isa<IndirectBrInst>(StopAt))462 Bonus = 8;463 }464 465 // Bump the threshold up so the early exit from the loop doesn't skip the466 // terminator-based Size adjustment at the end.467 Threshold += Bonus;468 469 // Sum up the cost of each instruction until we get to the terminator. Don't470 // include the terminator because the copy won't include it.471 unsigned Size = 0;472 for (; &*I != StopAt; ++I) {473 474 // Stop scanning the block if we've reached the threshold.475 if (Size > Threshold)476 return Size;477 478 // Bail out if this instruction gives back a token type, it is not possible479 // to duplicate it if it is used outside this BB.480 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))481 return ~0U;482 483 // Blocks with NoDuplicate are modelled as having infinite cost, so they484 // are never duplicated.485 if (const CallInst *CI = dyn_cast<CallInst>(I))486 if (CI->cannotDuplicate() || CI->isConvergent())487 return ~0U;488 489 if (TTI->getInstructionCost(&*I, TargetTransformInfo::TCK_SizeAndLatency) ==490 TargetTransformInfo::TCC_Free)491 continue;492 493 // All other instructions count for at least one unit.494 ++Size;495 496 // Calls are more expensive. If they are non-intrinsic calls, we model them497 // as having cost of 4. If they are a non-vector intrinsic, we model them498 // as having cost of 2 total, and if they are a vector intrinsic, we model499 // them as having cost 1.500 if (const CallInst *CI = dyn_cast<CallInst>(I)) {501 if (!isa<IntrinsicInst>(CI))502 Size += 3;503 else if (!CI->getType()->isVectorTy())504 Size += 1;505 }506 }507 508 return Size > Bonus ? Size - Bonus : 0;509}510 511/// findLoopHeaders - We do not want jump threading to turn proper loop512/// structures into irreducible loops. Doing this breaks up the loop nesting513/// hierarchy and pessimizes later transformations. To prevent this from514/// happening, we first have to find the loop headers. Here we approximate this515/// by finding targets of backedges in the CFG.516///517/// Note that there definitely are cases when we want to allow threading of518/// edges across a loop header. For example, threading a jump from outside the519/// loop (the preheader) to an exit block of the loop is definitely profitable.520/// It is also almost always profitable to thread backedges from within the loop521/// to exit blocks, and is often profitable to thread backedges to other blocks522/// within the loop (forming a nested loop). This simple analysis is not rich523/// enough to track all of these properties and keep it up-to-date as the CFG524/// mutates, so we don't allow any of these transformations.525void JumpThreadingPass::findLoopHeaders(Function &F) {526 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;527 FindFunctionBackedges(F, Edges);528 LoopHeaders.insert_range(llvm::make_second_range(Edges));529}530 531/// getKnownConstant - Helper method to determine if we can thread over a532/// terminator with the given value as its condition, and if so what value to533/// use for that. What kind of value this is depends on whether we want an534/// integer or a block address, but an undef is always accepted.535/// Returns null if Val is null or not an appropriate constant.536static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {537 if (!Val)538 return nullptr;539 540 // Undef is "known" enough.541 if (UndefValue *U = dyn_cast<UndefValue>(Val))542 return U;543 544 if (Preference == WantBlockAddress)545 return dyn_cast<BlockAddress>(Val->stripPointerCasts());546 547 return dyn_cast<ConstantInt>(Val);548}549 550/// computeValueKnownInPredecessors - Given a basic block BB and a value V, see551/// if we can infer that the value is a known ConstantInt/BlockAddress or undef552/// in any of our predecessors. If so, return the known list of value and pred553/// BB in the result vector.554///555/// This returns true if there were any known values.556bool JumpThreadingPass::computeValueKnownInPredecessorsImpl(557 Value *V, BasicBlock *BB, PredValueInfo &Result,558 ConstantPreference Preference, SmallPtrSet<Value *, 4> &RecursionSet,559 Instruction *CxtI) {560 const DataLayout &DL = BB->getDataLayout();561 562 // This method walks up use-def chains recursively. Because of this, we could563 // get into an infinite loop going around loops in the use-def chain. To564 // prevent this, keep track of what (value, block) pairs we've already visited565 // and terminate the search if we loop back to them566 if (!RecursionSet.insert(V).second)567 return false;568 569 // If V is a constant, then it is known in all predecessors.570 if (Constant *KC = getKnownConstant(V, Preference)) {571 for (BasicBlock *Pred : predecessors(BB))572 Result.emplace_back(KC, Pred);573 574 return !Result.empty();575 }576 577 // If V is a non-instruction value, or an instruction in a different block,578 // then it can't be derived from a PHI.579 Instruction *I = dyn_cast<Instruction>(V);580 if (!I || I->getParent() != BB) {581 582 // Okay, if this is a live-in value, see if it has a known value at the any583 // edge from our predecessors.584 for (BasicBlock *P : predecessors(BB)) {585 using namespace PatternMatch;586 // If the value is known by LazyValueInfo to be a constant in a587 // predecessor, use that information to try to thread this block.588 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);589 // If I is a non-local compare-with-constant instruction, use more-rich590 // 'getPredicateOnEdge' method. This would be able to handle value591 // inequalities better, for example if the compare is "X < 4" and "X < 3"592 // is known true but "X < 4" itself is not available.593 CmpPredicate Pred;594 Value *Val;595 Constant *Cst;596 if (!PredCst && match(V, m_Cmp(Pred, m_Value(Val), m_Constant(Cst))))597 PredCst = LVI->getPredicateOnEdge(Pred, Val, Cst, P, BB, CxtI);598 if (Constant *KC = getKnownConstant(PredCst, Preference))599 Result.emplace_back(KC, P);600 }601 602 return !Result.empty();603 }604 605 /// If I is a PHI node, then we know the incoming values for any constants.606 if (PHINode *PN = dyn_cast<PHINode>(I)) {607 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {608 Value *InVal = PN->getIncomingValue(i);609 if (Constant *KC = getKnownConstant(InVal, Preference)) {610 Result.emplace_back(KC, PN->getIncomingBlock(i));611 } else {612 Constant *CI = LVI->getConstantOnEdge(InVal,613 PN->getIncomingBlock(i),614 BB, CxtI);615 if (Constant *KC = getKnownConstant(CI, Preference))616 Result.emplace_back(KC, PN->getIncomingBlock(i));617 }618 }619 620 return !Result.empty();621 }622 623 // Handle Cast instructions.624 if (CastInst *CI = dyn_cast<CastInst>(I)) {625 Value *Source = CI->getOperand(0);626 PredValueInfoTy Vals;627 computeValueKnownInPredecessorsImpl(Source, BB, Vals, Preference,628 RecursionSet, CxtI);629 if (Vals.empty())630 return false;631 632 // Convert the known values.633 for (auto &Val : Vals)634 if (Constant *Folded = ConstantFoldCastOperand(CI->getOpcode(), Val.first,635 CI->getType(), DL))636 Result.emplace_back(Folded, Val.second);637 638 return !Result.empty();639 }640 641 if (FreezeInst *FI = dyn_cast<FreezeInst>(I)) {642 Value *Source = FI->getOperand(0);643 computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,644 RecursionSet, CxtI);645 646 erase_if(Result, [](auto &Pair) {647 return !isGuaranteedNotToBeUndefOrPoison(Pair.first);648 });649 650 return !Result.empty();651 }652 653 // Handle some boolean conditions.654 if (I->getType()->getPrimitiveSizeInBits() == 1) {655 using namespace PatternMatch;656 if (Preference != WantInteger)657 return false;658 // X | true -> true659 // X & false -> false660 Value *Op0, *Op1;661 if (match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1))) ||662 match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) {663 PredValueInfoTy LHSVals, RHSVals;664 665 computeValueKnownInPredecessorsImpl(Op0, BB, LHSVals, WantInteger,666 RecursionSet, CxtI);667 computeValueKnownInPredecessorsImpl(Op1, BB, RHSVals, WantInteger,668 RecursionSet, CxtI);669 670 if (LHSVals.empty() && RHSVals.empty())671 return false;672 673 ConstantInt *InterestingVal;674 if (match(I, m_LogicalOr()))675 InterestingVal = ConstantInt::getTrue(I->getContext());676 else677 InterestingVal = ConstantInt::getFalse(I->getContext());678 679 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;680 681 // Scan for the sentinel. If we find an undef, force it to the682 // interesting value: x|undef -> true and x&undef -> false.683 for (const auto &LHSVal : LHSVals)684 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {685 Result.emplace_back(InterestingVal, LHSVal.second);686 LHSKnownBBs.insert(LHSVal.second);687 }688 for (const auto &RHSVal : RHSVals)689 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {690 // If we already inferred a value for this block on the LHS, don't691 // re-add it.692 if (!LHSKnownBBs.count(RHSVal.second))693 Result.emplace_back(InterestingVal, RHSVal.second);694 }695 696 return !Result.empty();697 }698 699 // Handle the NOT form of XOR.700 if (I->getOpcode() == Instruction::Xor &&701 isa<ConstantInt>(I->getOperand(1)) &&702 cast<ConstantInt>(I->getOperand(1))->isOne()) {703 computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, Result,704 WantInteger, RecursionSet, CxtI);705 if (Result.empty())706 return false;707 708 // Invert the known values.709 for (auto &R : Result)710 R.first = ConstantExpr::getNot(R.first);711 712 return true;713 }714 715 // Try to simplify some other binary operator values.716 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {717 if (Preference != WantInteger)718 return false;719 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {720 PredValueInfoTy LHSVals;721 computeValueKnownInPredecessorsImpl(BO->getOperand(0), BB, LHSVals,722 WantInteger, RecursionSet, CxtI);723 724 // Try to use constant folding to simplify the binary operator.725 for (const auto &LHSVal : LHSVals) {726 Constant *V = LHSVal.first;727 Constant *Folded =728 ConstantFoldBinaryOpOperands(BO->getOpcode(), V, CI, DL);729 730 if (Constant *KC = getKnownConstant(Folded, WantInteger))731 Result.emplace_back(KC, LHSVal.second);732 }733 }734 735 return !Result.empty();736 }737 738 // Handle compare with phi operand, where the PHI is defined in this block.739 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {740 if (Preference != WantInteger)741 return false;742 Type *CmpType = Cmp->getType();743 Value *CmpLHS = Cmp->getOperand(0);744 Value *CmpRHS = Cmp->getOperand(1);745 CmpInst::Predicate Pred = Cmp->getPredicate();746 747 PHINode *PN = dyn_cast<PHINode>(CmpLHS);748 if (!PN)749 PN = dyn_cast<PHINode>(CmpRHS);750 // Do not perform phi translation across a loop header phi, because this751 // may result in comparison of values from two different loop iterations.752 // FIXME: This check is broken if LoopHeaders is not populated.753 if (PN && PN->getParent() == BB && !LoopHeaders.contains(BB)) {754 const DataLayout &DL = PN->getDataLayout();755 // We can do this simplification if any comparisons fold to true or false.756 // See if any do.757 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {758 BasicBlock *PredBB = PN->getIncomingBlock(i);759 Value *LHS, *RHS;760 if (PN == CmpLHS) {761 LHS = PN->getIncomingValue(i);762 RHS = CmpRHS->DoPHITranslation(BB, PredBB);763 } else {764 LHS = CmpLHS->DoPHITranslation(BB, PredBB);765 RHS = PN->getIncomingValue(i);766 }767 Value *Res = simplifyCmpInst(Pred, LHS, RHS, {DL});768 if (!Res) {769 if (!isa<Constant>(RHS))770 continue;771 772 // getPredicateOnEdge call will make no sense if LHS is defined in BB.773 auto LHSInst = dyn_cast<Instruction>(LHS);774 if (LHSInst && LHSInst->getParent() == BB)775 continue;776 777 Res = LVI->getPredicateOnEdge(Pred, LHS, cast<Constant>(RHS), PredBB,778 BB, CxtI ? CxtI : Cmp);779 }780 781 if (Constant *KC = getKnownConstant(Res, WantInteger))782 Result.emplace_back(KC, PredBB);783 }784 785 return !Result.empty();786 }787 788 // If comparing a live-in value against a constant, see if we know the789 // live-in value on any predecessors.790 if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {791 Constant *CmpConst = cast<Constant>(CmpRHS);792 793 if (!isa<Instruction>(CmpLHS) ||794 cast<Instruction>(CmpLHS)->getParent() != BB) {795 for (BasicBlock *P : predecessors(BB)) {796 // If the value is known by LazyValueInfo to be a constant in a797 // predecessor, use that information to try to thread this block.798 Constant *Res = LVI->getPredicateOnEdge(Pred, CmpLHS, CmpConst, P, BB,799 CxtI ? CxtI : Cmp);800 if (Constant *KC = getKnownConstant(Res, WantInteger))801 Result.emplace_back(KC, P);802 }803 804 return !Result.empty();805 }806 807 // InstCombine can fold some forms of constant range checks into808 // (icmp (add (x, C1)), C2). See if we have we have such a thing with809 // x as a live-in.810 {811 using namespace PatternMatch;812 813 Value *AddLHS;814 ConstantInt *AddConst;815 if (isa<ConstantInt>(CmpConst) &&816 match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {817 if (!isa<Instruction>(AddLHS) ||818 cast<Instruction>(AddLHS)->getParent() != BB) {819 for (BasicBlock *P : predecessors(BB)) {820 // If the value is known by LazyValueInfo to be a ConstantRange in821 // a predecessor, use that information to try to thread this822 // block.823 ConstantRange CR = LVI->getConstantRangeOnEdge(824 AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));825 // Propagate the range through the addition.826 CR = CR.add(AddConst->getValue());827 828 // Get the range where the compare returns true.829 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(830 Pred, cast<ConstantInt>(CmpConst)->getValue());831 832 Constant *ResC;833 if (CmpRange.contains(CR))834 ResC = ConstantInt::getTrue(CmpType);835 else if (CmpRange.inverse().contains(CR))836 ResC = ConstantInt::getFalse(CmpType);837 else838 continue;839 840 Result.emplace_back(ResC, P);841 }842 843 return !Result.empty();844 }845 }846 }847 848 // Try to find a constant value for the LHS of a comparison,849 // and evaluate it statically if we can.850 PredValueInfoTy LHSVals;851 computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,852 WantInteger, RecursionSet, CxtI);853 854 for (const auto &LHSVal : LHSVals) {855 Constant *V = LHSVal.first;856 Constant *Folded =857 ConstantFoldCompareInstOperands(Pred, V, CmpConst, DL);858 if (Constant *KC = getKnownConstant(Folded, WantInteger))859 Result.emplace_back(KC, LHSVal.second);860 }861 862 return !Result.empty();863 }864 }865 866 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {867 // Handle select instructions where at least one operand is a known constant868 // and we can figure out the condition value for any predecessor block.869 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);870 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);871 PredValueInfoTy Conds;872 if ((TrueVal || FalseVal) &&873 computeValueKnownInPredecessorsImpl(SI->getCondition(), BB, Conds,874 WantInteger, RecursionSet, CxtI)) {875 for (auto &C : Conds) {876 Constant *Cond = C.first;877 878 // Figure out what value to use for the condition.879 bool KnownCond;880 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {881 // A known boolean.882 KnownCond = CI->isOne();883 } else {884 assert(isa<UndefValue>(Cond) && "Unexpected condition value");885 // Either operand will do, so be sure to pick the one that's a known886 // constant.887 // FIXME: Do this more cleverly if both values are known constants?888 KnownCond = (TrueVal != nullptr);889 }890 891 // See if the select has a known constant value for this predecessor.892 if (Constant *Val = KnownCond ? TrueVal : FalseVal)893 Result.emplace_back(Val, C.second);894 }895 896 return !Result.empty();897 }898 }899 900 // If all else fails, see if LVI can figure out a constant value for us.901 assert(CxtI->getParent() == BB && "CxtI should be in BB");902 Constant *CI = LVI->getConstant(V, CxtI);903 if (Constant *KC = getKnownConstant(CI, Preference)) {904 for (BasicBlock *Pred : predecessors(BB))905 Result.emplace_back(KC, Pred);906 }907 908 return !Result.empty();909}910 911/// GetBestDestForBranchOnUndef - If we determine that the specified block ends912/// in an undefined jump, decide which block is best to revector to.913///914/// Since we can pick an arbitrary destination, we pick the successor with the915/// fewest predecessors. This should reduce the in-degree of the others.916static unsigned getBestDestForJumpOnUndef(BasicBlock *BB) {917 Instruction *BBTerm = BB->getTerminator();918 unsigned MinSucc = 0;919 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);920 // Compute the successor with the minimum number of predecessors.921 unsigned MinNumPreds = pred_size(TestBB);922 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {923 TestBB = BBTerm->getSuccessor(i);924 unsigned NumPreds = pred_size(TestBB);925 if (NumPreds < MinNumPreds) {926 MinSucc = i;927 MinNumPreds = NumPreds;928 }929 }930 931 return MinSucc;932}933 934static bool hasAddressTakenAndUsed(BasicBlock *BB) {935 if (!BB->hasAddressTaken()) return false;936 937 // If the block has its address taken, it may be a tree of dead constants938 // hanging off of it. These shouldn't keep the block alive.939 BlockAddress *BA = BlockAddress::get(BB);940 BA->removeDeadConstantUsers();941 return !BA->use_empty();942}943 944/// processBlock - If there are any predecessors whose control can be threaded945/// through to a successor, transform them now.946bool JumpThreadingPass::processBlock(BasicBlock *BB) {947 // If the block is trivially dead, just return and let the caller nuke it.948 // This simplifies other transformations.949 if (DTU->isBBPendingDeletion(BB) ||950 (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))951 return false;952 953 // If this block has a single predecessor, and if that pred has a single954 // successor, merge the blocks. This encourages recursive jump threading955 // because now the condition in this block can be threaded through956 // predecessors of our predecessor block.957 if (maybeMergeBasicBlockIntoOnlyPred(BB))958 return true;959 960 if (tryToUnfoldSelectInCurrBB(BB))961 return true;962 963 // Look if we can propagate guards to predecessors.964 if (HasGuards && processGuards(BB))965 return true;966 967 // What kind of constant we're looking for.968 ConstantPreference Preference = WantInteger;969 970 // Look to see if the terminator is a conditional branch, switch or indirect971 // branch, if not we can't thread it.972 Value *Condition;973 Instruction *Terminator = BB->getTerminator();974 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {975 // Can't thread an unconditional jump.976 if (BI->isUnconditional()) return false;977 Condition = BI->getCondition();978 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {979 Condition = SI->getCondition();980 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {981 // Can't thread indirect branch with no successors.982 if (IB->getNumSuccessors() == 0) return false;983 Condition = IB->getAddress()->stripPointerCasts();984 Preference = WantBlockAddress;985 } else {986 return false; // Must be an invoke or callbr.987 }988 989 // Keep track if we constant folded the condition in this invocation.990 bool ConstantFolded = false;991 992 // Run constant folding to see if we can reduce the condition to a simple993 // constant.994 if (Instruction *I = dyn_cast<Instruction>(Condition)) {995 Value *SimpleVal =996 ConstantFoldInstruction(I, BB->getDataLayout(), TLI);997 if (SimpleVal) {998 I->replaceAllUsesWith(SimpleVal);999 if (isInstructionTriviallyDead(I, TLI))1000 I->eraseFromParent();1001 Condition = SimpleVal;1002 ConstantFolded = true;1003 }1004 }1005 1006 // If the terminator is branching on an undef or freeze undef, we can pick any1007 // of the successors to branch to. Let getBestDestForJumpOnUndef decide.1008 auto *FI = dyn_cast<FreezeInst>(Condition);1009 if (isa<UndefValue>(Condition) ||1010 (FI && isa<UndefValue>(FI->getOperand(0)) && FI->hasOneUse())) {1011 unsigned BestSucc = getBestDestForJumpOnUndef(BB);1012 std::vector<DominatorTree::UpdateType> Updates;1013 1014 // Fold the branch/switch.1015 Instruction *BBTerm = BB->getTerminator();1016 Updates.reserve(BBTerm->getNumSuccessors());1017 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {1018 if (i == BestSucc) continue;1019 BasicBlock *Succ = BBTerm->getSuccessor(i);1020 Succ->removePredecessor(BB, true);1021 Updates.push_back({DominatorTree::Delete, BB, Succ});1022 }1023 1024 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()1025 << "' folding undef terminator: " << *BBTerm << '\n');1026 Instruction *NewBI = BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm->getIterator());1027 NewBI->setDebugLoc(BBTerm->getDebugLoc());1028 ++NumFolds;1029 BBTerm->eraseFromParent();1030 DTU->applyUpdatesPermissive(Updates);1031 if (FI)1032 FI->eraseFromParent();1033 return true;1034 }1035 1036 // If the terminator of this block is branching on a constant, simplify the1037 // terminator to an unconditional branch. This can occur due to threading in1038 // other blocks.1039 if (getKnownConstant(Condition, Preference)) {1040 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()1041 << "' folding terminator: " << *BB->getTerminator()1042 << '\n');1043 ++NumFolds;1044 ConstantFoldTerminator(BB, true, nullptr, DTU.get());1045 if (auto *BPI = getBPI())1046 BPI->eraseBlock(BB);1047 return true;1048 }1049 1050 Instruction *CondInst = dyn_cast<Instruction>(Condition);1051 1052 // All the rest of our checks depend on the condition being an instruction.1053 if (!CondInst) {1054 // FIXME: Unify this with code below.1055 if (processThreadableEdges(Condition, BB, Preference, Terminator))1056 return true;1057 return ConstantFolded;1058 }1059 1060 // Some of the following optimization can safely work on the unfrozen cond.1061 Value *CondWithoutFreeze = CondInst;1062 if (auto *FI = dyn_cast<FreezeInst>(CondInst))1063 CondWithoutFreeze = FI->getOperand(0);1064 1065 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondWithoutFreeze)) {1066 // If we're branching on a conditional, LVI might be able to determine1067 // it's value at the branch instruction. We only handle comparisons1068 // against a constant at this time.1069 if (Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1))) {1070 Constant *Res =1071 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),1072 CondConst, BB->getTerminator(),1073 /*UseBlockValue=*/false);1074 if (Res) {1075 // We can safely replace *some* uses of the CondInst if it has1076 // exactly one value as returned by LVI. RAUW is incorrect in the1077 // presence of guards and assumes, that have the `Cond` as the use. This1078 // is because we use the guards/assume to reason about the `Cond` value1079 // at the end of block, but RAUW unconditionally replaces all uses1080 // including the guards/assumes themselves and the uses before the1081 // guard/assume.1082 if (replaceFoldableUses(CondCmp, Res, BB))1083 return true;1084 }1085 1086 // We did not manage to simplify this branch, try to see whether1087 // CondCmp depends on a known phi-select pattern.1088 if (tryToUnfoldSelect(CondCmp, BB))1089 return true;1090 }1091 }1092 1093 if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))1094 if (tryToUnfoldSelect(SI, BB))1095 return true;1096 1097 // Check for some cases that are worth simplifying. Right now we want to look1098 // for loads that are used by a switch or by the condition for the branch. If1099 // we see one, check to see if it's partially redundant. If so, insert a PHI1100 // which can then be used to thread the values.1101 Value *SimplifyValue = CondWithoutFreeze;1102 1103 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))1104 if (isa<Constant>(CondCmp->getOperand(1)))1105 SimplifyValue = CondCmp->getOperand(0);1106 1107 // TODO: There are other places where load PRE would be profitable, such as1108 // more complex comparisons.1109 if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue))1110 if (simplifyPartiallyRedundantLoad(LoadI))1111 return true;1112 1113 // Before threading, try to propagate profile data backwards:1114 if (PHINode *PN = dyn_cast<PHINode>(CondInst))1115 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))1116 updatePredecessorProfileMetadata(PN, BB);1117 1118 // Handle a variety of cases where we are branching on something derived from1119 // a PHI node in the current block. If we can prove that any predecessors1120 // compute a predictable value based on a PHI node, thread those predecessors.1121 if (processThreadableEdges(CondInst, BB, Preference, Terminator))1122 return true;1123 1124 // If this is an otherwise-unfoldable branch on a phi node or freeze(phi) in1125 // the current block, see if we can simplify.1126 PHINode *PN = dyn_cast<PHINode>(CondWithoutFreeze);1127 if (PN && PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))1128 return processBranchOnPHI(PN);1129 1130 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.1131 if (CondInst->getOpcode() == Instruction::Xor &&1132 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))1133 return processBranchOnXOR(cast<BinaryOperator>(CondInst));1134 1135 // Search for a stronger dominating condition that can be used to simplify a1136 // conditional branch leaving BB.1137 if (processImpliedCondition(BB))1138 return true;1139 1140 return false;1141}1142 1143bool JumpThreadingPass::processImpliedCondition(BasicBlock *BB) {1144 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());1145 if (!BI || !BI->isConditional())1146 return false;1147 1148 Value *Cond = BI->getCondition();1149 // Assuming that predecessor's branch was taken, if pred's branch condition1150 // (V) implies Cond, Cond can be either true, undef, or poison. In this case,1151 // freeze(Cond) is either true or a nondeterministic value.1152 // If freeze(Cond) has only one use, we can freely fold freeze(Cond) to true1153 // without affecting other instructions.1154 auto *FICond = dyn_cast<FreezeInst>(Cond);1155 if (FICond && FICond->hasOneUse())1156 Cond = FICond->getOperand(0);1157 else1158 FICond = nullptr;1159 1160 BasicBlock *CurrentBB = BB;1161 BasicBlock *CurrentPred = BB->getSinglePredecessor();1162 unsigned Iter = 0;1163 1164 auto &DL = BB->getDataLayout();1165 1166 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {1167 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());1168 if (!PBI || !PBI->isConditional())1169 return false;1170 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)1171 return false;1172 1173 bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB;1174 std::optional<bool> Implication =1175 isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue);1176 1177 // If the branch condition of BB (which is Cond) and CurrentPred are1178 // exactly the same freeze instruction, Cond can be folded into CondIsTrue.1179 if (!Implication && FICond && isa<FreezeInst>(PBI->getCondition())) {1180 if (cast<FreezeInst>(PBI->getCondition())->getOperand(0) ==1181 FICond->getOperand(0))1182 Implication = CondIsTrue;1183 }1184 1185 if (Implication) {1186 BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1);1187 BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0);1188 RemoveSucc->removePredecessor(BB);1189 BranchInst *UncondBI = BranchInst::Create(KeepSucc, BI->getIterator());1190 UncondBI->setDebugLoc(BI->getDebugLoc());1191 ++NumFolds;1192 BI->eraseFromParent();1193 if (FICond)1194 FICond->eraseFromParent();1195 1196 DTU->applyUpdatesPermissive({{DominatorTree::Delete, BB, RemoveSucc}});1197 if (auto *BPI = getBPI())1198 BPI->eraseBlock(BB);1199 return true;1200 }1201 CurrentBB = CurrentPred;1202 CurrentPred = CurrentBB->getSinglePredecessor();1203 }1204 1205 return false;1206}1207 1208/// Return true if Op is an instruction defined in the given block.1209static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {1210 if (Instruction *OpInst = dyn_cast<Instruction>(Op))1211 if (OpInst->getParent() == BB)1212 return true;1213 return false;1214}1215 1216/// simplifyPartiallyRedundantLoad - If LoadI is an obviously partially1217/// redundant load instruction, eliminate it by replacing it with a PHI node.1218/// This is an important optimization that encourages jump threading, and needs1219/// to be run interlaced with other jump threading tasks.1220bool JumpThreadingPass::simplifyPartiallyRedundantLoad(LoadInst *LoadI) {1221 // Don't hack volatile and ordered loads.1222 if (!LoadI->isUnordered()) return false;1223 1224 // If the load is defined in a block with exactly one predecessor, it can't be1225 // partially redundant.1226 BasicBlock *LoadBB = LoadI->getParent();1227 if (LoadBB->getSinglePredecessor())1228 return false;1229 1230 // If the load is defined in an EH pad, it can't be partially redundant,1231 // because the edges between the invoke and the EH pad cannot have other1232 // instructions between them.1233 if (LoadBB->isEHPad())1234 return false;1235 1236 Value *LoadedPtr = LoadI->getOperand(0);1237 1238 // If the loaded operand is defined in the LoadBB and its not a phi,1239 // it can't be available in predecessors.1240 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))1241 return false;1242 1243 // Scan a few instructions up from the load, to see if it is obviously live at1244 // the entry to its block.1245 BasicBlock::iterator BBIt(LoadI);1246 bool IsLoadCSE;1247 BatchAAResults BatchAA(*AA);1248 // The dominator tree is updated lazily and may not be valid at this point.1249 BatchAA.disableDominatorTree();1250 if (Value *AvailableVal = FindAvailableLoadedValue(1251 LoadI, LoadBB, BBIt, DefMaxInstsToScan, &BatchAA, &IsLoadCSE)) {1252 // If the value of the load is locally available within the block, just use1253 // it. This frequently occurs for reg2mem'd allocas.1254 1255 if (IsLoadCSE) {1256 LoadInst *NLoadI = cast<LoadInst>(AvailableVal);1257 combineMetadataForCSE(NLoadI, LoadI, false);1258 LVI->forgetValue(NLoadI);1259 };1260 1261 // If the returned value is the load itself, replace with poison. This can1262 // only happen in dead loops.1263 if (AvailableVal == LoadI)1264 AvailableVal = PoisonValue::get(LoadI->getType());1265 if (AvailableVal->getType() != LoadI->getType()) {1266 AvailableVal = CastInst::CreateBitOrPointerCast(1267 AvailableVal, LoadI->getType(), "", LoadI->getIterator());1268 cast<Instruction>(AvailableVal)->setDebugLoc(LoadI->getDebugLoc());1269 }1270 LoadI->replaceAllUsesWith(AvailableVal);1271 LoadI->eraseFromParent();1272 return true;1273 }1274 1275 // Otherwise, if we scanned the whole block and got to the top of the block,1276 // we know the block is locally transparent to the load. If not, something1277 // might clobber its value.1278 if (BBIt != LoadBB->begin())1279 return false;1280 1281 // If all of the loads and stores that feed the value have the same AA tags,1282 // then we can propagate them onto any newly inserted loads.1283 AAMDNodes AATags = LoadI->getAAMetadata();1284 1285 SmallPtrSet<BasicBlock*, 8> PredsScanned;1286 1287 using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;1288 1289 AvailablePredsTy AvailablePreds;1290 BasicBlock *OneUnavailablePred = nullptr;1291 SmallVector<LoadInst*, 8> CSELoads;1292 1293 // If we got here, the loaded value is transparent through to the start of the1294 // block. Check to see if it is available in any of the predecessor blocks.1295 for (BasicBlock *PredBB : predecessors(LoadBB)) {1296 // If we already scanned this predecessor, skip it.1297 if (!PredsScanned.insert(PredBB).second)1298 continue;1299 1300 BBIt = PredBB->end();1301 unsigned NumScanedInst = 0;1302 Value *PredAvailable = nullptr;1303 // NOTE: We don't CSE load that is volatile or anything stronger than1304 // unordered, that should have been checked when we entered the function.1305 assert(LoadI->isUnordered() &&1306 "Attempting to CSE volatile or atomic loads");1307 // If this is a load on a phi pointer, phi-translate it and search1308 // for available load/store to the pointer in predecessors.1309 Type *AccessTy = LoadI->getType();1310 const auto &DL = LoadI->getDataLayout();1311 MemoryLocation Loc(LoadedPtr->DoPHITranslation(LoadBB, PredBB),1312 LocationSize::precise(DL.getTypeStoreSize(AccessTy)),1313 AATags);1314 PredAvailable = findAvailablePtrLoadStore(1315 Loc, AccessTy, LoadI->isAtomic(), PredBB, BBIt, DefMaxInstsToScan,1316 &BatchAA, &IsLoadCSE, &NumScanedInst);1317 1318 // If PredBB has a single predecessor, continue scanning through the1319 // single predecessor.1320 BasicBlock *SinglePredBB = PredBB;1321 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&1322 NumScanedInst < DefMaxInstsToScan) {1323 SinglePredBB = SinglePredBB->getSinglePredecessor();1324 if (SinglePredBB) {1325 BBIt = SinglePredBB->end();1326 PredAvailable = findAvailablePtrLoadStore(1327 Loc, AccessTy, LoadI->isAtomic(), SinglePredBB, BBIt,1328 (DefMaxInstsToScan - NumScanedInst), &BatchAA, &IsLoadCSE,1329 &NumScanedInst);1330 }1331 }1332 1333 if (!PredAvailable) {1334 OneUnavailablePred = PredBB;1335 continue;1336 }1337 1338 if (IsLoadCSE)1339 CSELoads.push_back(cast<LoadInst>(PredAvailable));1340 1341 // If so, this load is partially redundant. Remember this info so that we1342 // can create a PHI node.1343 AvailablePreds.emplace_back(PredBB, PredAvailable);1344 }1345 1346 // If the loaded value isn't available in any predecessor, it isn't partially1347 // redundant.1348 if (AvailablePreds.empty()) return false;1349 1350 // Okay, the loaded value is available in at least one (and maybe all!)1351 // predecessors. If the value is unavailable in more than one unique1352 // predecessor, we want to insert a merge block for those common predecessors.1353 // This ensures that we only have to insert one reload, thus not increasing1354 // code size.1355 BasicBlock *UnavailablePred = nullptr;1356 1357 // If the value is unavailable in one of predecessors, we will end up1358 // inserting a new instruction into them. It is only valid if all the1359 // instructions before LoadI are guaranteed to pass execution to its1360 // successor, or if LoadI is safe to speculate.1361 // TODO: If this logic becomes more complex, and we will perform PRE insertion1362 // farther than to a predecessor, we need to reuse the code from GVN's PRE.1363 // It requires domination tree analysis, so for this simple case it is an1364 // overkill.1365 if (PredsScanned.size() != AvailablePreds.size() &&1366 !isSafeToSpeculativelyExecute(LoadI))1367 for (auto I = LoadBB->begin(); &*I != LoadI; ++I)1368 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))1369 return false;1370 1371 // If there is exactly one predecessor where the value is unavailable, the1372 // already computed 'OneUnavailablePred' block is it. If it ends in an1373 // unconditional branch, we know that it isn't a critical edge.1374 if (PredsScanned.size() == AvailablePreds.size()+1 &&1375 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {1376 UnavailablePred = OneUnavailablePred;1377 } else if (PredsScanned.size() != AvailablePreds.size()) {1378 // Otherwise, we had multiple unavailable predecessors or we had a critical1379 // edge from the one.1380 SmallVector<BasicBlock*, 8> PredsToSplit;1381 SmallPtrSet<BasicBlock *, 8> AvailablePredSet(1382 llvm::from_range, llvm::make_first_range(AvailablePreds));1383 1384 // Add all the unavailable predecessors to the PredsToSplit list.1385 for (BasicBlock *P : predecessors(LoadBB)) {1386 // If the predecessor is an indirect goto, we can't split the edge.1387 if (isa<IndirectBrInst>(P->getTerminator()))1388 return false;1389 1390 if (!AvailablePredSet.count(P))1391 PredsToSplit.push_back(P);1392 }1393 1394 // Split them out to their own block.1395 UnavailablePred = splitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");1396 }1397 1398 // If the value isn't available in all predecessors, then there will be1399 // exactly one where it isn't available. Insert a load on that edge and add1400 // it to the AvailablePreds list.1401 if (UnavailablePred) {1402 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&1403 "Can't handle critical edge here!");1404 LoadInst *NewVal = new LoadInst(1405 LoadI->getType(), LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),1406 LoadI->getName() + ".pr", false, LoadI->getAlign(),1407 LoadI->getOrdering(), LoadI->getSyncScopeID(),1408 UnavailablePred->getTerminator()->getIterator());1409 NewVal->setDebugLoc(LoadI->getDebugLoc());1410 if (AATags)1411 NewVal->setAAMetadata(AATags);1412 1413 AvailablePreds.emplace_back(UnavailablePred, NewVal);1414 }1415 1416 // Now we know that each predecessor of this block has a value in1417 // AvailablePreds, sort them for efficient access as we're walking the preds.1418 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());1419 1420 // Create a PHI node at the start of the block for the PRE'd load value.1421 PHINode *PN = PHINode::Create(LoadI->getType(), pred_size(LoadBB), "");1422 PN->insertBefore(LoadBB->begin());1423 PN->takeName(LoadI);1424 PN->setDebugLoc(LoadI->getDebugLoc());1425 1426 // Insert new entries into the PHI for each predecessor. A single block may1427 // have multiple entries here.1428 for (BasicBlock *P : predecessors(LoadBB)) {1429 AvailablePredsTy::iterator I =1430 llvm::lower_bound(AvailablePreds, std::make_pair(P, (Value *)nullptr));1431 1432 assert(I != AvailablePreds.end() && I->first == P &&1433 "Didn't find entry for predecessor!");1434 1435 // If we have an available predecessor but it requires casting, insert the1436 // cast in the predecessor and use the cast. Note that we have to update the1437 // AvailablePreds vector as we go so that all of the PHI entries for this1438 // predecessor use the same bitcast.1439 Value *&PredV = I->second;1440 if (PredV->getType() != LoadI->getType()) {1441 PredV = CastInst::CreateBitOrPointerCast(1442 PredV, LoadI->getType(), "", P->getTerminator()->getIterator());1443 // The new cast is producing the value used to replace the load1444 // instruction, so uses the load's debug location. If P does not always1445 // branch to the load BB however then the debug location must be dropped,1446 // as it is hoisted past a conditional branch.1447 DebugLoc DL = P->getTerminator()->getNumSuccessors() == 11448 ? LoadI->getDebugLoc()1449 : DebugLoc::getDropped();1450 cast<CastInst>(PredV)->setDebugLoc(DL);1451 }1452 1453 PN->addIncoming(PredV, I->first);1454 }1455 1456 for (LoadInst *PredLoadI : CSELoads) {1457 combineMetadataForCSE(PredLoadI, LoadI, true);1458 LVI->forgetValue(PredLoadI);1459 }1460 1461 LoadI->replaceAllUsesWith(PN);1462 LoadI->eraseFromParent();1463 1464 return true;1465}1466 1467/// findMostPopularDest - The specified list contains multiple possible1468/// threadable destinations. Pick the one that occurs the most frequently in1469/// the list.1470static BasicBlock *1471findMostPopularDest(BasicBlock *BB,1472 const SmallVectorImpl<std::pair<BasicBlock *,1473 BasicBlock *>> &PredToDestList) {1474 assert(!PredToDestList.empty());1475 1476 // Determine popularity. If there are multiple possible destinations, we1477 // explicitly choose to ignore 'undef' destinations. We prefer to thread1478 // blocks with known and real destinations to threading undef. We'll handle1479 // them later if interesting.1480 MapVector<BasicBlock *, unsigned> DestPopularity;1481 1482 // Populate DestPopularity with the successors in the order they appear in the1483 // successor list. This way, we ensure determinism by iterating it in the1484 // same order in llvm::max_element below. We map nullptr to 0 so that we can1485 // return nullptr when PredToDestList contains nullptr only.1486 DestPopularity[nullptr] = 0;1487 for (auto *SuccBB : successors(BB))1488 DestPopularity[SuccBB] = 0;1489 1490 for (const auto &PredToDest : PredToDestList)1491 if (PredToDest.second)1492 DestPopularity[PredToDest.second]++;1493 1494 // Find the most popular dest.1495 auto MostPopular = llvm::max_element(DestPopularity, llvm::less_second());1496 1497 // Okay, we have finally picked the most popular destination.1498 return MostPopular->first;1499}1500 1501// Try to evaluate the value of V when the control flows from PredPredBB to1502// BB->getSinglePredecessor() and then on to BB.1503Constant *JumpThreadingPass::evaluateOnPredecessorEdge(BasicBlock *BB,1504 BasicBlock *PredPredBB,1505 Value *V,1506 const DataLayout &DL) {1507 SmallPtrSet<Value *, 8> Visited;1508 return evaluateOnPredecessorEdge(BB, PredPredBB, V, DL, Visited);1509}1510 1511Constant *JumpThreadingPass::evaluateOnPredecessorEdge(1512 BasicBlock *BB, BasicBlock *PredPredBB, Value *V, const DataLayout &DL,1513 SmallPtrSet<Value *, 8> &Visited) {1514 if (!Visited.insert(V).second)1515 return nullptr;1516 auto _ = make_scope_exit([&Visited, V]() { Visited.erase(V); });1517 1518 BasicBlock *PredBB = BB->getSinglePredecessor();1519 assert(PredBB && "Expected a single predecessor");1520 1521 if (Constant *Cst = dyn_cast<Constant>(V)) {1522 return Cst;1523 }1524 1525 // Consult LVI if V is not an instruction in BB or PredBB.1526 Instruction *I = dyn_cast<Instruction>(V);1527 if (!I || (I->getParent() != BB && I->getParent() != PredBB)) {1528 return LVI->getConstantOnEdge(V, PredPredBB, PredBB, nullptr);1529 }1530 1531 // Look into a PHI argument.1532 if (PHINode *PHI = dyn_cast<PHINode>(V)) {1533 if (PHI->getParent() == PredBB)1534 return dyn_cast<Constant>(PHI->getIncomingValueForBlock(PredPredBB));1535 return nullptr;1536 }1537 1538 // If we have a CmpInst, try to fold it for each incoming edge into PredBB.1539 // Note that during the execution of the pass, phi nodes may become constant1540 // and may be removed, which can lead to self-referencing instructions in1541 // code that becomes unreachable. Consequently, we need to handle those1542 // instructions in unreachable code and check before going into recursion.1543 if (CmpInst *CondCmp = dyn_cast<CmpInst>(V)) {1544 if (CondCmp->getParent() == BB) {1545 Constant *Op0 = evaluateOnPredecessorEdge(1546 BB, PredPredBB, CondCmp->getOperand(0), DL, Visited);1547 Constant *Op1 = evaluateOnPredecessorEdge(1548 BB, PredPredBB, CondCmp->getOperand(1), DL, Visited);1549 if (Op0 && Op1) {1550 return ConstantFoldCompareInstOperands(CondCmp->getPredicate(), Op0,1551 Op1, DL);1552 }1553 }1554 return nullptr;1555 }1556 1557 return nullptr;1558}1559 1560bool JumpThreadingPass::processThreadableEdges(Value *Cond, BasicBlock *BB,1561 ConstantPreference Preference,1562 Instruction *CxtI) {1563 // If threading this would thread across a loop header, don't even try to1564 // thread the edge.1565 if (LoopHeaders.count(BB))1566 return false;1567 1568 PredValueInfoTy PredValues;1569 if (!computeValueKnownInPredecessors(Cond, BB, PredValues, Preference,1570 CxtI)) {1571 // We don't have known values in predecessors. See if we can thread through1572 // BB and its sole predecessor.1573 return maybethreadThroughTwoBasicBlocks(BB, Cond);1574 }1575 1576 assert(!PredValues.empty() &&1577 "computeValueKnownInPredecessors returned true with no values");1578 1579 LLVM_DEBUG(dbgs() << "IN BB: " << *BB;1580 for (const auto &PredValue : PredValues) {1581 dbgs() << " BB '" << BB->getName()1582 << "': FOUND condition = " << *PredValue.first1583 << " for pred '" << PredValue.second->getName() << "'.\n";1584 });1585 1586 // Decide what we want to thread through. Convert our list of known values to1587 // a list of known destinations for each pred. This also discards duplicate1588 // predecessors and keeps track of the undefined inputs (which are represented1589 // as a null dest in the PredToDestList).1590 SmallPtrSet<BasicBlock*, 16> SeenPreds;1591 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;1592 1593 BasicBlock *OnlyDest = nullptr;1594 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;1595 Constant *OnlyVal = nullptr;1596 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;1597 1598 for (const auto &PredValue : PredValues) {1599 BasicBlock *Pred = PredValue.second;1600 if (!SeenPreds.insert(Pred).second)1601 continue; // Duplicate predecessor entry.1602 1603 Constant *Val = PredValue.first;1604 1605 BasicBlock *DestBB;1606 if (isa<UndefValue>(Val))1607 DestBB = nullptr;1608 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {1609 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");1610 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());1611 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {1612 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");1613 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();1614 } else {1615 assert(isa<IndirectBrInst>(BB->getTerminator())1616 && "Unexpected terminator");1617 assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress");1618 DestBB = cast<BlockAddress>(Val)->getBasicBlock();1619 }1620 1621 // If we have exactly one destination, remember it for efficiency below.1622 if (PredToDestList.empty()) {1623 OnlyDest = DestBB;1624 OnlyVal = Val;1625 } else {1626 if (OnlyDest != DestBB)1627 OnlyDest = MultipleDestSentinel;1628 // It possible we have same destination, but different value, e.g. default1629 // case in switchinst.1630 if (Val != OnlyVal)1631 OnlyVal = MultipleVal;1632 }1633 1634 // If the predecessor ends with an indirect goto, we can't change its1635 // destination.1636 if (isa<IndirectBrInst>(Pred->getTerminator()))1637 continue;1638 1639 PredToDestList.emplace_back(Pred, DestBB);1640 }1641 1642 // If all edges were unthreadable, we fail.1643 if (PredToDestList.empty())1644 return false;1645 1646 // If all the predecessors go to a single known successor, we want to fold,1647 // not thread. By doing so, we do not need to duplicate the current block and1648 // also miss potential opportunities in case we dont/cant duplicate.1649 if (OnlyDest && OnlyDest != MultipleDestSentinel) {1650 if (BB->hasNPredecessors(PredToDestList.size())) {1651 bool SeenFirstBranchToOnlyDest = false;1652 std::vector <DominatorTree::UpdateType> Updates;1653 Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);1654 for (BasicBlock *SuccBB : successors(BB)) {1655 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {1656 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.1657 } else {1658 SuccBB->removePredecessor(BB, true); // This is unreachable successor.1659 Updates.push_back({DominatorTree::Delete, BB, SuccBB});1660 }1661 }1662 1663 // Finally update the terminator.1664 Instruction *Term = BB->getTerminator();1665 Instruction *NewBI = BranchInst::Create(OnlyDest, Term->getIterator());1666 NewBI->setDebugLoc(Term->getDebugLoc());1667 ++NumFolds;1668 Term->eraseFromParent();1669 DTU->applyUpdatesPermissive(Updates);1670 if (auto *BPI = getBPI())1671 BPI->eraseBlock(BB);1672 1673 // If the condition is now dead due to the removal of the old terminator,1674 // erase it.1675 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {1676 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())1677 CondInst->eraseFromParent();1678 // We can safely replace *some* uses of the CondInst if it has1679 // exactly one value as returned by LVI. RAUW is incorrect in the1680 // presence of guards and assumes, that have the `Cond` as the use. This1681 // is because we use the guards/assume to reason about the `Cond` value1682 // at the end of block, but RAUW unconditionally replaces all uses1683 // including the guards/assumes themselves and the uses before the1684 // guard/assume.1685 else if (OnlyVal && OnlyVal != MultipleVal)1686 replaceFoldableUses(CondInst, OnlyVal, BB);1687 }1688 return true;1689 }1690 }1691 1692 // Determine which is the most common successor. If we have many inputs and1693 // this block is a switch, we want to start by threading the batch that goes1694 // to the most popular destination first. If we only know about one1695 // threadable destination (the common case) we can avoid this.1696 BasicBlock *MostPopularDest = OnlyDest;1697 1698 if (MostPopularDest == MultipleDestSentinel) {1699 // Remove any loop headers from the Dest list, threadEdge conservatively1700 // won't process them, but we might have other destination that are eligible1701 // and we still want to process.1702 erase_if(PredToDestList,1703 [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {1704 return LoopHeaders.contains(PredToDest.second);1705 });1706 1707 if (PredToDestList.empty())1708 return false;1709 1710 MostPopularDest = findMostPopularDest(BB, PredToDestList);1711 }1712 1713 // Now that we know what the most popular destination is, factor all1714 // predecessors that will jump to it into a single predecessor.1715 SmallVector<BasicBlock*, 16> PredsToFactor;1716 for (const auto &PredToDest : PredToDestList)1717 if (PredToDest.second == MostPopularDest) {1718 BasicBlock *Pred = PredToDest.first;1719 1720 // This predecessor may be a switch or something else that has multiple1721 // edges to the block. Factor each of these edges by listing them1722 // according to # occurrences in PredsToFactor.1723 for (BasicBlock *Succ : successors(Pred))1724 if (Succ == BB)1725 PredsToFactor.push_back(Pred);1726 }1727 1728 // If the threadable edges are branching on an undefined value, we get to pick1729 // the destination that these predecessors should get to.1730 if (!MostPopularDest)1731 MostPopularDest = BB->getTerminator()->1732 getSuccessor(getBestDestForJumpOnUndef(BB));1733 1734 // Ok, try to thread it!1735 return tryThreadEdge(BB, PredsToFactor, MostPopularDest);1736}1737 1738/// processBranchOnPHI - We have an otherwise unthreadable conditional branch on1739/// a PHI node (or freeze PHI) in the current block. See if there are any1740/// simplifications we can do based on inputs to the phi node.1741bool JumpThreadingPass::processBranchOnPHI(PHINode *PN) {1742 BasicBlock *BB = PN->getParent();1743 1744 // TODO: We could make use of this to do it once for blocks with common PHI1745 // values.1746 SmallVector<BasicBlock*, 1> PredBBs;1747 PredBBs.resize(1);1748 1749 // If any of the predecessor blocks end in an unconditional branch, we can1750 // *duplicate* the conditional branch into that block in order to further1751 // encourage jump threading and to eliminate cases where we have branch on a1752 // phi of an icmp (branch on icmp is much better).1753 // This is still beneficial when a frozen phi is used as the branch condition1754 // because it allows CodeGenPrepare to further canonicalize br(freeze(icmp))1755 // to br(icmp(freeze ...)).1756 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {1757 BasicBlock *PredBB = PN->getIncomingBlock(i);1758 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))1759 if (PredBr->isUnconditional()) {1760 PredBBs[0] = PredBB;1761 // Try to duplicate BB into PredBB.1762 if (duplicateCondBranchOnPHIIntoPred(BB, PredBBs))1763 return true;1764 }1765 }1766 1767 return false;1768}1769 1770/// processBranchOnXOR - We have an otherwise unthreadable conditional branch on1771/// a xor instruction in the current block. See if there are any1772/// simplifications we can do based on inputs to the xor.1773bool JumpThreadingPass::processBranchOnXOR(BinaryOperator *BO) {1774 BasicBlock *BB = BO->getParent();1775 1776 // If either the LHS or RHS of the xor is a constant, don't do this1777 // optimization.1778 if (isa<ConstantInt>(BO->getOperand(0)) ||1779 isa<ConstantInt>(BO->getOperand(1)))1780 return false;1781 1782 // If the first instruction in BB isn't a phi, we won't be able to infer1783 // anything special about any particular predecessor.1784 if (!isa<PHINode>(BB->front()))1785 return false;1786 1787 // If this BB is a landing pad, we won't be able to split the edge into it.1788 if (BB->isEHPad())1789 return false;1790 1791 // If we have a xor as the branch input to this block, and we know that the1792 // LHS or RHS of the xor in any predecessor is true/false, then we can clone1793 // the condition into the predecessor and fix that value to true, saving some1794 // logical ops on that path and encouraging other paths to simplify.1795 //1796 // This copies something like this:1797 //1798 // BB:1799 // %X = phi i1 [1], [%X']1800 // %Y = icmp eq i32 %A, %B1801 // %Z = xor i1 %X, %Y1802 // br i1 %Z, ...1803 //1804 // Into:1805 // BB':1806 // %Y = icmp ne i32 %A, %B1807 // br i1 %Y, ...1808 1809 PredValueInfoTy XorOpValues;1810 bool isLHS = true;1811 if (!computeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,1812 WantInteger, BO)) {1813 assert(XorOpValues.empty());1814 if (!computeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,1815 WantInteger, BO))1816 return false;1817 isLHS = false;1818 }1819 1820 assert(!XorOpValues.empty() &&1821 "computeValueKnownInPredecessors returned true with no values");1822 1823 // Scan the information to see which is most popular: true or false. The1824 // predecessors can be of the set true, false, or undef.1825 unsigned NumTrue = 0, NumFalse = 0;1826 for (const auto &XorOpValue : XorOpValues) {1827 if (isa<UndefValue>(XorOpValue.first))1828 // Ignore undefs for the count.1829 continue;1830 if (cast<ConstantInt>(XorOpValue.first)->isZero())1831 ++NumFalse;1832 else1833 ++NumTrue;1834 }1835 1836 // Determine which value to split on, true, false, or undef if neither.1837 ConstantInt *SplitVal = nullptr;1838 if (NumTrue > NumFalse)1839 SplitVal = ConstantInt::getTrue(BB->getContext());1840 else if (NumTrue != 0 || NumFalse != 0)1841 SplitVal = ConstantInt::getFalse(BB->getContext());1842 1843 // Collect all of the blocks that this can be folded into so that we can1844 // factor this once and clone it once.1845 SmallVector<BasicBlock*, 8> BlocksToFoldInto;1846 for (const auto &XorOpValue : XorOpValues) {1847 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))1848 continue;1849 1850 BlocksToFoldInto.push_back(XorOpValue.second);1851 }1852 1853 // If we inferred a value for all of the predecessors, then duplication won't1854 // help us. However, we can just replace the LHS or RHS with the constant.1855 if (BlocksToFoldInto.size() ==1856 cast<PHINode>(BB->front()).getNumIncomingValues()) {1857 if (!SplitVal) {1858 // If all preds provide undef, just nuke the xor, because it is undef too.1859 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));1860 BO->eraseFromParent();1861 } else if (SplitVal->isZero() && BO != BO->getOperand(isLHS)) {1862 // If all preds provide 0, replace the xor with the other input.1863 BO->replaceAllUsesWith(BO->getOperand(isLHS));1864 BO->eraseFromParent();1865 } else {1866 // If all preds provide 1, set the computed value to 1.1867 BO->setOperand(!isLHS, SplitVal);1868 }1869 1870 return true;1871 }1872 1873 // If any of predecessors end with an indirect goto, we can't change its1874 // destination.1875 if (any_of(BlocksToFoldInto, [](BasicBlock *Pred) {1876 return isa<IndirectBrInst>(Pred->getTerminator());1877 }))1878 return false;1879 1880 // Try to duplicate BB into PredBB.1881 return duplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);1882}1883 1884/// addPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new1885/// predecessor to the PHIBB block. If it has PHI nodes, add entries for1886/// NewPred using the entries from OldPred (suitably mapped).1887static void addPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,1888 BasicBlock *OldPred,1889 BasicBlock *NewPred,1890 ValueToValueMapTy &ValueMap) {1891 for (PHINode &PN : PHIBB->phis()) {1892 // Ok, we have a PHI node. Figure out what the incoming value was for the1893 // DestBlock.1894 Value *IV = PN.getIncomingValueForBlock(OldPred);1895 1896 // Remap the value if necessary.1897 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {1898 ValueToValueMapTy::iterator I = ValueMap.find(Inst);1899 if (I != ValueMap.end())1900 IV = I->second;1901 }1902 1903 PN.addIncoming(IV, NewPred);1904 }1905}1906 1907/// Merge basic block BB into its sole predecessor if possible.1908bool JumpThreadingPass::maybeMergeBasicBlockIntoOnlyPred(BasicBlock *BB) {1909 BasicBlock *SinglePred = BB->getSinglePredecessor();1910 if (!SinglePred)1911 return false;1912 1913 const Instruction *TI = SinglePred->getTerminator();1914 if (TI->isSpecialTerminator() || TI->getNumSuccessors() != 1 ||1915 SinglePred == BB || hasAddressTakenAndUsed(BB))1916 return false;1917 1918 // MergeBasicBlockIntoOnlyPred may delete SinglePred, we need to avoid1919 // deleting a BB pointer from Unreachable.1920 if (Unreachable.count(SinglePred))1921 return false;1922 1923 // If SinglePred was a loop header, BB becomes one.1924 if (LoopHeaders.erase(SinglePred))1925 LoopHeaders.insert(BB);1926 1927 LVI->eraseBlock(SinglePred);1928 MergeBasicBlockIntoOnlyPred(BB, DTU.get());1929 1930 // Now that BB is merged into SinglePred (i.e. SinglePred code followed by1931 // BB code within one basic block `BB`), we need to invalidate the LVI1932 // information associated with BB, because the LVI information need not be1933 // true for all of BB after the merge. For example,1934 // Before the merge, LVI info and code is as follows:1935 // SinglePred: <LVI info1 for %p val>1936 // %y = use of %p1937 // call @exit() // need not transfer execution to successor.1938 // assume(%p) // from this point on %p is true1939 // br label %BB1940 // BB: <LVI info2 for %p val, i.e. %p is true>1941 // %x = use of %p1942 // br label exit1943 //1944 // Note that this LVI info for blocks BB and SinglPred is correct for %p1945 // (info2 and info1 respectively). After the merge and the deletion of the1946 // LVI info1 for SinglePred. We have the following code:1947 // BB: <LVI info2 for %p val>1948 // %y = use of %p1949 // call @exit()1950 // assume(%p)1951 // %x = use of %p <-- LVI info2 is correct from here onwards.1952 // br label exit1953 // LVI info2 for BB is incorrect at the beginning of BB.1954 1955 // Invalidate LVI information for BB if the LVI is not provably true for1956 // all of BB.1957 if (!isGuaranteedToTransferExecutionToSuccessor(BB))1958 LVI->eraseBlock(BB);1959 return true;1960}1961 1962/// Update the SSA form. NewBB contains instructions that are copied from BB.1963/// ValueMapping maps old values in BB to new ones in NewBB.1964void JumpThreadingPass::updateSSA(BasicBlock *BB, BasicBlock *NewBB,1965 ValueToValueMapTy &ValueMapping) {1966 // If there were values defined in BB that are used outside the block, then we1967 // now have to update all uses of the value to use either the original value,1968 // the cloned value, or some PHI derived value. This can require arbitrary1969 // PHI insertion, of which we are prepared to do, clean these up now.1970 SSAUpdater SSAUpdate;1971 SmallVector<Use *, 16> UsesToRename;1972 SmallVector<DbgVariableRecord *, 4> DbgVariableRecords;1973 1974 for (Instruction &I : *BB) {1975 // Scan all uses of this instruction to see if it is used outside of its1976 // block, and if so, record them in UsesToRename.1977 for (Use &U : I.uses()) {1978 Instruction *User = cast<Instruction>(U.getUser());1979 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {1980 if (UserPN->getIncomingBlock(U) == BB)1981 continue;1982 } else if (User->getParent() == BB)1983 continue;1984 1985 UsesToRename.push_back(&U);1986 }1987 1988 // Find debug values outside of the block1989 findDbgValues(&I, DbgVariableRecords);1990 llvm::erase_if(DbgVariableRecords, [&](const DbgVariableRecord *DbgVarRec) {1991 return DbgVarRec->getParent() == BB;1992 });1993 1994 // If there are no uses outside the block, we're done with this instruction.1995 if (UsesToRename.empty() && DbgVariableRecords.empty())1996 continue;1997 LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");1998 1999 // We found a use of I outside of BB. Rename all uses of I that are outside2000 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks2001 // with the two values we know.2002 SSAUpdate.Initialize(I.getType(), I.getName());2003 SSAUpdate.AddAvailableValue(BB, &I);2004 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);2005 2006 while (!UsesToRename.empty())2007 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());2008 if (!DbgVariableRecords.empty()) {2009 SSAUpdate.UpdateDebugValues(&I, DbgVariableRecords);2010 DbgVariableRecords.clear();2011 }2012 2013 LLVM_DEBUG(dbgs() << "\n");2014 }2015}2016 2017static void remapSourceAtoms(ValueToValueMapTy &VM, BasicBlock::iterator Begin,2018 BasicBlock::iterator End) {2019 if (VM.AtomMap.empty())2020 return;2021 for (auto It = Begin; It != End; ++It)2022 RemapSourceAtom(&*It, VM);2023}2024 2025/// Clone instructions in range [BI, BE) to NewBB. For PHI nodes, we only clone2026/// arguments that come from PredBB. Return the map from the variables in the2027/// source basic block to the variables in the newly created basic block.2028 2029void JumpThreadingPass::cloneInstructions(ValueToValueMapTy &ValueMapping,2030 BasicBlock::iterator BI,2031 BasicBlock::iterator BE,2032 BasicBlock *NewBB,2033 BasicBlock *PredBB) {2034 // We are going to have to map operands from the source basic block to the new2035 // copy of the block 'NewBB'. If there are PHI nodes in the source basic2036 // block, evaluate them to account for entry from PredBB.2037 2038 // Retargets dbg.value to any renamed variables.2039 auto RetargetDbgVariableRecordIfPossible = [&](DbgVariableRecord *DVR) {2040 SmallSet<std::pair<Value *, Value *>, 16> OperandsToRemap;2041 for (auto *Op : DVR->location_ops()) {2042 Instruction *OpInst = dyn_cast<Instruction>(Op);2043 if (!OpInst)2044 continue;2045 2046 auto I = ValueMapping.find(OpInst);2047 if (I != ValueMapping.end())2048 OperandsToRemap.insert({OpInst, I->second});2049 }2050 2051 for (auto &[OldOp, MappedOp] : OperandsToRemap)2052 DVR->replaceVariableLocationOp(OldOp, MappedOp);2053 };2054 2055 BasicBlock *RangeBB = BI->getParent();2056 2057 // Clone the phi nodes of the source basic block into NewBB. The resulting2058 // phi nodes are trivial since NewBB only has one predecessor, but SSAUpdater2059 // might need to rewrite the operand of the cloned phi.2060 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {2061 PHINode *NewPN = PHINode::Create(PN->getType(), 1, PN->getName(), NewBB);2062 NewPN->addIncoming(PN->getIncomingValueForBlock(PredBB), PredBB);2063 ValueMapping[PN] = NewPN;2064 if (const DebugLoc &DL = PN->getDebugLoc())2065 mapAtomInstance(DL, ValueMapping);2066 }2067 2068 // Clone noalias scope declarations in the threaded block. When threading a2069 // loop exit, we would otherwise end up with two idential scope declarations2070 // visible at the same time.2071 SmallVector<MDNode *> NoAliasScopes;2072 DenseMap<MDNode *, MDNode *> ClonedScopes;2073 LLVMContext &Context = PredBB->getContext();2074 identifyNoAliasScopesToClone(BI, BE, NoAliasScopes);2075 cloneNoAliasScopes(NoAliasScopes, ClonedScopes, "thread", Context);2076 2077 auto CloneAndRemapDbgInfo = [&](Instruction *NewInst, Instruction *From) {2078 auto DVRRange = NewInst->cloneDebugInfoFrom(From);2079 for (DbgVariableRecord &DVR : filterDbgVars(DVRRange))2080 RetargetDbgVariableRecordIfPossible(&DVR);2081 };2082 2083 // Clone the non-phi instructions of the source basic block into NewBB,2084 // keeping track of the mapping and using it to remap operands in the cloned2085 // instructions.2086 for (; BI != BE; ++BI) {2087 Instruction *New = BI->clone();2088 New->setName(BI->getName());2089 New->insertInto(NewBB, NewBB->end());2090 ValueMapping[&*BI] = New;2091 adaptNoAliasScopes(New, ClonedScopes, Context);2092 2093 CloneAndRemapDbgInfo(New, &*BI);2094 if (const DebugLoc &DL = New->getDebugLoc())2095 mapAtomInstance(DL, ValueMapping);2096 2097 // Remap operands to patch up intra-block references.2098 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)2099 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {2100 ValueToValueMapTy::iterator I = ValueMapping.find(Inst);2101 if (I != ValueMapping.end())2102 New->setOperand(i, I->second);2103 }2104 }2105 2106 // There may be DbgVariableRecords on the terminator, clone directly from2107 // marker to marker as there isn't an instruction there.2108 if (BE != RangeBB->end() && BE->hasDbgRecords()) {2109 // Dump them at the end.2110 DbgMarker *Marker = RangeBB->getMarker(BE);2111 DbgMarker *EndMarker = NewBB->createMarker(NewBB->end());2112 auto DVRRange = EndMarker->cloneDebugInfoFrom(Marker, std::nullopt);2113 for (DbgVariableRecord &DVR : filterDbgVars(DVRRange))2114 RetargetDbgVariableRecordIfPossible(&DVR);2115 }2116}2117 2118/// Attempt to thread through two successive basic blocks.2119bool JumpThreadingPass::maybethreadThroughTwoBasicBlocks(BasicBlock *BB,2120 Value *Cond) {2121 // Consider:2122 //2123 // PredBB:2124 // %var = phi i32* [ null, %bb1 ], [ @a, %bb2 ]2125 // %tobool = icmp eq i32 %cond, 02126 // br i1 %tobool, label %BB, label ...2127 //2128 // BB:2129 // %cmp = icmp eq i32* %var, null2130 // br i1 %cmp, label ..., label ...2131 //2132 // We don't know the value of %var at BB even if we know which incoming edge2133 // we take to BB. However, once we duplicate PredBB for each of its incoming2134 // edges (say, PredBB1 and PredBB2), we know the value of %var in each copy of2135 // PredBB. Then we can thread edges PredBB1->BB and PredBB2->BB through BB.2136 2137 // Require that BB end with a Branch for simplicity.2138 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());2139 if (!CondBr)2140 return false;2141 2142 // BB must have exactly one predecessor.2143 BasicBlock *PredBB = BB->getSinglePredecessor();2144 if (!PredBB)2145 return false;2146 2147 // Require that PredBB end with a conditional Branch. If PredBB ends with an2148 // unconditional branch, we should be merging PredBB and BB instead. For2149 // simplicity, we don't deal with a switch.2150 BranchInst *PredBBBranch = dyn_cast<BranchInst>(PredBB->getTerminator());2151 if (!PredBBBranch || PredBBBranch->isUnconditional())2152 return false;2153 2154 // If PredBB has exactly one incoming edge, we don't gain anything by copying2155 // PredBB.2156 if (PredBB->getSinglePredecessor())2157 return false;2158 2159 // Don't thread through PredBB if it contains a successor edge to itself, in2160 // which case we would infinite loop. Suppose we are threading an edge from2161 // PredPredBB through PredBB and BB to SuccBB with PredBB containing a2162 // successor edge to itself. If we allowed jump threading in this case, we2163 // could duplicate PredBB and BB as, say, PredBB.thread and BB.thread. Since2164 // PredBB.thread has a successor edge to PredBB, we would immediately come up2165 // with another jump threading opportunity from PredBB.thread through PredBB2166 // and BB to SuccBB. This jump threading would repeatedly occur. That is, we2167 // would keep peeling one iteration from PredBB.2168 if (llvm::is_contained(successors(PredBB), PredBB))2169 return false;2170 2171 // Don't thread across a loop header.2172 if (LoopHeaders.count(PredBB))2173 return false;2174 2175 // Avoid complication with duplicating EH pads.2176 if (PredBB->isEHPad())2177 return false;2178 2179 // Find a predecessor that we can thread. For simplicity, we only consider a2180 // successor edge out of BB to which we thread exactly one incoming edge into2181 // PredBB.2182 unsigned ZeroCount = 0;2183 unsigned OneCount = 0;2184 BasicBlock *ZeroPred = nullptr;2185 BasicBlock *OnePred = nullptr;2186 const DataLayout &DL = BB->getDataLayout();2187 for (BasicBlock *P : predecessors(PredBB)) {2188 // If PredPred ends with IndirectBrInst, we can't handle it.2189 if (isa<IndirectBrInst>(P->getTerminator()))2190 continue;2191 if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>(2192 evaluateOnPredecessorEdge(BB, P, Cond, DL))) {2193 if (CI->isZero()) {2194 ZeroCount++;2195 ZeroPred = P;2196 } else if (CI->isOne()) {2197 OneCount++;2198 OnePred = P;2199 }2200 }2201 }2202 2203 // Disregard complicated cases where we have to thread multiple edges.2204 BasicBlock *PredPredBB;2205 if (ZeroCount == 1) {2206 PredPredBB = ZeroPred;2207 } else if (OneCount == 1) {2208 PredPredBB = OnePred;2209 } else {2210 return false;2211 }2212 2213 BasicBlock *SuccBB = CondBr->getSuccessor(PredPredBB == ZeroPred);2214 2215 // If threading to the same block as we come from, we would infinite loop.2216 if (SuccBB == BB) {2217 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()2218 << "' - would thread to self!\n");2219 return false;2220 }2221 2222 // If threading this would thread across a loop header, don't thread the edge.2223 // See the comments above findLoopHeaders for justifications and caveats.2224 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {2225 LLVM_DEBUG({2226 bool BBIsHeader = LoopHeaders.count(BB);2227 bool SuccIsHeader = LoopHeaders.count(SuccBB);2228 dbgs() << " Not threading across "2229 << (BBIsHeader ? "loop header BB '" : "block BB '")2230 << BB->getName() << "' to dest "2231 << (SuccIsHeader ? "loop header BB '" : "block BB '")2232 << SuccBB->getName()2233 << "' - it might create an irreducible loop!\n";2234 });2235 return false;2236 }2237 2238 // Compute the cost of duplicating BB and PredBB.2239 unsigned BBCost = getJumpThreadDuplicationCost(2240 TTI, BB, BB->getTerminator(), BBDupThreshold);2241 unsigned PredBBCost = getJumpThreadDuplicationCost(2242 TTI, PredBB, PredBB->getTerminator(), BBDupThreshold);2243 2244 // Give up if costs are too high. We need to check BBCost and PredBBCost2245 // individually before checking their sum because getJumpThreadDuplicationCost2246 // return (unsigned)~0 for those basic blocks that cannot be duplicated.2247 if (BBCost > BBDupThreshold || PredBBCost > BBDupThreshold ||2248 BBCost + PredBBCost > BBDupThreshold) {2249 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()2250 << "' - Cost is too high: " << PredBBCost2251 << " for PredBB, " << BBCost << "for BB\n");2252 return false;2253 }2254 2255 // Now we are ready to duplicate PredBB.2256 threadThroughTwoBasicBlocks(PredPredBB, PredBB, BB, SuccBB);2257 return true;2258}2259 2260void JumpThreadingPass::threadThroughTwoBasicBlocks(BasicBlock *PredPredBB,2261 BasicBlock *PredBB,2262 BasicBlock *BB,2263 BasicBlock *SuccBB) {2264 LLVM_DEBUG(dbgs() << " Threading through '" << PredBB->getName() << "' and '"2265 << BB->getName() << "'\n");2266 2267 // Build BPI/BFI before any changes are made to IR.2268 bool HasProfile = doesBlockHaveProfileData(BB);2269 auto *BFI = getOrCreateBFI(HasProfile);2270 auto *BPI = getOrCreateBPI(BFI != nullptr);2271 2272 BranchInst *CondBr = cast<BranchInst>(BB->getTerminator());2273 BranchInst *PredBBBranch = cast<BranchInst>(PredBB->getTerminator());2274 2275 BasicBlock *NewBB =2276 BasicBlock::Create(PredBB->getContext(), PredBB->getName() + ".thread",2277 PredBB->getParent(), PredBB);2278 NewBB->moveAfter(PredBB);2279 2280 // Set the block frequency of NewBB.2281 if (BFI) {2282 assert(BPI && "It's expected BPI to exist along with BFI");2283 auto NewBBFreq = BFI->getBlockFreq(PredPredBB) *2284 BPI->getEdgeProbability(PredPredBB, PredBB);2285 BFI->setBlockFreq(NewBB, NewBBFreq);2286 }2287 2288 // We are going to have to map operands from the original BB block to the new2289 // copy of the block 'NewBB'. If there are PHI nodes in PredBB, evaluate them2290 // to account for entry from PredPredBB.2291 ValueToValueMapTy ValueMapping;2292 cloneInstructions(ValueMapping, PredBB->begin(), PredBB->end(), NewBB,2293 PredPredBB);2294 2295 // Copy the edge probabilities from PredBB to NewBB.2296 if (BPI)2297 BPI->copyEdgeProbabilities(PredBB, NewBB);2298 2299 // Update the terminator of PredPredBB to jump to NewBB instead of PredBB.2300 // This eliminates predecessors from PredPredBB, which requires us to simplify2301 // any PHI nodes in PredBB.2302 Instruction *PredPredTerm = PredPredBB->getTerminator();2303 for (unsigned i = 0, e = PredPredTerm->getNumSuccessors(); i != e; ++i)2304 if (PredPredTerm->getSuccessor(i) == PredBB) {2305 PredBB->removePredecessor(PredPredBB, true);2306 PredPredTerm->setSuccessor(i, NewBB);2307 }2308 2309 addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(0), PredBB, NewBB,2310 ValueMapping);2311 addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(1), PredBB, NewBB,2312 ValueMapping);2313 2314 DTU->applyUpdatesPermissive(2315 {{DominatorTree::Insert, NewBB, CondBr->getSuccessor(0)},2316 {DominatorTree::Insert, NewBB, CondBr->getSuccessor(1)},2317 {DominatorTree::Insert, PredPredBB, NewBB},2318 {DominatorTree::Delete, PredPredBB, PredBB}});2319 2320 // Remap source location atoms beacuse we're duplicating control flow.2321 remapSourceAtoms(ValueMapping, NewBB->begin(), NewBB->end());2322 2323 updateSSA(PredBB, NewBB, ValueMapping);2324 2325 // Clean up things like PHI nodes with single operands, dead instructions,2326 // etc.2327 SimplifyInstructionsInBlock(NewBB, TLI);2328 SimplifyInstructionsInBlock(PredBB, TLI);2329 2330 SmallVector<BasicBlock *, 1> PredsToFactor;2331 PredsToFactor.push_back(NewBB);2332 threadEdge(BB, PredsToFactor, SuccBB);2333}2334 2335/// tryThreadEdge - Thread an edge if it's safe and profitable to do so.2336bool JumpThreadingPass::tryThreadEdge(2337 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs,2338 BasicBlock *SuccBB) {2339 // If threading to the same block as we come from, we would infinite loop.2340 if (SuccBB == BB) {2341 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()2342 << "' - would thread to self!\n");2343 return false;2344 }2345 2346 // If threading this would thread across a loop header, don't thread the edge.2347 // See the comments above findLoopHeaders for justifications and caveats.2348 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {2349 LLVM_DEBUG({2350 bool BBIsHeader = LoopHeaders.count(BB);2351 bool SuccIsHeader = LoopHeaders.count(SuccBB);2352 dbgs() << " Not threading across "2353 << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName()2354 << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '")2355 << SuccBB->getName() << "' - it might create an irreducible loop!\n";2356 });2357 return false;2358 }2359 2360 unsigned JumpThreadCost = getJumpThreadDuplicationCost(2361 TTI, BB, BB->getTerminator(), BBDupThreshold);2362 if (JumpThreadCost > BBDupThreshold) {2363 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()2364 << "' - Cost is too high: " << JumpThreadCost << "\n");2365 return false;2366 }2367 2368 threadEdge(BB, PredBBs, SuccBB);2369 return true;2370}2371 2372/// threadEdge - We have decided that it is safe and profitable to factor the2373/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB2374/// across BB. Transform the IR to reflect this change.2375void JumpThreadingPass::threadEdge(BasicBlock *BB,2376 const SmallVectorImpl<BasicBlock *> &PredBBs,2377 BasicBlock *SuccBB) {2378 assert(SuccBB != BB && "Don't create an infinite loop");2379 2380 assert(!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) &&2381 "Don't thread across loop headers");2382 2383 // Build BPI/BFI before any changes are made to IR.2384 bool HasProfile = doesBlockHaveProfileData(BB);2385 auto *BFI = getOrCreateBFI(HasProfile);2386 auto *BPI = getOrCreateBPI(BFI != nullptr);2387 2388 // And finally, do it! Start by factoring the predecessors if needed.2389 BasicBlock *PredBB;2390 if (PredBBs.size() == 1)2391 PredBB = PredBBs[0];2392 else {2393 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()2394 << " common predecessors.\n");2395 PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm");2396 }2397 2398 // And finally, do it!2399 LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName()2400 << "' to '" << SuccBB->getName()2401 << ", across block:\n " << *BB << "\n");2402 2403 LVI->threadEdge(PredBB, BB, SuccBB);2404 2405 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),2406 BB->getName()+".thread",2407 BB->getParent(), BB);2408 NewBB->moveAfter(PredBB);2409 2410 // Set the block frequency of NewBB.2411 if (BFI) {2412 assert(BPI && "It's expected BPI to exist along with BFI");2413 auto NewBBFreq =2414 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);2415 BFI->setBlockFreq(NewBB, NewBBFreq);2416 }2417 2418 // Copy all the instructions from BB to NewBB except the terminator.2419 ValueToValueMapTy ValueMapping;2420 cloneInstructions(ValueMapping, BB->begin(), std::prev(BB->end()), NewBB,2421 PredBB);2422 2423 // We didn't copy the terminator from BB over to NewBB, because there is now2424 // an unconditional jump to SuccBB. Insert the unconditional jump.2425 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);2426 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());2427 2428 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the2429 // PHI nodes for NewBB now.2430 addPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);2431 2432 // Update the terminator of PredBB to jump to NewBB instead of BB. This2433 // eliminates predecessors from BB, which requires us to simplify any PHI2434 // nodes in BB.2435 Instruction *PredTerm = PredBB->getTerminator();2436 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)2437 if (PredTerm->getSuccessor(i) == BB) {2438 BB->removePredecessor(PredBB, true);2439 PredTerm->setSuccessor(i, NewBB);2440 }2441 2442 // Enqueue required DT updates.2443 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, SuccBB},2444 {DominatorTree::Insert, PredBB, NewBB},2445 {DominatorTree::Delete, PredBB, BB}});2446 2447 remapSourceAtoms(ValueMapping, NewBB->begin(), NewBB->end());2448 updateSSA(BB, NewBB, ValueMapping);2449 2450 // At this point, the IR is fully up to date and consistent. Do a quick scan2451 // over the new instructions and zap any that are constants or dead. This2452 // frequently happens because of phi translation.2453 SimplifyInstructionsInBlock(NewBB, TLI);2454 2455 // Update the edge weight from BB to SuccBB, which should be less than before.2456 updateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB, BFI, BPI, HasProfile);2457 2458 // Threaded an edge!2459 ++NumThreads;2460}2461 2462/// Create a new basic block that will be the predecessor of BB and successor of2463/// all blocks in Preds. When profile data is available, update the frequency of2464/// this new block.2465BasicBlock *JumpThreadingPass::splitBlockPreds(BasicBlock *BB,2466 ArrayRef<BasicBlock *> Preds,2467 const char *Suffix) {2468 SmallVector<BasicBlock *, 2> NewBBs;2469 2470 // Collect the frequencies of all predecessors of BB, which will be used to2471 // update the edge weight of the result of splitting predecessors.2472 DenseMap<BasicBlock *, BlockFrequency> FreqMap;2473 auto *BFI = getBFI();2474 if (BFI) {2475 auto *BPI = getOrCreateBPI(true);2476 for (auto *Pred : Preds)2477 FreqMap.insert(std::make_pair(2478 Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));2479 }2480 2481 // In the case when BB is a LandingPad block we create 2 new predecessors2482 // instead of just one.2483 if (BB->isLandingPad()) {2484 std::string NewName = std::string(Suffix) + ".split-lp";2485 SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);2486 } else {2487 NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));2488 }2489 2490 std::vector<DominatorTree::UpdateType> Updates;2491 Updates.reserve((2 * Preds.size()) + NewBBs.size());2492 for (auto *NewBB : NewBBs) {2493 BlockFrequency NewBBFreq(0);2494 Updates.push_back({DominatorTree::Insert, NewBB, BB});2495 for (auto *Pred : predecessors(NewBB)) {2496 Updates.push_back({DominatorTree::Delete, Pred, BB});2497 Updates.push_back({DominatorTree::Insert, Pred, NewBB});2498 if (BFI) // Update frequencies between Pred -> NewBB.2499 NewBBFreq += FreqMap.lookup(Pred);2500 }2501 if (BFI) // Apply the summed frequency to NewBB.2502 BFI->setBlockFreq(NewBB, NewBBFreq);2503 }2504 2505 DTU->applyUpdatesPermissive(Updates);2506 return NewBBs[0];2507}2508 2509bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {2510 const Instruction *TI = BB->getTerminator();2511 if (!TI || TI->getNumSuccessors() < 2)2512 return false;2513 2514 return hasValidBranchWeightMD(*TI);2515}2516 2517/// Update the block frequency of BB and branch weight and the metadata on the2518/// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -2519/// Freq(PredBB->BB) / Freq(BB->SuccBB).2520void JumpThreadingPass::updateBlockFreqAndEdgeWeight(BasicBlock *PredBB,2521 BasicBlock *BB,2522 BasicBlock *NewBB,2523 BasicBlock *SuccBB,2524 BlockFrequencyInfo *BFI,2525 BranchProbabilityInfo *BPI,2526 bool HasProfile) {2527 assert(((BFI && BPI) || (!BFI && !BFI)) &&2528 "Both BFI & BPI should either be set or unset");2529 2530 if (!BFI) {2531 assert(!HasProfile &&2532 "It's expected to have BFI/BPI when profile info exists");2533 return;2534 }2535 2536 // As the edge from PredBB to BB is deleted, we have to update the block2537 // frequency of BB.2538 auto BBOrigFreq = BFI->getBlockFreq(BB);2539 auto NewBBFreq = BFI->getBlockFreq(NewBB);2540 auto BBNewFreq = BBOrigFreq - NewBBFreq;2541 BFI->setBlockFreq(BB, BBNewFreq);2542 2543 // Collect updated outgoing edges' frequencies from BB and use them to update2544 // edge probabilities.2545 SmallVector<uint64_t, 4> BBSuccFreq;2546 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {2547 auto BB2SuccBBFreq =2548 BBOrigFreq * BPI->getEdgeProbability(BB, I.getSuccessorIndex());2549 auto SuccFreq = (*I == SuccBB) ? BB2SuccBBFreq - NewBBFreq : BB2SuccBBFreq;2550 BBSuccFreq.push_back(SuccFreq.getFrequency());2551 }2552 2553 uint64_t MaxBBSuccFreq = *llvm::max_element(BBSuccFreq);2554 2555 SmallVector<BranchProbability, 4> BBSuccProbs;2556 if (MaxBBSuccFreq == 0)2557 BBSuccProbs.assign(BBSuccFreq.size(),2558 {1, static_cast<unsigned>(BBSuccFreq.size())});2559 else {2560 for (uint64_t Freq : BBSuccFreq)2561 BBSuccProbs.push_back(2562 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));2563 // Normalize edge probabilities so that they sum up to one.2564 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),2565 BBSuccProbs.end());2566 }2567 2568 // Update edge probabilities in BPI.2569 BPI->setEdgeProbability(BB, BBSuccProbs);2570 2571 // Update the profile metadata as well.2572 //2573 // Don't do this if the profile of the transformed blocks was statically2574 // estimated. (This could occur despite the function having an entry2575 // frequency in completely cold parts of the CFG.)2576 //2577 // In this case we don't want to suggest to subsequent passes that the2578 // calculated weights are fully consistent. Consider this graph:2579 //2580 // check_12581 // 50% / |2582 // eq_1 | 50%2583 // \ |2584 // check_22585 // 50% / |2586 // eq_2 | 50%2587 // \ |2588 // check_32589 // 50% / |2590 // eq_3 | 50%2591 // \ |2592 //2593 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,2594 // the overall probabilities are inconsistent; the total probability that the2595 // value is either 1, 2 or 3 is 150%.2596 //2597 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_32598 // becomes 0%. This is even worse if the edge whose probability becomes 0% is2599 // the loop exit edge. Then based solely on static estimation we would assume2600 // the loop was extremely hot.2601 //2602 // FIXME this locally as well so that BPI and BFI are consistent as well. We2603 // shouldn't make edges extremely likely or unlikely based solely on static2604 // estimation.2605 if (BBSuccProbs.size() >= 2 && HasProfile) {2606 SmallVector<uint32_t, 4> Weights;2607 for (auto Prob : BBSuccProbs)2608 Weights.push_back(Prob.getNumerator());2609 2610 auto TI = BB->getTerminator();2611 setBranchWeights(*TI, Weights, hasBranchWeightOrigin(*TI));2612 }2613}2614 2615/// duplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch2616/// to BB which contains an i1 PHI node and a conditional branch on that PHI.2617/// If we can duplicate the contents of BB up into PredBB do so now, this2618/// improves the odds that the branch will be on an analyzable instruction like2619/// a compare.2620bool JumpThreadingPass::duplicateCondBranchOnPHIIntoPred(2621 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {2622 assert(!PredBBs.empty() && "Can't handle an empty set");2623 2624 // If BB is a loop header, then duplicating this block outside the loop would2625 // cause us to transform this into an irreducible loop, don't do this.2626 // See the comments above findLoopHeaders for justifications and caveats.2627 if (LoopHeaders.count(BB)) {2628 LLVM_DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()2629 << "' into predecessor block '" << PredBBs[0]->getName()2630 << "' - it might create an irreducible loop!\n");2631 return false;2632 }2633 2634 unsigned DuplicationCost = getJumpThreadDuplicationCost(2635 TTI, BB, BB->getTerminator(), BBDupThreshold);2636 if (DuplicationCost > BBDupThreshold) {2637 LLVM_DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()2638 << "' - Cost is too high: " << DuplicationCost << "\n");2639 return false;2640 }2641 2642 // And finally, do it! Start by factoring the predecessors if needed.2643 std::vector<DominatorTree::UpdateType> Updates;2644 BasicBlock *PredBB;2645 if (PredBBs.size() == 1)2646 PredBB = PredBBs[0];2647 else {2648 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()2649 << " common predecessors.\n");2650 PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm");2651 }2652 Updates.push_back({DominatorTree::Delete, PredBB, BB});2653 2654 // Okay, we decided to do this! Clone all the instructions in BB onto the end2655 // of PredBB.2656 LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName()2657 << "' into end of '" << PredBB->getName()2658 << "' to eliminate branch on phi. Cost: "2659 << DuplicationCost << " block is:" << *BB << "\n");2660 2661 // Unless PredBB ends with an unconditional branch, split the edge so that we2662 // can just clone the bits from BB into the end of the new PredBB.2663 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());2664 2665 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {2666 BasicBlock *OldPredBB = PredBB;2667 PredBB = SplitEdge(OldPredBB, BB);2668 Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});2669 Updates.push_back({DominatorTree::Insert, PredBB, BB});2670 Updates.push_back({DominatorTree::Delete, OldPredBB, BB});2671 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());2672 }2673 2674 // We are going to have to map operands from the original BB block into the2675 // PredBB block. Evaluate PHI nodes in BB.2676 ValueToValueMapTy ValueMapping;2677 2678 // Remember the position before the inserted instructions.2679 auto RItBeforeInsertPt = std::next(OldPredBranch->getReverseIterator());2680 2681 BasicBlock::iterator BI = BB->begin();2682 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)2683 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);2684 // Clone the non-phi instructions of BB into PredBB, keeping track of the2685 // mapping and using it to remap operands in the cloned instructions.2686 for (; BI != BB->end(); ++BI) {2687 Instruction *New = BI->clone();2688 New->insertInto(PredBB, OldPredBranch->getIterator());2689 2690 // Remap operands to patch up intra-block references.2691 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)2692 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {2693 ValueToValueMapTy::iterator I = ValueMapping.find(Inst);2694 if (I != ValueMapping.end())2695 New->setOperand(i, I->second);2696 }2697 2698 // Remap debug variable operands.2699 remapDebugVariable(ValueMapping, New);2700 if (const DebugLoc &DL = New->getDebugLoc())2701 mapAtomInstance(DL, ValueMapping);2702 2703 // If this instruction can be simplified after the operands are updated,2704 // just use the simplified value instead. This frequently happens due to2705 // phi translation.2706 if (Value *IV = simplifyInstruction(2707 New,2708 {BB->getDataLayout(), TLI, nullptr, nullptr, New})) {2709 ValueMapping[&*BI] = IV;2710 if (!New->mayHaveSideEffects()) {2711 New->eraseFromParent();2712 New = nullptr;2713 // Clone debug-info on the elided instruction to the destination2714 // position.2715 OldPredBranch->cloneDebugInfoFrom(&*BI, std::nullopt, true);2716 }2717 } else {2718 ValueMapping[&*BI] = New;2719 }2720 if (New) {2721 // Otherwise, insert the new instruction into the block.2722 New->setName(BI->getName());2723 // Clone across any debug-info attached to the old instruction.2724 New->cloneDebugInfoFrom(&*BI);2725 // Update Dominance from simplified New instruction operands.2726 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)2727 if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))2728 Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});2729 }2730 }2731 2732 // Check to see if the targets of the branch had PHI nodes. If so, we need to2733 // add entries to the PHI nodes for branch from PredBB now.2734 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());2735 addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,2736 ValueMapping);2737 addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,2738 ValueMapping);2739 2740 // KeyInstructions: Remap the cloned instructions' atoms only.2741 remapSourceAtoms(ValueMapping, std::prev(RItBeforeInsertPt)->getIterator(),2742 OldPredBranch->getIterator());2743 2744 updateSSA(BB, PredBB, ValueMapping);2745 2746 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge2747 // that we nuked.2748 BB->removePredecessor(PredBB, true);2749 2750 // Remove the unconditional branch at the end of the PredBB block.2751 OldPredBranch->eraseFromParent();2752 if (auto *BPI = getBPI())2753 BPI->copyEdgeProbabilities(BB, PredBB);2754 DTU->applyUpdatesPermissive(Updates);2755 2756 ++NumDupes;2757 return true;2758}2759 2760// Pred is a predecessor of BB with an unconditional branch to BB. SI is2761// a Select instruction in Pred. BB has other predecessors and SI is used in2762// a PHI node in BB. SI has no other use.2763// A new basic block, NewBB, is created and SI is converted to compare and2764// conditional branch. SI is erased from parent.2765void JumpThreadingPass::unfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB,2766 SelectInst *SI, PHINode *SIUse,2767 unsigned Idx) {2768 // Expand the select.2769 //2770 // Pred --2771 // | v2772 // | NewBB2773 // | |2774 // |-----2775 // v2776 // BB2777 BranchInst *PredTerm = cast<BranchInst>(Pred->getTerminator());2778 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",2779 BB->getParent(), BB);2780 // Move the unconditional branch to NewBB.2781 PredTerm->removeFromParent();2782 PredTerm->insertInto(NewBB, NewBB->end());2783 // Create a conditional branch and update PHI nodes.2784 auto *BI = BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);2785 BI->applyMergedLocation(PredTerm->getDebugLoc(), SI->getDebugLoc());2786 BI->copyMetadata(*SI, {LLVMContext::MD_prof});2787 SIUse->setIncomingValue(Idx, SI->getFalseValue());2788 SIUse->addIncoming(SI->getTrueValue(), NewBB);2789 2790 uint64_t TrueWeight = 1;2791 uint64_t FalseWeight = 1;2792 // Copy probabilities from 'SI' to created conditional branch in 'Pred'.2793 if (extractBranchWeights(*SI, TrueWeight, FalseWeight) &&2794 (TrueWeight + FalseWeight) != 0) {2795 SmallVector<BranchProbability, 2> BP;2796 BP.emplace_back(BranchProbability::getBranchProbability(2797 TrueWeight, TrueWeight + FalseWeight));2798 BP.emplace_back(BranchProbability::getBranchProbability(2799 FalseWeight, TrueWeight + FalseWeight));2800 // Update BPI if exists.2801 if (auto *BPI = getBPI())2802 BPI->setEdgeProbability(Pred, BP);2803 }2804 // Set the block frequency of NewBB.2805 if (auto *BFI = getBFI()) {2806 if ((TrueWeight + FalseWeight) == 0) {2807 TrueWeight = 1;2808 FalseWeight = 1;2809 }2810 BranchProbability PredToNewBBProb = BranchProbability::getBranchProbability(2811 TrueWeight, TrueWeight + FalseWeight);2812 auto NewBBFreq = BFI->getBlockFreq(Pred) * PredToNewBBProb;2813 BFI->setBlockFreq(NewBB, NewBBFreq);2814 }2815 2816 // The select is now dead.2817 SI->eraseFromParent();2818 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, BB},2819 {DominatorTree::Insert, Pred, NewBB}});2820 2821 // Update any other PHI nodes in BB.2822 for (BasicBlock::iterator BI = BB->begin();2823 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)2824 if (Phi != SIUse)2825 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);2826}2827 2828bool JumpThreadingPass::tryToUnfoldSelect(SwitchInst *SI, BasicBlock *BB) {2829 PHINode *CondPHI = dyn_cast<PHINode>(SI->getCondition());2830 2831 if (!CondPHI || CondPHI->getParent() != BB)2832 return false;2833 2834 for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) {2835 BasicBlock *Pred = CondPHI->getIncomingBlock(I);2836 SelectInst *PredSI = dyn_cast<SelectInst>(CondPHI->getIncomingValue(I));2837 2838 // The second and third condition can be potentially relaxed. Currently2839 // the conditions help to simplify the code and allow us to reuse existing2840 // code, developed for tryToUnfoldSelect(CmpInst *, BasicBlock *)2841 if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse())2842 continue;2843 2844 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());2845 if (!PredTerm || !PredTerm->isUnconditional())2846 continue;2847 2848 unfoldSelectInstr(Pred, BB, PredSI, CondPHI, I);2849 return true;2850 }2851 return false;2852}2853 2854/// tryToUnfoldSelect - Look for blocks of the form2855/// bb1:2856/// %a = select2857/// br bb22858///2859/// bb2:2860/// %p = phi [%a, %bb1] ...2861/// %c = icmp %p2862/// br i1 %c2863///2864/// And expand the select into a branch structure if one of its arms allows %c2865/// to be folded. This later enables threading from bb1 over bb2.2866bool JumpThreadingPass::tryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {2867 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());2868 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));2869 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));2870 2871 if (!CondBr || !CondBr->isConditional() || !CondLHS ||2872 CondLHS->getParent() != BB)2873 return false;2874 2875 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {2876 BasicBlock *Pred = CondLHS->getIncomingBlock(I);2877 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));2878 2879 // Look if one of the incoming values is a select in the corresponding2880 // predecessor.2881 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())2882 continue;2883 2884 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());2885 if (!PredTerm || !PredTerm->isUnconditional())2886 continue;2887 2888 // Now check if one of the select values would allow us to constant fold the2889 // terminator in BB. We don't do the transform if both sides fold, those2890 // cases will be threaded in any case.2891 Constant *LHSRes =2892 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),2893 CondRHS, Pred, BB, CondCmp);2894 Constant *RHSRes =2895 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),2896 CondRHS, Pred, BB, CondCmp);2897 if ((LHSRes || RHSRes) && LHSRes != RHSRes) {2898 unfoldSelectInstr(Pred, BB, SI, CondLHS, I);2899 return true;2900 }2901 }2902 return false;2903}2904 2905/// tryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the2906/// same BB in the form2907/// bb:2908/// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...2909/// %s = select %p, trueval, falseval2910///2911/// or2912///2913/// bb:2914/// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...2915/// %c = cmp %p, 02916/// %s = select %c, trueval, falseval2917///2918/// And expand the select into a branch structure. This later enables2919/// jump-threading over bb in this pass.2920///2921/// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold2922/// select if the associated PHI has at least one constant. If the unfolded2923/// select is not jump-threaded, it will be folded again in the later2924/// optimizations.2925bool JumpThreadingPass::tryToUnfoldSelectInCurrBB(BasicBlock *BB) {2926 // This transform would reduce the quality of msan diagnostics.2927 // Disable this transform under MemorySanitizer.2928 if (BB->getParent()->hasFnAttribute(Attribute::SanitizeMemory))2929 return false;2930 2931 // If threading this would thread across a loop header, don't thread the edge.2932 // See the comments above findLoopHeaders for justifications and caveats.2933 if (LoopHeaders.count(BB))2934 return false;2935 2936 for (BasicBlock::iterator BI = BB->begin();2937 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {2938 // Look for a Phi having at least one constant incoming value.2939 if (llvm::all_of(PN->incoming_values(),2940 [](Value *V) { return !isa<ConstantInt>(V); }))2941 continue;2942 2943 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {2944 using namespace PatternMatch;2945 2946 // Check if SI is in BB and use V as condition.2947 if (SI->getParent() != BB)2948 return false;2949 Value *Cond = SI->getCondition();2950 bool IsAndOr = match(SI, m_CombineOr(m_LogicalAnd(), m_LogicalOr()));2951 return Cond && Cond == V && Cond->getType()->isIntegerTy(1) && !IsAndOr;2952 };2953 2954 SelectInst *SI = nullptr;2955 for (Use &U : PN->uses()) {2956 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {2957 // Look for a ICmp in BB that compares PN with a constant and is the2958 // condition of a Select.2959 if (Cmp->getParent() == BB && Cmp->hasOneUse() &&2960 isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))2961 if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))2962 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {2963 SI = SelectI;2964 break;2965 }2966 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {2967 // Look for a Select in BB that uses PN as condition.2968 if (isUnfoldCandidate(SelectI, U.get())) {2969 SI = SelectI;2970 break;2971 }2972 }2973 }2974 2975 if (!SI)2976 continue;2977 // Expand the select.2978 Value *Cond = SI->getCondition();2979 if (!isGuaranteedNotToBeUndefOrPoison(Cond, nullptr, SI)) {2980 Cond = new FreezeInst(Cond, "cond.fr", SI->getIterator());2981 cast<FreezeInst>(Cond)->setDebugLoc(DebugLoc::getTemporary());2982 }2983 MDNode *BranchWeights = getBranchWeightMDNode(*SI);2984 Instruction *Term =2985 SplitBlockAndInsertIfThen(Cond, SI, false, BranchWeights);2986 BasicBlock *SplitBB = SI->getParent();2987 BasicBlock *NewBB = Term->getParent();2988 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI->getIterator());2989 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());2990 NewPN->addIncoming(SI->getFalseValue(), BB);2991 NewPN->setDebugLoc(SI->getDebugLoc());2992 SI->replaceAllUsesWith(NewPN);2993 SI->eraseFromParent();2994 // NewBB and SplitBB are newly created blocks which require insertion.2995 std::vector<DominatorTree::UpdateType> Updates;2996 Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);2997 Updates.push_back({DominatorTree::Insert, BB, SplitBB});2998 Updates.push_back({DominatorTree::Insert, BB, NewBB});2999 Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});3000 // BB's successors were moved to SplitBB, update DTU accordingly.3001 for (auto *Succ : successors(SplitBB)) {3002 Updates.push_back({DominatorTree::Delete, BB, Succ});3003 Updates.push_back({DominatorTree::Insert, SplitBB, Succ});3004 }3005 DTU->applyUpdatesPermissive(Updates);3006 return true;3007 }3008 return false;3009}3010 3011/// Try to propagate a guard from the current BB into one of its predecessors3012/// in case if another branch of execution implies that the condition of this3013/// guard is always true. Currently we only process the simplest case that3014/// looks like:3015///3016/// Start:3017/// %cond = ...3018/// br i1 %cond, label %T1, label %F13019/// T1:3020/// br label %Merge3021/// F1:3022/// br label %Merge3023/// Merge:3024/// %condGuard = ...3025/// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]3026///3027/// And cond either implies condGuard or !condGuard. In this case all the3028/// instructions before the guard can be duplicated in both branches, and the3029/// guard is then threaded to one of them.3030bool JumpThreadingPass::processGuards(BasicBlock *BB) {3031 using namespace PatternMatch;3032 3033 // We only want to deal with two predecessors.3034 BasicBlock *Pred1, *Pred2;3035 auto PI = pred_begin(BB), PE = pred_end(BB);3036 if (PI == PE)3037 return false;3038 Pred1 = *PI++;3039 if (PI == PE)3040 return false;3041 Pred2 = *PI++;3042 if (PI != PE)3043 return false;3044 if (Pred1 == Pred2)3045 return false;3046 3047 // Try to thread one of the guards of the block.3048 // TODO: Look up deeper than to immediate predecessor?3049 auto *Parent = Pred1->getSinglePredecessor();3050 if (!Parent || Parent != Pred2->getSinglePredecessor())3051 return false;3052 3053 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))3054 for (auto &I : *BB)3055 if (isGuard(&I) && threadGuard(BB, cast<IntrinsicInst>(&I), BI))3056 return true;3057 3058 return false;3059}3060 3061/// Try to propagate the guard from BB which is the lower block of a diamond3062/// to one of its branches, in case if diamond's condition implies guard's3063/// condition.3064bool JumpThreadingPass::threadGuard(BasicBlock *BB, IntrinsicInst *Guard,3065 BranchInst *BI) {3066 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?");3067 assert(BI->isConditional() && "Unconditional branch has 2 successors?");3068 Value *GuardCond = Guard->getArgOperand(0);3069 Value *BranchCond = BI->getCondition();3070 BasicBlock *TrueDest = BI->getSuccessor(0);3071 BasicBlock *FalseDest = BI->getSuccessor(1);3072 3073 auto &DL = BB->getDataLayout();3074 bool TrueDestIsSafe = false;3075 bool FalseDestIsSafe = false;3076 3077 // True dest is safe if BranchCond => GuardCond.3078 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);3079 if (Impl && *Impl)3080 TrueDestIsSafe = true;3081 else {3082 // False dest is safe if !BranchCond => GuardCond.3083 Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);3084 if (Impl && *Impl)3085 FalseDestIsSafe = true;3086 }3087 3088 if (!TrueDestIsSafe && !FalseDestIsSafe)3089 return false;3090 3091 BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;3092 BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;3093 3094 ValueToValueMapTy UnguardedMapping, GuardedMapping;3095 Instruction *AfterGuard = Guard->getNextNode();3096 unsigned Cost =3097 getJumpThreadDuplicationCost(TTI, BB, AfterGuard, BBDupThreshold);3098 if (Cost > BBDupThreshold)3099 return false;3100 // Duplicate all instructions before the guard and the guard itself to the3101 // branch where implication is not proved.3102 BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween(3103 BB, PredGuardedBlock, AfterGuard, GuardedMapping, *DTU);3104 assert(GuardedBlock && "Could not create the guarded block?");3105 // Duplicate all instructions before the guard in the unguarded branch.3106 // Since we have successfully duplicated the guarded block and this block3107 // has fewer instructions, we expect it to succeed.3108 BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween(3109 BB, PredUnguardedBlock, Guard, UnguardedMapping, *DTU);3110 assert(UnguardedBlock && "Could not create the unguarded block?");3111 LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block "3112 << GuardedBlock->getName() << "\n");3113 // Some instructions before the guard may still have uses. For them, we need3114 // to create Phi nodes merging their copies in both guarded and unguarded3115 // branches. Those instructions that have no uses can be just removed.3116 SmallVector<Instruction *, 4> ToRemove;3117 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)3118 if (!isa<PHINode>(&*BI))3119 ToRemove.push_back(&*BI);3120 3121 BasicBlock::iterator InsertionPoint = BB->getFirstInsertionPt();3122 assert(InsertionPoint != BB->end() && "Empty block?");3123 // Substitute with Phis & remove.3124 for (auto *Inst : reverse(ToRemove)) {3125 if (!Inst->use_empty()) {3126 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);3127 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);3128 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);3129 NewPN->setDebugLoc(Inst->getDebugLoc());3130 NewPN->insertBefore(InsertionPoint);3131 Inst->replaceAllUsesWith(NewPN);3132 }3133 Inst->dropDbgRecords();3134 Inst->eraseFromParent();3135 }3136 return true;3137}3138 3139PreservedAnalyses JumpThreadingPass::getPreservedAnalysis() const {3140 PreservedAnalyses PA;3141 PA.preserve<LazyValueAnalysis>();3142 PA.preserve<DominatorTreeAnalysis>();3143 3144 // TODO: We would like to preserve BPI/BFI. Enable once all paths update them.3145 // TODO: Would be nice to verify BPI/BFI consistency as well.3146 return PA;3147}3148 3149template <typename AnalysisT>3150typename AnalysisT::Result *JumpThreadingPass::runExternalAnalysis() {3151 assert(FAM && "Can't run external analysis without FunctionAnalysisManager");3152 3153 // If there were no changes since last call to 'runExternalAnalysis' then all3154 // analysis is either up to date or explicitly invalidated. Just go ahead and3155 // run the "external" analysis.3156 if (!ChangedSinceLastAnalysisUpdate) {3157 assert(!DTU->hasPendingUpdates() &&3158 "Lost update of 'ChangedSinceLastAnalysisUpdate'?");3159 // Run the "external" analysis.3160 return &FAM->getResult<AnalysisT>(*F);3161 }3162 ChangedSinceLastAnalysisUpdate = false;3163 3164 auto PA = getPreservedAnalysis();3165 // TODO: This shouldn't be needed once 'getPreservedAnalysis' reports BPI/BFI3166 // as preserved.3167 PA.preserve<BranchProbabilityAnalysis>();3168 PA.preserve<BlockFrequencyAnalysis>();3169 // Report everything except explicitly preserved as invalid.3170 FAM->invalidate(*F, PA);3171 // Update DT/PDT.3172 DTU->flush();3173 // Make sure DT/PDT are valid before running "external" analysis.3174 assert(DTU->getDomTree().verify(DominatorTree::VerificationLevel::Fast));3175 assert((!DTU->hasPostDomTree() ||3176 DTU->getPostDomTree().verify(3177 PostDominatorTree::VerificationLevel::Fast)));3178 // Run the "external" analysis.3179 auto *Result = &FAM->getResult<AnalysisT>(*F);3180 // Update analysis JumpThreading depends on and not explicitly preserved.3181 TTI = &FAM->getResult<TargetIRAnalysis>(*F);3182 TLI = &FAM->getResult<TargetLibraryAnalysis>(*F);3183 AA = &FAM->getResult<AAManager>(*F);3184 3185 return Result;3186}3187 3188BranchProbabilityInfo *JumpThreadingPass::getBPI() {3189 if (!BPI) {3190 assert(FAM && "Can't create BPI without FunctionAnalysisManager");3191 BPI = FAM->getCachedResult<BranchProbabilityAnalysis>(*F);3192 }3193 return BPI;3194}3195 3196BlockFrequencyInfo *JumpThreadingPass::getBFI() {3197 if (!BFI) {3198 assert(FAM && "Can't create BFI without FunctionAnalysisManager");3199 BFI = FAM->getCachedResult<BlockFrequencyAnalysis>(*F);3200 }3201 return BFI;3202}3203 3204// Important note on validity of BPI/BFI. JumpThreading tries to preserve3205// BPI/BFI as it goes. Thus if cached instance exists it will be updated.3206// Otherwise, new instance of BPI/BFI is created (up to date by definition).3207BranchProbabilityInfo *JumpThreadingPass::getOrCreateBPI(bool Force) {3208 auto *Res = getBPI();3209 if (Res)3210 return Res;3211 3212 if (Force)3213 BPI = runExternalAnalysis<BranchProbabilityAnalysis>();3214 3215 return BPI;3216}3217 3218BlockFrequencyInfo *JumpThreadingPass::getOrCreateBFI(bool Force) {3219 auto *Res = getBFI();3220 if (Res)3221 return Res;3222 3223 if (Force)3224 BFI = runExternalAnalysis<BlockFrequencyAnalysis>();3225 3226 return BFI;3227}3228