3774 lines · cpp
1///===- SimpleLoopUnswitch.cpp - Hoist loop-invariant control flow ---------===//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#include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h"10#include "llvm/ADT/DenseMap.h"11#include "llvm/ADT/STLExtras.h"12#include "llvm/ADT/Sequence.h"13#include "llvm/ADT/SetVector.h"14#include "llvm/ADT/SmallPtrSet.h"15#include "llvm/ADT/SmallVector.h"16#include "llvm/ADT/Statistic.h"17#include "llvm/ADT/Twine.h"18#include "llvm/Analysis/AssumptionCache.h"19#include "llvm/Analysis/BlockFrequencyInfo.h"20#include "llvm/Analysis/CFG.h"21#include "llvm/Analysis/CodeMetrics.h"22#include "llvm/Analysis/DomTreeUpdater.h"23#include "llvm/Analysis/GuardUtils.h"24#include "llvm/Analysis/LoopAnalysisManager.h"25#include "llvm/Analysis/LoopInfo.h"26#include "llvm/Analysis/LoopIterator.h"27#include "llvm/Analysis/MemorySSA.h"28#include "llvm/Analysis/MemorySSAUpdater.h"29#include "llvm/Analysis/MustExecute.h"30#include "llvm/Analysis/ScalarEvolution.h"31#include "llvm/Analysis/TargetTransformInfo.h"32#include "llvm/Analysis/ValueTracking.h"33#include "llvm/IR/BasicBlock.h"34#include "llvm/IR/Constant.h"35#include "llvm/IR/Constants.h"36#include "llvm/IR/Dominators.h"37#include "llvm/IR/Function.h"38#include "llvm/IR/IRBuilder.h"39#include "llvm/IR/InstrTypes.h"40#include "llvm/IR/Instruction.h"41#include "llvm/IR/Instructions.h"42#include "llvm/IR/IntrinsicInst.h"43#include "llvm/IR/MDBuilder.h"44#include "llvm/IR/Module.h"45#include "llvm/IR/PatternMatch.h"46#include "llvm/IR/ProfDataUtils.h"47#include "llvm/IR/Use.h"48#include "llvm/IR/Value.h"49#include "llvm/Support/Casting.h"50#include "llvm/Support/CommandLine.h"51#include "llvm/Support/Debug.h"52#include "llvm/Support/ErrorHandling.h"53#include "llvm/Support/GenericDomTree.h"54#include "llvm/Support/InstructionCost.h"55#include "llvm/Support/raw_ostream.h"56#include "llvm/Transforms/Scalar/LoopPassManager.h"57#include "llvm/Transforms/Utils/BasicBlockUtils.h"58#include "llvm/Transforms/Utils/Cloning.h"59#include "llvm/Transforms/Utils/Local.h"60#include "llvm/Transforms/Utils/LoopUtils.h"61#include "llvm/Transforms/Utils/ValueMapper.h"62#include <algorithm>63#include <cassert>64#include <iterator>65#include <numeric>66#include <optional>67#include <utility>68 69#define DEBUG_TYPE "simple-loop-unswitch"70 71using namespace llvm;72using namespace llvm::PatternMatch;73 74STATISTIC(NumBranches, "Number of branches unswitched");75STATISTIC(NumSwitches, "Number of switches unswitched");76STATISTIC(NumSelects, "Number of selects turned into branches for unswitching");77STATISTIC(NumGuards, "Number of guards turned into branches for unswitching");78STATISTIC(NumTrivial, "Number of unswitches that are trivial");79STATISTIC(80 NumCostMultiplierSkipped,81 "Number of unswitch candidates that had their cost multiplier skipped");82STATISTIC(NumInvariantConditionsInjected,83 "Number of invariant conditions injected and unswitched");84 85namespace llvm {86static cl::opt<bool> EnableNonTrivialUnswitch(87 "enable-nontrivial-unswitch", cl::init(false), cl::Hidden,88 cl::desc("Forcibly enables non-trivial loop unswitching rather than "89 "following the configuration passed into the pass."));90 91static cl::opt<int>92 UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden,93 cl::desc("The cost threshold for unswitching a loop."));94 95static cl::opt<bool> EnableUnswitchCostMultiplier(96 "enable-unswitch-cost-multiplier", cl::init(true), cl::Hidden,97 cl::desc("Enable unswitch cost multiplier that prohibits exponential "98 "explosion in nontrivial unswitch."));99static cl::opt<int> UnswitchSiblingsToplevelDiv(100 "unswitch-siblings-toplevel-div", cl::init(2), cl::Hidden,101 cl::desc("Toplevel siblings divisor for cost multiplier."));102static cl::opt<int> UnswitchParentBlocksDiv(103 "unswitch-parent-blocks-div", cl::init(8), cl::Hidden,104 cl::desc("Outer loop size divisor for cost multiplier."));105static cl::opt<int> UnswitchNumInitialUnscaledCandidates(106 "unswitch-num-initial-unscaled-candidates", cl::init(8), cl::Hidden,107 cl::desc("Number of unswitch candidates that are ignored when calculating "108 "cost multiplier."));109static cl::opt<bool> UnswitchGuards(110 "simple-loop-unswitch-guards", cl::init(true), cl::Hidden,111 cl::desc("If enabled, simple loop unswitching will also consider "112 "llvm.experimental.guard intrinsics as unswitch candidates."));113static cl::opt<bool> DropNonTrivialImplicitNullChecks(114 "simple-loop-unswitch-drop-non-trivial-implicit-null-checks",115 cl::init(false), cl::Hidden,116 cl::desc("If enabled, drop make.implicit metadata in unswitched implicit "117 "null checks to save time analyzing if we can keep it."));118static cl::opt<unsigned>119 MSSAThreshold("simple-loop-unswitch-memoryssa-threshold",120 cl::desc("Max number of memory uses to explore during "121 "partial unswitching analysis"),122 cl::init(100), cl::Hidden);123static cl::opt<bool> FreezeLoopUnswitchCond(124 "freeze-loop-unswitch-cond", cl::init(true), cl::Hidden,125 cl::desc("If enabled, the freeze instruction will be added to condition "126 "of loop unswitch to prevent miscompilation."));127 128static cl::opt<bool> InjectInvariantConditions(129 "simple-loop-unswitch-inject-invariant-conditions", cl::Hidden,130 cl::desc("Whether we should inject new invariants and unswitch them to "131 "eliminate some existing (non-invariant) conditions."),132 cl::init(true));133 134static cl::opt<unsigned> InjectInvariantConditionHotnesThreshold(135 "simple-loop-unswitch-inject-invariant-condition-hotness-threshold",136 cl::Hidden,137 cl::desc("Only try to inject loop invariant conditions and "138 "unswitch on them to eliminate branches that are "139 "not-taken 1/<this option> times or less."),140 cl::init(16));141 142static cl::opt<bool> EstimateProfile("simple-loop-unswitch-estimate-profile",143 cl::Hidden, cl::init(true));144extern cl::opt<bool> ProfcheckDisableMetadataFixes;145} // namespace llvm146 147AnalysisKey ShouldRunExtraSimpleLoopUnswitch::Key;148namespace {149struct CompareDesc {150 BranchInst *Term;151 Value *Invariant;152 BasicBlock *InLoopSucc;153 154 CompareDesc(BranchInst *Term, Value *Invariant, BasicBlock *InLoopSucc)155 : Term(Term), Invariant(Invariant), InLoopSucc(InLoopSucc) {}156};157 158struct InjectedInvariant {159 ICmpInst::Predicate Pred;160 Value *LHS;161 Value *RHS;162 BasicBlock *InLoopSucc;163 164 InjectedInvariant(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,165 BasicBlock *InLoopSucc)166 : Pred(Pred), LHS(LHS), RHS(RHS), InLoopSucc(InLoopSucc) {}167};168 169struct NonTrivialUnswitchCandidate {170 Instruction *TI = nullptr;171 TinyPtrVector<Value *> Invariants;172 std::optional<InstructionCost> Cost;173 std::optional<InjectedInvariant> PendingInjection;174 NonTrivialUnswitchCandidate(175 Instruction *TI, ArrayRef<Value *> Invariants,176 std::optional<InstructionCost> Cost = std::nullopt,177 std::optional<InjectedInvariant> PendingInjection = std::nullopt)178 : TI(TI), Invariants(Invariants), Cost(Cost),179 PendingInjection(PendingInjection) {};180 181 bool hasPendingInjection() const { return PendingInjection.has_value(); }182};183} // end anonymous namespace.184 185// Helper to skip (select x, true, false), which matches both a logical AND and186// OR and can confuse code that tries to determine if \p Cond is either a187// logical AND or OR but not both.188static Value *skipTrivialSelect(Value *Cond) {189 Value *CondNext;190 while (match(Cond, m_Select(m_Value(CondNext), m_One(), m_Zero())))191 Cond = CondNext;192 return Cond;193}194 195/// Collect all of the loop invariant input values transitively used by the196/// homogeneous instruction graph from a given root.197///198/// This essentially walks from a root recursively through loop variant operands199/// which have perform the same logical operation (AND or OR) and finds all200/// inputs which are loop invariant. For some operations these can be201/// re-associated and unswitched out of the loop entirely.202static TinyPtrVector<Value *>203collectHomogenousInstGraphLoopInvariants(const Loop &L, Instruction &Root,204 const LoopInfo &LI) {205 assert(!L.isLoopInvariant(&Root) &&206 "Only need to walk the graph if root itself is not invariant.");207 TinyPtrVector<Value *> Invariants;208 209 bool IsRootAnd = match(&Root, m_LogicalAnd());210 bool IsRootOr = match(&Root, m_LogicalOr());211 212 // Build a worklist and recurse through operators collecting invariants.213 SmallVector<Instruction *, 4> Worklist;214 SmallPtrSet<Instruction *, 8> Visited;215 Worklist.push_back(&Root);216 Visited.insert(&Root);217 do {218 Instruction &I = *Worklist.pop_back_val();219 for (Value *OpV : I.operand_values()) {220 // Skip constants as unswitching isn't interesting for them.221 if (isa<Constant>(OpV))222 continue;223 224 // Add it to our result if loop invariant.225 if (L.isLoopInvariant(OpV)) {226 Invariants.push_back(OpV);227 continue;228 }229 230 // If not an instruction with the same opcode, nothing we can do.231 Instruction *OpI = dyn_cast<Instruction>(skipTrivialSelect(OpV));232 233 if (OpI && ((IsRootAnd && match(OpI, m_LogicalAnd())) ||234 (IsRootOr && match(OpI, m_LogicalOr())))) {235 // Visit this operand.236 if (Visited.insert(OpI).second)237 Worklist.push_back(OpI);238 }239 }240 } while (!Worklist.empty());241 242 return Invariants;243}244 245static void replaceLoopInvariantUses(const Loop &L, Value *Invariant,246 Constant &Replacement) {247 assert(!isa<Constant>(Invariant) && "Why are we unswitching on a constant?");248 249 // Replace uses of LIC in the loop with the given constant.250 // We use make_early_inc_range as set invalidates the iterator.251 for (Use &U : llvm::make_early_inc_range(Invariant->uses())) {252 Instruction *UserI = dyn_cast<Instruction>(U.getUser());253 254 // Replace this use within the loop body.255 if (UserI && L.contains(UserI))256 U.set(&Replacement);257 }258}259 260/// Check that all the LCSSA PHI nodes in the loop exit block have trivial261/// incoming values along this edge.262static bool areLoopExitPHIsLoopInvariant(const Loop &L,263 const BasicBlock &ExitingBB,264 const BasicBlock &ExitBB) {265 for (const Instruction &I : ExitBB) {266 auto *PN = dyn_cast<PHINode>(&I);267 if (!PN)268 // No more PHIs to check.269 return true;270 271 // If the incoming value for this edge isn't loop invariant the unswitch272 // won't be trivial.273 if (!L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB)))274 return false;275 }276 llvm_unreachable("Basic blocks should never be empty!");277}278 279/// Copy a set of loop invariant values \p Invariants and insert them at the280/// end of \p BB and conditionally branch on the copied condition. We only281/// branch on a single value.282/// We attempt to estimate the profile of the resulting conditional branch from283/// \p ComputeProfFrom, which is the original conditional branch we're284/// unswitching.285/// When \p Direction is true, the \p Invariants form a disjunction, and the286/// branch conditioned on it exits the loop on the "true" case. When \p287/// Direction is false, the \p Invariants form a conjunction and the branch288/// exits on the "false" case.289static void buildPartialUnswitchConditionalBranch(290 BasicBlock &BB, ArrayRef<Value *> Invariants, bool Direction,291 BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, bool InsertFreeze,292 const Instruction *I, AssumptionCache *AC, const DominatorTree &DT,293 const BranchInst &ComputeProfFrom) {294 295 SmallVector<uint32_t> BranchWeights;296 bool HasBranchWeights = EstimateProfile && !ProfcheckDisableMetadataFixes &&297 extractBranchWeights(ComputeProfFrom, BranchWeights);298 // If Direction is true, that means we had a disjunction and that the "true"299 // case exits. The probability of the disjunction of the subset of terms is at300 // most as high as the original one. So, if the probability is higher than the301 // one we'd assign in absence of a profile (i.e. 0.5), we will use 0.5,302 // but if it's lower, we will use the original probability.303 // Conversely, if Direction is false, that means we had a conjunction, and the304 // probability of exiting is captured in the second branch weight. That305 // probability is a disjunction (of the negation of the original terms). The306 // same reasoning applies as above.307 // Issue #165649: should we expect BFI to conserve, and use that to calculate308 // the branch weights?309 if (HasBranchWeights &&310 static_cast<double>(BranchWeights[Direction ? 0 : 1]) /311 static_cast<double>(sum_of(BranchWeights)) >312 0.5)313 HasBranchWeights = false;314 315 IRBuilder<> IRB(&BB);316 IRB.SetCurrentDebugLocation(DebugLoc::getCompilerGenerated());317 318 SmallVector<Value *> FrozenInvariants;319 for (Value *Inv : Invariants) {320 if (InsertFreeze && !isGuaranteedNotToBeUndefOrPoison(Inv, AC, I, &DT))321 Inv = IRB.CreateFreeze(Inv, Inv->getName() + ".fr");322 FrozenInvariants.push_back(Inv);323 }324 325 Value *Cond = Direction ? IRB.CreateOr(FrozenInvariants)326 : IRB.CreateAnd(FrozenInvariants);327 auto *BR = IRB.CreateCondBr(328 Cond, Direction ? &UnswitchedSucc : &NormalSucc,329 Direction ? &NormalSucc : &UnswitchedSucc,330 HasBranchWeights ? ComputeProfFrom.getMetadata(LLVMContext::MD_prof)331 : nullptr);332 if (!HasBranchWeights)333 setExplicitlyUnknownBranchWeightsIfProfiled(*BR, DEBUG_TYPE);334}335 336/// Copy a set of loop invariant values, and conditionally branch on them.337static void buildPartialInvariantUnswitchConditionalBranch(338 BasicBlock &BB, ArrayRef<Value *> ToDuplicate, bool Direction,339 BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, Loop &L,340 MemorySSAUpdater *MSSAU, const BranchInst &OriginalBranch) {341 ValueToValueMapTy VMap;342 for (auto *Val : reverse(ToDuplicate)) {343 Instruction *Inst = cast<Instruction>(Val);344 Instruction *NewInst = Inst->clone();345 346 if (const DebugLoc &DL = Inst->getDebugLoc())347 mapAtomInstance(DL, VMap);348 349 NewInst->insertInto(&BB, BB.end());350 RemapInstruction(NewInst, VMap,351 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);352 VMap[Val] = NewInst;353 354 if (!MSSAU)355 continue;356 357 MemorySSA *MSSA = MSSAU->getMemorySSA();358 if (auto *MemUse =359 dyn_cast_or_null<MemoryUse>(MSSA->getMemoryAccess(Inst))) {360 auto *DefiningAccess = MemUse->getDefiningAccess();361 // Get the first defining access before the loop.362 while (L.contains(DefiningAccess->getBlock())) {363 // If the defining access is a MemoryPhi, get the incoming364 // value for the pre-header as defining access.365 if (auto *MemPhi = dyn_cast<MemoryPhi>(DefiningAccess))366 DefiningAccess =367 MemPhi->getIncomingValueForBlock(L.getLoopPreheader());368 else369 DefiningAccess = cast<MemoryDef>(DefiningAccess)->getDefiningAccess();370 }371 MSSAU->createMemoryAccessInBB(NewInst, DefiningAccess,372 NewInst->getParent(),373 MemorySSA::BeforeTerminator);374 }375 }376 377 IRBuilder<> IRB(&BB);378 IRB.SetCurrentDebugLocation(DebugLoc::getCompilerGenerated());379 Value *Cond = VMap[ToDuplicate[0]];380 // The expectation is that ToDuplicate[0] is the condition used by the381 // OriginalBranch, case in which we can clone the profile metadata from there.382 auto *ProfData =383 !ProfcheckDisableMetadataFixes &&384 ToDuplicate[0] == skipTrivialSelect(OriginalBranch.getCondition())385 ? OriginalBranch.getMetadata(LLVMContext::MD_prof)386 : nullptr;387 auto *BR =388 IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc,389 Direction ? &NormalSucc : &UnswitchedSucc, ProfData);390 if (!ProfData)391 setExplicitlyUnknownBranchWeightsIfProfiled(*BR, DEBUG_TYPE);392}393 394/// Rewrite the PHI nodes in an unswitched loop exit basic block.395///396/// Requires that the loop exit and unswitched basic block are the same, and397/// that the exiting block was a unique predecessor of that block. Rewrites the398/// PHI nodes in that block such that what were LCSSA PHI nodes become trivial399/// PHI nodes from the old preheader that now contains the unswitched400/// terminator.401static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB,402 BasicBlock &OldExitingBB,403 BasicBlock &OldPH) {404 for (PHINode &PN : UnswitchedBB.phis()) {405 // When the loop exit is directly unswitched we just need to update the406 // incoming basic block. We loop to handle weird cases with repeated407 // incoming blocks, but expect to typically only have one operand here.408 for (auto i : seq<int>(0, PN.getNumOperands())) {409 assert(PN.getIncomingBlock(i) == &OldExitingBB &&410 "Found incoming block different from unique predecessor!");411 PN.setIncomingBlock(i, &OldPH);412 }413 }414}415 416/// Rewrite the PHI nodes in the loop exit basic block and the split off417/// unswitched block.418///419/// Because the exit block remains an exit from the loop, this rewrites the420/// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI421/// nodes into the unswitched basic block to select between the value in the422/// old preheader and the loop exit.423static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB,424 BasicBlock &UnswitchedBB,425 BasicBlock &OldExitingBB,426 BasicBlock &OldPH,427 bool FullUnswitch) {428 assert(&ExitBB != &UnswitchedBB &&429 "Must have different loop exit and unswitched blocks!");430 BasicBlock::iterator InsertPt = UnswitchedBB.begin();431 for (PHINode &PN : ExitBB.phis()) {432 auto *NewPN = PHINode::Create(PN.getType(), /*NumReservedValues*/ 2,433 PN.getName() + ".split");434 NewPN->insertBefore(InsertPt);435 436 // Walk backwards over the old PHI node's inputs to minimize the cost of437 // removing each one. We have to do this weird loop manually so that we438 // create the same number of new incoming edges in the new PHI as we expect439 // each case-based edge to be included in the unswitched switch in some440 // cases.441 // FIXME: This is really, really gross. It would be much cleaner if LLVM442 // allowed us to create a single entry for a predecessor block without443 // having separate entries for each "edge" even though these edges are444 // required to produce identical results.445 for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) {446 if (PN.getIncomingBlock(i) != &OldExitingBB)447 continue;448 449 Value *Incoming = PN.getIncomingValue(i);450 if (FullUnswitch)451 // No more edge from the old exiting block to the exit block.452 PN.removeIncomingValue(i);453 454 NewPN->addIncoming(Incoming, &OldPH);455 }456 457 // Now replace the old PHI with the new one and wire the old one in as an458 // input to the new one.459 PN.replaceAllUsesWith(NewPN);460 NewPN->addIncoming(&PN, &ExitBB);461 }462}463 464/// Hoist the current loop up to the innermost loop containing a remaining exit.465///466/// Because we've removed an exit from the loop, we may have changed the set of467/// loops reachable and need to move the current loop up the loop nest or even468/// to an entirely separate nest.469static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader,470 DominatorTree &DT, LoopInfo &LI,471 MemorySSAUpdater *MSSAU, ScalarEvolution *SE) {472 // If the loop is already at the top level, we can't hoist it anywhere.473 Loop *OldParentL = L.getParentLoop();474 if (!OldParentL)475 return;476 477 SmallVector<BasicBlock *, 4> Exits;478 L.getExitBlocks(Exits);479 Loop *NewParentL = nullptr;480 for (auto *ExitBB : Exits)481 if (Loop *ExitL = LI.getLoopFor(ExitBB))482 if (!NewParentL || NewParentL->contains(ExitL))483 NewParentL = ExitL;484 485 if (NewParentL == OldParentL)486 return;487 488 // The new parent loop (if different) should always contain the old one.489 if (NewParentL)490 assert(NewParentL->contains(OldParentL) &&491 "Can only hoist this loop up the nest!");492 493 // The preheader will need to move with the body of this loop. However,494 // because it isn't in this loop we also need to update the primary loop map.495 assert(OldParentL == LI.getLoopFor(&Preheader) &&496 "Parent loop of this loop should contain this loop's preheader!");497 LI.changeLoopFor(&Preheader, NewParentL);498 499 // Remove this loop from its old parent.500 OldParentL->removeChildLoop(&L);501 502 // Add the loop either to the new parent or as a top-level loop.503 if (NewParentL)504 NewParentL->addChildLoop(&L);505 else506 LI.addTopLevelLoop(&L);507 508 // Remove this loops blocks from the old parent and every other loop up the509 // nest until reaching the new parent. Also update all of these510 // no-longer-containing loops to reflect the nesting change.511 for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL;512 OldContainingL = OldContainingL->getParentLoop()) {513 llvm::erase_if(OldContainingL->getBlocksVector(),514 [&](const BasicBlock *BB) {515 return BB == &Preheader || L.contains(BB);516 });517 518 OldContainingL->getBlocksSet().erase(&Preheader);519 for (BasicBlock *BB : L.blocks())520 OldContainingL->getBlocksSet().erase(BB);521 522 // Because we just hoisted a loop out of this one, we have essentially523 // created new exit paths from it. That means we need to form LCSSA PHI524 // nodes for values used in the no-longer-nested loop.525 formLCSSA(*OldContainingL, DT, &LI, SE);526 527 // We shouldn't need to form dedicated exits because the exit introduced528 // here is the (just split by unswitching) preheader. However, after trivial529 // unswitching it is possible to get new non-dedicated exits out of parent530 // loop so let's conservatively form dedicated exit blocks and figure out531 // if we can optimize later.532 formDedicatedExitBlocks(OldContainingL, &DT, &LI, MSSAU,533 /*PreserveLCSSA*/ true);534 }535}536 537// Return the top-most loop containing ExitBB and having ExitBB as exiting block538// or the loop containing ExitBB, if there is no parent loop containing ExitBB539// as exiting block.540static Loop *getTopMostExitingLoop(const BasicBlock *ExitBB,541 const LoopInfo &LI) {542 Loop *TopMost = LI.getLoopFor(ExitBB);543 Loop *Current = TopMost;544 while (Current) {545 if (Current->isLoopExiting(ExitBB))546 TopMost = Current;547 Current = Current->getParentLoop();548 }549 return TopMost;550}551 552/// Unswitch a trivial branch if the condition is loop invariant.553///554/// This routine should only be called when loop code leading to the branch has555/// been validated as trivial (no side effects). This routine checks if the556/// condition is invariant and one of the successors is a loop exit. This557/// allows us to unswitch without duplicating the loop, making it trivial.558///559/// If this routine fails to unswitch the branch it returns false.560///561/// If the branch can be unswitched, this routine splits the preheader and562/// hoists the branch above that split. Preserves loop simplified form563/// (splitting the exit block as necessary). It simplifies the branch within564/// the loop to an unconditional branch but doesn't remove it entirely. Further565/// cleanup can be done with some simplifycfg like pass.566///567/// If `SE` is not null, it will be updated based on the potential loop SCEVs568/// invalidated by this.569static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT,570 LoopInfo &LI, ScalarEvolution *SE,571 MemorySSAUpdater *MSSAU) {572 assert(BI.isConditional() && "Can only unswitch a conditional branch!");573 LLVM_DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n");574 575 // The loop invariant values that we want to unswitch.576 TinyPtrVector<Value *> Invariants;577 578 // When true, we're fully unswitching the branch rather than just unswitching579 // some input conditions to the branch.580 bool FullUnswitch = false;581 582 Value *Cond = skipTrivialSelect(BI.getCondition());583 if (L.isLoopInvariant(Cond)) {584 Invariants.push_back(Cond);585 FullUnswitch = true;586 } else {587 if (auto *CondInst = dyn_cast<Instruction>(Cond))588 Invariants = collectHomogenousInstGraphLoopInvariants(L, *CondInst, LI);589 if (Invariants.empty()) {590 LLVM_DEBUG(dbgs() << " Couldn't find invariant inputs!\n");591 return false;592 }593 }594 595 // Check that one of the branch's successors exits, and which one.596 bool ExitDirection = true;597 int LoopExitSuccIdx = 0;598 auto *LoopExitBB = BI.getSuccessor(0);599 if (L.contains(LoopExitBB)) {600 ExitDirection = false;601 LoopExitSuccIdx = 1;602 LoopExitBB = BI.getSuccessor(1);603 if (L.contains(LoopExitBB)) {604 LLVM_DEBUG(dbgs() << " Branch doesn't exit the loop!\n");605 return false;606 }607 }608 auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx);609 auto *ParentBB = BI.getParent();610 if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB)) {611 LLVM_DEBUG(dbgs() << " Loop exit PHI's aren't loop-invariant!\n");612 return false;613 }614 615 // When unswitching only part of the branch's condition, we need the exit616 // block to be reached directly from the partially unswitched input. This can617 // be done when the exit block is along the true edge and the branch condition618 // is a graph of `or` operations, or the exit block is along the false edge619 // and the condition is a graph of `and` operations.620 if (!FullUnswitch) {621 if (ExitDirection ? !match(Cond, m_LogicalOr())622 : !match(Cond, m_LogicalAnd())) {623 LLVM_DEBUG(dbgs() << " Branch condition is in improper form for "624 "non-full unswitch!\n");625 return false;626 }627 }628 629 LLVM_DEBUG({630 dbgs() << " unswitching trivial invariant conditions for: " << BI631 << "\n";632 for (Value *Invariant : Invariants) {633 dbgs() << " " << *Invariant << " == true";634 if (Invariant != Invariants.back())635 dbgs() << " ||";636 dbgs() << "\n";637 }638 });639 640 // If we have scalar evolutions, we need to invalidate them including this641 // loop, the loop containing the exit block and the topmost parent loop642 // exiting via LoopExitBB.643 if (SE) {644 if (const Loop *ExitL = getTopMostExitingLoop(LoopExitBB, LI))645 SE->forgetLoop(ExitL);646 else647 // Forget the entire nest as this exits the entire nest.648 SE->forgetTopmostLoop(&L);649 SE->forgetBlockAndLoopDispositions();650 }651 652 if (MSSAU && VerifyMemorySSA)653 MSSAU->getMemorySSA()->verifyMemorySSA();654 655 // Split the preheader, so that we know that there is a safe place to insert656 // the conditional branch. We will change the preheader to have a conditional657 // branch on LoopCond.658 BasicBlock *OldPH = L.getLoopPreheader();659 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU);660 661 // Now that we have a place to insert the conditional branch, create a place662 // to branch to: this is the exit block out of the loop that we are663 // unswitching. We need to split this if there are other loop predecessors.664 // Because the loop is in simplified form, *any* other predecessor is enough.665 BasicBlock *UnswitchedBB;666 if (FullUnswitch && LoopExitBB->getUniquePredecessor()) {667 assert(LoopExitBB->getUniquePredecessor() == BI.getParent() &&668 "A branch's parent isn't a predecessor!");669 UnswitchedBB = LoopExitBB;670 } else {671 UnswitchedBB =672 SplitBlock(LoopExitBB, LoopExitBB->begin(), &DT, &LI, MSSAU, "", false);673 }674 675 if (MSSAU && VerifyMemorySSA)676 MSSAU->getMemorySSA()->verifyMemorySSA();677 678 // Actually move the invariant uses into the unswitched position. If possible,679 // we do this by moving the instructions, but when doing partial unswitching680 // we do it by building a new merge of the values in the unswitched position.681 OldPH->getTerminator()->eraseFromParent();682 if (FullUnswitch) {683 // If fully unswitching, we can use the existing branch instruction.684 // Splice it into the old PH to gate reaching the new preheader and re-point685 // its successors.686 BI.moveBefore(*OldPH, OldPH->end());687 BI.setCondition(Cond);688 if (MSSAU) {689 // Temporarily clone the terminator, to make MSSA update cheaper by690 // separating "insert edge" updates from "remove edge" ones.691 BI.clone()->insertInto(ParentBB, ParentBB->end());692 } else {693 // Create a new unconditional branch that will continue the loop as a new694 // terminator.695 Instruction *NewBI = BranchInst::Create(ContinueBB, ParentBB);696 NewBI->setDebugLoc(BI.getDebugLoc());697 }698 BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB);699 BI.setSuccessor(1 - LoopExitSuccIdx, NewPH);700 } else {701 // Only unswitching a subset of inputs to the condition, so we will need to702 // build a new branch that merges the invariant inputs.703 if (ExitDirection)704 assert(match(skipTrivialSelect(BI.getCondition()), m_LogicalOr()) &&705 "Must have an `or` of `i1`s or `select i1 X, true, Y`s for the "706 "condition!");707 else708 assert(match(skipTrivialSelect(BI.getCondition()), m_LogicalAnd()) &&709 "Must have an `and` of `i1`s or `select i1 X, Y, false`s for the"710 " condition!");711 buildPartialUnswitchConditionalBranch(712 *OldPH, Invariants, ExitDirection, *UnswitchedBB, *NewPH,713 FreezeLoopUnswitchCond, OldPH->getTerminator(), nullptr, DT, BI);714 }715 716 // Update the dominator tree with the added edge.717 DT.insertEdge(OldPH, UnswitchedBB);718 719 // After the dominator tree was updated with the added edge, update MemorySSA720 // if available.721 if (MSSAU) {722 SmallVector<CFGUpdate, 1> Updates;723 Updates.push_back({cfg::UpdateKind::Insert, OldPH, UnswitchedBB});724 MSSAU->applyInsertUpdates(Updates, DT);725 }726 727 // Finish updating dominator tree and memory ssa for full unswitch.728 if (FullUnswitch) {729 if (MSSAU) {730 Instruction *Term = ParentBB->getTerminator();731 // Remove the cloned branch instruction and create unconditional branch732 // now.733 Instruction *NewBI = BranchInst::Create(ContinueBB, ParentBB);734 NewBI->setDebugLoc(Term->getDebugLoc());735 Term->eraseFromParent();736 MSSAU->removeEdge(ParentBB, LoopExitBB);737 }738 DT.deleteEdge(ParentBB, LoopExitBB);739 }740 741 if (MSSAU && VerifyMemorySSA)742 MSSAU->getMemorySSA()->verifyMemorySSA();743 744 // Rewrite the relevant PHI nodes.745 if (UnswitchedBB == LoopExitBB)746 rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH);747 else748 rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB,749 *ParentBB, *OldPH, FullUnswitch);750 751 // The constant we can replace all of our invariants with inside the loop752 // body. If any of the invariants have a value other than this the loop won't753 // be entered.754 ConstantInt *Replacement = ExitDirection755 ? ConstantInt::getFalse(BI.getContext())756 : ConstantInt::getTrue(BI.getContext());757 758 // Since this is an i1 condition we can also trivially replace uses of it759 // within the loop with a constant.760 for (Value *Invariant : Invariants)761 replaceLoopInvariantUses(L, Invariant, *Replacement);762 763 // If this was full unswitching, we may have changed the nesting relationship764 // for this loop so hoist it to its correct parent if needed.765 if (FullUnswitch)766 hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE);767 768 if (MSSAU && VerifyMemorySSA)769 MSSAU->getMemorySSA()->verifyMemorySSA();770 771 LLVM_DEBUG(dbgs() << " done: unswitching trivial branch...\n");772 ++NumTrivial;773 ++NumBranches;774 return true;775}776 777/// Unswitch a trivial switch if the condition is loop invariant.778///779/// This routine should only be called when loop code leading to the switch has780/// been validated as trivial (no side effects). This routine checks if the781/// condition is invariant and that at least one of the successors is a loop782/// exit. This allows us to unswitch without duplicating the loop, making it783/// trivial.784///785/// If this routine fails to unswitch the switch it returns false.786///787/// If the switch can be unswitched, this routine splits the preheader and788/// copies the switch above that split. If the default case is one of the789/// exiting cases, it copies the non-exiting cases and points them at the new790/// preheader. If the default case is not exiting, it copies the exiting cases791/// and points the default at the preheader. It preserves loop simplified form792/// (splitting the exit blocks as necessary). It simplifies the switch within793/// the loop by removing now-dead cases. If the default case is one of those794/// unswitched, it replaces its destination with a new basic block containing795/// only unreachable. Such basic blocks, while technically loop exits, are not796/// considered for unswitching so this is a stable transform and the same797/// switch will not be revisited. If after unswitching there is only a single798/// in-loop successor, the switch is further simplified to an unconditional799/// branch. Still more cleanup can be done with some simplifycfg like pass.800///801/// If `SE` is not null, it will be updated based on the potential loop SCEVs802/// invalidated by this.803static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT,804 LoopInfo &LI, ScalarEvolution *SE,805 MemorySSAUpdater *MSSAU) {806 LLVM_DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n");807 Value *LoopCond = SI.getCondition();808 809 // If this isn't switching on an invariant condition, we can't unswitch it.810 if (!L.isLoopInvariant(LoopCond))811 return false;812 813 auto *ParentBB = SI.getParent();814 815 // The same check must be used both for the default and the exit cases. We816 // should never leave edges from the switch instruction to a basic block that817 // we are unswitching, hence the condition used to determine the default case818 // needs to also be used to populate ExitCaseIndices, which is then used to819 // remove cases from the switch.820 auto IsTriviallyUnswitchableExitBlock = [&](BasicBlock &BBToCheck) {821 // BBToCheck is not an exit block if it is inside loop L.822 if (L.contains(&BBToCheck))823 return false;824 // BBToCheck is not trivial to unswitch if its phis aren't loop invariant.825 if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, BBToCheck))826 return false;827 // We do not unswitch a block that only has an unreachable statement, as828 // it's possible this is a previously unswitched block. Only unswitch if829 // either the terminator is not unreachable, or, if it is, it's not the only830 // instruction in the block.831 auto *TI = BBToCheck.getTerminator();832 bool isUnreachable = isa<UnreachableInst>(TI);833 return !isUnreachable || &*BBToCheck.getFirstNonPHIOrDbg() != TI;834 };835 836 SmallVector<int, 4> ExitCaseIndices;837 for (auto Case : SI.cases())838 if (IsTriviallyUnswitchableExitBlock(*Case.getCaseSuccessor()))839 ExitCaseIndices.push_back(Case.getCaseIndex());840 BasicBlock *DefaultExitBB = nullptr;841 SwitchInstProfUpdateWrapper::CaseWeightOpt DefaultCaseWeight =842 SwitchInstProfUpdateWrapper::getSuccessorWeight(SI, 0);843 if (IsTriviallyUnswitchableExitBlock(*SI.getDefaultDest())) {844 DefaultExitBB = SI.getDefaultDest();845 } else if (ExitCaseIndices.empty())846 return false;847 848 LLVM_DEBUG(dbgs() << " unswitching trivial switch...\n");849 850 if (MSSAU && VerifyMemorySSA)851 MSSAU->getMemorySSA()->verifyMemorySSA();852 853 // We may need to invalidate SCEVs for the outermost loop reached by any of854 // the exits.855 Loop *OuterL = &L;856 857 if (DefaultExitBB) {858 // Check the loop containing this exit.859 Loop *ExitL = getTopMostExitingLoop(DefaultExitBB, LI);860 if (!ExitL || ExitL->contains(OuterL))861 OuterL = ExitL;862 }863 for (unsigned Index : ExitCaseIndices) {864 auto CaseI = SI.case_begin() + Index;865 // Compute the outer loop from this exit.866 Loop *ExitL = getTopMostExitingLoop(CaseI->getCaseSuccessor(), LI);867 if (!ExitL || ExitL->contains(OuterL))868 OuterL = ExitL;869 }870 871 if (SE) {872 if (OuterL)873 SE->forgetLoop(OuterL);874 else875 SE->forgetTopmostLoop(&L);876 }877 878 if (DefaultExitBB) {879 // Clear out the default destination temporarily to allow accurate880 // predecessor lists to be examined below.881 SI.setDefaultDest(nullptr);882 }883 884 // Store the exit cases into a separate data structure and remove them from885 // the switch.886 SmallVector<std::tuple<ConstantInt *, BasicBlock *,887 SwitchInstProfUpdateWrapper::CaseWeightOpt>,888 4> ExitCases;889 ExitCases.reserve(ExitCaseIndices.size());890 SwitchInstProfUpdateWrapper SIW(SI);891 // We walk the case indices backwards so that we remove the last case first892 // and don't disrupt the earlier indices.893 for (unsigned Index : reverse(ExitCaseIndices)) {894 auto CaseI = SI.case_begin() + Index;895 // Save the value of this case.896 auto W = SIW.getSuccessorWeight(CaseI->getSuccessorIndex());897 ExitCases.emplace_back(CaseI->getCaseValue(), CaseI->getCaseSuccessor(), W);898 // Delete the unswitched cases.899 SIW.removeCase(CaseI);900 }901 902 // Check if after this all of the remaining cases point at the same903 // successor.904 BasicBlock *CommonSuccBB = nullptr;905 if (SI.getNumCases() > 0 &&906 all_of(drop_begin(SI.cases()), [&SI](const SwitchInst::CaseHandle &Case) {907 return Case.getCaseSuccessor() == SI.case_begin()->getCaseSuccessor();908 }))909 CommonSuccBB = SI.case_begin()->getCaseSuccessor();910 if (!DefaultExitBB) {911 // If we're not unswitching the default, we need it to match any cases to912 // have a common successor or if we have no cases it is the common913 // successor.914 if (SI.getNumCases() == 0)915 CommonSuccBB = SI.getDefaultDest();916 else if (SI.getDefaultDest() != CommonSuccBB)917 CommonSuccBB = nullptr;918 }919 920 // Split the preheader, so that we know that there is a safe place to insert921 // the switch.922 BasicBlock *OldPH = L.getLoopPreheader();923 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU);924 OldPH->getTerminator()->eraseFromParent();925 926 // Now add the unswitched switch. This new switch instruction inherits the927 // debug location of the old switch, because it semantically replace the old928 // one.929 auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH);930 NewSI->setDebugLoc(SIW->getDebugLoc());931 SwitchInstProfUpdateWrapper NewSIW(*NewSI);932 933 // Rewrite the IR for the unswitched basic blocks. This requires two steps.934 // First, we split any exit blocks with remaining in-loop predecessors. Then935 // we update the PHIs in one of two ways depending on if there was a split.936 // We walk in reverse so that we split in the same order as the cases937 // appeared. This is purely for convenience of reading the resulting IR, but938 // it doesn't cost anything really.939 SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs;940 SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap;941 // Handle the default exit if necessary.942 // FIXME: It'd be great if we could merge this with the loop below but LLVM's943 // ranges aren't quite powerful enough yet.944 if (DefaultExitBB) {945 if (pred_empty(DefaultExitBB)) {946 UnswitchedExitBBs.insert(DefaultExitBB);947 rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH);948 } else {949 auto *SplitBB =950 SplitBlock(DefaultExitBB, DefaultExitBB->begin(), &DT, &LI, MSSAU);951 rewritePHINodesForExitAndUnswitchedBlocks(*DefaultExitBB, *SplitBB,952 *ParentBB, *OldPH,953 /*FullUnswitch*/ true);954 DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB;955 }956 }957 // Note that we must use a reference in the for loop so that we update the958 // container.959 for (auto &ExitCase : reverse(ExitCases)) {960 // Grab a reference to the exit block in the pair so that we can update it.961 BasicBlock *ExitBB = std::get<1>(ExitCase);962 963 // If this case is the last edge into the exit block, we can simply reuse it964 // as it will no longer be a loop exit. No mapping necessary.965 if (pred_empty(ExitBB)) {966 // Only rewrite once.967 if (UnswitchedExitBBs.insert(ExitBB).second)968 rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH);969 continue;970 }971 972 // Otherwise we need to split the exit block so that we retain an exit973 // block from the loop and a target for the unswitched condition.974 BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB];975 if (!SplitExitBB) {976 // If this is the first time we see this, do the split and remember it.977 SplitExitBB = SplitBlock(ExitBB, ExitBB->begin(), &DT, &LI, MSSAU);978 rewritePHINodesForExitAndUnswitchedBlocks(*ExitBB, *SplitExitBB,979 *ParentBB, *OldPH,980 /*FullUnswitch*/ true);981 }982 // Update the case pair to point to the split block.983 std::get<1>(ExitCase) = SplitExitBB;984 }985 986 // Now add the unswitched cases. We do this in reverse order as we built them987 // in reverse order.988 for (auto &ExitCase : reverse(ExitCases)) {989 ConstantInt *CaseVal = std::get<0>(ExitCase);990 BasicBlock *UnswitchedBB = std::get<1>(ExitCase);991 992 NewSIW.addCase(CaseVal, UnswitchedBB, std::get<2>(ExitCase));993 }994 995 // If the default was unswitched, re-point it and add explicit cases for996 // entering the loop.997 if (DefaultExitBB) {998 NewSIW->setDefaultDest(DefaultExitBB);999 NewSIW.setSuccessorWeight(0, DefaultCaseWeight);1000 1001 // We removed all the exit cases, so we just copy the cases to the1002 // unswitched switch.1003 for (const auto &Case : SI.cases())1004 NewSIW.addCase(Case.getCaseValue(), NewPH,1005 SIW.getSuccessorWeight(Case.getSuccessorIndex()));1006 } else if (DefaultCaseWeight) {1007 // We have to set branch weight of the default case.1008 uint64_t SW = *DefaultCaseWeight;1009 for (const auto &Case : SI.cases()) {1010 auto W = SIW.getSuccessorWeight(Case.getSuccessorIndex());1011 assert(W &&1012 "case weight must be defined as default case weight is defined");1013 SW += *W;1014 }1015 NewSIW.setSuccessorWeight(0, SW);1016 }1017 1018 // If we ended up with a common successor for every path through the switch1019 // after unswitching, rewrite it to an unconditional branch to make it easy1020 // to recognize. Otherwise we potentially have to recognize the default case1021 // pointing at unreachable and other complexity.1022 if (CommonSuccBB) {1023 BasicBlock *BB = SI.getParent();1024 // We may have had multiple edges to this common successor block, so remove1025 // them as predecessors. We skip the first one, either the default or the1026 // actual first case.1027 bool SkippedFirst = DefaultExitBB == nullptr;1028 for (auto Case : SI.cases()) {1029 assert(Case.getCaseSuccessor() == CommonSuccBB &&1030 "Non-common successor!");1031 (void)Case;1032 if (!SkippedFirst) {1033 SkippedFirst = true;1034 continue;1035 }1036 CommonSuccBB->removePredecessor(BB,1037 /*KeepOneInputPHIs*/ true);1038 }1039 // Now nuke the switch and replace it with a direct branch.1040 Instruction *NewBI = BranchInst::Create(CommonSuccBB, BB);1041 NewBI->setDebugLoc(SIW->getDebugLoc());1042 SIW.eraseFromParent();1043 } else if (DefaultExitBB) {1044 assert(SI.getNumCases() > 0 &&1045 "If we had no cases we'd have a common successor!");1046 // Move the last case to the default successor. This is valid as if the1047 // default got unswitched it cannot be reached. This has the advantage of1048 // being simple and keeping the number of edges from this switch to1049 // successors the same, and avoiding any PHI update complexity.1050 auto LastCaseI = std::prev(SI.case_end());1051 1052 SI.setDefaultDest(LastCaseI->getCaseSuccessor());1053 SIW.setSuccessorWeight(1054 0, SIW.getSuccessorWeight(LastCaseI->getSuccessorIndex()));1055 SIW.removeCase(LastCaseI);1056 }1057 1058 // Walk the unswitched exit blocks and the unswitched split blocks and update1059 // the dominator tree based on the CFG edits. While we are walking unordered1060 // containers here, the API for applyUpdates takes an unordered list of1061 // updates and requires them to not contain duplicates.1062 SmallVector<DominatorTree::UpdateType, 4> DTUpdates;1063 for (auto *UnswitchedExitBB : UnswitchedExitBBs) {1064 DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedExitBB});1065 DTUpdates.push_back({DT.Insert, OldPH, UnswitchedExitBB});1066 }1067 for (auto SplitUnswitchedPair : SplitExitBBMap) {1068 DTUpdates.push_back({DT.Delete, ParentBB, SplitUnswitchedPair.first});1069 DTUpdates.push_back({DT.Insert, OldPH, SplitUnswitchedPair.second});1070 }1071 1072 if (MSSAU) {1073 MSSAU->applyUpdates(DTUpdates, DT, /*UpdateDT=*/true);1074 if (VerifyMemorySSA)1075 MSSAU->getMemorySSA()->verifyMemorySSA();1076 } else {1077 DT.applyUpdates(DTUpdates);1078 }1079 1080 assert(DT.verify(DominatorTree::VerificationLevel::Fast));1081 1082 // We may have changed the nesting relationship for this loop so hoist it to1083 // its correct parent if needed.1084 hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE);1085 1086 if (MSSAU && VerifyMemorySSA)1087 MSSAU->getMemorySSA()->verifyMemorySSA();1088 1089 ++NumTrivial;1090 ++NumSwitches;1091 LLVM_DEBUG(dbgs() << " done: unswitching trivial switch...\n");1092 return true;1093}1094 1095/// This routine scans the loop to find a branch or switch which occurs before1096/// any side effects occur. These can potentially be unswitched without1097/// duplicating the loop. If a branch or switch is successfully unswitched the1098/// scanning continues to see if subsequent branches or switches have become1099/// trivial. Once all trivial candidates have been unswitched, this routine1100/// returns.1101///1102/// The return value indicates whether anything was unswitched (and therefore1103/// changed).1104///1105/// If `SE` is not null, it will be updated based on the potential loop SCEVs1106/// invalidated by this.1107static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT,1108 LoopInfo &LI, ScalarEvolution *SE,1109 MemorySSAUpdater *MSSAU) {1110 bool Changed = false;1111 1112 // If loop header has only one reachable successor we should keep looking for1113 // trivial condition candidates in the successor as well. An alternative is1114 // to constant fold conditions and merge successors into loop header (then we1115 // only need to check header's terminator). The reason for not doing this in1116 // LoopUnswitch pass is that it could potentially break LoopPassManager's1117 // invariants. Folding dead branches could either eliminate the current loop1118 // or make other loops unreachable. LCSSA form might also not be preserved1119 // after deleting branches. The following code keeps traversing loop header's1120 // successors until it finds the trivial condition candidate (condition that1121 // is not a constant). Since unswitching generates branches with constant1122 // conditions, this scenario could be very common in practice.1123 BasicBlock *CurrentBB = L.getHeader();1124 SmallPtrSet<BasicBlock *, 8> Visited;1125 Visited.insert(CurrentBB);1126 do {1127 // Check if there are any side-effecting instructions (e.g. stores, calls,1128 // volatile loads) in the part of the loop that the code *would* execute1129 // without unswitching.1130 if (MSSAU) // Possible early exit with MSSA1131 if (auto *Defs = MSSAU->getMemorySSA()->getBlockDefs(CurrentBB))1132 if (!isa<MemoryPhi>(*Defs->begin()) || (++Defs->begin() != Defs->end()))1133 return Changed;1134 if (llvm::any_of(*CurrentBB,1135 [](Instruction &I) { return I.mayHaveSideEffects(); }))1136 return Changed;1137 1138 Instruction *CurrentTerm = CurrentBB->getTerminator();1139 1140 if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) {1141 // Don't bother trying to unswitch past a switch with a constant1142 // condition. This should be removed prior to running this pass by1143 // simplifycfg.1144 if (isa<Constant>(SI->getCondition()))1145 return Changed;1146 1147 if (!unswitchTrivialSwitch(L, *SI, DT, LI, SE, MSSAU))1148 // Couldn't unswitch this one so we're done.1149 return Changed;1150 1151 // Mark that we managed to unswitch something.1152 Changed = true;1153 1154 // If unswitching turned the terminator into an unconditional branch then1155 // we can continue. The unswitching logic specifically works to fold any1156 // cases it can into an unconditional branch to make it easier to1157 // recognize here.1158 auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator());1159 if (!BI || BI->isConditional())1160 return Changed;1161 1162 CurrentBB = BI->getSuccessor(0);1163 continue;1164 }1165 1166 auto *BI = dyn_cast<BranchInst>(CurrentTerm);1167 if (!BI)1168 // We do not understand other terminator instructions.1169 return Changed;1170 1171 // Don't bother trying to unswitch past an unconditional branch or a branch1172 // with a constant value. These should be removed by simplifycfg prior to1173 // running this pass.1174 if (!BI->isConditional() ||1175 isa<Constant>(skipTrivialSelect(BI->getCondition())))1176 return Changed;1177 1178 // Found a trivial condition candidate: non-foldable conditional branch. If1179 // we fail to unswitch this, we can't do anything else that is trivial.1180 if (!unswitchTrivialBranch(L, *BI, DT, LI, SE, MSSAU))1181 return Changed;1182 1183 // Mark that we managed to unswitch something.1184 Changed = true;1185 1186 // If we only unswitched some of the conditions feeding the branch, we won't1187 // have collapsed it to a single successor.1188 BI = cast<BranchInst>(CurrentBB->getTerminator());1189 if (BI->isConditional())1190 return Changed;1191 1192 // Follow the newly unconditional branch into its successor.1193 CurrentBB = BI->getSuccessor(0);1194 1195 // When continuing, if we exit the loop or reach a previous visited block,1196 // then we can not reach any trivial condition candidates (unfoldable1197 // branch instructions or switch instructions) and no unswitch can happen.1198 } while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second);1199 1200 return Changed;1201}1202 1203/// Build the cloned blocks for an unswitched copy of the given loop.1204///1205/// The cloned blocks are inserted before the loop preheader (`LoopPH`) and1206/// after the split block (`SplitBB`) that will be used to select between the1207/// cloned and original loop.1208///1209/// This routine handles cloning all of the necessary loop blocks and exit1210/// blocks including rewriting their instructions and the relevant PHI nodes.1211/// Any loop blocks or exit blocks which are dominated by a different successor1212/// than the one for this clone of the loop blocks can be trivially skipped. We1213/// use the `DominatingSucc` map to determine whether a block satisfies that1214/// property with a simple map lookup.1215///1216/// It also correctly creates the unconditional branch in the cloned1217/// unswitched parent block to only point at the unswitched successor.1218///1219/// This does not handle most of the necessary updates to `LoopInfo`. Only exit1220/// block splitting is correctly reflected in `LoopInfo`, essentially all of1221/// the cloned blocks (and their loops) are left without full `LoopInfo`1222/// updates. This also doesn't fully update `DominatorTree`. It adds the cloned1223/// blocks to them but doesn't create the cloned `DominatorTree` structure and1224/// instead the caller must recompute an accurate DT. It *does* correctly1225/// update the `AssumptionCache` provided in `AC`.1226static BasicBlock *buildClonedLoopBlocks(1227 Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB,1228 ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB,1229 BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB,1230 const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc,1231 ValueToValueMapTy &VMap,1232 SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC,1233 DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU,1234 ScalarEvolution *SE) {1235 SmallVector<BasicBlock *, 4> NewBlocks;1236 NewBlocks.reserve(L.getNumBlocks() + ExitBlocks.size());1237 1238 // We will need to clone a bunch of blocks, wrap up the clone operation in1239 // a helper.1240 auto CloneBlock = [&](BasicBlock *OldBB) {1241 // Clone the basic block and insert it before the new preheader.1242 BasicBlock *NewBB = CloneBasicBlock(OldBB, VMap, ".us", OldBB->getParent());1243 NewBB->moveBefore(LoopPH);1244 1245 // Record this block and the mapping.1246 NewBlocks.push_back(NewBB);1247 VMap[OldBB] = NewBB;1248 1249 return NewBB;1250 };1251 1252 // We skip cloning blocks when they have a dominating succ that is not the1253 // succ we are cloning for.1254 auto SkipBlock = [&](BasicBlock *BB) {1255 auto It = DominatingSucc.find(BB);1256 return It != DominatingSucc.end() && It->second != UnswitchedSuccBB;1257 };1258 1259 // First, clone the preheader.1260 auto *ClonedPH = CloneBlock(LoopPH);1261 1262 // Then clone all the loop blocks, skipping the ones that aren't necessary.1263 for (auto *LoopBB : L.blocks())1264 if (!SkipBlock(LoopBB))1265 CloneBlock(LoopBB);1266 1267 // Split all the loop exit edges so that when we clone the exit blocks, if1268 // any of the exit blocks are *also* a preheader for some other loop, we1269 // don't create multiple predecessors entering the loop header.1270 for (auto *ExitBB : ExitBlocks) {1271 if (SkipBlock(ExitBB))1272 continue;1273 1274 // When we are going to clone an exit, we don't need to clone all the1275 // instructions in the exit block and we want to ensure we have an easy1276 // place to merge the CFG, so split the exit first. This is always safe to1277 // do because there cannot be any non-loop predecessors of a loop exit in1278 // loop simplified form.1279 auto *MergeBB = SplitBlock(ExitBB, ExitBB->begin(), &DT, &LI, MSSAU);1280 1281 // Rearrange the names to make it easier to write test cases by having the1282 // exit block carry the suffix rather than the merge block carrying the1283 // suffix.1284 MergeBB->takeName(ExitBB);1285 ExitBB->setName(Twine(MergeBB->getName()) + ".split");1286 1287 // Now clone the original exit block.1288 auto *ClonedExitBB = CloneBlock(ExitBB);1289 assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 &&1290 "Exit block should have been split to have one successor!");1291 assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB &&1292 "Cloned exit block has the wrong successor!");1293 1294 // Remap any cloned instructions and create a merge phi node for them.1295 for (auto ZippedInsts : llvm::zip_first(1296 llvm::make_range(ExitBB->begin(), std::prev(ExitBB->end())),1297 llvm::make_range(ClonedExitBB->begin(),1298 std::prev(ClonedExitBB->end())))) {1299 Instruction &I = std::get<0>(ZippedInsts);1300 Instruction &ClonedI = std::get<1>(ZippedInsts);1301 1302 // The only instructions in the exit block should be PHI nodes and1303 // potentially a landing pad.1304 assert(1305 (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) &&1306 "Bad instruction in exit block!");1307 // We should have a value map between the instruction and its clone.1308 assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!");1309 1310 // Forget SCEVs based on exit phis in case SCEV looked through the phi.1311 if (SE)1312 if (auto *PN = dyn_cast<PHINode>(&I))1313 SE->forgetLcssaPhiWithNewPredecessor(&L, PN);1314 1315 BasicBlock::iterator InsertPt = MergeBB->getFirstInsertionPt();1316 1317 auto *MergePN =1318 PHINode::Create(I.getType(), /*NumReservedValues*/ 2, ".us-phi");1319 MergePN->insertBefore(InsertPt);1320 MergePN->setDebugLoc(InsertPt->getDebugLoc());1321 I.replaceAllUsesWith(MergePN);1322 MergePN->addIncoming(&I, ExitBB);1323 MergePN->addIncoming(&ClonedI, ClonedExitBB);1324 }1325 }1326 1327 // Rewrite the instructions in the cloned blocks to refer to the instructions1328 // in the cloned blocks. We have to do this as a second pass so that we have1329 // everything available. Also, we have inserted new instructions which may1330 // include assume intrinsics, so we update the assumption cache while1331 // processing this.1332 Module *M = ClonedPH->getParent()->getParent();1333 for (auto *ClonedBB : NewBlocks)1334 for (Instruction &I : *ClonedBB) {1335 RemapDbgRecordRange(M, I.getDbgRecordRange(), VMap,1336 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);1337 RemapInstruction(&I, VMap,1338 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);1339 if (auto *II = dyn_cast<AssumeInst>(&I))1340 AC.registerAssumption(II);1341 }1342 1343 // Update any PHI nodes in the cloned successors of the skipped blocks to not1344 // have spurious incoming values.1345 for (auto *LoopBB : L.blocks())1346 if (SkipBlock(LoopBB))1347 for (auto *SuccBB : successors(LoopBB))1348 if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB)))1349 for (PHINode &PN : ClonedSuccBB->phis())1350 PN.removeIncomingValue(LoopBB, /*DeletePHIIfEmpty*/ false);1351 1352 // Remove the cloned parent as a predecessor of any successor we ended up1353 // cloning other than the unswitched one.1354 auto *ClonedParentBB = cast<BasicBlock>(VMap.lookup(ParentBB));1355 for (auto *SuccBB : successors(ParentBB)) {1356 if (SuccBB == UnswitchedSuccBB)1357 continue;1358 1359 auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB));1360 if (!ClonedSuccBB)1361 continue;1362 1363 ClonedSuccBB->removePredecessor(ClonedParentBB,1364 /*KeepOneInputPHIs*/ true);1365 }1366 1367 // Replace the cloned branch with an unconditional branch to the cloned1368 // unswitched successor.1369 auto *ClonedSuccBB = cast<BasicBlock>(VMap.lookup(UnswitchedSuccBB));1370 Instruction *ClonedTerminator = ClonedParentBB->getTerminator();1371 // Trivial Simplification. If Terminator is a conditional branch and1372 // condition becomes dead - erase it.1373 Value *ClonedConditionToErase = nullptr;1374 if (auto *BI = dyn_cast<BranchInst>(ClonedTerminator))1375 ClonedConditionToErase = BI->getCondition();1376 else if (auto *SI = dyn_cast<SwitchInst>(ClonedTerminator))1377 ClonedConditionToErase = SI->getCondition();1378 1379 Instruction *BI = BranchInst::Create(ClonedSuccBB, ClonedParentBB);1380 BI->setDebugLoc(ClonedTerminator->getDebugLoc());1381 ClonedTerminator->eraseFromParent();1382 1383 if (ClonedConditionToErase)1384 RecursivelyDeleteTriviallyDeadInstructions(ClonedConditionToErase, nullptr,1385 MSSAU);1386 1387 // If there are duplicate entries in the PHI nodes because of multiple edges1388 // to the unswitched successor, we need to nuke all but one as we replaced it1389 // with a direct branch.1390 for (PHINode &PN : ClonedSuccBB->phis()) {1391 bool Found = false;1392 // Loop over the incoming operands backwards so we can easily delete as we1393 // go without invalidating the index.1394 for (int i = PN.getNumOperands() - 1; i >= 0; --i) {1395 if (PN.getIncomingBlock(i) != ClonedParentBB)1396 continue;1397 if (!Found) {1398 Found = true;1399 continue;1400 }1401 PN.removeIncomingValue(i, /*DeletePHIIfEmpty*/ false);1402 }1403 }1404 1405 // Record the domtree updates for the new blocks.1406 SmallPtrSet<BasicBlock *, 4> SuccSet;1407 for (auto *ClonedBB : NewBlocks) {1408 for (auto *SuccBB : successors(ClonedBB))1409 if (SuccSet.insert(SuccBB).second)1410 DTUpdates.push_back({DominatorTree::Insert, ClonedBB, SuccBB});1411 SuccSet.clear();1412 }1413 1414 return ClonedPH;1415}1416 1417/// Recursively clone the specified loop and all of its children.1418///1419/// The target parent loop for the clone should be provided, or can be null if1420/// the clone is a top-level loop. While cloning, all the blocks are mapped1421/// with the provided value map. The entire original loop must be present in1422/// the value map. The cloned loop is returned.1423static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL,1424 const ValueToValueMapTy &VMap, LoopInfo &LI) {1425 auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) {1426 assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!");1427 ClonedL.reserveBlocks(OrigL.getNumBlocks());1428 for (auto *BB : OrigL.blocks()) {1429 auto *ClonedBB = cast<BasicBlock>(VMap.lookup(BB));1430 ClonedL.addBlockEntry(ClonedBB);1431 if (LI.getLoopFor(BB) == &OrigL)1432 LI.changeLoopFor(ClonedBB, &ClonedL);1433 }1434 };1435 1436 // We specially handle the first loop because it may get cloned into1437 // a different parent and because we most commonly are cloning leaf loops.1438 Loop *ClonedRootL = LI.AllocateLoop();1439 if (RootParentL)1440 RootParentL->addChildLoop(ClonedRootL);1441 else1442 LI.addTopLevelLoop(ClonedRootL);1443 AddClonedBlocksToLoop(OrigRootL, *ClonedRootL);1444 1445 if (OrigRootL.isInnermost())1446 return ClonedRootL;1447 1448 // If we have a nest, we can quickly clone the entire loop nest using an1449 // iterative approach because it is a tree. We keep the cloned parent in the1450 // data structure to avoid repeatedly querying through a map to find it.1451 SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone;1452 // Build up the loops to clone in reverse order as we'll clone them from the1453 // back.1454 for (Loop *ChildL : llvm::reverse(OrigRootL))1455 LoopsToClone.push_back({ClonedRootL, ChildL});1456 do {1457 Loop *ClonedParentL, *L;1458 std::tie(ClonedParentL, L) = LoopsToClone.pop_back_val();1459 Loop *ClonedL = LI.AllocateLoop();1460 ClonedParentL->addChildLoop(ClonedL);1461 AddClonedBlocksToLoop(*L, *ClonedL);1462 for (Loop *ChildL : llvm::reverse(*L))1463 LoopsToClone.push_back({ClonedL, ChildL});1464 } while (!LoopsToClone.empty());1465 1466 return ClonedRootL;1467}1468 1469/// Build the cloned loops of an original loop from unswitching.1470///1471/// Because unswitching simplifies the CFG of the loop, this isn't a trivial1472/// operation. We need to re-verify that there even is a loop (as the backedge1473/// may not have been cloned), and even if there are remaining backedges the1474/// backedge set may be different. However, we know that each child loop is1475/// undisturbed, we only need to find where to place each child loop within1476/// either any parent loop or within a cloned version of the original loop.1477///1478/// Because child loops may end up cloned outside of any cloned version of the1479/// original loop, multiple cloned sibling loops may be created. All of them1480/// are returned so that the newly introduced loop nest roots can be1481/// identified.1482static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks,1483 const ValueToValueMapTy &VMap, LoopInfo &LI,1484 SmallVectorImpl<Loop *> &NonChildClonedLoops) {1485 Loop *ClonedL = nullptr;1486 1487 auto *OrigPH = OrigL.getLoopPreheader();1488 auto *OrigHeader = OrigL.getHeader();1489 1490 auto *ClonedPH = cast<BasicBlock>(VMap.lookup(OrigPH));1491 auto *ClonedHeader = cast<BasicBlock>(VMap.lookup(OrigHeader));1492 1493 // We need to know the loops of the cloned exit blocks to even compute the1494 // accurate parent loop. If we only clone exits to some parent of the1495 // original parent, we want to clone into that outer loop. We also keep track1496 // of the loops that our cloned exit blocks participate in.1497 Loop *ParentL = nullptr;1498 SmallVector<BasicBlock *, 4> ClonedExitsInLoops;1499 SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap;1500 ClonedExitsInLoops.reserve(ExitBlocks.size());1501 for (auto *ExitBB : ExitBlocks)1502 if (auto *ClonedExitBB = cast_or_null<BasicBlock>(VMap.lookup(ExitBB)))1503 if (Loop *ExitL = LI.getLoopFor(ExitBB)) {1504 ExitLoopMap[ClonedExitBB] = ExitL;1505 ClonedExitsInLoops.push_back(ClonedExitBB);1506 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))1507 ParentL = ExitL;1508 }1509 assert((!ParentL || ParentL == OrigL.getParentLoop() ||1510 ParentL->contains(OrigL.getParentLoop())) &&1511 "The computed parent loop should always contain (or be) the parent of "1512 "the original loop.");1513 1514 // We build the set of blocks dominated by the cloned header from the set of1515 // cloned blocks out of the original loop. While not all of these will1516 // necessarily be in the cloned loop, it is enough to establish that they1517 // aren't in unreachable cycles, etc.1518 SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks;1519 for (auto *BB : OrigL.blocks())1520 if (auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB)))1521 ClonedLoopBlocks.insert(ClonedBB);1522 1523 // Rebuild the set of blocks that will end up in the cloned loop. We may have1524 // skipped cloning some region of this loop which can in turn skip some of1525 // the backedges so we have to rebuild the blocks in the loop based on the1526 // backedges that remain after cloning.1527 SmallVector<BasicBlock *, 16> Worklist;1528 SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop;1529 for (auto *Pred : predecessors(ClonedHeader)) {1530 // The only possible non-loop header predecessor is the preheader because1531 // we know we cloned the loop in simplified form.1532 if (Pred == ClonedPH)1533 continue;1534 1535 // Because the loop was in simplified form, the only non-loop predecessor1536 // should be the preheader.1537 assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop "1538 "header other than the preheader "1539 "that is not part of the loop!");1540 1541 // Insert this block into the loop set and on the first visit (and if it1542 // isn't the header we're currently walking) put it into the worklist to1543 // recurse through.1544 if (BlocksInClonedLoop.insert(Pred).second && Pred != ClonedHeader)1545 Worklist.push_back(Pred);1546 }1547 1548 // If we had any backedges then there *is* a cloned loop. Put the header into1549 // the loop set and then walk the worklist backwards to find all the blocks1550 // that remain within the loop after cloning.1551 if (!BlocksInClonedLoop.empty()) {1552 BlocksInClonedLoop.insert(ClonedHeader);1553 1554 while (!Worklist.empty()) {1555 BasicBlock *BB = Worklist.pop_back_val();1556 assert(BlocksInClonedLoop.count(BB) &&1557 "Didn't put block into the loop set!");1558 1559 // Insert any predecessors that are in the possible set into the cloned1560 // set, and if the insert is successful, add them to the worklist. Note1561 // that we filter on the blocks that are definitely reachable via the1562 // backedge to the loop header so we may prune out dead code within the1563 // cloned loop.1564 for (auto *Pred : predecessors(BB))1565 if (ClonedLoopBlocks.count(Pred) &&1566 BlocksInClonedLoop.insert(Pred).second)1567 Worklist.push_back(Pred);1568 }1569 1570 ClonedL = LI.AllocateLoop();1571 if (ParentL) {1572 ParentL->addBasicBlockToLoop(ClonedPH, LI);1573 ParentL->addChildLoop(ClonedL);1574 } else {1575 LI.addTopLevelLoop(ClonedL);1576 }1577 NonChildClonedLoops.push_back(ClonedL);1578 1579 ClonedL->reserveBlocks(BlocksInClonedLoop.size());1580 // We don't want to just add the cloned loop blocks based on how we1581 // discovered them. The original order of blocks was carefully built in1582 // a way that doesn't rely on predecessor ordering. Rather than re-invent1583 // that logic, we just re-walk the original blocks (and those of the child1584 // loops) and filter them as we add them into the cloned loop.1585 for (auto *BB : OrigL.blocks()) {1586 auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB));1587 if (!ClonedBB || !BlocksInClonedLoop.count(ClonedBB))1588 continue;1589 1590 // Directly add the blocks that are only in this loop.1591 if (LI.getLoopFor(BB) == &OrigL) {1592 ClonedL->addBasicBlockToLoop(ClonedBB, LI);1593 continue;1594 }1595 1596 // We want to manually add it to this loop and parents.1597 // Registering it with LoopInfo will happen when we clone the top1598 // loop for this block.1599 for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop())1600 PL->addBlockEntry(ClonedBB);1601 }1602 1603 // Now add each child loop whose header remains within the cloned loop. All1604 // of the blocks within the loop must satisfy the same constraints as the1605 // header so once we pass the header checks we can just clone the entire1606 // child loop nest.1607 for (Loop *ChildL : OrigL) {1608 auto *ClonedChildHeader =1609 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));1610 if (!ClonedChildHeader || !BlocksInClonedLoop.count(ClonedChildHeader))1611 continue;1612 1613#ifndef NDEBUG1614 // We should never have a cloned child loop header but fail to have1615 // all of the blocks for that child loop.1616 for (auto *ChildLoopBB : ChildL->blocks())1617 assert(BlocksInClonedLoop.count(1618 cast<BasicBlock>(VMap.lookup(ChildLoopBB))) &&1619 "Child cloned loop has a header within the cloned outer "1620 "loop but not all of its blocks!");1621#endif1622 1623 cloneLoopNest(*ChildL, ClonedL, VMap, LI);1624 }1625 }1626 1627 // Now that we've handled all the components of the original loop that were1628 // cloned into a new loop, we still need to handle anything from the original1629 // loop that wasn't in a cloned loop.1630 1631 // Figure out what blocks are left to place within any loop nest containing1632 // the unswitched loop. If we never formed a loop, the cloned PH is one of1633 // them.1634 SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet;1635 if (BlocksInClonedLoop.empty())1636 UnloopedBlockSet.insert(ClonedPH);1637 for (auto *ClonedBB : ClonedLoopBlocks)1638 if (!BlocksInClonedLoop.count(ClonedBB))1639 UnloopedBlockSet.insert(ClonedBB);1640 1641 // Copy the cloned exits and sort them in ascending loop depth, we'll work1642 // backwards across these to process them inside out. The order shouldn't1643 // matter as we're just trying to build up the map from inside-out; we use1644 // the map in a more stably ordered way below.1645 auto OrderedClonedExitsInLoops = ClonedExitsInLoops;1646 llvm::sort(OrderedClonedExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) {1647 return ExitLoopMap.lookup(LHS)->getLoopDepth() <1648 ExitLoopMap.lookup(RHS)->getLoopDepth();1649 });1650 1651 // Populate the existing ExitLoopMap with everything reachable from each1652 // exit, starting from the inner most exit.1653 while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) {1654 assert(Worklist.empty() && "Didn't clear worklist!");1655 1656 BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val();1657 Loop *ExitL = ExitLoopMap.lookup(ExitBB);1658 1659 // Walk the CFG back until we hit the cloned PH adding everything reachable1660 // and in the unlooped set to this exit block's loop.1661 Worklist.push_back(ExitBB);1662 do {1663 BasicBlock *BB = Worklist.pop_back_val();1664 // We can stop recursing at the cloned preheader (if we get there).1665 if (BB == ClonedPH)1666 continue;1667 1668 for (BasicBlock *PredBB : predecessors(BB)) {1669 // If this pred has already been moved to our set or is part of some1670 // (inner) loop, no update needed.1671 if (!UnloopedBlockSet.erase(PredBB)) {1672 assert(1673 (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) &&1674 "Predecessor not mapped to a loop!");1675 continue;1676 }1677 1678 // We just insert into the loop set here. We'll add these blocks to the1679 // exit loop after we build up the set in an order that doesn't rely on1680 // predecessor order (which in turn relies on use list order).1681 bool Inserted = ExitLoopMap.insert({PredBB, ExitL}).second;1682 (void)Inserted;1683 assert(Inserted && "Should only visit an unlooped block once!");1684 1685 // And recurse through to its predecessors.1686 Worklist.push_back(PredBB);1687 }1688 } while (!Worklist.empty());1689 }1690 1691 // Now that the ExitLoopMap gives as mapping for all the non-looping cloned1692 // blocks to their outer loops, walk the cloned blocks and the cloned exits1693 // in their original order adding them to the correct loop.1694 1695 // We need a stable insertion order. We use the order of the original loop1696 // order and map into the correct parent loop.1697 for (auto *BB : llvm::concat<BasicBlock *const>(1698 ArrayRef(ClonedPH), ClonedLoopBlocks, ClonedExitsInLoops))1699 if (Loop *OuterL = ExitLoopMap.lookup(BB))1700 OuterL->addBasicBlockToLoop(BB, LI);1701 1702#ifndef NDEBUG1703 for (auto &BBAndL : ExitLoopMap) {1704 auto *BB = BBAndL.first;1705 auto *OuterL = BBAndL.second;1706 assert(LI.getLoopFor(BB) == OuterL &&1707 "Failed to put all blocks into outer loops!");1708 }1709#endif1710 1711 // Now that all the blocks are placed into the correct containing loop in the1712 // absence of child loops, find all the potentially cloned child loops and1713 // clone them into whatever outer loop we placed their header into.1714 for (Loop *ChildL : OrigL) {1715 auto *ClonedChildHeader =1716 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));1717 if (!ClonedChildHeader || BlocksInClonedLoop.count(ClonedChildHeader))1718 continue;1719 1720#ifndef NDEBUG1721 for (auto *ChildLoopBB : ChildL->blocks())1722 assert(VMap.count(ChildLoopBB) &&1723 "Cloned a child loop header but not all of that loops blocks!");1724#endif1725 1726 NonChildClonedLoops.push_back(cloneLoopNest(1727 *ChildL, ExitLoopMap.lookup(ClonedChildHeader), VMap, LI));1728 }1729}1730 1731static void1732deleteDeadClonedBlocks(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,1733 ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps,1734 DominatorTree &DT, MemorySSAUpdater *MSSAU) {1735 // Find all the dead clones, and remove them from their successors.1736 SmallVector<BasicBlock *, 16> DeadBlocks;1737 for (BasicBlock *BB : llvm::concat<BasicBlock *const>(L.blocks(), ExitBlocks))1738 for (const auto &VMap : VMaps)1739 if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(VMap->lookup(BB)))1740 if (!DT.isReachableFromEntry(ClonedBB)) {1741 for (BasicBlock *SuccBB : successors(ClonedBB))1742 SuccBB->removePredecessor(ClonedBB);1743 DeadBlocks.push_back(ClonedBB);1744 }1745 1746 // Remove all MemorySSA in the dead blocks1747 if (MSSAU) {1748 SmallSetVector<BasicBlock *, 8> DeadBlockSet(DeadBlocks.begin(),1749 DeadBlocks.end());1750 MSSAU->removeBlocks(DeadBlockSet);1751 }1752 1753 // Drop any remaining references to break cycles.1754 for (BasicBlock *BB : DeadBlocks)1755 BB->dropAllReferences();1756 // Erase them from the IR.1757 for (BasicBlock *BB : DeadBlocks)1758 BB->eraseFromParent();1759}1760 1761static void deleteDeadBlocksFromLoop(Loop &L,1762 SmallVectorImpl<BasicBlock *> &ExitBlocks,1763 DominatorTree &DT, LoopInfo &LI,1764 MemorySSAUpdater *MSSAU,1765 ScalarEvolution *SE,1766 LPMUpdater &LoopUpdater) {1767 // Find all the dead blocks tied to this loop, and remove them from their1768 // successors.1769 SmallSetVector<BasicBlock *, 8> DeadBlockSet;1770 1771 // Start with loop/exit blocks and get a transitive closure of reachable dead1772 // blocks.1773 SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(),1774 ExitBlocks.end());1775 DeathCandidates.append(L.blocks().begin(), L.blocks().end());1776 while (!DeathCandidates.empty()) {1777 auto *BB = DeathCandidates.pop_back_val();1778 if (!DeadBlockSet.count(BB) && !DT.isReachableFromEntry(BB)) {1779 for (BasicBlock *SuccBB : successors(BB)) {1780 SuccBB->removePredecessor(BB);1781 DeathCandidates.push_back(SuccBB);1782 }1783 DeadBlockSet.insert(BB);1784 }1785 }1786 1787 // Remove all MemorySSA in the dead blocks1788 if (MSSAU)1789 MSSAU->removeBlocks(DeadBlockSet);1790 1791 // Filter out the dead blocks from the exit blocks list so that it can be1792 // used in the caller.1793 llvm::erase_if(ExitBlocks,1794 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });1795 1796 // Walk from this loop up through its parents removing all of the dead blocks.1797 for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) {1798 for (auto *BB : DeadBlockSet)1799 ParentL->getBlocksSet().erase(BB);1800 llvm::erase_if(ParentL->getBlocksVector(),1801 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });1802 }1803 1804 // Now delete the dead child loops. This raw delete will clear them1805 // recursively.1806 llvm::erase_if(L.getSubLoopsVector(), [&](Loop *ChildL) {1807 if (!DeadBlockSet.count(ChildL->getHeader()))1808 return false;1809 1810 assert(llvm::all_of(ChildL->blocks(),1811 [&](BasicBlock *ChildBB) {1812 return DeadBlockSet.count(ChildBB);1813 }) &&1814 "If the child loop header is dead all blocks in the child loop must "1815 "be dead as well!");1816 LoopUpdater.markLoopAsDeleted(*ChildL, ChildL->getName());1817 if (SE)1818 SE->forgetBlockAndLoopDispositions();1819 LI.destroy(ChildL);1820 return true;1821 });1822 1823 // Remove the loop mappings for the dead blocks and drop all the references1824 // from these blocks to others to handle cyclic references as we start1825 // deleting the blocks themselves.1826 for (auto *BB : DeadBlockSet) {1827 // Check that the dominator tree has already been updated.1828 assert(!DT.getNode(BB) && "Should already have cleared domtree!");1829 LI.changeLoopFor(BB, nullptr);1830 // Drop all uses of the instructions to make sure we won't have dangling1831 // uses in other blocks.1832 for (auto &I : *BB)1833 if (!I.use_empty())1834 I.replaceAllUsesWith(PoisonValue::get(I.getType()));1835 BB->dropAllReferences();1836 }1837 1838 // Actually delete the blocks now that they've been fully unhooked from the1839 // IR.1840 for (auto *BB : DeadBlockSet)1841 BB->eraseFromParent();1842}1843 1844/// Recompute the set of blocks in a loop after unswitching.1845///1846/// This walks from the original headers predecessors to rebuild the loop. We1847/// take advantage of the fact that new blocks can't have been added, and so we1848/// filter by the original loop's blocks. This also handles potentially1849/// unreachable code that we don't want to explore but might be found examining1850/// the predecessors of the header.1851///1852/// If the original loop is no longer a loop, this will return an empty set. If1853/// it remains a loop, all the blocks within it will be added to the set1854/// (including those blocks in inner loops).1855static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L,1856 LoopInfo &LI) {1857 SmallPtrSet<const BasicBlock *, 16> LoopBlockSet;1858 1859 auto *PH = L.getLoopPreheader();1860 auto *Header = L.getHeader();1861 1862 // A worklist to use while walking backwards from the header.1863 SmallVector<BasicBlock *, 16> Worklist;1864 1865 // First walk the predecessors of the header to find the backedges. This will1866 // form the basis of our walk.1867 for (auto *Pred : predecessors(Header)) {1868 // Skip the preheader.1869 if (Pred == PH)1870 continue;1871 1872 // Because the loop was in simplified form, the only non-loop predecessor1873 // is the preheader.1874 assert(L.contains(Pred) && "Found a predecessor of the loop header other "1875 "than the preheader that is not part of the "1876 "loop!");1877 1878 // Insert this block into the loop set and on the first visit and, if it1879 // isn't the header we're currently walking, put it into the worklist to1880 // recurse through.1881 if (LoopBlockSet.insert(Pred).second && Pred != Header)1882 Worklist.push_back(Pred);1883 }1884 1885 // If no backedges were found, we're done.1886 if (LoopBlockSet.empty())1887 return LoopBlockSet;1888 1889 // We found backedges, recurse through them to identify the loop blocks.1890 while (!Worklist.empty()) {1891 BasicBlock *BB = Worklist.pop_back_val();1892 assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!");1893 1894 // No need to walk past the header.1895 if (BB == Header)1896 continue;1897 1898 // Because we know the inner loop structure remains valid we can use the1899 // loop structure to jump immediately across the entire nested loop.1900 // Further, because it is in loop simplified form, we can directly jump1901 // to its preheader afterward.1902 if (Loop *InnerL = LI.getLoopFor(BB))1903 if (InnerL != &L) {1904 assert(L.contains(InnerL) &&1905 "Should not reach a loop *outside* this loop!");1906 // The preheader is the only possible predecessor of the loop so1907 // insert it into the set and check whether it was already handled.1908 auto *InnerPH = InnerL->getLoopPreheader();1909 assert(L.contains(InnerPH) && "Cannot contain an inner loop block "1910 "but not contain the inner loop "1911 "preheader!");1912 if (!LoopBlockSet.insert(InnerPH).second)1913 // The only way to reach the preheader is through the loop body1914 // itself so if it has been visited the loop is already handled.1915 continue;1916 1917 // Insert all of the blocks (other than those already present) into1918 // the loop set. We expect at least the block that led us to find the1919 // inner loop to be in the block set, but we may also have other loop1920 // blocks if they were already enqueued as predecessors of some other1921 // outer loop block.1922 for (auto *InnerBB : InnerL->blocks()) {1923 if (InnerBB == BB) {1924 assert(LoopBlockSet.count(InnerBB) &&1925 "Block should already be in the set!");1926 continue;1927 }1928 1929 LoopBlockSet.insert(InnerBB);1930 }1931 1932 // Add the preheader to the worklist so we will continue past the1933 // loop body.1934 Worklist.push_back(InnerPH);1935 continue;1936 }1937 1938 // Insert any predecessors that were in the original loop into the new1939 // set, and if the insert is successful, add them to the worklist.1940 for (auto *Pred : predecessors(BB))1941 if (L.contains(Pred) && LoopBlockSet.insert(Pred).second)1942 Worklist.push_back(Pred);1943 }1944 1945 assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!");1946 1947 // We've found all the blocks participating in the loop, return our completed1948 // set.1949 return LoopBlockSet;1950}1951 1952/// Rebuild a loop after unswitching removes some subset of blocks and edges.1953///1954/// The removal may have removed some child loops entirely but cannot have1955/// disturbed any remaining child loops. However, they may need to be hoisted1956/// to the parent loop (or to be top-level loops). The original loop may be1957/// completely removed.1958///1959/// The sibling loops resulting from this update are returned. If the original1960/// loop remains a valid loop, it will be the first entry in this list with all1961/// of the newly sibling loops following it.1962///1963/// Returns true if the loop remains a loop after unswitching, and false if it1964/// is no longer a loop after unswitching (and should not continue to be1965/// referenced).1966static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,1967 LoopInfo &LI,1968 SmallVectorImpl<Loop *> &HoistedLoops,1969 ScalarEvolution *SE) {1970 auto *PH = L.getLoopPreheader();1971 1972 // Compute the actual parent loop from the exit blocks. Because we may have1973 // pruned some exits the loop may be different from the original parent.1974 Loop *ParentL = nullptr;1975 SmallVector<Loop *, 4> ExitLoops;1976 SmallVector<BasicBlock *, 4> ExitsInLoops;1977 ExitsInLoops.reserve(ExitBlocks.size());1978 for (auto *ExitBB : ExitBlocks)1979 if (Loop *ExitL = LI.getLoopFor(ExitBB)) {1980 ExitLoops.push_back(ExitL);1981 ExitsInLoops.push_back(ExitBB);1982 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))1983 ParentL = ExitL;1984 }1985 1986 // Recompute the blocks participating in this loop. This may be empty if it1987 // is no longer a loop.1988 auto LoopBlockSet = recomputeLoopBlockSet(L, LI);1989 1990 // If we still have a loop, we need to re-set the loop's parent as the exit1991 // block set changing may have moved it within the loop nest. Note that this1992 // can only happen when this loop has a parent as it can only hoist the loop1993 // *up* the nest.1994 if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) {1995 // Remove this loop's (original) blocks from all of the intervening loops.1996 for (Loop *IL = L.getParentLoop(); IL != ParentL;1997 IL = IL->getParentLoop()) {1998 IL->getBlocksSet().erase(PH);1999 for (auto *BB : L.blocks())2000 IL->getBlocksSet().erase(BB);2001 llvm::erase_if(IL->getBlocksVector(), [&](BasicBlock *BB) {2002 return BB == PH || L.contains(BB);2003 });2004 }2005 2006 LI.changeLoopFor(PH, ParentL);2007 L.getParentLoop()->removeChildLoop(&L);2008 if (ParentL)2009 ParentL->addChildLoop(&L);2010 else2011 LI.addTopLevelLoop(&L);2012 }2013 2014 // Now we update all the blocks which are no longer within the loop.2015 auto &Blocks = L.getBlocksVector();2016 auto BlocksSplitI =2017 LoopBlockSet.empty()2018 ? Blocks.begin()2019 : std::stable_partition(2020 Blocks.begin(), Blocks.end(),2021 [&](BasicBlock *BB) { return LoopBlockSet.count(BB); });2022 2023 // Before we erase the list of unlooped blocks, build a set of them.2024 SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end());2025 if (LoopBlockSet.empty())2026 UnloopedBlocks.insert(PH);2027 2028 // Now erase these blocks from the loop.2029 for (auto *BB : make_range(BlocksSplitI, Blocks.end()))2030 L.getBlocksSet().erase(BB);2031 Blocks.erase(BlocksSplitI, Blocks.end());2032 2033 // Sort the exits in ascending loop depth, we'll work backwards across these2034 // to process them inside out.2035 llvm::stable_sort(ExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) {2036 return LI.getLoopDepth(LHS) < LI.getLoopDepth(RHS);2037 });2038 2039 // We'll build up a set for each exit loop.2040 SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks;2041 Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop.2042 2043 auto RemoveUnloopedBlocksFromLoop =2044 [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) {2045 for (auto *BB : UnloopedBlocks)2046 L.getBlocksSet().erase(BB);2047 llvm::erase_if(L.getBlocksVector(), [&](BasicBlock *BB) {2048 return UnloopedBlocks.count(BB);2049 });2050 };2051 2052 SmallVector<BasicBlock *, 16> Worklist;2053 while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) {2054 assert(Worklist.empty() && "Didn't clear worklist!");2055 assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!");2056 2057 // Grab the next exit block, in decreasing loop depth order.2058 BasicBlock *ExitBB = ExitsInLoops.pop_back_val();2059 Loop &ExitL = *LI.getLoopFor(ExitBB);2060 assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!");2061 2062 // Erase all of the unlooped blocks from the loops between the previous2063 // exit loop and this exit loop. This works because the ExitInLoops list is2064 // sorted in increasing order of loop depth and thus we visit loops in2065 // decreasing order of loop depth.2066 for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop())2067 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);2068 2069 // Walk the CFG back until we hit the cloned PH adding everything reachable2070 // and in the unlooped set to this exit block's loop.2071 Worklist.push_back(ExitBB);2072 do {2073 BasicBlock *BB = Worklist.pop_back_val();2074 // We can stop recursing at the cloned preheader (if we get there).2075 if (BB == PH)2076 continue;2077 2078 for (BasicBlock *PredBB : predecessors(BB)) {2079 // If this pred has already been moved to our set or is part of some2080 // (inner) loop, no update needed.2081 if (!UnloopedBlocks.erase(PredBB)) {2082 assert((NewExitLoopBlocks.count(PredBB) ||2083 ExitL.contains(LI.getLoopFor(PredBB))) &&2084 "Predecessor not in a nested loop (or already visited)!");2085 continue;2086 }2087 2088 // We just insert into the loop set here. We'll add these blocks to the2089 // exit loop after we build up the set in a deterministic order rather2090 // than the predecessor-influenced visit order.2091 bool Inserted = NewExitLoopBlocks.insert(PredBB).second;2092 (void)Inserted;2093 assert(Inserted && "Should only visit an unlooped block once!");2094 2095 // And recurse through to its predecessors.2096 Worklist.push_back(PredBB);2097 }2098 } while (!Worklist.empty());2099 2100 // If blocks in this exit loop were directly part of the original loop (as2101 // opposed to a child loop) update the map to point to this exit loop. This2102 // just updates a map and so the fact that the order is unstable is fine.2103 for (auto *BB : NewExitLoopBlocks)2104 if (Loop *BBL = LI.getLoopFor(BB))2105 if (BBL == &L || !L.contains(BBL))2106 LI.changeLoopFor(BB, &ExitL);2107 2108 // We will remove the remaining unlooped blocks from this loop in the next2109 // iteration or below.2110 NewExitLoopBlocks.clear();2111 }2112 2113 // Any remaining unlooped blocks are no longer part of any loop unless they2114 // are part of some child loop.2115 for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop())2116 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);2117 for (auto *BB : UnloopedBlocks)2118 if (Loop *BBL = LI.getLoopFor(BB))2119 if (BBL == &L || !L.contains(BBL))2120 LI.changeLoopFor(BB, nullptr);2121 2122 // Sink all the child loops whose headers are no longer in the loop set to2123 // the parent (or to be top level loops). We reach into the loop and directly2124 // update its subloop vector to make this batch update efficient.2125 auto &SubLoops = L.getSubLoopsVector();2126 auto SubLoopsSplitI =2127 LoopBlockSet.empty()2128 ? SubLoops.begin()2129 : std::stable_partition(2130 SubLoops.begin(), SubLoops.end(), [&](Loop *SubL) {2131 return LoopBlockSet.count(SubL->getHeader());2132 });2133 for (auto *HoistedL : make_range(SubLoopsSplitI, SubLoops.end())) {2134 HoistedLoops.push_back(HoistedL);2135 HoistedL->setParentLoop(nullptr);2136 2137 // To compute the new parent of this hoisted loop we look at where we2138 // placed the preheader above. We can't lookup the header itself because we2139 // retained the mapping from the header to the hoisted loop. But the2140 // preheader and header should have the exact same new parent computed2141 // based on the set of exit blocks from the original loop as the preheader2142 // is a predecessor of the header and so reached in the reverse walk. And2143 // because the loops were all in simplified form the preheader of the2144 // hoisted loop can't be part of some *other* loop.2145 if (auto *NewParentL = LI.getLoopFor(HoistedL->getLoopPreheader()))2146 NewParentL->addChildLoop(HoistedL);2147 else2148 LI.addTopLevelLoop(HoistedL);2149 }2150 SubLoops.erase(SubLoopsSplitI, SubLoops.end());2151 2152 // Actually delete the loop if nothing remained within it.2153 if (Blocks.empty()) {2154 assert(SubLoops.empty() &&2155 "Failed to remove all subloops from the original loop!");2156 if (Loop *ParentL = L.getParentLoop())2157 ParentL->removeChildLoop(llvm::find(*ParentL, &L));2158 else2159 LI.removeLoop(llvm::find(LI, &L));2160 // markLoopAsDeleted for L should be triggered by the caller (it is2161 // typically done within postUnswitch).2162 if (SE)2163 SE->forgetBlockAndLoopDispositions();2164 LI.destroy(&L);2165 return false;2166 }2167 2168 return true;2169}2170 2171/// Helper to visit a dominator subtree, invoking a callable on each node.2172///2173/// Returning false at any point will stop walking past that node of the tree.2174template <typename CallableT>2175void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) {2176 SmallVector<DomTreeNode *, 4> DomWorklist;2177 DomWorklist.push_back(DT[BB]);2178#ifndef NDEBUG2179 SmallPtrSet<DomTreeNode *, 4> Visited;2180 Visited.insert(DT[BB]);2181#endif2182 do {2183 DomTreeNode *N = DomWorklist.pop_back_val();2184 2185 // Visit this node.2186 if (!Callable(N->getBlock()))2187 continue;2188 2189 // Accumulate the child nodes.2190 for (DomTreeNode *ChildN : *N) {2191 assert(Visited.insert(ChildN).second &&2192 "Cannot visit a node twice when walking a tree!");2193 DomWorklist.push_back(ChildN);2194 }2195 } while (!DomWorklist.empty());2196}2197 2198void postUnswitch(Loop &L, LPMUpdater &U, StringRef LoopName,2199 bool CurrentLoopValid, bool PartiallyInvariant,2200 bool InjectedCondition, ArrayRef<Loop *> NewLoops) {2201 // If we did a non-trivial unswitch, we have added new (cloned) loops.2202 if (!NewLoops.empty())2203 U.addSiblingLoops(NewLoops);2204 2205 // If the current loop remains valid, we should revisit it to catch any2206 // other unswitch opportunities. Otherwise, we need to mark it as deleted.2207 if (CurrentLoopValid) {2208 if (PartiallyInvariant) {2209 // Mark the new loop as partially unswitched, to avoid unswitching on2210 // the same condition again.2211 auto &Context = L.getHeader()->getContext();2212 MDNode *DisableUnswitchMD = MDNode::get(2213 Context,2214 MDString::get(Context, "llvm.loop.unswitch.partial.disable"));2215 MDNode *NewLoopID = makePostTransformationMetadata(2216 Context, L.getLoopID(), {"llvm.loop.unswitch.partial"},2217 {DisableUnswitchMD});2218 L.setLoopID(NewLoopID);2219 } else if (InjectedCondition) {2220 // Do the same for injection of invariant conditions.2221 auto &Context = L.getHeader()->getContext();2222 MDNode *DisableUnswitchMD = MDNode::get(2223 Context,2224 MDString::get(Context, "llvm.loop.unswitch.injection.disable"));2225 MDNode *NewLoopID = makePostTransformationMetadata(2226 Context, L.getLoopID(), {"llvm.loop.unswitch.injection"},2227 {DisableUnswitchMD});2228 L.setLoopID(NewLoopID);2229 } else2230 U.revisitCurrentLoop();2231 } else2232 U.markLoopAsDeleted(L, LoopName);2233}2234 2235static void unswitchNontrivialInvariants(2236 Loop &L, Instruction &TI, ArrayRef<Value *> Invariants,2237 IVConditionInfo &PartialIVInfo, DominatorTree &DT, LoopInfo &LI,2238 AssumptionCache &AC, ScalarEvolution *SE, MemorySSAUpdater *MSSAU,2239 LPMUpdater &LoopUpdater, bool InsertFreeze, bool InjectedCondition) {2240 auto *ParentBB = TI.getParent();2241 BranchInst *BI = dyn_cast<BranchInst>(&TI);2242 SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(&TI);2243 2244 // Save the current loop name in a variable so that we can report it even2245 // after it has been deleted.2246 std::string LoopName(L.getName());2247 2248 // We can only unswitch switches, conditional branches with an invariant2249 // condition, or combining invariant conditions with an instruction or2250 // partially invariant instructions.2251 assert((SI || (BI && BI->isConditional())) &&2252 "Can only unswitch switches and conditional branch!");2253 bool PartiallyInvariant = !PartialIVInfo.InstToDuplicate.empty();2254 bool FullUnswitch =2255 SI || (skipTrivialSelect(BI->getCondition()) == Invariants[0] &&2256 !PartiallyInvariant);2257 if (FullUnswitch)2258 assert(Invariants.size() == 1 &&2259 "Cannot have other invariants with full unswitching!");2260 else2261 assert(isa<Instruction>(skipTrivialSelect(BI->getCondition())) &&2262 "Partial unswitching requires an instruction as the condition!");2263 2264 if (MSSAU && VerifyMemorySSA)2265 MSSAU->getMemorySSA()->verifyMemorySSA();2266 2267 // Constant and BBs tracking the cloned and continuing successor. When we are2268 // unswitching the entire condition, this can just be trivially chosen to2269 // unswitch towards `true`. However, when we are unswitching a set of2270 // invariants combined with `and` or `or` or partially invariant instructions,2271 // the combining operation determines the best direction to unswitch: we want2272 // to unswitch the direction that will collapse the branch.2273 bool Direction = true;2274 int ClonedSucc = 0;2275 if (!FullUnswitch) {2276 Value *Cond = skipTrivialSelect(BI->getCondition());2277 (void)Cond;2278 assert(((match(Cond, m_LogicalAnd()) ^ match(Cond, m_LogicalOr())) ||2279 PartiallyInvariant) &&2280 "Only `or`, `and`, an `select`, partially invariant instructions "2281 "can combine invariants being unswitched.");2282 if (!match(Cond, m_LogicalOr())) {2283 if (match(Cond, m_LogicalAnd()) ||2284 (PartiallyInvariant && !PartialIVInfo.KnownValue->isOneValue())) {2285 Direction = false;2286 ClonedSucc = 1;2287 }2288 }2289 }2290 2291 BasicBlock *RetainedSuccBB =2292 BI ? BI->getSuccessor(1 - ClonedSucc) : SI->getDefaultDest();2293 SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs;2294 if (BI)2295 UnswitchedSuccBBs.insert(BI->getSuccessor(ClonedSucc));2296 else2297 for (auto Case : SI->cases())2298 if (Case.getCaseSuccessor() != RetainedSuccBB)2299 UnswitchedSuccBBs.insert(Case.getCaseSuccessor());2300 2301 assert(!UnswitchedSuccBBs.count(RetainedSuccBB) &&2302 "Should not unswitch the same successor we are retaining!");2303 2304 // The branch should be in this exact loop. Any inner loop's invariant branch2305 // should be handled by unswitching that inner loop. The caller of this2306 // routine should filter out any candidates that remain (but were skipped for2307 // whatever reason).2308 assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!");2309 2310 // Compute the parent loop now before we start hacking on things.2311 Loop *ParentL = L.getParentLoop();2312 // Get blocks in RPO order for MSSA update, before changing the CFG.2313 LoopBlocksRPO LBRPO(&L);2314 if (MSSAU)2315 LBRPO.perform(&LI);2316 2317 // Compute the outer-most loop containing one of our exit blocks. This is the2318 // furthest up our loopnest which can be mutated, which we will use below to2319 // update things.2320 Loop *OuterExitL = &L;2321 SmallVector<BasicBlock *, 4> ExitBlocks;2322 L.getUniqueExitBlocks(ExitBlocks);2323 for (auto *ExitBB : ExitBlocks) {2324 // ExitBB can be an exit block for several levels in the loop nest. Make2325 // sure we find the top most.2326 Loop *NewOuterExitL = getTopMostExitingLoop(ExitBB, LI);2327 if (!NewOuterExitL) {2328 // We exited the entire nest with this block, so we're done.2329 OuterExitL = nullptr;2330 break;2331 }2332 if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(OuterExitL))2333 OuterExitL = NewOuterExitL;2334 }2335 2336 // At this point, we're definitely going to unswitch something so invalidate2337 // any cached information in ScalarEvolution for the outer most loop2338 // containing an exit block and all nested loops.2339 if (SE) {2340 if (OuterExitL)2341 SE->forgetLoop(OuterExitL);2342 else2343 SE->forgetTopmostLoop(&L);2344 SE->forgetBlockAndLoopDispositions();2345 }2346 2347 // If the edge from this terminator to a successor dominates that successor,2348 // store a map from each block in its dominator subtree to it. This lets us2349 // tell when cloning for a particular successor if a block is dominated by2350 // some *other* successor with a single data structure. We use this to2351 // significantly reduce cloning.2352 SmallDenseMap<BasicBlock *, BasicBlock *, 16> DominatingSucc;2353 for (auto *SuccBB : llvm::concat<BasicBlock *const>(ArrayRef(RetainedSuccBB),2354 UnswitchedSuccBBs))2355 if (SuccBB->getUniquePredecessor() ||2356 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {2357 return PredBB == ParentBB || DT.dominates(SuccBB, PredBB);2358 }))2359 visitDomSubTree(DT, SuccBB, [&](BasicBlock *BB) {2360 DominatingSucc[BB] = SuccBB;2361 return true;2362 });2363 2364 // Split the preheader, so that we know that there is a safe place to insert2365 // the conditional branch. We will change the preheader to have a conditional2366 // branch on LoopCond. The original preheader will become the split point2367 // between the unswitched versions, and we will have a new preheader for the2368 // original loop.2369 BasicBlock *SplitBB = L.getLoopPreheader();2370 BasicBlock *LoopPH = SplitEdge(SplitBB, L.getHeader(), &DT, &LI, MSSAU);2371 2372 // Keep track of the dominator tree updates needed.2373 SmallVector<DominatorTree::UpdateType, 4> DTUpdates;2374 2375 // Clone the loop for each unswitched successor.2376 SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps;2377 VMaps.reserve(UnswitchedSuccBBs.size());2378 SmallDenseMap<BasicBlock *, BasicBlock *, 4> ClonedPHs;2379 for (auto *SuccBB : UnswitchedSuccBBs) {2380 VMaps.emplace_back(new ValueToValueMapTy());2381 ClonedPHs[SuccBB] = buildClonedLoopBlocks(2382 L, LoopPH, SplitBB, ExitBlocks, ParentBB, SuccBB, RetainedSuccBB,2383 DominatingSucc, *VMaps.back(), DTUpdates, AC, DT, LI, MSSAU, SE);2384 }2385 2386 // Drop metadata if we may break its semantics by moving this instr into the2387 // split block.2388 if (TI.getMetadata(LLVMContext::MD_make_implicit)) {2389 if (DropNonTrivialImplicitNullChecks)2390 // Do not spend time trying to understand if we can keep it, just drop it2391 // to save compile time.2392 TI.setMetadata(LLVMContext::MD_make_implicit, nullptr);2393 else {2394 // It is only legal to preserve make.implicit metadata if we are2395 // guaranteed no reach implicit null check after following this branch.2396 ICFLoopSafetyInfo SafetyInfo;2397 SafetyInfo.computeLoopSafetyInfo(&L);2398 if (!SafetyInfo.isGuaranteedToExecute(TI, &DT, &L))2399 TI.setMetadata(LLVMContext::MD_make_implicit, nullptr);2400 }2401 }2402 2403 // The stitching of the branched code back together depends on whether we're2404 // doing full unswitching or not with the exception that we always want to2405 // nuke the initial terminator placed in the split block.2406 SplitBB->getTerminator()->eraseFromParent();2407 if (FullUnswitch) {2408 // Keep a clone of the terminator for MSSA updates.2409 Instruction *NewTI = TI.clone();2410 NewTI->insertInto(ParentBB, ParentBB->end());2411 2412 // Splice the terminator from the original loop and rewrite its2413 // successors.2414 TI.moveBefore(*SplitBB, SplitBB->end());2415 TI.dropLocation();2416 2417 // First wire up the moved terminator to the preheaders.2418 if (BI) {2419 BasicBlock *ClonedPH = ClonedPHs.begin()->second;2420 BI->setSuccessor(ClonedSucc, ClonedPH);2421 BI->setSuccessor(1 - ClonedSucc, LoopPH);2422 Value *Cond = skipTrivialSelect(BI->getCondition());2423 if (InsertFreeze) {2424 // We don't give any debug location to the new freeze, because the2425 // BI (`dyn_cast<BranchInst>(TI)`) is an in-loop instruction hoisted2426 // out of the loop.2427 Cond = new FreezeInst(Cond, Cond->getName() + ".fr", BI->getIterator());2428 cast<Instruction>(Cond)->setDebugLoc(DebugLoc::getDropped());2429 }2430 BI->setCondition(Cond);2431 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});2432 } else {2433 assert(SI && "Must either be a branch or switch!");2434 2435 // Walk the cases and directly update their successors.2436 assert(SI->getDefaultDest() == RetainedSuccBB &&2437 "Not retaining default successor!");2438 SI->setDefaultDest(LoopPH);2439 for (const auto &Case : SI->cases())2440 if (Case.getCaseSuccessor() == RetainedSuccBB)2441 Case.setSuccessor(LoopPH);2442 else2443 Case.setSuccessor(ClonedPHs.find(Case.getCaseSuccessor())->second);2444 2445 if (InsertFreeze)2446 SI->setCondition(new FreezeInst(SI->getCondition(),2447 SI->getCondition()->getName() + ".fr",2448 SI->getIterator()));2449 2450 // We need to use the set to populate domtree updates as even when there2451 // are multiple cases pointing at the same successor we only want to2452 // remove and insert one edge in the domtree.2453 for (BasicBlock *SuccBB : UnswitchedSuccBBs)2454 DTUpdates.push_back(2455 {DominatorTree::Insert, SplitBB, ClonedPHs.find(SuccBB)->second});2456 }2457 2458 if (MSSAU) {2459 DT.applyUpdates(DTUpdates);2460 DTUpdates.clear();2461 2462 // Remove all but one edge to the retained block and all unswitched2463 // blocks. This is to avoid having duplicate entries in the cloned Phis,2464 // when we know we only keep a single edge for each case.2465 MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, RetainedSuccBB);2466 for (BasicBlock *SuccBB : UnswitchedSuccBBs)2467 MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, SuccBB);2468 2469 for (auto &VMap : VMaps)2470 MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap,2471 /*IgnoreIncomingWithNoClones=*/true);2472 MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);2473 2474 // Remove all edges to unswitched blocks.2475 for (BasicBlock *SuccBB : UnswitchedSuccBBs)2476 MSSAU->removeEdge(ParentBB, SuccBB);2477 }2478 2479 // Now unhook the successor relationship as we'll be replacing2480 // the terminator with a direct branch. This is much simpler for branches2481 // than switches so we handle those first.2482 if (BI) {2483 // Remove the parent as a predecessor of the unswitched successor.2484 assert(UnswitchedSuccBBs.size() == 1 &&2485 "Only one possible unswitched block for a branch!");2486 BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin();2487 UnswitchedSuccBB->removePredecessor(ParentBB,2488 /*KeepOneInputPHIs*/ true);2489 DTUpdates.push_back({DominatorTree::Delete, ParentBB, UnswitchedSuccBB});2490 } else {2491 // Note that we actually want to remove the parent block as a predecessor2492 // of *every* case successor. The case successor is either unswitched,2493 // completely eliminating an edge from the parent to that successor, or it2494 // is a duplicate edge to the retained successor as the retained successor2495 // is always the default successor and as we'll replace this with a direct2496 // branch we no longer need the duplicate entries in the PHI nodes.2497 SwitchInst *NewSI = cast<SwitchInst>(NewTI);2498 assert(NewSI->getDefaultDest() == RetainedSuccBB &&2499 "Not retaining default successor!");2500 for (const auto &Case : NewSI->cases())2501 Case.getCaseSuccessor()->removePredecessor(2502 ParentBB,2503 /*KeepOneInputPHIs*/ true);2504 2505 // We need to use the set to populate domtree updates as even when there2506 // are multiple cases pointing at the same successor we only want to2507 // remove and insert one edge in the domtree.2508 for (BasicBlock *SuccBB : UnswitchedSuccBBs)2509 DTUpdates.push_back({DominatorTree::Delete, ParentBB, SuccBB});2510 }2511 2512 // Create a new unconditional branch to the continuing block (as opposed to2513 // the one cloned).2514 Instruction *NewBI = BranchInst::Create(RetainedSuccBB, ParentBB);2515 NewBI->setDebugLoc(NewTI->getDebugLoc());2516 2517 // After MSSAU update, remove the cloned terminator instruction NewTI.2518 NewTI->eraseFromParent();2519 } else {2520 assert(BI && "Only branches have partial unswitching.");2521 assert(UnswitchedSuccBBs.size() == 1 &&2522 "Only one possible unswitched block for a branch!");2523 BasicBlock *ClonedPH = ClonedPHs.begin()->second;2524 // When doing a partial unswitch, we have to do a bit more work to build up2525 // the branch in the split block.2526 if (PartiallyInvariant)2527 buildPartialInvariantUnswitchConditionalBranch(2528 *SplitBB, Invariants, Direction, *ClonedPH, *LoopPH, L, MSSAU, *BI);2529 else {2530 buildPartialUnswitchConditionalBranch(2531 *SplitBB, Invariants, Direction, *ClonedPH, *LoopPH,2532 FreezeLoopUnswitchCond, BI, &AC, DT, *BI);2533 }2534 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});2535 2536 if (MSSAU) {2537 DT.applyUpdates(DTUpdates);2538 DTUpdates.clear();2539 2540 // Perform MSSA cloning updates.2541 for (auto &VMap : VMaps)2542 MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap,2543 /*IgnoreIncomingWithNoClones=*/true);2544 MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);2545 }2546 }2547 2548 // Apply the updates accumulated above to get an up-to-date dominator tree.2549 DT.applyUpdates(DTUpdates);2550 2551 // Now that we have an accurate dominator tree, first delete the dead cloned2552 // blocks so that we can accurately build any cloned loops. It is important to2553 // not delete the blocks from the original loop yet because we still want to2554 // reference the original loop to understand the cloned loop's structure.2555 deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT, MSSAU);2556 2557 // Build the cloned loop structure itself. This may be substantially2558 // different from the original structure due to the simplified CFG. This also2559 // handles inserting all the cloned blocks into the correct loops.2560 SmallVector<Loop *, 4> NonChildClonedLoops;2561 for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps)2562 buildClonedLoops(L, ExitBlocks, *VMap, LI, NonChildClonedLoops);2563 2564 // Now that our cloned loops have been built, we can update the original loop.2565 // First we delete the dead blocks from it and then we rebuild the loop2566 // structure taking these deletions into account.2567 deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI, MSSAU, SE, LoopUpdater);2568 2569 if (MSSAU && VerifyMemorySSA)2570 MSSAU->getMemorySSA()->verifyMemorySSA();2571 2572 SmallVector<Loop *, 4> HoistedLoops;2573 bool IsStillLoop =2574 rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops, SE);2575 2576 if (MSSAU && VerifyMemorySSA)2577 MSSAU->getMemorySSA()->verifyMemorySSA();2578 2579 // This transformation has a high risk of corrupting the dominator tree, and2580 // the below steps to rebuild loop structures will result in hard to debug2581 // errors in that case so verify that the dominator tree is sane first.2582 // FIXME: Remove this when the bugs stop showing up and rely on existing2583 // verification steps.2584 assert(DT.verify(DominatorTree::VerificationLevel::Fast));2585 2586 if (BI && !PartiallyInvariant) {2587 // If we unswitched a branch which collapses the condition to a known2588 // constant we want to replace all the uses of the invariants within both2589 // the original and cloned blocks. We do this here so that we can use the2590 // now updated dominator tree to identify which side the users are on.2591 assert(UnswitchedSuccBBs.size() == 1 &&2592 "Only one possible unswitched block for a branch!");2593 BasicBlock *ClonedPH = ClonedPHs.begin()->second;2594 2595 // When considering multiple partially-unswitched invariants2596 // we cant just go replace them with constants in both branches.2597 //2598 // For 'AND' we infer that true branch ("continue") means true2599 // for each invariant operand.2600 // For 'OR' we can infer that false branch ("continue") means false2601 // for each invariant operand.2602 // So it happens that for multiple-partial case we dont replace2603 // in the unswitched branch.2604 bool ReplaceUnswitched =2605 FullUnswitch || (Invariants.size() == 1) || PartiallyInvariant;2606 2607 ConstantInt *UnswitchedReplacement =2608 Direction ? ConstantInt::getTrue(BI->getContext())2609 : ConstantInt::getFalse(BI->getContext());2610 ConstantInt *ContinueReplacement =2611 Direction ? ConstantInt::getFalse(BI->getContext())2612 : ConstantInt::getTrue(BI->getContext());2613 for (Value *Invariant : Invariants) {2614 assert(!isa<Constant>(Invariant) &&2615 "Should not be replacing constant values!");2616 // Use make_early_inc_range here as set invalidates the iterator.2617 for (Use &U : llvm::make_early_inc_range(Invariant->uses())) {2618 Instruction *UserI = dyn_cast<Instruction>(U.getUser());2619 if (!UserI)2620 continue;2621 2622 // Replace it with the 'continue' side if in the main loop body, and the2623 // unswitched if in the cloned blocks.2624 if (DT.dominates(LoopPH, UserI->getParent()))2625 U.set(ContinueReplacement);2626 else if (ReplaceUnswitched &&2627 DT.dominates(ClonedPH, UserI->getParent()))2628 U.set(UnswitchedReplacement);2629 }2630 }2631 }2632 2633 // We can change which blocks are exit blocks of all the cloned sibling2634 // loops, the current loop, and any parent loops which shared exit blocks2635 // with the current loop. As a consequence, we need to re-form LCSSA for2636 // them. But we shouldn't need to re-form LCSSA for any child loops.2637 // FIXME: This could be made more efficient by tracking which exit blocks are2638 // new, and focusing on them, but that isn't likely to be necessary.2639 //2640 // In order to reasonably rebuild LCSSA we need to walk inside-out across the2641 // loop nest and update every loop that could have had its exits changed. We2642 // also need to cover any intervening loops. We add all of these loops to2643 // a list and sort them by loop depth to achieve this without updating2644 // unnecessary loops.2645 auto UpdateLoop = [&](Loop &UpdateL) {2646#ifndef NDEBUG2647 UpdateL.verifyLoop();2648 for (Loop *ChildL : UpdateL) {2649 ChildL->verifyLoop();2650 assert(ChildL->isRecursivelyLCSSAForm(DT, LI) &&2651 "Perturbed a child loop's LCSSA form!");2652 }2653#endif2654 // First build LCSSA for this loop so that we can preserve it when2655 // forming dedicated exits. We don't want to perturb some other loop's2656 // LCSSA while doing that CFG edit.2657 formLCSSA(UpdateL, DT, &LI, SE);2658 2659 // For loops reached by this loop's original exit blocks we may2660 // introduced new, non-dedicated exits. At least try to re-form dedicated2661 // exits for these loops. This may fail if they couldn't have dedicated2662 // exits to start with.2663 formDedicatedExitBlocks(&UpdateL, &DT, &LI, MSSAU, /*PreserveLCSSA*/ true);2664 };2665 2666 // For non-child cloned loops and hoisted loops, we just need to update LCSSA2667 // and we can do it in any order as they don't nest relative to each other.2668 //2669 // Also check if any of the loops we have updated have become top-level loops2670 // as that will necessitate widening the outer loop scope.2671 for (Loop *UpdatedL :2672 llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) {2673 UpdateLoop(*UpdatedL);2674 if (UpdatedL->isOutermost())2675 OuterExitL = nullptr;2676 }2677 if (IsStillLoop) {2678 UpdateLoop(L);2679 if (L.isOutermost())2680 OuterExitL = nullptr;2681 }2682 2683 // If the original loop had exit blocks, walk up through the outer most loop2684 // of those exit blocks to update LCSSA and form updated dedicated exits.2685 if (OuterExitL != &L)2686 for (Loop *OuterL = ParentL; OuterL != OuterExitL;2687 OuterL = OuterL->getParentLoop())2688 UpdateLoop(*OuterL);2689 2690#ifndef NDEBUG2691 // Verify the entire loop structure to catch any incorrect updates before we2692 // progress in the pass pipeline.2693 LI.verify(DT);2694#endif2695 2696 // Now that we've unswitched something, make callbacks to report the changes.2697 // For that we need to merge together the updated loops and the cloned loops2698 // and check whether the original loop survived.2699 SmallVector<Loop *, 4> SibLoops;2700 for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops))2701 if (UpdatedL->getParentLoop() == ParentL)2702 SibLoops.push_back(UpdatedL);2703 postUnswitch(L, LoopUpdater, LoopName, IsStillLoop, PartiallyInvariant,2704 InjectedCondition, SibLoops);2705 2706 if (MSSAU && VerifyMemorySSA)2707 MSSAU->getMemorySSA()->verifyMemorySSA();2708 2709 if (BI)2710 ++NumBranches;2711 else2712 ++NumSwitches;2713}2714 2715/// Recursively compute the cost of a dominator subtree based on the per-block2716/// cost map provided.2717///2718/// The recursive computation is memozied into the provided DT-indexed cost map2719/// to allow querying it for most nodes in the domtree without it becoming2720/// quadratic.2721static InstructionCost computeDomSubtreeCost(2722 DomTreeNode &N,2723 const SmallDenseMap<BasicBlock *, InstructionCost, 4> &BBCostMap,2724 SmallDenseMap<DomTreeNode *, InstructionCost, 4> &DTCostMap) {2725 // Don't accumulate cost (or recurse through) blocks not in our block cost2726 // map and thus not part of the duplication cost being considered.2727 auto BBCostIt = BBCostMap.find(N.getBlock());2728 if (BBCostIt == BBCostMap.end())2729 return 0;2730 2731 // Lookup this node to see if we already computed its cost.2732 auto DTCostIt = DTCostMap.find(&N);2733 if (DTCostIt != DTCostMap.end())2734 return DTCostIt->second;2735 2736 // If not, we have to compute it. We can't use insert above and update2737 // because computing the cost may insert more things into the map.2738 InstructionCost Cost = std::accumulate(2739 N.begin(), N.end(), BBCostIt->second,2740 [&](InstructionCost Sum, DomTreeNode *ChildN) -> InstructionCost {2741 return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap);2742 });2743 bool Inserted = DTCostMap.insert({&N, Cost}).second;2744 (void)Inserted;2745 assert(Inserted && "Should not insert a node while visiting children!");2746 return Cost;2747}2748 2749/// Turns a select instruction into implicit control flow branch,2750/// making the following replacement:2751///2752/// head:2753/// --code before select--2754/// select %cond, %trueval, %falseval2755/// --code after select--2756///2757/// into2758///2759/// head:2760/// --code before select--2761/// br i1 %cond, label %then, label %tail2762///2763/// then:2764/// br %tail2765///2766/// tail:2767/// phi [ %trueval, %then ], [ %falseval, %head]2768/// unreachable2769///2770/// It also makes all relevant DT and LI updates, so that all structures are in2771/// valid state after this transform.2772static BranchInst *turnSelectIntoBranch(SelectInst *SI, DominatorTree &DT,2773 LoopInfo &LI, MemorySSAUpdater *MSSAU,2774 AssumptionCache *AC) {2775 LLVM_DEBUG(dbgs() << "Turning " << *SI << " into a branch.\n");2776 BasicBlock *HeadBB = SI->getParent();2777 2778 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);2779 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false,2780 SI->getMetadata(LLVMContext::MD_prof), &DTU, &LI);2781 auto *CondBr = cast<BranchInst>(HeadBB->getTerminator());2782 BasicBlock *ThenBB = CondBr->getSuccessor(0),2783 *TailBB = CondBr->getSuccessor(1);2784 if (MSSAU)2785 MSSAU->moveAllAfterSpliceBlocks(HeadBB, TailBB, SI);2786 2787 PHINode *Phi =2788 PHINode::Create(SI->getType(), 2, "unswitched.select", SI->getIterator());2789 Phi->addIncoming(SI->getTrueValue(), ThenBB);2790 Phi->addIncoming(SI->getFalseValue(), HeadBB);2791 Phi->setDebugLoc(SI->getDebugLoc());2792 SI->replaceAllUsesWith(Phi);2793 SI->eraseFromParent();2794 2795 if (MSSAU && VerifyMemorySSA)2796 MSSAU->getMemorySSA()->verifyMemorySSA();2797 2798 ++NumSelects;2799 return CondBr;2800}2801 2802/// Turns a llvm.experimental.guard intrinsic into implicit control flow branch,2803/// making the following replacement:2804///2805/// --code before guard--2806/// call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]2807/// --code after guard--2808///2809/// into2810///2811/// --code before guard--2812/// br i1 %cond, label %guarded, label %deopt2813///2814/// guarded:2815/// --code after guard--2816///2817/// deopt:2818/// call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]2819/// unreachable2820///2821/// It also makes all relevant DT and LI updates, so that all structures are in2822/// valid state after this transform.2823static BranchInst *turnGuardIntoBranch(IntrinsicInst *GI, Loop &L,2824 DominatorTree &DT, LoopInfo &LI,2825 MemorySSAUpdater *MSSAU) {2826 LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n");2827 BasicBlock *CheckBB = GI->getParent();2828 2829 if (MSSAU && VerifyMemorySSA)2830 MSSAU->getMemorySSA()->verifyMemorySSA();2831 2832 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);2833 // llvm.experimental.guard doesn't have branch weights. We can assume,2834 // however, that the deopt path is unlikely.2835 Instruction *DeoptBlockTerm = SplitBlockAndInsertIfThen(2836 GI->getArgOperand(0), GI, true,2837 !ProfcheckDisableMetadataFixes && EstimateProfile2838 ? MDBuilder(GI->getContext()).createUnlikelyBranchWeights()2839 : nullptr,2840 &DTU, &LI);2841 BranchInst *CheckBI = cast<BranchInst>(CheckBB->getTerminator());2842 // SplitBlockAndInsertIfThen inserts control flow that branches to2843 // DeoptBlockTerm if the condition is true. We want the opposite.2844 CheckBI->swapSuccessors();2845 2846 BasicBlock *GuardedBlock = CheckBI->getSuccessor(0);2847 GuardedBlock->setName("guarded");2848 CheckBI->getSuccessor(1)->setName("deopt");2849 BasicBlock *DeoptBlock = CheckBI->getSuccessor(1);2850 2851 if (MSSAU)2852 MSSAU->moveAllAfterSpliceBlocks(CheckBB, GuardedBlock, GI);2853 2854 GI->moveBefore(DeoptBlockTerm->getIterator());2855 GI->setArgOperand(0, ConstantInt::getFalse(GI->getContext()));2856 2857 if (MSSAU) {2858 MemoryDef *MD = cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(GI));2859 MSSAU->moveToPlace(MD, DeoptBlock, MemorySSA::BeforeTerminator);2860 if (VerifyMemorySSA)2861 MSSAU->getMemorySSA()->verifyMemorySSA();2862 }2863 2864 if (VerifyLoopInfo)2865 LI.verify(DT);2866 ++NumGuards;2867 return CheckBI;2868}2869 2870/// Cost multiplier is a way to limit potentially exponential behavior2871/// of loop-unswitch. Cost is multiplied in proportion of 2^number of unswitch2872/// candidates available. Also consider the number of "sibling" loops with2873/// the idea of accounting for previous unswitches that already happened on this2874/// cluster of loops. There was an attempt to keep this formula simple,2875/// just enough to limit the worst case behavior. Even if it is not that simple2876/// now it is still not an attempt to provide a detailed heuristic size2877/// prediction.2878///2879/// TODO: Make a proper accounting of "explosion" effect for all kinds of2880/// unswitch candidates, making adequate predictions instead of wild guesses.2881/// That requires knowing not just the number of "remaining" candidates but2882/// also costs of unswitching for each of these candidates.2883static int CalculateUnswitchCostMultiplier(2884 const Instruction &TI, const Loop &L, const LoopInfo &LI,2885 const DominatorTree &DT,2886 ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates) {2887 2888 // Guards and other exiting conditions do not contribute to exponential2889 // explosion as soon as they dominate the latch (otherwise there might be2890 // another path to the latch remaining that does not allow to eliminate the2891 // loop copy on unswitch).2892 const BasicBlock *Latch = L.getLoopLatch();2893 const BasicBlock *CondBlock = TI.getParent();2894 if (DT.dominates(CondBlock, Latch) &&2895 (isGuard(&TI) ||2896 (TI.isTerminator() &&2897 llvm::count_if(successors(&TI), [&L](const BasicBlock *SuccBB) {2898 return L.contains(SuccBB);2899 }) <= 1))) {2900 NumCostMultiplierSkipped++;2901 return 1;2902 }2903 2904 // Each invariant non-trivial condition, after being unswitched, is supposed2905 // to have its own specialized sibling loop (the invariant condition has been2906 // hoisted out of the child loop into a newly-cloned loop). When unswitching2907 // conditions in nested loops, the basic block size of the outer loop should2908 // not be altered. If such a size significantly increases across unswitching2909 // invocations, something may be wrong; so adjust the final cost taking this2910 // into account.2911 auto *ParentL = L.getParentLoop();2912 int ParentLoopSizeMultiplier = 1;2913 if (ParentL)2914 ParentLoopSizeMultiplier =2915 std::max<int>(ParentL->getNumBlocks() / UnswitchParentBlocksDiv, 1);2916 2917 int SiblingsCount =2918 (ParentL ? ParentL->getSubLoopsVector().size() : llvm::size(LI));2919 // Count amount of clones that all the candidates might cause during2920 // unswitching. Branch/guard/select counts as 1, switch counts as log2 of its2921 // cases.2922 int UnswitchedClones = 0;2923 for (const auto &Candidate : UnswitchCandidates) {2924 const Instruction *CI = Candidate.TI;2925 const BasicBlock *CondBlock = CI->getParent();2926 bool SkipExitingSuccessors = DT.dominates(CondBlock, Latch);2927 if (isa<SelectInst>(CI)) {2928 UnswitchedClones++;2929 continue;2930 }2931 if (isGuard(CI)) {2932 if (!SkipExitingSuccessors)2933 UnswitchedClones++;2934 continue;2935 }2936 int NonExitingSuccessors =2937 llvm::count_if(successors(CondBlock),2938 [SkipExitingSuccessors, &L](const BasicBlock *SuccBB) {2939 return !SkipExitingSuccessors || L.contains(SuccBB);2940 });2941 UnswitchedClones += Log2_32(NonExitingSuccessors);2942 }2943 2944 // Ignore up to the "unscaled candidates" number of unswitch candidates2945 // when calculating the power-of-two scaling of the cost. The main idea2946 // with this control is to allow a small number of unswitches to happen2947 // and rely more on siblings multiplier (see below) when the number2948 // of candidates is small.2949 unsigned ClonesPower =2950 std::max(UnswitchedClones - (int)UnswitchNumInitialUnscaledCandidates, 0);2951 2952 // Allowing top-level loops to spread a bit more than nested ones.2953 int SiblingsMultiplier =2954 std::max((ParentL ? SiblingsCount2955 : SiblingsCount / (int)UnswitchSiblingsToplevelDiv),2956 1);2957 // Compute the cost multiplier in a way that won't overflow by saturating2958 // at an upper bound.2959 int CostMultiplier;2960 if (ClonesPower > Log2_32(UnswitchThreshold) ||2961 SiblingsMultiplier > UnswitchThreshold ||2962 ParentLoopSizeMultiplier > UnswitchThreshold)2963 CostMultiplier = UnswitchThreshold;2964 else2965 CostMultiplier = std::min(SiblingsMultiplier * (1 << ClonesPower),2966 (int)UnswitchThreshold);2967 2968 LLVM_DEBUG(dbgs() << " Computed multiplier " << CostMultiplier2969 << " (siblings " << SiblingsMultiplier << " * parent size "2970 << ParentLoopSizeMultiplier << " * clones "2971 << (1 << ClonesPower) << ")"2972 << " for unswitch candidate: " << TI << "\n");2973 return CostMultiplier;2974}2975 2976static bool collectUnswitchCandidates(2977 SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates,2978 IVConditionInfo &PartialIVInfo, Instruction *&PartialIVCondBranch,2979 const Loop &L, const LoopInfo &LI, AAResults &AA,2980 const MemorySSAUpdater *MSSAU) {2981 assert(UnswitchCandidates.empty() && "Should be!");2982 2983 auto AddUnswitchCandidatesForInst = [&](Instruction *I, Value *Cond) {2984 Cond = skipTrivialSelect(Cond);2985 if (isa<Constant>(Cond))2986 return;2987 if (L.isLoopInvariant(Cond)) {2988 UnswitchCandidates.push_back({I, {Cond}});2989 return;2990 }2991 if (match(Cond, m_CombineOr(m_LogicalAnd(), m_LogicalOr()))) {2992 TinyPtrVector<Value *> Invariants =2993 collectHomogenousInstGraphLoopInvariants(2994 L, *static_cast<Instruction *>(Cond), LI);2995 if (!Invariants.empty())2996 UnswitchCandidates.push_back({I, std::move(Invariants)});2997 }2998 };2999 3000 // Whether or not we should also collect guards in the loop.3001 bool CollectGuards = false;3002 if (UnswitchGuards) {3003 auto *GuardDecl = Intrinsic::getDeclarationIfExists(3004 L.getHeader()->getParent()->getParent(), Intrinsic::experimental_guard);3005 if (GuardDecl && !GuardDecl->use_empty())3006 CollectGuards = true;3007 }3008 3009 for (auto *BB : L.blocks()) {3010 if (LI.getLoopFor(BB) != &L)3011 continue;3012 3013 for (auto &I : *BB) {3014 if (auto *SI = dyn_cast<SelectInst>(&I)) {3015 auto *Cond = SI->getCondition();3016 // Do not unswitch vector selects and logical and/or selects3017 if (Cond->getType()->isIntegerTy(1) && !SI->getType()->isIntegerTy(1))3018 AddUnswitchCandidatesForInst(SI, Cond);3019 } else if (CollectGuards && isGuard(&I)) {3020 auto *Cond =3021 skipTrivialSelect(cast<IntrinsicInst>(&I)->getArgOperand(0));3022 // TODO: Support AND, OR conditions and partial unswitching.3023 if (!isa<Constant>(Cond) && L.isLoopInvariant(Cond))3024 UnswitchCandidates.push_back({&I, {Cond}});3025 }3026 }3027 3028 if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {3029 // We can only consider fully loop-invariant switch conditions as we need3030 // to completely eliminate the switch after unswitching.3031 if (!isa<Constant>(SI->getCondition()) &&3032 L.isLoopInvariant(SI->getCondition()) && !BB->getUniqueSuccessor())3033 UnswitchCandidates.push_back({SI, {SI->getCondition()}});3034 continue;3035 }3036 3037 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());3038 if (!BI || !BI->isConditional() ||3039 BI->getSuccessor(0) == BI->getSuccessor(1))3040 continue;3041 3042 AddUnswitchCandidatesForInst(BI, BI->getCondition());3043 }3044 3045 if (MSSAU && !findOptionMDForLoop(&L, "llvm.loop.unswitch.partial.disable") &&3046 !any_of(UnswitchCandidates, [&L](auto &TerminatorAndInvariants) {3047 return TerminatorAndInvariants.TI == L.getHeader()->getTerminator();3048 })) {3049 MemorySSA *MSSA = MSSAU->getMemorySSA();3050 if (auto Info = hasPartialIVCondition(L, MSSAThreshold, *MSSA, AA)) {3051 LLVM_DEBUG(3052 dbgs() << "simple-loop-unswitch: Found partially invariant condition "3053 << *Info->InstToDuplicate[0] << "\n");3054 PartialIVInfo = *Info;3055 PartialIVCondBranch = L.getHeader()->getTerminator();3056 TinyPtrVector<Value *> ValsToDuplicate;3057 llvm::append_range(ValsToDuplicate, Info->InstToDuplicate);3058 UnswitchCandidates.push_back(3059 {L.getHeader()->getTerminator(), std::move(ValsToDuplicate)});3060 }3061 }3062 return !UnswitchCandidates.empty();3063}3064 3065/// Tries to canonicalize condition described by:3066///3067/// br (LHS pred RHS), label IfTrue, label IfFalse3068///3069/// into its equivalent where `Pred` is something that we support for injected3070/// invariants (so far it is limited to ult), LHS in canonicalized form is3071/// non-invariant and RHS is an invariant.3072static void canonicalizeForInvariantConditionInjection(CmpPredicate &Pred,3073 Value *&LHS, Value *&RHS,3074 BasicBlock *&IfTrue,3075 BasicBlock *&IfFalse,3076 const Loop &L) {3077 if (!L.contains(IfTrue)) {3078 Pred = ICmpInst::getInversePredicate(Pred);3079 std::swap(IfTrue, IfFalse);3080 }3081 3082 // Move loop-invariant argument to RHS position.3083 if (L.isLoopInvariant(LHS)) {3084 Pred = ICmpInst::getSwappedPredicate(Pred);3085 std::swap(LHS, RHS);3086 }3087 3088 if (Pred == ICmpInst::ICMP_SGE && match(RHS, m_Zero())) {3089 // Turn "x >=s 0" into "x <u UMIN_INT"3090 Pred = ICmpInst::ICMP_ULT;3091 RHS = ConstantInt::get(3092 RHS->getContext(),3093 APInt::getSignedMinValue(RHS->getType()->getIntegerBitWidth()));3094 }3095}3096 3097/// Returns true, if predicate described by ( \p Pred, \p LHS, \p RHS )3098/// succeeding into blocks ( \p IfTrue, \p IfFalse) can be optimized by3099/// injecting a loop-invariant condition.3100static bool shouldTryInjectInvariantCondition(3101 const ICmpInst::Predicate Pred, const Value *LHS, const Value *RHS,3102 const BasicBlock *IfTrue, const BasicBlock *IfFalse, const Loop &L) {3103 if (L.isLoopInvariant(LHS) || !L.isLoopInvariant(RHS))3104 return false;3105 // TODO: Support other predicates.3106 if (Pred != ICmpInst::ICMP_ULT)3107 return false;3108 // TODO: Support non-loop-exiting branches?3109 if (!L.contains(IfTrue) || L.contains(IfFalse))3110 return false;3111 // FIXME: For some reason this causes problems with MSSA updates, need to3112 // investigate why. So far, just don't unswitch latch.3113 if (L.getHeader() == IfTrue)3114 return false;3115 return true;3116}3117 3118/// Returns true, if metadata on \p BI allows us to optimize branching into \p3119/// TakenSucc via injection of invariant conditions. The branch should be not3120/// enough and not previously unswitched, the information about this comes from3121/// the metadata.3122bool shouldTryInjectBasingOnMetadata(const BranchInst *BI,3123 const BasicBlock *TakenSucc) {3124 SmallVector<uint32_t> Weights;3125 if (!extractBranchWeights(*BI, Weights))3126 return false;3127 unsigned T = InjectInvariantConditionHotnesThreshold;3128 BranchProbability LikelyTaken(T - 1, T);3129 3130 assert(Weights.size() == 2 && "Unexpected profile data!");3131 size_t Idx = BI->getSuccessor(0) == TakenSucc ? 0 : 1;3132 auto Num = Weights[Idx];3133 auto Denom = Weights[0] + Weights[1];3134 // Degenerate or overflowed metadata.3135 if (Denom == 0 || Num > Denom)3136 return false;3137 BranchProbability ActualTaken(Num, Denom);3138 if (LikelyTaken > ActualTaken)3139 return false;3140 return true;3141}3142 3143/// Materialize pending invariant condition of the given candidate into IR. The3144/// injected loop-invariant condition implies the original loop-variant branch3145/// condition, so the materialization turns3146///3147/// loop_block:3148/// ...3149/// br i1 %variant_cond, label InLoopSucc, label OutOfLoopSucc3150///3151/// into3152///3153/// preheader:3154/// %invariant_cond = LHS pred RHS3155/// ...3156/// loop_block:3157/// br i1 %invariant_cond, label InLoopSucc, label OriginalCheck3158/// OriginalCheck:3159/// br i1 %variant_cond, label InLoopSucc, label OutOfLoopSucc3160/// ...3161static NonTrivialUnswitchCandidate3162injectPendingInvariantConditions(NonTrivialUnswitchCandidate Candidate, Loop &L,3163 DominatorTree &DT, LoopInfo &LI,3164 AssumptionCache &AC, MemorySSAUpdater *MSSAU) {3165 assert(Candidate.hasPendingInjection() && "Nothing to inject!");3166 BasicBlock *Preheader = L.getLoopPreheader();3167 assert(Preheader && "Loop is not in simplified form?");3168 assert(LI.getLoopFor(Candidate.TI->getParent()) == &L &&3169 "Unswitching branch of inner loop!");3170 3171 auto Pred = Candidate.PendingInjection->Pred;3172 auto *LHS = Candidate.PendingInjection->LHS;3173 auto *RHS = Candidate.PendingInjection->RHS;3174 auto *InLoopSucc = Candidate.PendingInjection->InLoopSucc;3175 auto *TI = cast<BranchInst>(Candidate.TI);3176 auto *BB = Candidate.TI->getParent();3177 auto *OutOfLoopSucc = InLoopSucc == TI->getSuccessor(0) ? TI->getSuccessor(1)3178 : TI->getSuccessor(0);3179 // FIXME: Remove this once limitation on successors is lifted.3180 assert(L.contains(InLoopSucc) && "Not supported yet!");3181 assert(!L.contains(OutOfLoopSucc) && "Not supported yet!");3182 auto &Ctx = BB->getContext();3183 3184 IRBuilder<> Builder(Preheader->getTerminator());3185 assert(ICmpInst::isUnsigned(Pred) && "Not supported yet!");3186 if (LHS->getType() != RHS->getType()) {3187 if (LHS->getType()->getIntegerBitWidth() <3188 RHS->getType()->getIntegerBitWidth())3189 LHS = Builder.CreateZExt(LHS, RHS->getType(), LHS->getName() + ".wide");3190 else3191 RHS = Builder.CreateZExt(RHS, LHS->getType(), RHS->getName() + ".wide");3192 }3193 // Do not use builder here: CreateICmp may simplify this into a constant and3194 // unswitching will break. Better optimize it away later.3195 auto *InjectedCond =3196 ICmpInst::Create(Instruction::ICmp, Pred, LHS, RHS, "injected.cond",3197 Preheader->getTerminator()->getIterator());3198 3199 BasicBlock *CheckBlock = BasicBlock::Create(Ctx, BB->getName() + ".check",3200 BB->getParent(), InLoopSucc);3201 Builder.SetInsertPoint(TI);3202 auto *InvariantBr =3203 Builder.CreateCondBr(InjectedCond, InLoopSucc, CheckBlock);3204 // We don't know anything about the relation between the limits.3205 setExplicitlyUnknownBranchWeightsIfProfiled(*InvariantBr, DEBUG_TYPE);3206 3207 Builder.SetInsertPoint(CheckBlock);3208 Builder.CreateCondBr(3209 TI->getCondition(), TI->getSuccessor(0), TI->getSuccessor(1),3210 !ProfcheckDisableMetadataFixes ? TI->getMetadata(LLVMContext::MD_prof)3211 : nullptr);3212 TI->eraseFromParent();3213 3214 // Fixup phis.3215 for (auto &I : *InLoopSucc) {3216 auto *PN = dyn_cast<PHINode>(&I);3217 if (!PN)3218 break;3219 auto *Inc = PN->getIncomingValueForBlock(BB);3220 PN->addIncoming(Inc, CheckBlock);3221 }3222 OutOfLoopSucc->replacePhiUsesWith(BB, CheckBlock);3223 3224 SmallVector<DominatorTree::UpdateType, 4> DTUpdates = {3225 { DominatorTree::Insert, BB, CheckBlock },3226 { DominatorTree::Insert, CheckBlock, InLoopSucc },3227 { DominatorTree::Insert, CheckBlock, OutOfLoopSucc },3228 { DominatorTree::Delete, BB, OutOfLoopSucc }3229 };3230 3231 DT.applyUpdates(DTUpdates);3232 if (MSSAU)3233 MSSAU->applyUpdates(DTUpdates, DT);3234 L.addBasicBlockToLoop(CheckBlock, LI);3235 3236#ifndef NDEBUG3237 DT.verify();3238 LI.verify(DT);3239 if (MSSAU && VerifyMemorySSA)3240 MSSAU->getMemorySSA()->verifyMemorySSA();3241#endif3242 3243 // TODO: In fact, cost of unswitching a new invariant candidate is *slightly*3244 // higher because we have just inserted a new block. Need to think how to3245 // adjust the cost of injected candidates when it was first computed.3246 LLVM_DEBUG(dbgs() << "Injected a new loop-invariant branch " << *InvariantBr3247 << " and considering it for unswitching.");3248 ++NumInvariantConditionsInjected;3249 return NonTrivialUnswitchCandidate(InvariantBr, { InjectedCond },3250 Candidate.Cost);3251}3252 3253/// Given chain of loop branch conditions looking like:3254/// br (Variant < Invariant1)3255/// br (Variant < Invariant2)3256/// br (Variant < Invariant3)3257/// ...3258/// collect set of invariant conditions on which we want to unswitch, which3259/// look like:3260/// Invariant1 <= Invariant23261/// Invariant2 <= Invariant33262/// ...3263/// Though they might not immediately exist in the IR, we can still inject them.3264static bool insertCandidatesWithPendingInjections(3265 SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates, Loop &L,3266 ICmpInst::Predicate Pred, ArrayRef<CompareDesc> Compares,3267 const DominatorTree &DT) {3268 3269 assert(ICmpInst::isRelational(Pred));3270 assert(ICmpInst::isStrictPredicate(Pred));3271 if (Compares.size() < 2)3272 return false;3273 ICmpInst::Predicate NonStrictPred = ICmpInst::getNonStrictPredicate(Pred);3274 for (auto Prev = Compares.begin(), Next = Compares.begin() + 1;3275 Next != Compares.end(); ++Prev, ++Next) {3276 Value *LHS = Next->Invariant;3277 Value *RHS = Prev->Invariant;3278 BasicBlock *InLoopSucc = Prev->InLoopSucc;3279 InjectedInvariant ToInject(NonStrictPred, LHS, RHS, InLoopSucc);3280 NonTrivialUnswitchCandidate Candidate(Prev->Term, { LHS, RHS },3281 std::nullopt, std::move(ToInject));3282 UnswitchCandidates.push_back(std::move(Candidate));3283 }3284 return true;3285}3286 3287/// Collect unswitch candidates by invariant conditions that are not immediately3288/// present in the loop. However, they can be injected into the code if we3289/// decide it's profitable.3290/// An example of such conditions is following:3291///3292/// for (...) {3293/// x = load ...3294/// if (! x <u C1) break;3295/// if (! x <u C2) break;3296/// <do something>3297/// }3298///3299/// We can unswitch by condition "C1 <=u C2". If that is true, then "x <u C1 <=3300/// C2" automatically implies "x <u C2", so we can get rid of one of3301/// loop-variant checks in unswitched loop version.3302static bool collectUnswitchCandidatesWithInjections(3303 SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates,3304 IVConditionInfo &PartialIVInfo, Instruction *&PartialIVCondBranch, Loop &L,3305 const DominatorTree &DT, const LoopInfo &LI, AAResults &AA,3306 const MemorySSAUpdater *MSSAU) {3307 if (!InjectInvariantConditions)3308 return false;3309 3310 if (!DT.isReachableFromEntry(L.getHeader()))3311 return false;3312 auto *Latch = L.getLoopLatch();3313 // Need to have a single latch and a preheader.3314 if (!Latch)3315 return false;3316 assert(L.getLoopPreheader() && "Must have a preheader!");3317 3318 DenseMap<Value *, SmallVector<CompareDesc, 4> > CandidatesULT;3319 // Traverse the conditions that dominate latch (and therefore dominate each3320 // other).3321 for (auto *DTN = DT.getNode(Latch); L.contains(DTN->getBlock());3322 DTN = DTN->getIDom()) {3323 CmpPredicate Pred;3324 Value *LHS = nullptr, *RHS = nullptr;3325 BasicBlock *IfTrue = nullptr, *IfFalse = nullptr;3326 auto *BB = DTN->getBlock();3327 // Ignore inner loops.3328 if (LI.getLoopFor(BB) != &L)3329 continue;3330 auto *Term = BB->getTerminator();3331 if (!match(Term, m_Br(m_ICmp(Pred, m_Value(LHS), m_Value(RHS)),3332 m_BasicBlock(IfTrue), m_BasicBlock(IfFalse))))3333 continue;3334 if (!LHS->getType()->isIntegerTy())3335 continue;3336 canonicalizeForInvariantConditionInjection(Pred, LHS, RHS, IfTrue, IfFalse,3337 L);3338 if (!shouldTryInjectInvariantCondition(Pred, LHS, RHS, IfTrue, IfFalse, L))3339 continue;3340 if (!shouldTryInjectBasingOnMetadata(cast<BranchInst>(Term), IfTrue))3341 continue;3342 // Strip ZEXT for unsigned predicate.3343 // TODO: once signed predicates are supported, also strip SEXT.3344 CompareDesc Desc(cast<BranchInst>(Term), RHS, IfTrue);3345 while (auto *Zext = dyn_cast<ZExtInst>(LHS))3346 LHS = Zext->getOperand(0);3347 CandidatesULT[LHS].push_back(Desc);3348 }3349 3350 bool Found = false;3351 for (auto &It : CandidatesULT)3352 Found |= insertCandidatesWithPendingInjections(3353 UnswitchCandidates, L, ICmpInst::ICMP_ULT, It.second, DT);3354 return Found;3355}3356 3357static bool isSafeForNoNTrivialUnswitching(Loop &L, LoopInfo &LI) {3358 if (!L.isSafeToClone())3359 return false;3360 for (auto *BB : L.blocks())3361 for (auto &I : *BB) {3362 if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB))3363 return false;3364 if (auto *CB = dyn_cast<CallBase>(&I)) {3365 assert(!CB->cannotDuplicate() && "Checked by L.isSafeToClone().");3366 if (CB->isConvergent())3367 return false;3368 }3369 }3370 3371 // Check if there are irreducible CFG cycles in this loop. If so, we cannot3372 // easily unswitch non-trivial edges out of the loop. Doing so might turn the3373 // irreducible control flow into reducible control flow and introduce new3374 // loops "out of thin air". If we ever discover important use cases for doing3375 // this, we can add support to loop unswitch, but it is a lot of complexity3376 // for what seems little or no real world benefit.3377 LoopBlocksRPO RPOT(&L);3378 RPOT.perform(&LI);3379 if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI))3380 return false;3381 3382 SmallVector<BasicBlock *, 4> ExitBlocks;3383 L.getUniqueExitBlocks(ExitBlocks);3384 // We cannot unswitch if exit blocks contain a cleanuppad/catchswitch3385 // instruction as we don't know how to split those exit blocks.3386 // FIXME: We should teach SplitBlock to handle this and remove this3387 // restriction.3388 for (auto *ExitBB : ExitBlocks) {3389 auto It = ExitBB->getFirstNonPHIIt();3390 if (isa<CleanupPadInst>(It) || isa<CatchSwitchInst>(It)) {3391 LLVM_DEBUG(dbgs() << "Cannot unswitch because of cleanuppad/catchswitch "3392 "in exit block\n");3393 return false;3394 }3395 }3396 3397 return true;3398}3399 3400static NonTrivialUnswitchCandidate findBestNonTrivialUnswitchCandidate(3401 ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates, const Loop &L,3402 const DominatorTree &DT, const LoopInfo &LI, AssumptionCache &AC,3403 const TargetTransformInfo &TTI, const IVConditionInfo &PartialIVInfo) {3404 // Given that unswitching these terminators will require duplicating parts of3405 // the loop, so we need to be able to model that cost. Compute the ephemeral3406 // values and set up a data structure to hold per-BB costs. We cache each3407 // block's cost so that we don't recompute this when considering different3408 // subsets of the loop for duplication during unswitching.3409 SmallPtrSet<const Value *, 4> EphValues;3410 CodeMetrics::collectEphemeralValues(&L, &AC, EphValues);3411 SmallDenseMap<BasicBlock *, InstructionCost, 4> BBCostMap;3412 3413 // Compute the cost of each block, as well as the total loop cost. Also, bail3414 // out if we see instructions which are incompatible with loop unswitching3415 // (convergent, noduplicate, or cross-basic-block tokens).3416 // FIXME: We might be able to safely handle some of these in non-duplicated3417 // regions.3418 TargetTransformInfo::TargetCostKind CostKind =3419 L.getHeader()->getParent()->hasMinSize()3420 ? TargetTransformInfo::TCK_CodeSize3421 : TargetTransformInfo::TCK_SizeAndLatency;3422 InstructionCost LoopCost = 0;3423 for (auto *BB : L.blocks()) {3424 InstructionCost Cost = 0;3425 for (auto &I : *BB) {3426 if (EphValues.count(&I))3427 continue;3428 Cost += TTI.getInstructionCost(&I, CostKind);3429 }3430 assert(Cost >= 0 && "Must not have negative costs!");3431 LoopCost += Cost;3432 assert(LoopCost >= 0 && "Must not have negative loop costs!");3433 BBCostMap[BB] = Cost;3434 }3435 LLVM_DEBUG(dbgs() << " Total loop cost: " << LoopCost << "\n");3436 3437 // Now we find the best candidate by searching for the one with the following3438 // properties in order:3439 //3440 // 1) An unswitching cost below the threshold3441 // 2) The smallest number of duplicated unswitch candidates (to avoid3442 // creating redundant subsequent unswitching)3443 // 3) The smallest cost after unswitching.3444 //3445 // We prioritize reducing fanout of unswitch candidates provided the cost3446 // remains below the threshold because this has a multiplicative effect.3447 //3448 // This requires memoizing each dominator subtree to avoid redundant work.3449 //3450 // FIXME: Need to actually do the number of candidates part above.3451 SmallDenseMap<DomTreeNode *, InstructionCost, 4> DTCostMap;3452 // Given a terminator which might be unswitched, computes the non-duplicated3453 // cost for that terminator.3454 auto ComputeUnswitchedCost = [&](Instruction &TI,3455 bool FullUnswitch) -> InstructionCost {3456 // Unswitching selects unswitches the entire loop.3457 if (isa<SelectInst>(TI))3458 return LoopCost;3459 3460 BasicBlock &BB = *TI.getParent();3461 SmallPtrSet<BasicBlock *, 4> Visited;3462 3463 InstructionCost Cost = 0;3464 for (BasicBlock *SuccBB : successors(&BB)) {3465 // Don't count successors more than once.3466 if (!Visited.insert(SuccBB).second)3467 continue;3468 3469 // If this is a partial unswitch candidate, then it must be a conditional3470 // branch with a condition of either `or`, `and`, their corresponding3471 // select forms or partially invariant instructions. In that case, one of3472 // the successors is necessarily duplicated, so don't even try to remove3473 // its cost.3474 if (!FullUnswitch) {3475 auto &BI = cast<BranchInst>(TI);3476 Value *Cond = skipTrivialSelect(BI.getCondition());3477 if (match(Cond, m_LogicalAnd())) {3478 if (SuccBB == BI.getSuccessor(1))3479 continue;3480 } else if (match(Cond, m_LogicalOr())) {3481 if (SuccBB == BI.getSuccessor(0))3482 continue;3483 } else if ((PartialIVInfo.KnownValue->isOneValue() &&3484 SuccBB == BI.getSuccessor(0)) ||3485 (!PartialIVInfo.KnownValue->isOneValue() &&3486 SuccBB == BI.getSuccessor(1)))3487 continue;3488 }3489 3490 // This successor's domtree will not need to be duplicated after3491 // unswitching if the edge to the successor dominates it (and thus the3492 // entire tree). This essentially means there is no other path into this3493 // subtree and so it will end up live in only one clone of the loop.3494 if (SuccBB->getUniquePredecessor() ||3495 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {3496 return PredBB == &BB || DT.dominates(SuccBB, PredBB);3497 })) {3498 Cost += computeDomSubtreeCost(*DT[SuccBB], BBCostMap, DTCostMap);3499 assert(Cost <= LoopCost &&3500 "Non-duplicated cost should never exceed total loop cost!");3501 }3502 }3503 3504 // Now scale the cost by the number of unique successors minus one. We3505 // subtract one because there is already at least one copy of the entire3506 // loop. This is computing the new cost of unswitching a condition.3507 // Note that guards always have 2 unique successors that are implicit and3508 // will be materialized if we decide to unswitch it.3509 int SuccessorsCount = isGuard(&TI) ? 2 : Visited.size();3510 assert(SuccessorsCount > 1 &&3511 "Cannot unswitch a condition without multiple distinct successors!");3512 return (LoopCost - Cost) * (SuccessorsCount - 1);3513 };3514 3515 std::optional<NonTrivialUnswitchCandidate> Best;3516 for (auto &Candidate : UnswitchCandidates) {3517 Instruction &TI = *Candidate.TI;3518 ArrayRef<Value *> Invariants = Candidate.Invariants;3519 BranchInst *BI = dyn_cast<BranchInst>(&TI);3520 bool FullUnswitch =3521 !BI || Candidate.hasPendingInjection() ||3522 (Invariants.size() == 1 &&3523 Invariants[0] == skipTrivialSelect(BI->getCondition()));3524 InstructionCost CandidateCost = ComputeUnswitchedCost(TI, FullUnswitch);3525 // Calculate cost multiplier which is a tool to limit potentially3526 // exponential behavior of loop-unswitch.3527 if (EnableUnswitchCostMultiplier) {3528 int CostMultiplier =3529 CalculateUnswitchCostMultiplier(TI, L, LI, DT, UnswitchCandidates);3530 assert(3531 (CostMultiplier > 0 && CostMultiplier <= UnswitchThreshold) &&3532 "cost multiplier needs to be in the range of 1..UnswitchThreshold");3533 CandidateCost *= CostMultiplier;3534 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost3535 << " (multiplier: " << CostMultiplier << ")"3536 << " for unswitch candidate: " << TI << "\n");3537 } else {3538 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost3539 << " for unswitch candidate: " << TI << "\n");3540 }3541 3542 if (!Best || CandidateCost < Best->Cost) {3543 Best = Candidate;3544 Best->Cost = CandidateCost;3545 }3546 }3547 assert(Best && "Must be!");3548 return *Best;3549}3550 3551// Insert a freeze on an unswitched branch if all is true:3552// 1. freeze-loop-unswitch-cond option is true3553// 2. The branch may not execute in the loop pre-transformation. If a branch may3554// not execute and could cause UB, it would always cause UB if it is hoisted outside3555// of the loop. Insert a freeze to prevent this case.3556// 3. The branch condition may be poison or undef3557static bool shouldInsertFreeze(Loop &L, Instruction &TI, DominatorTree &DT,3558 AssumptionCache &AC) {3559 assert(isa<BranchInst>(TI) || isa<SwitchInst>(TI));3560 if (!FreezeLoopUnswitchCond)3561 return false;3562 3563 ICFLoopSafetyInfo SafetyInfo;3564 SafetyInfo.computeLoopSafetyInfo(&L);3565 if (SafetyInfo.isGuaranteedToExecute(TI, &DT, &L))3566 return false;3567 3568 Value *Cond;3569 if (BranchInst *BI = dyn_cast<BranchInst>(&TI))3570 Cond = skipTrivialSelect(BI->getCondition());3571 else3572 Cond = skipTrivialSelect(cast<SwitchInst>(&TI)->getCondition());3573 return !isGuaranteedNotToBeUndefOrPoison(3574 Cond, &AC, L.getLoopPreheader()->getTerminator(), &DT);3575}3576 3577static bool unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI,3578 AssumptionCache &AC, AAResults &AA,3579 TargetTransformInfo &TTI, ScalarEvolution *SE,3580 MemorySSAUpdater *MSSAU,3581 LPMUpdater &LoopUpdater) {3582 // Collect all invariant conditions within this loop (as opposed to an inner3583 // loop which would be handled when visiting that inner loop).3584 SmallVector<NonTrivialUnswitchCandidate, 4> UnswitchCandidates;3585 IVConditionInfo PartialIVInfo;3586 Instruction *PartialIVCondBranch = nullptr;3587 collectUnswitchCandidates(UnswitchCandidates, PartialIVInfo,3588 PartialIVCondBranch, L, LI, AA, MSSAU);3589 if (!findOptionMDForLoop(&L, "llvm.loop.unswitch.injection.disable"))3590 collectUnswitchCandidatesWithInjections(UnswitchCandidates, PartialIVInfo,3591 PartialIVCondBranch, L, DT, LI, AA,3592 MSSAU);3593 // If we didn't find any candidates, we're done.3594 if (UnswitchCandidates.empty())3595 return false;3596 3597 LLVM_DEBUG(3598 dbgs() << "Considering " << UnswitchCandidates.size()3599 << " non-trivial loop invariant conditions for unswitching.\n");3600 3601 NonTrivialUnswitchCandidate Best = findBestNonTrivialUnswitchCandidate(3602 UnswitchCandidates, L, DT, LI, AC, TTI, PartialIVInfo);3603 3604 assert(Best.TI && "Failed to find loop unswitch candidate");3605 assert(Best.Cost && "Failed to compute cost");3606 3607 if (*Best.Cost >= UnswitchThreshold) {3608 LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: " << *Best.Cost3609 << "\n");3610 return false;3611 }3612 3613 bool InjectedCondition = false;3614 if (Best.hasPendingInjection()) {3615 Best = injectPendingInvariantConditions(Best, L, DT, LI, AC, MSSAU);3616 InjectedCondition = true;3617 }3618 assert(!Best.hasPendingInjection() &&3619 "All injections should have been done by now!");3620 3621 if (Best.TI != PartialIVCondBranch)3622 PartialIVInfo.InstToDuplicate.clear();3623 3624 bool InsertFreeze;3625 if (auto *SI = dyn_cast<SelectInst>(Best.TI)) {3626 // If the best candidate is a select, turn it into a branch. Select3627 // instructions with a poison conditional do not propagate poison, but3628 // branching on poison causes UB. Insert a freeze on the select3629 // conditional to prevent UB after turning the select into a branch.3630 InsertFreeze = !isGuaranteedNotToBeUndefOrPoison(3631 SI->getCondition(), &AC, L.getLoopPreheader()->getTerminator(), &DT);3632 Best.TI = turnSelectIntoBranch(SI, DT, LI, MSSAU, &AC);3633 } else {3634 // If the best candidate is a guard, turn it into a branch.3635 if (isGuard(Best.TI))3636 Best.TI =3637 turnGuardIntoBranch(cast<IntrinsicInst>(Best.TI), L, DT, LI, MSSAU);3638 InsertFreeze = shouldInsertFreeze(L, *Best.TI, DT, AC);3639 }3640 3641 LLVM_DEBUG(dbgs() << " Unswitching non-trivial (cost = " << Best.Cost3642 << ") terminator: " << *Best.TI << "\n");3643 unswitchNontrivialInvariants(L, *Best.TI, Best.Invariants, PartialIVInfo, DT,3644 LI, AC, SE, MSSAU, LoopUpdater, InsertFreeze,3645 InjectedCondition);3646 return true;3647}3648 3649/// Unswitch control flow predicated on loop invariant conditions.3650///3651/// This first hoists all branches or switches which are trivial (IE, do not3652/// require duplicating any part of the loop) out of the loop body. It then3653/// looks at other loop invariant control flows and tries to unswitch those as3654/// well by cloning the loop if the result is small enough.3655///3656/// The `DT`, `LI`, `AC`, `AA`, `TTI` parameters are required analyses that are3657/// also updated based on the unswitch. The `MSSA` analysis is also updated if3658/// valid (i.e. its use is enabled).3659///3660/// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is3661/// true, we will attempt to do non-trivial unswitching as well as trivial3662/// unswitching.3663///3664/// The `postUnswitch` function will be run after unswitching is complete3665/// with information on whether or not the provided loop remains a loop and3666/// a list of new sibling loops created.3667///3668/// If `SE` is non-null, we will update that analysis based on the unswitching3669/// done.3670static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI,3671 AssumptionCache &AC, AAResults &AA,3672 TargetTransformInfo &TTI, bool Trivial,3673 bool NonTrivial, ScalarEvolution *SE,3674 MemorySSAUpdater *MSSAU, LPMUpdater &LoopUpdater) {3675 assert(L.isRecursivelyLCSSAForm(DT, LI) &&3676 "Loops must be in LCSSA form before unswitching.");3677 3678 // Must be in loop simplified form: we need a preheader and dedicated exits.3679 if (!L.isLoopSimplifyForm())3680 return false;3681 3682 // Try trivial unswitch first before loop over other basic blocks in the loop.3683 if (Trivial && unswitchAllTrivialConditions(L, DT, LI, SE, MSSAU)) {3684 // If we unswitched successfully we will want to clean up the loop before3685 // processing it further so just mark it as unswitched and return.3686 postUnswitch(L, LoopUpdater, L.getName(),3687 /*CurrentLoopValid*/ true, /*PartiallyInvariant*/ false,3688 /*InjectedCondition*/ false, {});3689 return true;3690 }3691 3692 const Function *F = L.getHeader()->getParent();3693 3694 // Check whether we should continue with non-trivial conditions.3695 // EnableNonTrivialUnswitch: Global variable that forces non-trivial3696 // unswitching for testing and debugging.3697 // NonTrivial: Parameter that enables non-trivial unswitching for this3698 // invocation of the transform. But this should be allowed only3699 // for targets without branch divergence.3700 //3701 // FIXME: If divergence analysis becomes available to a loop3702 // transform, we should allow unswitching for non-trivial uniform3703 // branches even on targets that have divergence.3704 // https://bugs.llvm.org/show_bug.cgi?id=488193705 bool ContinueWithNonTrivial =3706 EnableNonTrivialUnswitch || (NonTrivial && !TTI.hasBranchDivergence(F));3707 if (!ContinueWithNonTrivial)3708 return false;3709 3710 // Skip non-trivial unswitching for optsize functions.3711 if (F->hasOptSize())3712 return false;3713 3714 // Perform legality checks.3715 if (!isSafeForNoNTrivialUnswitching(L, LI))3716 return false;3717 3718 // For non-trivial unswitching, because it often creates new loops, we rely on3719 // the pass manager to iterate on the loops rather than trying to immediately3720 // reach a fixed point. There is no substantial advantage to iterating3721 // internally, and if any of the new loops are simplified enough to contain3722 // trivial unswitching we want to prefer those.3723 3724 // Try to unswitch the best invariant condition. We prefer this full unswitch to3725 // a partial unswitch when possible below the threshold.3726 if (unswitchBestCondition(L, DT, LI, AC, AA, TTI, SE, MSSAU, LoopUpdater))3727 return true;3728 3729 // No other opportunities to unswitch.3730 return false;3731}3732 3733PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM,3734 LoopStandardAnalysisResults &AR,3735 LPMUpdater &U) {3736 Function &F = *L.getHeader()->getParent();3737 (void)F;3738 LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L3739 << "\n");3740 3741 std::optional<MemorySSAUpdater> MSSAU;3742 if (AR.MSSA) {3743 MSSAU = MemorySSAUpdater(AR.MSSA);3744 if (VerifyMemorySSA)3745 AR.MSSA->verifyMemorySSA();3746 }3747 if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC, AR.AA, AR.TTI, Trivial, NonTrivial,3748 &AR.SE, MSSAU ? &*MSSAU : nullptr, U))3749 return PreservedAnalyses::all();3750 3751 if (AR.MSSA && VerifyMemorySSA)3752 AR.MSSA->verifyMemorySSA();3753 3754 // Historically this pass has had issues with the dominator tree so verify it3755 // in asserts builds.3756 assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast));3757 3758 auto PA = getLoopPassPreservedAnalyses();3759 if (AR.MSSA)3760 PA.preserve<MemorySSAAnalysis>();3761 return PA;3762}3763 3764void SimpleLoopUnswitchPass::printPipeline(3765 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {3766 static_cast<PassInfoMixin<SimpleLoopUnswitchPass> *>(this)->printPipeline(3767 OS, MapClassName2PassName);3768 3769 OS << '<';3770 OS << (NonTrivial ? "" : "no-") << "nontrivial;";3771 OS << (Trivial ? "" : "no-") << "trivial";3772 OS << '>';3773}3774