blob: e8dc879a184b451a0e5df89ed0337a6713534304 [file] [log] [blame]
//===- LoopLoadElimination.cpp - Loop Load Elimination Pass ---------------===//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
// This file implement a loop-aware load elimination pass.
// It uses LoopAccessAnalysis to identify loop-carried dependences with a
// distance of one between stores and loads. These form the candidates for the
// transformation. The source value of each store then propagated to the user
// of the corresponding load. This makes the load dead.
// The pass can also version the loop and add memchecks in order to prove that
// may-aliasing stores can't change the value in memory before it's read by the
// load.
#include "llvm/Transforms/Scalar/LoopLoadElimination.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DepthFirstIterator.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/BlockFrequencyInfo.h"
#include "llvm/Analysis/GlobalsModRef.h"
#include "llvm/Analysis/LazyBlockFrequencyInfo.h"
#include "llvm/Analysis/LoopAccessAnalysis.h"
#include "llvm/Analysis/LoopAnalysisManager.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/MemorySSA.h"
#include "llvm/Analysis/ProfileSummaryInfo.h"
#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/Analysis/ScalarEvolutionExpander.h"
#include "llvm/Analysis/ScalarEvolutionExpressions.h"
#include "llvm/Analysis/TargetLibraryInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/PassManager.h"
#include "llvm/IR/Type.h"
#include "llvm/IR/Value.h"
#include "llvm/Pass.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Transforms/Utils.h"
#include "llvm/Transforms/Utils/LoopVersioning.h"
#include "llvm/Transforms/Utils/SizeOpts.h"
#include <algorithm>
#include <cassert>
#include <forward_list>
#include <set>
#include <tuple>
#include <utility>
using namespace llvm;
#define LLE_OPTION "loop-load-elim"
static cl::opt<unsigned> CheckPerElim(
"runtime-check-per-loop-load-elim", cl::Hidden,
cl::desc("Max number of memchecks allowed per eliminated load on average"),
static cl::opt<unsigned> LoadElimSCEVCheckThreshold(
"loop-load-elimination-scev-check-threshold", cl::init(8), cl::Hidden,
cl::desc("The maximum number of SCEV checks allowed for Loop "
"Load Elimination"));
STATISTIC(NumLoopLoadEliminted, "Number of loads eliminated by LLE");
namespace {
/// Represent a store-to-forwarding candidate.
struct StoreToLoadForwardingCandidate {
LoadInst *Load;
StoreInst *Store;
StoreToLoadForwardingCandidate(LoadInst *Load, StoreInst *Store)
: Load(Load), Store(Store) {}
/// Return true if the dependence from the store to the load has a
/// distance of one. E.g. A[i+1] = A[i]
bool isDependenceDistanceOfOne(PredicatedScalarEvolution &PSE,
Loop *L) const {
Value *LoadPtr = Load->getPointerOperand();
Value *StorePtr = Store->getPointerOperand();
Type *LoadPtrType = LoadPtr->getType();
Type *LoadType = LoadPtrType->getPointerElementType();
assert(LoadPtrType->getPointerAddressSpace() ==
StorePtr->getType()->getPointerAddressSpace() &&
LoadType == StorePtr->getType()->getPointerElementType() &&
"Should be a known dependence");
// Currently we only support accesses with unit stride. FIXME: we should be
// able to handle non unit stirde as well as long as the stride is equal to
// the dependence distance.
if (getPtrStride(PSE, LoadPtr, L) != 1 ||
getPtrStride(PSE, StorePtr, L) != 1)
return false;
auto &DL = Load->getParent()->getModule()->getDataLayout();
unsigned TypeByteSize = DL.getTypeAllocSize(const_cast<Type *>(LoadType));
auto *LoadPtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(LoadPtr));
auto *StorePtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(StorePtr));
// We don't need to check non-wrapping here because forward/backward
// dependence wouldn't be valid if these weren't monotonic accesses.
auto *Dist = cast<SCEVConstant>(
PSE.getSE()->getMinusSCEV(StorePtrSCEV, LoadPtrSCEV));
const APInt &Val = Dist->getAPInt();
return Val == TypeByteSize;
Value *getLoadPtr() const { return Load->getPointerOperand(); }
#ifndef NDEBUG
friend raw_ostream &operator<<(raw_ostream &OS,
const StoreToLoadForwardingCandidate &Cand) {
OS << *Cand.Store << " -->\n";
OS.indent(2) << *Cand.Load << "\n";
return OS;
} // end anonymous namespace
/// Check if the store dominates all latches, so as long as there is no
/// intervening store this value will be loaded in the next iteration.
static bool doesStoreDominatesAllLatches(BasicBlock *StoreBlock, Loop *L,
DominatorTree *DT) {
SmallVector<BasicBlock *, 8> Latches;
return llvm::all_of(Latches, [&](const BasicBlock *Latch) {
return DT->dominates(StoreBlock, Latch);
/// Return true if the load is not executed on all paths in the loop.
static bool isLoadConditional(LoadInst *Load, Loop *L) {
return Load->getParent() != L->getHeader();
namespace {
/// The per-loop class that does most of the work.
class LoadEliminationForLoop {
LoadEliminationForLoop(Loop *L, LoopInfo *LI, const LoopAccessInfo &LAI,
DominatorTree *DT, BlockFrequencyInfo *BFI,
ProfileSummaryInfo* PSI)
: L(L), LI(LI), LAI(LAI), DT(DT), BFI(BFI), PSI(PSI), PSE(LAI.getPSE()) {}
/// Look through the loop-carried and loop-independent dependences in
/// this loop and find store->load dependences.
/// Note that no candidate is returned if LAA has failed to analyze the loop
/// (e.g. if it's not bottom-tested, contains volatile memops, etc.)
findStoreToLoadDependences(const LoopAccessInfo &LAI) {
std::forward_list<StoreToLoadForwardingCandidate> Candidates;
const auto *Deps = LAI.getDepChecker().getDependences();
if (!Deps)
return Candidates;
// Find store->load dependences (consequently true dep). Both lexically
// forward and backward dependences qualify. Disqualify loads that have
// other unknown dependences.
SmallPtrSet<Instruction *, 4> LoadsWithUnknownDepedence;
for (const auto &Dep : *Deps) {
Instruction *Source = Dep.getSource(LAI);
Instruction *Destination = Dep.getDestination(LAI);
if (Dep.Type == MemoryDepChecker::Dependence::Unknown) {
if (isa<LoadInst>(Source))
if (isa<LoadInst>(Destination))
if (Dep.isBackward())
// Note that the designations source and destination follow the program
// order, i.e. source is always first. (The direction is given by the
// DepType.)
std::swap(Source, Destination);
assert(Dep.isForward() && "Needs to be a forward dependence");
auto *Store = dyn_cast<StoreInst>(Source);
if (!Store)
auto *Load = dyn_cast<LoadInst>(Destination);
if (!Load)
// Only progagate the value if they are of the same type.
if (Store->getPointerOperandType() != Load->getPointerOperandType())
Candidates.emplace_front(Load, Store);
if (!LoadsWithUnknownDepedence.empty())
Candidates.remove_if([&](const StoreToLoadForwardingCandidate &C) {
return LoadsWithUnknownDepedence.count(C.Load);
return Candidates;
/// Return the index of the instruction according to program order.
unsigned getInstrIndex(Instruction *Inst) {
auto I = InstOrder.find(Inst);
assert(I != InstOrder.end() && "No index for instruction");
return I->second;
/// If a load has multiple candidates associated (i.e. different
/// stores), it means that it could be forwarding from multiple stores
/// depending on control flow. Remove these candidates.
/// Here, we rely on LAA to include the relevant loop-independent dependences.
/// LAA is known to omit these in the very simple case when the read and the
/// write within an alias set always takes place using the *same* pointer.
/// However, we know that this is not the case here, i.e. we can rely on LAA
/// to provide us with loop-independent dependences for the cases we're
/// interested. Consider the case for example where a loop-independent
/// dependece S1->S2 invalidates the forwarding S3->S2.
/// A[i] = ... (S1)
/// ... = A[i] (S2)
/// A[i+1] = ... (S3)
/// LAA will perform dependence analysis here because there are two
/// *different* pointers involved in the same alias set (&A[i] and &A[i+1]).
void removeDependencesFromMultipleStores(
std::forward_list<StoreToLoadForwardingCandidate> &Candidates) {
// If Store is nullptr it means that we have multiple stores forwarding to
// this store.
using LoadToSingleCandT =
DenseMap<LoadInst *, const StoreToLoadForwardingCandidate *>;
LoadToSingleCandT LoadToSingleCand;
for (const auto &Cand : Candidates) {
bool NewElt;
LoadToSingleCandT::iterator Iter;
std::tie(Iter, NewElt) =
LoadToSingleCand.insert(std::make_pair(Cand.Load, &Cand));
if (!NewElt) {
const StoreToLoadForwardingCandidate *&OtherCand = Iter->second;
// Already multiple stores forward to this load.
if (OtherCand == nullptr)
// Handle the very basic case when the two stores are in the same block
// so deciding which one forwards is easy. The later one forwards as
// long as they both have a dependence distance of one to the load.
if (Cand.Store->getParent() == OtherCand->Store->getParent() &&
Cand.isDependenceDistanceOfOne(PSE, L) &&
OtherCand->isDependenceDistanceOfOne(PSE, L)) {
// They are in the same block, the later one will forward to the load.
if (getInstrIndex(OtherCand->Store) < getInstrIndex(Cand.Store))
OtherCand = &Cand;
} else
OtherCand = nullptr;
Candidates.remove_if([&](const StoreToLoadForwardingCandidate &Cand) {
if (LoadToSingleCand[Cand.Load] != &Cand) {
dbgs() << "Removing from candidates: \n"
<< Cand
<< " The load may have multiple stores forwarding to "
<< "it\n");
return true;
return false;
/// Given two pointers operations by their RuntimePointerChecking
/// indices, return true if they require an alias check.
/// We need a check if one is a pointer for a candidate load and the other is
/// a pointer for a possibly intervening store.
bool needsChecking(unsigned PtrIdx1, unsigned PtrIdx2,
const SmallPtrSet<Value *, 4> &PtrsWrittenOnFwdingPath,
const std::set<Value *> &CandLoadPtrs) {
Value *Ptr1 =
Value *Ptr2 =
return ((PtrsWrittenOnFwdingPath.count(Ptr1) && CandLoadPtrs.count(Ptr2)) ||
(PtrsWrittenOnFwdingPath.count(Ptr2) && CandLoadPtrs.count(Ptr1)));
/// Return pointers that are possibly written to on the path from a
/// forwarding store to a load.
/// These pointers need to be alias-checked against the forwarding candidates.
SmallPtrSet<Value *, 4> findPointersWrittenOnForwardingPath(
const SmallVectorImpl<StoreToLoadForwardingCandidate> &Candidates) {
// From FirstStore to LastLoad neither of the elimination candidate loads
// should overlap with any of the stores.
// E.g.:
// st1 C[i]
// ld1 B[i] <-------,
// ld0 A[i] <----, | * LastLoad
// ... | |
// st2 E[i] | |
// st3 B[i+1] -- | -' * FirstStore
// st0 A[i+1] ---'
// st4 D[i]
// st0 forwards to ld0 if the accesses in st4 and st1 don't overlap with
// ld0.
LoadInst *LastLoad =
std::max_element(Candidates.begin(), Candidates.end(),
[&](const StoreToLoadForwardingCandidate &A,
const StoreToLoadForwardingCandidate &B) {
return getInstrIndex(A.Load) < getInstrIndex(B.Load);
StoreInst *FirstStore =
std::min_element(Candidates.begin(), Candidates.end(),
[&](const StoreToLoadForwardingCandidate &A,
const StoreToLoadForwardingCandidate &B) {
return getInstrIndex(A.Store) <
// We're looking for stores after the first forwarding store until the end
// of the loop, then from the beginning of the loop until the last
// forwarded-to load. Collect the pointer for the stores.
SmallPtrSet<Value *, 4> PtrsWrittenOnFwdingPath;
auto InsertStorePtr = [&](Instruction *I) {
if (auto *S = dyn_cast<StoreInst>(I))
const auto &MemInstrs = LAI.getDepChecker().getMemoryInstructions();
std::for_each(MemInstrs.begin() + getInstrIndex(FirstStore) + 1,
MemInstrs.end(), InsertStorePtr);
std::for_each(MemInstrs.begin(), &MemInstrs[getInstrIndex(LastLoad)],
return PtrsWrittenOnFwdingPath;
/// Determine the pointer alias checks to prove that there are no
/// intervening stores.
SmallVector<RuntimePointerChecking::PointerCheck, 4> collectMemchecks(
const SmallVectorImpl<StoreToLoadForwardingCandidate> &Candidates) {
SmallPtrSet<Value *, 4> PtrsWrittenOnFwdingPath =
// Collect the pointers of the candidate loads.
// FIXME: SmallPtrSet does not work with std::inserter.
std::set<Value *> CandLoadPtrs;
std::inserter(CandLoadPtrs, CandLoadPtrs.begin()),
const auto &AllChecks = LAI.getRuntimePointerChecking()->getChecks();
SmallVector<RuntimePointerChecking::PointerCheck, 4> Checks;
copy_if(AllChecks, std::back_inserter(Checks),
[&](const RuntimePointerChecking::PointerCheck &Check) {
for (auto PtrIdx1 : Check.first->Members)
for (auto PtrIdx2 : Check.second->Members)
if (needsChecking(PtrIdx1, PtrIdx2, PtrsWrittenOnFwdingPath,
return true;
return false;
LLVM_DEBUG(dbgs() << "\nPointer Checks (count: " << Checks.size()
<< "):\n");
LLVM_DEBUG(LAI.getRuntimePointerChecking()->printChecks(dbgs(), Checks));
return Checks;
/// Perform the transformation for a candidate.
propagateStoredValueToLoadUsers(const StoreToLoadForwardingCandidate &Cand,
SCEVExpander &SEE) {
// loop:
// %x = load %gep_i
// = ... %x
// store %y, %gep_i_plus_1
// =>
// ph:
// %x.initial = load %gep_0
// loop:
// %x.storeforward = phi [%x.initial, %ph] [%y, %loop]
// %x = load %gep_i <---- now dead
// = ... %x.storeforward
// store %y, %gep_i_plus_1
Value *Ptr = Cand.Load->getPointerOperand();
auto *PtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(Ptr));
auto *PH = L->getLoopPreheader();
Value *InitialPtr = SEE.expandCodeFor(PtrSCEV->getStart(), Ptr->getType(),
Value *Initial = new LoadInst(
Cand.Load->getType(), InitialPtr, "load_initial",
/* isVolatile */ false, MaybeAlign(Cand.Load->getAlignment()),
PHINode *PHI = PHINode::Create(Initial->getType(), 2, "store_forwarded",
PHI->addIncoming(Initial, PH);
PHI->addIncoming(Cand.Store->getOperand(0), L->getLoopLatch());
/// Top-level driver for each loop: find store->load forwarding
/// candidates, add run-time checks and perform transformation.
bool processLoop() {
LLVM_DEBUG(dbgs() << "\nIn \"" << L->getHeader()->getParent()->getName()
<< "\" checking " << *L << "\n");
// Look for store-to-load forwarding cases across the
// backedge. E.g.:
// loop:
// %x = load %gep_i
// = ... %x
// store %y, %gep_i_plus_1
// =>
// ph:
// %x.initial = load %gep_0
// loop:
// %x.storeforward = phi [%x.initial, %ph] [%y, %loop]
// %x = load %gep_i <---- now dead
// = ... %x.storeforward
// store %y, %gep_i_plus_1
// First start with store->load dependences.
auto StoreToLoadDependences = findStoreToLoadDependences(LAI);
if (StoreToLoadDependences.empty())
return false;
// Generate an index for each load and store according to the original
// program order. This will be used later.
InstOrder = LAI.getDepChecker().generateInstructionOrderMap();
// To keep things simple for now, remove those where the load is potentially
// fed by multiple stores.
if (StoreToLoadDependences.empty())
return false;
// Filter the candidates further.
SmallVector<StoreToLoadForwardingCandidate, 4> Candidates;
unsigned NumForwarding = 0;
for (const StoreToLoadForwardingCandidate Cand : StoreToLoadDependences) {
LLVM_DEBUG(dbgs() << "Candidate " << Cand);
// Make sure that the stored values is available everywhere in the loop in
// the next iteration.
if (!doesStoreDominatesAllLatches(Cand.Store->getParent(), L, DT))
// If the load is conditional we can't hoist its 0-iteration instance to
// the preheader because that would make it unconditional. Thus we would
// access a memory location that the original loop did not access.
if (isLoadConditional(Cand.Load, L))
// Check whether the SCEV difference is the same as the induction step,
// thus we load the value in the next iteration.
if (!Cand.isDependenceDistanceOfOne(PSE, L))
<< NumForwarding
<< ". Valid store-to-load forwarding across the loop backedge\n");
if (Candidates.empty())
return false;
// Check intervening may-alias stores. These need runtime checks for alias
// disambiguation.
SmallVector<RuntimePointerChecking::PointerCheck, 4> Checks =
// Too many checks are likely to outweigh the benefits of forwarding.
if (Checks.size() > Candidates.size() * CheckPerElim) {
LLVM_DEBUG(dbgs() << "Too many run-time checks needed.\n");
return false;
if (LAI.getPSE().getUnionPredicate().getComplexity() >
LoadElimSCEVCheckThreshold) {
LLVM_DEBUG(dbgs() << "Too many SCEV run-time checks needed.\n");
return false;
if (!Checks.empty() || !LAI.getPSE().getUnionPredicate().isAlwaysTrue()) {
if (LAI.hasConvergentOp()) {
LLVM_DEBUG(dbgs() << "Versioning is needed but not allowed with "
"convergent calls\n");
return false;
auto *HeaderBB = L->getHeader();
auto *F = HeaderBB->getParent();
bool OptForSize = F->hasOptSize() ||
llvm::shouldOptimizeForSize(HeaderBB, PSI, BFI);
if (OptForSize) {
dbgs() << "Versioning is needed but not allowed when optimizing "
"for size.\n");
return false;
if (!L->isLoopSimplifyForm()) {
LLVM_DEBUG(dbgs() << "Loop is not is loop-simplify form");
return false;
// Point of no-return, start the transformation. First, version the loop
// if necessary.
LoopVersioning LV(LAI, L, LI, DT, PSE.getSE(), false);
// Next, propagate the value stored by the store to the users of the load.
// Also for the first iteration, generate the initial value of the load.
SCEVExpander SEE(*PSE.getSE(), L->getHeader()->getModule()->getDataLayout(),
for (const auto &Cand : Candidates)
propagateStoredValueToLoadUsers(Cand, SEE);
NumLoopLoadEliminted += NumForwarding;
return true;
Loop *L;
/// Maps the load/store instructions to their index according to
/// program order.
DenseMap<Instruction *, unsigned> InstOrder;
// Analyses used.
LoopInfo *LI;
const LoopAccessInfo &LAI;
DominatorTree *DT;
BlockFrequencyInfo *BFI;
ProfileSummaryInfo *PSI;
PredicatedScalarEvolution PSE;
} // end anonymous namespace
static bool
eliminateLoadsAcrossLoops(Function &F, LoopInfo &LI, DominatorTree &DT,
BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI,
function_ref<const LoopAccessInfo &(Loop &)> GetLAI) {
// Build up a worklist of inner-loops to transform to avoid iterator
// invalidation.
// FIXME: This logic comes from other passes that actually change the loop
// nest structure. It isn't clear this is necessary (or useful) for a pass
// which merely optimizes the use of loads in a loop.
SmallVector<Loop *, 8> Worklist;
for (Loop *TopLevelLoop : LI)
for (Loop *L : depth_first(TopLevelLoop))
// We only handle inner-most loops.
if (L->empty())
// Now walk the identified inner loops.
bool Changed = false;
for (Loop *L : Worklist) {
// The actual work is performed by LoadEliminationForLoop.
LoadEliminationForLoop LEL(L, &LI, GetLAI(*L), &DT, BFI, PSI);
Changed |= LEL.processLoop();
return Changed;
namespace {
/// The pass. Most of the work is delegated to the per-loop
/// LoadEliminationForLoop class.
class LoopLoadElimination : public FunctionPass {
static char ID;
LoopLoadElimination() : FunctionPass(ID) {
bool runOnFunction(Function &F) override {
if (skipFunction(F))
return false;
auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
auto &LAA = getAnalysis<LoopAccessLegacyAnalysis>();
auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
auto *BFI = (PSI && PSI->hasProfileSummary()) ?
&getAnalysis<LazyBlockFrequencyInfoPass>().getBFI() :
// Process each loop nest in the function.
return eliminateLoadsAcrossLoops(
[&LAA](Loop &L) -> const LoopAccessInfo & { return LAA.getInfo(&L); });
void getAnalysisUsage(AnalysisUsage &AU) const override {
} // end anonymous namespace
char LoopLoadElimination::ID;
static const char LLE_name[] = "Loop Load Elimination";
INITIALIZE_PASS_BEGIN(LoopLoadElimination, LLE_OPTION, LLE_name, false, false)
INITIALIZE_PASS_END(LoopLoadElimination, LLE_OPTION, LLE_name, false, false)
FunctionPass *llvm::createLoopLoadEliminationPass() {
return new LoopLoadElimination();
PreservedAnalyses LoopLoadEliminationPass::run(Function &F,
FunctionAnalysisManager &AM) {
auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
auto &LI = AM.getResult<LoopAnalysis>(F);
auto &TTI = AM.getResult<TargetIRAnalysis>(F);
auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
auto &AA = AM.getResult<AAManager>(F);
auto &AC = AM.getResult<AssumptionAnalysis>(F);
auto &MAM = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F).getManager();
auto *PSI = MAM.getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
auto *BFI = (PSI && PSI->hasProfileSummary()) ?
&AM.getResult<BlockFrequencyAnalysis>(F) : nullptr;
MemorySSA *MSSA = EnableMSSALoopDependency
? &AM.getResult<MemorySSAAnalysis>(F).getMSSA()
: nullptr;
auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager();
bool Changed = eliminateLoadsAcrossLoops(
F, LI, DT, BFI, PSI, [&](Loop &L) -> const LoopAccessInfo & {
LoopStandardAnalysisResults AR = {AA, AC, DT, LI, SE, TLI, TTI, MSSA};
return LAM.getResult<LoopAccessAnalysis>(L, AR);
if (!Changed)
return PreservedAnalyses::all();
PreservedAnalyses PA;
return PA;