//===----------------------------------------------------------------------===//
#include "llvm/Analysis/DependenceAnalysis.h"
+#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/ScalarEvolutionExpressions.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/IR/InstIterator.h"
+#include "llvm/IR/Module.h"
#include "llvm/IR/Operator.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
INITIALIZE_PASS_BEGIN(DependenceAnalysis, "da",
"Dependence Analysis", true, true)
-INITIALIZE_PASS_DEPENDENCY(LoopInfo)
+INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
INITIALIZE_PASS_END(DependenceAnalysis, "da",
this->F = &F;
AA = &getAnalysis<AliasAnalysis>();
SE = &getAnalysis<ScalarEvolution>();
- LI = &getAnalysis<LoopInfo>();
+ LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
return false;
}
AU.setPreservesAll();
AU.addRequiredTransitive<AliasAnalysis>();
AU.addRequiredTransitive<ScalarEvolution>();
- AU.addRequiredTransitive<LoopInfo>();
+ AU.addRequiredTransitive<LoopInfoWrapperPass>();
}
DstI != DstE; ++DstI) {
if (isa<StoreInst>(*DstI) || isa<LoadInst>(*DstI)) {
OS << "da analyze - ";
- if (Dependence *D = DA->depends(&*SrcI, &*DstI, true)) {
+ if (auto D = DA->depends(&*SrcI, &*DstI, true)) {
D->dump(OS);
for (unsigned Level = 1; Level <= D->getLevels(); Level++) {
if (D->isSplitable(Level)) {
OS << "da analyze - split level = " << Level;
- OS << ", iteration = " << *DA->getSplitIteration(D, Level);
+ OS << ", iteration = " << *DA->getSplitIteration(*D, Level);
OS << "!\n";
}
}
- delete D;
}
else
OS << "none!\n";
//===----------------------------------------------------------------------===//
// FullDependence methods
-FullDependence::FullDependence(Instruction *Source,
- Instruction *Destination,
+FullDependence::FullDependence(Instruction *Source, Instruction *Destination,
bool PossiblyLoopIndependent,
- unsigned CommonLevels) :
- Dependence(Source, Destination),
- Levels(CommonLevels),
- LoopIndependent(PossiblyLoopIndependent) {
+ unsigned CommonLevels)
+ : Dependence(Source, Destination), Levels(CommonLevels),
+ LoopIndependent(PossiblyLoopIndependent) {
Consistent = true;
- DV = CommonLevels ? new DVEntry[CommonLevels] : nullptr;
+ if (CommonLevels)
+ DV = make_unique<DVEntry[]>(CommonLevels);
}
// The rest are simple getters that hide the implementation.
OS << "!\n";
}
-
-
-static
-AliasAnalysis::AliasResult underlyingObjectsAlias(AliasAnalysis *AA,
- const Value *A,
- const Value *B) {
- const Value *AObj = GetUnderlyingObject(A);
- const Value *BObj = GetUnderlyingObject(B);
- return AA->alias(AObj, AA->getTypeStoreSize(AObj->getType()),
- BObj, AA->getTypeStoreSize(BObj->getType()));
+static AliasResult underlyingObjectsAlias(AliasAnalysis *AA,
+ const DataLayout &DL, const Value *A,
+ const Value *B) {
+ const Value *AObj = GetUnderlyingObject(A, DL);
+ const Value *BObj = GetUnderlyingObject(B, DL);
+ return AA->alias(AObj, DL.getTypeStoreSize(AObj->getType()),
+ BObj, DL.getTypeStoreSize(BObj->getType()));
}
}
}
+void DependenceAnalysis::unifySubscriptType(ArrayRef<Subscript *> Pairs) {
+
+ unsigned widestWidthSeen = 0;
+ Type *widestType;
+
+ // Go through each pair and find the widest bit to which we need
+ // to extend all of them.
+ for (unsigned i = 0; i < Pairs.size(); i++) {
+ const SCEV *Src = Pairs[i]->Src;
+ const SCEV *Dst = Pairs[i]->Dst;
+ IntegerType *SrcTy = dyn_cast<IntegerType>(Src->getType());
+ IntegerType *DstTy = dyn_cast<IntegerType>(Dst->getType());
+ if (SrcTy == nullptr || DstTy == nullptr) {
+ assert(SrcTy == DstTy && "This function only unify integer types and "
+ "expect Src and Dst share the same type "
+ "otherwise.");
+ continue;
+ }
+ if (SrcTy->getBitWidth() > widestWidthSeen) {
+ widestWidthSeen = SrcTy->getBitWidth();
+ widestType = SrcTy;
+ }
+ if (DstTy->getBitWidth() > widestWidthSeen) {
+ widestWidthSeen = DstTy->getBitWidth();
+ widestType = DstTy;
+ }
+ }
+
+
+ assert(widestWidthSeen > 0);
+
+ // Now extend each pair to the widest seen.
+ for (unsigned i = 0; i < Pairs.size(); i++) {
+ const SCEV *Src = Pairs[i]->Src;
+ const SCEV *Dst = Pairs[i]->Dst;
+ IntegerType *SrcTy = dyn_cast<IntegerType>(Src->getType());
+ IntegerType *DstTy = dyn_cast<IntegerType>(Dst->getType());
+ if (SrcTy == nullptr || DstTy == nullptr) {
+ assert(SrcTy == DstTy && "This function only unify integer types and "
+ "expect Src and Dst share the same type "
+ "otherwise.");
+ continue;
+ }
+ if (SrcTy->getBitWidth() < widestWidthSeen)
+ // Sign-extend Src to widestType
+ Pairs[i]->Src = SE->getSignExtendExpr(Src, widestType);
+ if (DstTy->getBitWidth() < widestWidthSeen) {
+ // Sign-extend Dst to widestType
+ Pairs[i]->Dst = SE->getSignExtendExpr(Dst, widestType);
+ }
+ }
+}
// removeMatchingExtensions - Examines a subscript pair.
// If the source and destination are identically sign (or zero)
(isa<SCEVSignExtendExpr>(Src) && isa<SCEVSignExtendExpr>(Dst))) {
const SCEVCastExpr *SrcCast = cast<SCEVCastExpr>(Src);
const SCEVCastExpr *DstCast = cast<SCEVCastExpr>(Dst);
- if (SrcCast->getType() == DstCast->getType()) {
- Pair->Src = SrcCast->getOperand();
- Pair->Dst = DstCast->getOperand();
+ const SCEV *SrcCastOp = SrcCast->getOperand();
+ const SCEV *DstCastOp = DstCast->getOperand();
+ if (SrcCastOp->getType() == DstCastOp->getType()) {
+ Pair->Src = SrcCastOp;
+ Pair->Dst = DstCastOp;
}
}
}
return isLoopInvariant(Src, LoopNest);
const SCEV *Start = AddRec->getStart();
const SCEV *Step = AddRec->getStepRecurrence(*SE);
+ const SCEV *UB = SE->getBackedgeTakenCount(AddRec->getLoop());
+ if (!isa<SCEVCouldNotCompute>(UB)) {
+ if (SE->getTypeSizeInBits(Start->getType()) <
+ SE->getTypeSizeInBits(UB->getType())) {
+ if (!AddRec->getNoWrapFlags())
+ return false;
+ }
+ }
if (!isLoopInvariant(Step, LoopNest))
return false;
Loops.set(mapSrcLoop(AddRec->getLoop()));
return isLoopInvariant(Dst, LoopNest);
const SCEV *Start = AddRec->getStart();
const SCEV *Step = AddRec->getStepRecurrence(*SE);
+ const SCEV *UB = SE->getBackedgeTakenCount(AddRec->getLoop());
+ if (!isa<SCEVCouldNotCompute>(UB)) {
+ if (SE->getTypeSizeInBits(Start->getType()) <
+ SE->getTypeSizeInBits(UB->getType())) {
+ if (!AddRec->getNoWrapFlags())
+ return false;
+ }
+ }
if (!isLoopInvariant(Step, LoopNest))
return false;
Loops.set(mapDstLoop(AddRec->getLoop()));
// All subscripts are all the same type.
// Loop bound may be smaller (e.g., a char).
// Should zero extend loop bound, since it's always >= 0.
-// This routine collects upper bound and extends if needed.
+// This routine collects upper bound and extends or truncates if needed.
+// Truncating is safe when subscripts are known not to wrap. Cases without
+// nowrap flags should have been rejected earlier.
// Return null if no bound available.
const SCEV *DependenceAnalysis::collectUpperBound(const Loop *L,
Type *T) const {
if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
const SCEV *UB = SE->getBackedgeTakenCount(L);
- return SE->getNoopOrZeroExtend(UB, T);
+ return SE->getTruncateOrZeroExtend(UB, T);
}
return nullptr;
}
// return the coefficient (the step)
// corresponding to the specified loop.
// If there isn't one, return 0.
-// For example, given a*i + b*j + c*k, zeroing the coefficient
+// For example, given a*i + b*j + c*k, finding the coefficient
// corresponding to the j loop would yield b.
const SCEV *DependenceAnalysis::findCoefficient(const SCEV *Expr,
const Loop *TargetLoop) const {
AddRec->getNoWrapFlags());
}
if (SE->isLoopInvariant(AddRec, TargetLoop))
- return SE->getAddRecExpr(AddRec,
- Value,
- TargetLoop,
- SCEV::FlagAnyWrap);
- return SE->getAddRecExpr(addToCoefficient(AddRec->getStart(),
- TargetLoop, Value),
- AddRec->getStepRecurrence(*SE),
- AddRec->getLoop(),
- AddRec->getNoWrapFlags());
+ return SE->getAddRecExpr(AddRec, Value, TargetLoop, SCEV::FlagAnyWrap);
+ return SE->getAddRecExpr(
+ addToCoefficient(AddRec->getStart(), TargetLoop, Value),
+ AddRec->getStepRecurrence(*SE), AddRec->getLoop(),
+ AddRec->getNoWrapFlags());
}
/// source and destination array references are recurrences on a nested loop,
/// this function flattens the nested recurrences into separate recurrences
/// for each loop level.
-bool
-DependenceAnalysis::tryDelinearize(const SCEV *SrcSCEV, const SCEV *DstSCEV,
- SmallVectorImpl<Subscript> &Pair) const {
+bool DependenceAnalysis::tryDelinearize(const SCEV *SrcSCEV,
+ const SCEV *DstSCEV,
+ SmallVectorImpl<Subscript> &Pair,
+ const SCEV *ElementSize) {
+ const SCEVUnknown *SrcBase =
+ dyn_cast<SCEVUnknown>(SE->getPointerBase(SrcSCEV));
+ const SCEVUnknown *DstBase =
+ dyn_cast<SCEVUnknown>(SE->getPointerBase(DstSCEV));
+
+ if (!SrcBase || !DstBase || SrcBase != DstBase)
+ return false;
+
+ SrcSCEV = SE->getMinusSCEV(SrcSCEV, SrcBase);
+ DstSCEV = SE->getMinusSCEV(DstSCEV, DstBase);
+
const SCEVAddRecExpr *SrcAR = dyn_cast<SCEVAddRecExpr>(SrcSCEV);
const SCEVAddRecExpr *DstAR = dyn_cast<SCEVAddRecExpr>(DstSCEV);
if (!SrcAR || !DstAR || !SrcAR->isAffine() || !DstAR->isAffine())
return false;
- SmallVector<const SCEV *, 4> SrcSubscripts, DstSubscripts, SrcSizes, DstSizes;
- const SCEV *RemainderS = SrcAR->delinearize(*SE, SrcSubscripts, SrcSizes);
- const SCEV *RemainderD = DstAR->delinearize(*SE, DstSubscripts, DstSizes);
+ // First step: collect parametric terms in both array references.
+ SmallVector<const SCEV *, 4> Terms;
+ SE->collectParametricTerms(SrcAR, Terms);
+ SE->collectParametricTerms(DstAR, Terms);
- int size = SrcSubscripts.size();
- // Fail when there is only a subscript: that's a linearized access function.
- if (size < 2)
- return false;
+ // Second step: find subscript sizes.
+ SmallVector<const SCEV *, 4> Sizes;
+ SE->findArrayDimensions(Terms, Sizes, ElementSize);
- int dstSize = DstSubscripts.size();
- // Fail when the number of subscripts in Src and Dst differ.
- if (size != dstSize)
- return false;
+ // Third step: compute the access functions for each subscript.
+ SmallVector<const SCEV *, 4> SrcSubscripts, DstSubscripts;
+ SE->computeAccessFunctions(SrcAR, SrcSubscripts, Sizes);
+ SE->computeAccessFunctions(DstAR, DstSubscripts, Sizes);
- // Fail when the size of any of the subscripts in Src and Dst differs: the
- // dependence analysis assumes that elements in the same array have same size.
- // SCEV delinearization does not have a context based on which it would decide
- // globally the size of subscripts that would best fit all the array accesses.
- for (int i = 0; i < size; ++i)
- if (SrcSizes[i] != DstSizes[i])
- return false;
-
- // When the difference in remainders is different than a constant it might be
- // that the base address of the arrays is not the same.
- const SCEV *DiffRemainders = SE->getMinusSCEV(RemainderS, RemainderD);
- if (!isa<SCEVConstant>(DiffRemainders))
+ // Fail when there is only a subscript: that's a linearized access function.
+ if (SrcSubscripts.size() < 2 || DstSubscripts.size() < 2 ||
+ SrcSubscripts.size() != DstSubscripts.size())
return false;
- // Normalize the last dimension: integrate the size of the "scalar dimension"
- // and the remainder of the delinearization.
- DstSubscripts[size-1] = SE->getMulExpr(DstSubscripts[size-1],
- DstSizes[size-1]);
- SrcSubscripts[size-1] = SE->getMulExpr(SrcSubscripts[size-1],
- SrcSizes[size-1]);
- SrcSubscripts[size-1] = SE->getAddExpr(SrcSubscripts[size-1], RemainderS);
- DstSubscripts[size-1] = SE->getAddExpr(DstSubscripts[size-1], RemainderD);
+ int size = SrcSubscripts.size();
-#ifndef NDEBUG
- DEBUG(errs() << "\nSrcSubscripts: ");
- for (int i = 0; i < size; i++)
- DEBUG(errs() << *SrcSubscripts[i]);
- DEBUG(errs() << "\nDstSubscripts: ");
- for (int i = 0; i < size; i++)
- DEBUG(errs() << *DstSubscripts[i]);
-#endif
+ DEBUG({
+ dbgs() << "\nSrcSubscripts: ";
+ for (int i = 0; i < size; i++)
+ dbgs() << *SrcSubscripts[i];
+ dbgs() << "\nDstSubscripts: ";
+ for (int i = 0; i < size; i++)
+ dbgs() << *DstSubscripts[i];
+ });
// The delinearization transforms a single-subscript MIV dependence test into
// a multi-subscript SIV dependence test that is easier to compute. So we
for (int i = 0; i < size; ++i) {
Pair[i].Src = SrcSubscripts[i];
Pair[i].Dst = DstSubscripts[i];
+ unifySubscriptType(&Pair[i]);
// FIXME: we should record the bounds SrcSizes[i] and DstSizes[i] that the
// delinearization has found, and add these constraints to the dependence
//
// Care is required to keep the routine below, getSplitIteration(),
// up to date with respect to this routine.
-Dependence *DependenceAnalysis::depends(Instruction *Src,
- Instruction *Dst,
- bool PossiblyLoopIndependent) {
+std::unique_ptr<Dependence>
+DependenceAnalysis::depends(Instruction *Src, Instruction *Dst,
+ bool PossiblyLoopIndependent) {
if (Src == Dst)
PossiblyLoopIndependent = false;
if (!isLoadOrStore(Src) || !isLoadOrStore(Dst)) {
// can only analyze simple loads and stores, i.e., no calls, invokes, etc.
DEBUG(dbgs() << "can only handle simple loads and stores\n");
- return new Dependence(Src, Dst);
+ return make_unique<Dependence>(Src, Dst);
}
Value *SrcPtr = getPointerOperand(Src);
Value *DstPtr = getPointerOperand(Dst);
- switch (underlyingObjectsAlias(AA, DstPtr, SrcPtr)) {
- case AliasAnalysis::MayAlias:
- case AliasAnalysis::PartialAlias:
+ switch (underlyingObjectsAlias(AA, F->getParent()->getDataLayout(), DstPtr,
+ SrcPtr)) {
+ case MayAlias:
+ case PartialAlias:
// cannot analyse objects if we don't understand their aliasing.
DEBUG(dbgs() << "can't analyze may or partial alias\n");
- return new Dependence(Src, Dst);
- case AliasAnalysis::NoAlias:
+ return make_unique<Dependence>(Src, Dst);
+ case NoAlias:
// If the objects noalias, they are distinct, accesses are independent.
DEBUG(dbgs() << "no alias\n");
return nullptr;
- case AliasAnalysis::MustAlias:
+ case MustAlias:
break; // The underlying objects alias; test accesses for dependence.
}
DEBUG(dbgs() << " SrcPtrSCEV = " << *SrcPtrSCEV << "\n");
DEBUG(dbgs() << " DstPtrSCEV = " << *DstPtrSCEV << "\n");
- UsefulGEP =
- isLoopInvariant(SrcPtrSCEV, LI->getLoopFor(Src->getParent())) &&
- isLoopInvariant(DstPtrSCEV, LI->getLoopFor(Dst->getParent()));
+ UsefulGEP = isLoopInvariant(SrcPtrSCEV, LI->getLoopFor(Src->getParent())) &&
+ isLoopInvariant(DstPtrSCEV, LI->getLoopFor(Dst->getParent())) &&
+ (SrcGEP->getNumOperands() == DstGEP->getNumOperands());
}
unsigned Pairs = UsefulGEP ? SrcGEP->idx_end() - SrcGEP->idx_begin() : 1;
SmallVector<Subscript, 4> Pair(Pairs);
++SrcIdx, ++DstIdx, ++P) {
Pair[P].Src = SE->getSCEV(*SrcIdx);
Pair[P].Dst = SE->getSCEV(*DstIdx);
+ unifySubscriptType(&Pair[P]);
}
}
else {
}
if (Delinearize && Pairs == 1 && CommonLevels > 1 &&
- tryDelinearize(Pair[0].Src, Pair[0].Dst, Pair)) {
+ tryDelinearize(Pair[0].Src, Pair[0].Dst, Pair, SE->getElementSize(Src))) {
DEBUG(dbgs() << " delinerized GEP\n");
Pairs = Pair.size();
}
LI->getLoopFor(Dst->getParent()),
Pair[SI].Loops);
Result.Consistent = false;
- }
- else if (Pair[SI].Classification == Subscript::ZIV) {
+ } else if (Pair[SI].Classification == Subscript::ZIV) {
// always separable
Separable.set(SI);
}
DEBUG(dbgs() << ", SIV\n");
unsigned Level;
const SCEV *SplitIter = nullptr;
- if (testSIV(Pair[SI].Src, Pair[SI].Dst, Level,
- Result, NewConstraint, SplitIter))
+ if (testSIV(Pair[SI].Src, Pair[SI].Dst, Level, Result, NewConstraint,
+ SplitIter))
return nullptr;
break;
}
SmallBitVector Sivs(Pairs);
SmallBitVector Mivs(Pairs);
SmallBitVector ConstrainedLevels(MaxLevels + 1);
+ SmallVector<Subscript *, 4> PairsInGroup;
for (int SJ = Group.find_first(); SJ >= 0; SJ = Group.find_next(SJ)) {
DEBUG(dbgs() << SJ << " ");
if (Pair[SJ].Classification == Subscript::SIV)
Sivs.set(SJ);
else
Mivs.set(SJ);
+ PairsInGroup.push_back(&Pair[SJ]);
}
+ unifySubscriptType(PairsInGroup);
DEBUG(dbgs() << "}\n");
while (Sivs.any()) {
bool Changed = false;
unsigned Level;
const SCEV *SplitIter = nullptr;
DEBUG(dbgs() << "SIV\n");
- if (testSIV(Pair[SJ].Src, Pair[SJ].Dst, Level,
- Result, NewConstraint, SplitIter))
+ if (testSIV(Pair[SJ].Src, Pair[SJ].Dst, Level, Result, NewConstraint,
+ SplitIter))
return nullptr;
ConstrainedLevels.set(Level);
if (intersectConstraints(&Constraints[Level], &NewConstraint)) {
// update Result.DV from constraint vector
DEBUG(dbgs() << " updating\n");
- for (int SJ = ConstrainedLevels.find_first();
- SJ >= 0; SJ = ConstrainedLevels.find_next(SJ)) {
+ for (int SJ = ConstrainedLevels.find_first(); SJ >= 0;
+ SJ = ConstrainedLevels.find_next(SJ)) {
+ if (SJ > (int)CommonLevels)
+ break;
updateDirection(Result.DV[SJ - 1], Constraints[SJ]);
if (Result.DV[SJ - 1].Direction == Dependence::DVEntry::NONE)
return nullptr;
return nullptr;
}
- FullDependence *Final = new FullDependence(Result);
- Result.DV = nullptr;
- return Final;
+ return make_unique<FullDependence>(std::move(Result));
}
//
// breaks the dependence and allows us to vectorize/parallelize
// both loops.
-const SCEV *DependenceAnalysis::getSplitIteration(const Dependence *Dep,
+const SCEV *DependenceAnalysis::getSplitIteration(const Dependence &Dep,
unsigned SplitLevel) {
- assert(Dep && "expected a pointer to a Dependence");
- assert(Dep->isSplitable(SplitLevel) &&
+ assert(Dep.isSplitable(SplitLevel) &&
"Dep should be splitable at SplitLevel");
- Instruction *Src = Dep->getSrc();
- Instruction *Dst = Dep->getDst();
+ Instruction *Src = Dep.getSrc();
+ Instruction *Dst = Dep.getDst();
assert(Src->mayReadFromMemory() || Src->mayWriteToMemory());
assert(Dst->mayReadFromMemory() || Dst->mayWriteToMemory());
assert(isLoadOrStore(Src));
assert(isLoadOrStore(Dst));
Value *SrcPtr = getPointerOperand(Src);
Value *DstPtr = getPointerOperand(Dst);
- assert(underlyingObjectsAlias(AA, DstPtr, SrcPtr) ==
- AliasAnalysis::MustAlias);
+ assert(underlyingObjectsAlias(AA, F->getParent()->getDataLayout(), DstPtr,
+ SrcPtr) == MustAlias);
// establish loop nesting levels
establishNestingLevels(Src, Dst);
SrcGEP->getPointerOperandType() == DstGEP->getPointerOperandType()) {
const SCEV *SrcPtrSCEV = SE->getSCEV(SrcGEP->getPointerOperand());
const SCEV *DstPtrSCEV = SE->getSCEV(DstGEP->getPointerOperand());
- UsefulGEP =
- isLoopInvariant(SrcPtrSCEV, LI->getLoopFor(Src->getParent())) &&
- isLoopInvariant(DstPtrSCEV, LI->getLoopFor(Dst->getParent()));
+ UsefulGEP = isLoopInvariant(SrcPtrSCEV, LI->getLoopFor(Src->getParent())) &&
+ isLoopInvariant(DstPtrSCEV, LI->getLoopFor(Dst->getParent())) &&
+ (SrcGEP->getNumOperands() == DstGEP->getNumOperands());
}
unsigned Pairs = UsefulGEP ? SrcGEP->idx_end() - SrcGEP->idx_begin() : 1;
SmallVector<Subscript, 4> Pair(Pairs);
}
if (Delinearize && Pairs == 1 && CommonLevels > 1 &&
- tryDelinearize(Pair[0].Src, Pair[0].Dst, Pair)) {
+ tryDelinearize(Pair[0].Src, Pair[0].Dst, Pair, SE->getElementSize(Src))) {
DEBUG(dbgs() << " delinerized GEP\n");
Pairs = Pair.size();
}