#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(LoopInfoWrapperPass)
-INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
-INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
+INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
+INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
INITIALIZE_PASS_END(DependenceAnalysis, "da",
"Dependence Analysis", true, true)
bool DependenceAnalysis::runOnFunction(Function &F) {
this->F = &F;
- AA = &getAnalysis<AliasAnalysis>();
- SE = &getAnalysis<ScalarEvolution>();
+ AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
+ SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
return false;
}
void DependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesAll();
- AU.addRequiredTransitive<AliasAnalysis>();
- AU.addRequiredTransitive<ScalarEvolution>();
+ AU.addRequiredTransitive<AAResultsWrapperPass>();
+ AU.addRequiredTransitive<ScalarEvolutionWrapperPass>();
AU.addRequiredTransitive<LoopInfoWrapperPass>();
}
//===----------------------------------------------------------------------===//
// 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(Subscript *Pair) {
- const SCEV *Src = Pair->Src;
- const SCEV *Dst = Pair->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.");
- return;
+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;
+ }
}
- if (SrcTy->getBitWidth() > DstTy->getBitWidth()) {
- // Sign-extend Dst to typeof(Src) if typeof(Src) is wider than typeof(Dst).
- Pair->Dst = SE->getSignExtendExpr(Dst, SrcTy);
- } else if (SrcTy->getBitWidth() < DstTy->getBitWidth()) {
- // Sign-extend Src to typeof(Dst) if typeof(Dst) is wider than typeof(Src).
- Pair->Src = SE->getSignExtendExpr(Src, 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);
+ }
}
}
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 {
/// 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 SCEV *ElementSize) {
+bool DependenceAnalysis::tryDelinearize(Instruction *Src,
+ Instruction *Dst,
+ SmallVectorImpl<Subscript> &Pair)
+{
+ Value *SrcPtr = getPointerOperand(Src);
+ Value *DstPtr = getPointerOperand(Dst);
+
+ Loop *SrcLoop = LI->getLoopFor(Src->getParent());
+ Loop *DstLoop = LI->getLoopFor(Dst->getParent());
+
+ // Below code mimics the code in Delinearization.cpp
+ const SCEV *SrcAccessFn =
+ SE->getSCEVAtScope(SrcPtr, SrcLoop);
+ const SCEV *DstAccessFn =
+ SE->getSCEVAtScope(DstPtr, DstLoop);
+
const SCEVUnknown *SrcBase =
- dyn_cast<SCEVUnknown>(SE->getPointerBase(SrcSCEV));
+ dyn_cast<SCEVUnknown>(SE->getPointerBase(SrcAccessFn));
const SCEVUnknown *DstBase =
- dyn_cast<SCEVUnknown>(SE->getPointerBase(DstSCEV));
+ dyn_cast<SCEVUnknown>(SE->getPointerBase(DstAccessFn));
if (!SrcBase || !DstBase || SrcBase != DstBase)
return false;
- SrcSCEV = SE->getMinusSCEV(SrcSCEV, SrcBase);
- DstSCEV = SE->getMinusSCEV(DstSCEV, DstBase);
+ const SCEV *ElementSize = SE->getElementSize(Src);
+ if (ElementSize != SE->getElementSize(Dst))
+ return false;
+
+ const SCEV *SrcSCEV = SE->getMinusSCEV(SrcAccessFn, SrcBase);
+ const SCEV *DstSCEV = SE->getMinusSCEV(DstAccessFn, DstBase);
const SCEVAddRecExpr *SrcAR = dyn_cast<SCEVAddRecExpr>(SrcSCEV);
const SCEVAddRecExpr *DstAR = dyn_cast<SCEVAddRecExpr>(DstSCEV);
// First step: collect parametric terms in both array references.
SmallVector<const SCEV *, 4> Terms;
- SrcAR->collectParametricTerms(*SE, Terms);
- DstAR->collectParametricTerms(*SE, Terms);
+ SE->collectParametricTerms(SrcAR, Terms);
+ SE->collectParametricTerms(DstAR, Terms);
// Second step: find subscript sizes.
SmallVector<const SCEV *, 4> Sizes;
// Third step: compute the access functions for each subscript.
SmallVector<const SCEV *, 4> SrcSubscripts, DstSubscripts;
- SrcAR->computeAccessFunctions(*SE, SrcSubscripts, Sizes);
- DstAR->computeAccessFunctions(*SE, DstSubscripts, Sizes);
+ SE->computeAccessFunctions(SrcAR, SrcSubscripts, Sizes);
+ SE->computeAccessFunctions(DstAR, DstSubscripts, Sizes);
// Fail when there is only a subscript: that's a linearized access function.
if (SrcSubscripts.size() < 2 || DstSubscripts.size() < 2 ||
}
#endif
-
// depends -
// Returns NULL if there is no dependence.
// Otherwise, return a Dependence with as many details as possible.
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 make_unique<Dependence>(Src, Dst);
- case AliasAnalysis::NoAlias:
+ 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() << " common nesting levels = " << CommonLevels << "\n");
DEBUG(dbgs() << " maximum nesting levels = " << MaxLevels << "\n");
- auto Result = llvm::make_unique<FullDependence>(
- Src, Dst, PossiblyLoopIndependent, CommonLevels);
+ FullDependence Result(Src, Dst, PossiblyLoopIndependent, CommonLevels);
++TotalArrayPairs;
// See if there are GEPs we can use.
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);
Pair[0].Dst = DstSCEV;
}
- if (Delinearize && Pairs == 1 && CommonLevels > 1 &&
- tryDelinearize(Pair[0].Src, Pair[0].Dst, Pair, SE->getElementSize(Src))) {
- DEBUG(dbgs() << " delinerized GEP\n");
- Pairs = Pair.size();
+ if (Delinearize && CommonLevels > 1) {
+ if (tryDelinearize(Src, Dst, Pair)) {
+ DEBUG(dbgs() << " delinerized GEP\n");
+ Pairs = Pair.size();
+ }
}
for (unsigned P = 0; P < Pairs; ++P) {
collectCommonLoops(Pair[SI].Dst,
LI->getLoopFor(Dst->getParent()),
Pair[SI].Loops);
- Result->Consistent = false;
- }
- else if (Pair[SI].Classification == Subscript::ZIV) {
+ Result.Consistent = false;
+ } else if (Pair[SI].Classification == Subscript::ZIV) {
// always separable
Separable.set(SI);
}
switch (Pair[SI].Classification) {
case Subscript::ZIV:
DEBUG(dbgs() << ", ZIV\n");
- if (testZIV(Pair[SI].Src, Pair[SI].Dst, *Result))
+ if (testZIV(Pair[SI].Src, Pair[SI].Dst, Result))
return nullptr;
break;
case Subscript::SIV: {
DEBUG(dbgs() << ", SIV\n");
unsigned Level;
const SCEV *SplitIter = nullptr;
- if (testSIV(Pair[SI].Src, Pair[SI].Dst, Level, *Result, NewConstraint,
+ if (testSIV(Pair[SI].Src, Pair[SI].Dst, Level, Result, NewConstraint,
SplitIter))
return nullptr;
break;
}
case Subscript::RDIV:
DEBUG(dbgs() << ", RDIV\n");
- if (testRDIV(Pair[SI].Src, Pair[SI].Dst, *Result))
+ if (testRDIV(Pair[SI].Src, Pair[SI].Dst, Result))
return nullptr;
break;
case Subscript::MIV:
DEBUG(dbgs() << ", MIV\n");
- if (testMIV(Pair[SI].Src, Pair[SI].Dst, Pair[SI].Loops, *Result))
+ if (testMIV(Pair[SI].Src, Pair[SI].Dst, Pair[SI].Loops, Result))
return nullptr;
break;
default:
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,
+ if (testSIV(Pair[SJ].Src, Pair[SJ].Dst, Level, Result, NewConstraint,
SplitIter))
return nullptr;
ConstrainedLevels.set(Level);
// SJ is an MIV subscript that's part of the current coupled group
DEBUG(dbgs() << "\tSJ = " << SJ << "\n");
if (propagate(Pair[SJ].Src, Pair[SJ].Dst, Pair[SJ].Loops,
- Constraints, Result->Consistent)) {
+ Constraints, Result.Consistent)) {
DEBUG(dbgs() << "\t Changed\n");
++DeltaPropagations;
Pair[SJ].Classification =
switch (Pair[SJ].Classification) {
case Subscript::ZIV:
DEBUG(dbgs() << "ZIV\n");
- if (testZIV(Pair[SJ].Src, Pair[SJ].Dst, *Result))
+ if (testZIV(Pair[SJ].Src, Pair[SJ].Dst, Result))
return nullptr;
Mivs.reset(SJ);
break;
for (int SJ = Mivs.find_first(); SJ >= 0; SJ = Mivs.find_next(SJ)) {
if (Pair[SJ].Classification == Subscript::RDIV) {
DEBUG(dbgs() << "RDIV test\n");
- if (testRDIV(Pair[SJ].Src, Pair[SJ].Dst, *Result))
+ if (testRDIV(Pair[SJ].Src, Pair[SJ].Dst, Result))
return nullptr;
// I don't yet understand how to propagate RDIV results
Mivs.reset(SJ);
for (int SJ = Mivs.find_first(); SJ >= 0; SJ = Mivs.find_next(SJ)) {
if (Pair[SJ].Classification == Subscript::MIV) {
DEBUG(dbgs() << "MIV test\n");
- if (testMIV(Pair[SJ].Src, Pair[SJ].Dst, Pair[SJ].Loops, *Result))
+ if (testMIV(Pair[SJ].Src, Pair[SJ].Dst, Pair[SJ].Loops, Result))
return nullptr;
}
else
llvm_unreachable("expected only MIV subscripts at this point");
}
- // update Result->DV from constraint vector
+ // update Result.DV from constraint vector
DEBUG(dbgs() << " updating\n");
- for (int SJ = ConstrainedLevels.find_first();
- SJ >= 0; SJ = ConstrainedLevels.find_next(SJ)) {
- updateDirection(Result->DV[SJ - 1], Constraints[SJ]);
- if (Result->DV[SJ - 1].Direction == Dependence::DVEntry::NONE)
+ 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;
}
}
CompleteLoops |= Pair[SI].Loops;
for (unsigned II = 1; II <= CommonLevels; ++II)
if (CompleteLoops[II])
- Result->DV[II - 1].Scalar = false;
+ Result.DV[II - 1].Scalar = false;
if (PossiblyLoopIndependent) {
// Make sure the LoopIndependent flag is set correctly.
// All directions must include equal, otherwise no
// loop-independent dependence is possible.
for (unsigned II = 1; II <= CommonLevels; ++II) {
- if (!(Result->getDirection(II) & Dependence::DVEntry::EQ)) {
- Result->LoopIndependent = false;
+ if (!(Result.getDirection(II) & Dependence::DVEntry::EQ)) {
+ Result.LoopIndependent = false;
break;
}
}
// loop-independent dependence possible, then no dependence exists.
bool AllEqual = true;
for (unsigned II = 1; II <= CommonLevels; ++II) {
- if (Result->getDirection(II) != Dependence::DVEntry::EQ) {
+ if (Result.getDirection(II) != Dependence::DVEntry::EQ) {
AllEqual = false;
break;
}
return nullptr;
}
- return std::move(Result);
+ return make_unique<FullDependence>(std::move(Result));
}
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);
Pair[0].Dst = DstSCEV;
}
- if (Delinearize && Pairs == 1 && CommonLevels > 1 &&
- tryDelinearize(Pair[0].Src, Pair[0].Dst, Pair, SE->getElementSize(Src))) {
- DEBUG(dbgs() << " delinerized GEP\n");
- Pairs = Pair.size();
+ if (Delinearize && CommonLevels > 1) {
+ if (tryDelinearize(Src, Dst, Pair)) {
+ DEBUG(dbgs() << " delinerized GEP\n");
+ Pairs = Pair.size();
+ }
}
for (unsigned P = 0; P < Pairs; ++P) {