//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "lda"
+#include "llvm/ADT/DenseSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/LoopDependenceAnalysis.h"
#include "llvm/Analysis/LoopPass.h"
#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/Assembly/Writer.h"
#include "llvm/Instructions.h"
#include "llvm/Operator.h"
#include "llvm/Support/Allocator.h"
return new LoopDependenceAnalysis();
}
-static RegisterPass<LoopDependenceAnalysis>
-R("lda", "Loop Dependence Analysis", false, true);
+INITIALIZE_PASS_BEGIN(LoopDependenceAnalysis, "lda",
+ "Loop Dependence Analysis", false, true)
+INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
+INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
+INITIALIZE_PASS_END(LoopDependenceAnalysis, "lda",
+ "Loop Dependence Analysis", false, true)
char LoopDependenceAnalysis::ID = 0;
//===----------------------------------------------------------------------===//
}
static bool IsLoadOrStoreInst(Value *I) {
- return isa<LoadInst>(I) || isa<StoreInst>(I);
+ // Returns true if the load or store can be analyzed. Atomic and volatile
+ // operations have properties which this analysis does not understand.
+ if (LoadInst *LI = dyn_cast<LoadInst>(I))
+ return LI->isUnordered();
+ else if (StoreInst *SI = dyn_cast<StoreInst>(I))
+ return SI->isUnordered();
+ return false;
}
static Value *GetPointerOperand(Value *I) {
static AliasAnalysis::AliasResult UnderlyingObjectsAlias(AliasAnalysis *AA,
const Value *A,
const Value *B) {
- const Value *aObj = A->getUnderlyingObject();
- const Value *bObj = B->getUnderlyingObject();
+ const Value *aObj = GetUnderlyingObject(A);
+ const Value *bObj = GetUnderlyingObject(B);
return AA->alias(aObj, AA->getTypeStoreSize(aObj->getType()),
bObj, AA->getTypeStoreSize(bObj->getType()));
}
static inline const SCEV *GetZeroSCEV(ScalarEvolution *SE) {
- return SE->getConstant(Type::Int32Ty, 0L);
+ return SE->getConstant(Type::getInt32Ty(SE->getContext()), 0L);
}
//===----------------------------------------------------------------------===//
P = Pairs.FindNodeOrInsertPos(id, insertPos);
if (P) return true;
- P = PairAllocator.Allocate<DependencePair>();
- new (P) DependencePair(id, A, B);
+ P = new (PairAllocator) DependencePair(id, A, B);
Pairs.InsertNode(P, insertPos);
return false;
}
-bool LoopDependenceAnalysis::isLoopInvariant(const SCEV *S) const {
+void LoopDependenceAnalysis::getLoops(const SCEV *S,
+ DenseSet<const Loop*>* Loops) const {
+ // Refactor this into an SCEVVisitor, if efficiency becomes a concern.
for (const Loop *L = this->L; L != 0; L = L->getParentLoop())
- if (!S->isLoopInvariant(L))
- return false;
- return true;
+ if (!SE->isLoopInvariant(S, L))
+ Loops->insert(L);
+}
+
+bool LoopDependenceAnalysis::isLoopInvariant(const SCEV *S) const {
+ DenseSet<const Loop*> loops;
+ getLoops(S, &loops);
+ return loops.empty();
}
bool LoopDependenceAnalysis::isAffine(const SCEV *S) const {
return isLoopInvariant(S) || (rec && rec->isAffine());
}
+bool LoopDependenceAnalysis::isZIVPair(const SCEV *A, const SCEV *B) const {
+ return isLoopInvariant(A) && isLoopInvariant(B);
+}
+
+bool LoopDependenceAnalysis::isSIVPair(const SCEV *A, const SCEV *B) const {
+ DenseSet<const Loop*> loops;
+ getLoops(A, &loops);
+ getLoops(B, &loops);
+ return loops.size() == 1;
+}
+
+LoopDependenceAnalysis::DependenceResult
+LoopDependenceAnalysis::analyseZIV(const SCEV *A,
+ const SCEV *B,
+ Subscript *S) const {
+ assert(isZIVPair(A, B) && "Attempted to ZIV-test non-ZIV SCEVs!");
+ return A == B ? Dependent : Independent;
+}
+
+LoopDependenceAnalysis::DependenceResult
+LoopDependenceAnalysis::analyseSIV(const SCEV *A,
+ const SCEV *B,
+ Subscript *S) const {
+ return Unknown; // TODO: Implement.
+}
+
+LoopDependenceAnalysis::DependenceResult
+LoopDependenceAnalysis::analyseMIV(const SCEV *A,
+ const SCEV *B,
+ Subscript *S) const {
+ return Unknown; // TODO: Implement.
+}
+
LoopDependenceAnalysis::DependenceResult
LoopDependenceAnalysis::analyseSubscript(const SCEV *A,
const SCEV *B,
Subscript *S) const {
- DEBUG(errs() << " Testing subscript: " << *A << ", " << *B << "\n");
+ DEBUG(dbgs() << " Testing subscript: " << *A << ", " << *B << "\n");
if (A == B) {
- DEBUG(errs() << " -> [D] same SCEV\n");
+ DEBUG(dbgs() << " -> [D] same SCEV\n");
return Dependent;
}
if (!isAffine(A) || !isAffine(B)) {
- DEBUG(errs() << " -> [?] not affine\n");
+ DEBUG(dbgs() << " -> [?] not affine\n");
return Unknown;
}
- // TODO: Implement ZIV/SIV/MIV testers.
+ if (isZIVPair(A, B))
+ return analyseZIV(A, B, S);
+
+ if (isSIVPair(A, B))
+ return analyseSIV(A, B, S);
- DEBUG(errs() << " -> [?] cannot analyse subscript\n");
- return Unknown;
+ return analyseMIV(A, B, S);
}
LoopDependenceAnalysis::DependenceResult
LoopDependenceAnalysis::analysePair(DependencePair *P) const {
- DEBUG(errs() << "Analysing:\n" << *P->A << "\n" << *P->B << "\n");
+ DEBUG(dbgs() << "Analysing:\n" << *P->A << "\n" << *P->B << "\n");
// We only analyse loads and stores but no possible memory accesses by e.g.
// free, call, or invoke instructions.
if (!IsLoadOrStoreInst(P->A) || !IsLoadOrStoreInst(P->B)) {
- DEBUG(errs() << "--> [?] no load/store\n");
+ DEBUG(dbgs() << "--> [?] no load/store\n");
return Unknown;
}
switch (UnderlyingObjectsAlias(AA, aPtr, bPtr)) {
case AliasAnalysis::MayAlias:
+ case AliasAnalysis::PartialAlias:
// We can not analyse objects if we do not know about their aliasing.
- DEBUG(errs() << "---> [?] may alias\n");
+ DEBUG(dbgs() << "---> [?] may alias\n");
return Unknown;
case AliasAnalysis::NoAlias:
// If the objects noalias, they are distinct, accesses are independent.
- DEBUG(errs() << "---> [I] no alias\n");
+ DEBUG(dbgs() << "---> [I] no alias\n");
return Independent;
case AliasAnalysis::MustAlias:
// FIXME: Is filtering coupled subscripts necessary?
- // Analyse indices pairwise (FIXME: use GetGEPOperands from BasicAA), adding
+ // Collect GEP operand pairs (FIXME: use GetGEPOperands from BasicAA), adding
// trailing zeroes to the smaller GEP, if needed.
- GEPOperator::const_op_iterator aIdx = aGEP->idx_begin(),
- aEnd = aGEP->idx_end(),
- bIdx = bGEP->idx_begin(),
- bEnd = bGEP->idx_end();
- while (aIdx != aEnd && bIdx != bEnd) {
+ typedef SmallVector<std::pair<const SCEV*, const SCEV*>, 4> GEPOpdPairsTy;
+ GEPOpdPairsTy opds;
+ for(GEPOperator::const_op_iterator aIdx = aGEP->idx_begin(),
+ aEnd = aGEP->idx_end(),
+ bIdx = bGEP->idx_begin(),
+ bEnd = bGEP->idx_end();
+ aIdx != aEnd && bIdx != bEnd;
+ aIdx += (aIdx != aEnd), bIdx += (bIdx != bEnd)) {
const SCEV* aSCEV = (aIdx != aEnd) ? SE->getSCEV(*aIdx) : GetZeroSCEV(SE);
const SCEV* bSCEV = (bIdx != bEnd) ? SE->getSCEV(*bIdx) : GetZeroSCEV(SE);
+ opds.push_back(std::make_pair(aSCEV, bSCEV));
+ }
+
+ if (!opds.empty() && opds[0].first != opds[0].second) {
+ // We cannot (yet) handle arbitrary GEP pointer offsets. By limiting
+ //
+ // TODO: this could be relaxed by adding the size of the underlying object
+ // to the first subscript. If we have e.g. (GEP x,0,i; GEP x,2,-i) and we
+ // know that x is a [100 x i8]*, we could modify the first subscript to be
+ // (i, 200-i) instead of (i, -i).
+ return Unknown;
+ }
+
+ // Now analyse the collected operand pairs (skipping the GEP ptr offsets).
+ for (GEPOpdPairsTy::const_iterator i = opds.begin() + 1, end = opds.end();
+ i != end; ++i) {
Subscript subscript;
- DependenceResult result = analyseSubscript(aSCEV, bSCEV, &subscript);
+ DependenceResult result = analyseSubscript(i->first, i->second, &subscript);
if (result != Dependent) {
// We either proved independence or failed to analyse this subscript.
// Further subscripts will not improve the situation, so abort early.
return result;
}
P->Subscripts.push_back(subscript);
- if (aIdx != aEnd) ++aIdx;
- if (bIdx != bEnd) ++bIdx;
}
- // Either there were no subscripts or all subscripts were analysed to be
- // dependent; in both cases we know the accesses are dependent.
+ // We successfully analysed all subscripts but failed to prove independence.
return Dependent;
}
// TODO: doc why const_cast is safe
PrintLoopInfo(OS, const_cast<LoopDependenceAnalysis*>(this), this->L);
}
-
-void LoopDependenceAnalysis::print(std::ostream &OS, const Module *M) const {
- raw_os_ostream os(OS);
- print(os, M);
-}