mirror of
https://github.com/RPCS3/llvm-mirror.git
synced 2024-10-19 02:52:53 +02:00
Remove LoopDependenceAnalysis.
It was unmaintained and not much more than a stub. The new DependenceAnalysis pass is both more general and complete. llvm-svn: 166810
This commit is contained in:
Benjamin Kramer
parent
e6c62b245f
commit
0f18b5e49c
@ -77,6 +77,7 @@ perl -e '$/ = undef; for (split(/\n/, <>)) { s:^ *///? ?::; print " <p>\n" if !
|
||||
<tr><td><a href="#basicaa">-basicaa</a></td><td>Basic Alias Analysis (stateless AA impl)</td></tr>
|
||||
<tr><td><a href="#basiccg">-basiccg</a></td><td>Basic CallGraph Construction</td></tr>
|
||||
<tr><td><a href="#count-aa">-count-aa</a></td><td>Count Alias Analysis Query Responses</td></tr>
|
||||
<tr><td><a href="#da">-da</a></td><td>Dependence Analysis</td></tr>
|
||||
<tr><td><a href="#debug-aa">-debug-aa</a></td><td>AA use debugger</td></tr>
|
||||
<tr><td><a href="#domfrontier">-domfrontier</a></td><td>Dominance Frontier Construction</td></tr>
|
||||
<tr><td><a href="#domtree">-domtree</a></td><td>Dominator Tree Construction</td></tr>
|
||||
@ -92,7 +93,6 @@ perl -e '$/ = undef; for (split(/\n/, <>)) { s:^ *///? ?::; print " <p>\n" if !
|
||||
<tr><td><a href="#intervals">-intervals</a></td><td>Interval Partition Construction</td></tr>
|
||||
<tr><td><a href="#iv-users">-iv-users</a></td><td>Induction Variable Users</td></tr>
|
||||
<tr><td><a href="#lazy-value-info">-lazy-value-info</a></td><td>Lazy Value Information Analysis</td></tr>
|
||||
<tr><td><a href="#lda">-lda</a></td><td>Loop Dependence Analysis</td></tr>
|
||||
<tr><td><a href="#libcall-aa">-libcall-aa</a></td><td>LibCall Alias Analysis</td></tr>
|
||||
<tr><td><a href="#lint">-lint</a></td><td>Statically lint-checks LLVM IR</td></tr>
|
||||
<tr><td><a href="#loops">-loops</a></td><td>Natural Loop Information</td></tr>
|
||||
@ -249,6 +249,15 @@ perl -e '$/ = undef; for (split(/\n/, <>)) { s:^ *///? ?::; print " <p>\n" if !
|
||||
</p>
|
||||
</div>
|
||||
|
||||
<!-------------------------------------------------------------------------- -->
|
||||
<h3>
|
||||
<a name="da">-da: Dependence Analysis</a>
|
||||
</h3>
|
||||
<div>
|
||||
<p>Dependence analysis framework, which is used to detect dependences in
|
||||
memory accesses.</p>
|
||||
</div>
|
||||
|
||||
<!-------------------------------------------------------------------------- -->
|
||||
<h3>
|
||||
<a name="debug-aa">-debug-aa: AA use debugger</a>
|
||||
@ -431,15 +440,6 @@ perl -e '$/ = undef; for (split(/\n/, <>)) { s:^ *///? ?::; print " <p>\n" if !
|
||||
<p>Interface for lazy computation of value constraint information.</p>
|
||||
</div>
|
||||
|
||||
<!-------------------------------------------------------------------------- -->
|
||||
<h3>
|
||||
<a name="lda">-lda: Loop Dependence Analysis</a>
|
||||
</h3>
|
||||
<div>
|
||||
<p>Loop dependence analysis framework, which is used to detect dependences in
|
||||
memory accesses in loops.</p>
|
||||
</div>
|
||||
|
||||
<!-------------------------------------------------------------------------- -->
|
||||
<h3>
|
||||
<a name="libcall-aa">-libcall-aa: LibCall Alias Analysis</a>
|
||||
|
||||
@ -1,124 +0,0 @@
|
||||
//===- llvm/Analysis/LoopDependenceAnalysis.h --------------- -*- C++ -*---===//
|
||||
//
|
||||
// The LLVM Compiler Infrastructure
|
||||
//
|
||||
// This file is distributed under the University of Illinois Open Source
|
||||
// License. See LICENSE.TXT for details.
|
||||
//
|
||||
//===----------------------------------------------------------------------===//
|
||||
//
|
||||
// LoopDependenceAnalysis is an LLVM pass that analyses dependences in memory
|
||||
// accesses in loops.
|
||||
//
|
||||
// Please note that this is work in progress and the interface is subject to
|
||||
// change.
|
||||
//
|
||||
// TODO: adapt as interface progresses
|
||||
//
|
||||
//===----------------------------------------------------------------------===//
|
||||
|
||||
#ifndef LLVM_ANALYSIS_LOOP_DEPENDENCE_ANALYSIS_H
|
||||
#define LLVM_ANALYSIS_LOOP_DEPENDENCE_ANALYSIS_H
|
||||
|
||||
#include "llvm/ADT/DenseSet.h"
|
||||
#include "llvm/ADT/FoldingSet.h"
|
||||
#include "llvm/ADT/SmallVector.h"
|
||||
#include "llvm/Analysis/LoopPass.h"
|
||||
#include "llvm/Support/Allocator.h"
|
||||
|
||||
namespace llvm {
|
||||
|
||||
class AliasAnalysis;
|
||||
class AnalysisUsage;
|
||||
class ScalarEvolution;
|
||||
class SCEV;
|
||||
class Value;
|
||||
class raw_ostream;
|
||||
|
||||
class LoopDependenceAnalysis : public LoopPass {
|
||||
AliasAnalysis *AA;
|
||||
ScalarEvolution *SE;
|
||||
|
||||
/// L - The loop we are currently analysing.
|
||||
Loop *L;
|
||||
|
||||
/// TODO: doc
|
||||
enum DependenceResult { Independent = 0, Dependent = 1, Unknown = 2 };
|
||||
|
||||
/// TODO: doc
|
||||
struct Subscript {
|
||||
/// TODO: Add distance, direction, breaking conditions, ...
|
||||
};
|
||||
|
||||
/// DependencePair - Represents a data dependence relation between to memory
|
||||
/// reference instructions.
|
||||
struct DependencePair : public FastFoldingSetNode {
|
||||
Value *A;
|
||||
Value *B;
|
||||
DependenceResult Result;
|
||||
SmallVector<Subscript, 4> Subscripts;
|
||||
|
||||
DependencePair(const FoldingSetNodeID &ID, Value *a, Value *b) :
|
||||
FastFoldingSetNode(ID), A(a), B(b), Result(Unknown), Subscripts() {}
|
||||
};
|
||||
|
||||
/// findOrInsertDependencePair - Return true if a DependencePair for the
|
||||
/// given Values already exists, false if a new DependencePair had to be
|
||||
/// created. The third argument is set to the pair found or created.
|
||||
bool findOrInsertDependencePair(Value*, Value*, DependencePair*&);
|
||||
|
||||
/// getLoops - Collect all loops of the loop nest L in which
|
||||
/// a given SCEV is variant.
|
||||
void getLoops(const SCEV*, DenseSet<const Loop*>*) const;
|
||||
|
||||
/// isLoopInvariant - True if a given SCEV is invariant in all loops of the
|
||||
/// loop nest starting at the innermost loop L.
|
||||
bool isLoopInvariant(const SCEV*) const;
|
||||
|
||||
/// isAffine - An SCEV is affine with respect to the loop nest starting at
|
||||
/// the innermost loop L if it is of the form A+B*X where A, B are invariant
|
||||
/// in the loop nest and X is a induction variable in the loop nest.
|
||||
bool isAffine(const SCEV*) const;
|
||||
|
||||
/// TODO: doc
|
||||
bool isZIVPair(const SCEV*, const SCEV*) const;
|
||||
bool isSIVPair(const SCEV*, const SCEV*) const;
|
||||
DependenceResult analyseZIV(const SCEV*, const SCEV*, Subscript*) const;
|
||||
DependenceResult analyseSIV(const SCEV*, const SCEV*, Subscript*) const;
|
||||
DependenceResult analyseMIV(const SCEV*, const SCEV*, Subscript*) const;
|
||||
DependenceResult analyseSubscript(const SCEV*, const SCEV*, Subscript*) const;
|
||||
DependenceResult analysePair(DependencePair*) const;
|
||||
|
||||
public:
|
||||
static char ID; // Class identification, replacement for typeinfo
|
||||
LoopDependenceAnalysis() : LoopPass(ID) {
|
||||
initializeLoopDependenceAnalysisPass(*PassRegistry::getPassRegistry());
|
||||
}
|
||||
|
||||
/// isDependencePair - Check whether two values can possibly give rise to
|
||||
/// a data dependence: that is the case if both are instructions accessing
|
||||
/// memory and at least one of those accesses is a write.
|
||||
bool isDependencePair(const Value*, const Value*) const;
|
||||
|
||||
/// depends - Return a boolean indicating if there is a data dependence
|
||||
/// between two instructions.
|
||||
bool depends(Value*, Value*);
|
||||
|
||||
bool runOnLoop(Loop*, LPPassManager&);
|
||||
virtual void releaseMemory();
|
||||
virtual void getAnalysisUsage(AnalysisUsage&) const;
|
||||
void print(raw_ostream&, const Module* = 0) const;
|
||||
|
||||
private:
|
||||
FoldingSet<DependencePair> Pairs;
|
||||
BumpPtrAllocator PairAllocator;
|
||||
}; // class LoopDependenceAnalysis
|
||||
|
||||
// createLoopDependenceAnalysisPass - This creates an instance of the
|
||||
// LoopDependenceAnalysis pass.
|
||||
//
|
||||
LoopPass *createLoopDependenceAnalysisPass();
|
||||
|
||||
} // namespace llvm
|
||||
|
||||
#endif /* LLVM_ANALYSIS_LOOP_DEPENDENCE_ANALYSIS_H */
|
||||
@ -185,13 +185,6 @@ namespace llvm {
|
||||
//
|
||||
FunctionPass *createDependenceAnalysisPass();
|
||||
|
||||
//===--------------------------------------------------------------------===//
|
||||
//
|
||||
// createLoopDependenceAnalysisPass - This creates an instance of the
|
||||
// LoopDependenceAnalysis pass.
|
||||
//
|
||||
LoopPass *createLoopDependenceAnalysisPass();
|
||||
|
||||
//===--------------------------------------------------------------------===//
|
||||
//
|
||||
// Minor pass prototypes, allowing us to expose them through bugpoint and
|
||||
|
||||
@ -147,7 +147,6 @@ void initializeProfileMetadataLoaderPassPass(PassRegistry&);
|
||||
void initializePathProfileLoaderPassPass(PassRegistry&);
|
||||
void initializeLocalStackSlotPassPass(PassRegistry&);
|
||||
void initializeLoopDeletionPass(PassRegistry&);
|
||||
void initializeLoopDependenceAnalysisPass(PassRegistry&);
|
||||
void initializeLoopExtractorPass(PassRegistry&);
|
||||
void initializeLoopInfoPass(PassRegistry&);
|
||||
void initializeLoopInstSimplifyPass(PassRegistry&);
|
||||
|
||||
@ -86,7 +86,6 @@ namespace {
|
||||
(void) llvm::createLCSSAPass();
|
||||
(void) llvm::createLICMPass();
|
||||
(void) llvm::createLazyValueInfoPass();
|
||||
(void) llvm::createLoopDependenceAnalysisPass();
|
||||
(void) llvm::createLoopExtractorPass();
|
||||
(void) llvm::createLoopSimplifyPass();
|
||||
(void) llvm::createLoopStrengthReducePass();
|
||||
|
||||
@ -47,7 +47,6 @@ void llvm::initializeAnalysis(PassRegistry &Registry) {
|
||||
initializeLazyValueInfoPass(Registry);
|
||||
initializeLibCallAliasAnalysisPass(Registry);
|
||||
initializeLintPass(Registry);
|
||||
initializeLoopDependenceAnalysisPass(Registry);
|
||||
initializeLoopInfoPass(Registry);
|
||||
initializeMemDepPrinterPass(Registry);
|
||||
initializeMemoryDependenceAnalysisPass(Registry);
|
||||
|
||||
@ -27,7 +27,6 @@ add_llvm_library(LLVMAnalysis
|
||||
LibCallSemantics.cpp
|
||||
Lint.cpp
|
||||
Loads.cpp
|
||||
LoopDependenceAnalysis.cpp
|
||||
LoopInfo.cpp
|
||||
LoopPass.cpp
|
||||
MemDepPrinter.cpp
|
||||
|
||||
@ -1,362 +0,0 @@
|
||||
//===- LoopDependenceAnalysis.cpp - LDA Implementation ----------*- C++ -*-===//
|
||||
//
|
||||
// The LLVM Compiler Infrastructure
|
||||
//
|
||||
// This file is distributed under the University of Illinois Open Source
|
||||
// License. See LICENSE.TXT for details.
|
||||
//
|
||||
//===----------------------------------------------------------------------===//
|
||||
//
|
||||
// This is the (beginning) of an implementation of a loop dependence analysis
|
||||
// framework, which is used to detect dependences in memory accesses in loops.
|
||||
//
|
||||
// Please note that this is work in progress and the interface is subject to
|
||||
// change.
|
||||
//
|
||||
// TODO: adapt as implementation progresses.
|
||||
//
|
||||
// TODO: document lingo (pair, subscript, index)
|
||||
//
|
||||
//===----------------------------------------------------------------------===//
|
||||
|
||||
#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"
|
||||
#include "llvm/Support/Debug.h"
|
||||
#include "llvm/Support/ErrorHandling.h"
|
||||
#include "llvm/Support/raw_ostream.h"
|
||||
#include "llvm/DataLayout.h"
|
||||
using namespace llvm;
|
||||
|
||||
STATISTIC(NumAnswered, "Number of dependence queries answered");
|
||||
STATISTIC(NumAnalysed, "Number of distinct dependence pairs analysed");
|
||||
STATISTIC(NumDependent, "Number of pairs with dependent accesses");
|
||||
STATISTIC(NumIndependent, "Number of pairs with independent accesses");
|
||||
STATISTIC(NumUnknown, "Number of pairs with unknown accesses");
|
||||
|
||||
LoopPass *llvm::createLoopDependenceAnalysisPass() {
|
||||
return new LoopDependenceAnalysis();
|
||||
}
|
||||
|
||||
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;
|
||||
|
||||
//===----------------------------------------------------------------------===//
|
||||
// Utility Functions
|
||||
//===----------------------------------------------------------------------===//
|
||||
|
||||
static inline bool IsMemRefInstr(const Value *V) {
|
||||
const Instruction *I = dyn_cast<const Instruction>(V);
|
||||
return I && (I->mayReadFromMemory() || I->mayWriteToMemory());
|
||||
}
|
||||
|
||||
static void GetMemRefInstrs(const Loop *L,
|
||||
SmallVectorImpl<Instruction*> &Memrefs) {
|
||||
for (Loop::block_iterator b = L->block_begin(), be = L->block_end();
|
||||
b != be; ++b)
|
||||
for (BasicBlock::iterator i = (*b)->begin(), ie = (*b)->end();
|
||||
i != ie; ++i)
|
||||
if (IsMemRefInstr(i))
|
||||
Memrefs.push_back(i);
|
||||
}
|
||||
|
||||
static bool IsLoadOrStoreInst(Value *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) {
|
||||
if (LoadInst *i = dyn_cast<LoadInst>(I))
|
||||
return i->getPointerOperand();
|
||||
if (StoreInst *i = dyn_cast<StoreInst>(I))
|
||||
return i->getPointerOperand();
|
||||
llvm_unreachable("Value is no load or store instruction!");
|
||||
}
|
||||
|
||||
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 inline const SCEV *GetZeroSCEV(ScalarEvolution *SE) {
|
||||
return SE->getConstant(Type::getInt32Ty(SE->getContext()), 0L);
|
||||
}
|
||||
|
||||
//===----------------------------------------------------------------------===//
|
||||
// Dependence Testing
|
||||
//===----------------------------------------------------------------------===//
|
||||
|
||||
bool LoopDependenceAnalysis::isDependencePair(const Value *A,
|
||||
const Value *B) const {
|
||||
return IsMemRefInstr(A) &&
|
||||
IsMemRefInstr(B) &&
|
||||
(cast<const Instruction>(A)->mayWriteToMemory() ||
|
||||
cast<const Instruction>(B)->mayWriteToMemory());
|
||||
}
|
||||
|
||||
bool LoopDependenceAnalysis::findOrInsertDependencePair(Value *A,
|
||||
Value *B,
|
||||
DependencePair *&P) {
|
||||
void *insertPos = 0;
|
||||
FoldingSetNodeID id;
|
||||
id.AddPointer(A);
|
||||
id.AddPointer(B);
|
||||
|
||||
P = Pairs.FindNodeOrInsertPos(id, insertPos);
|
||||
if (P) return true;
|
||||
|
||||
P = new (PairAllocator) DependencePair(id, A, B);
|
||||
Pairs.InsertNode(P, insertPos);
|
||||
return false;
|
||||
}
|
||||
|
||||
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 (!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 {
|
||||
const SCEVAddRecExpr *rec = dyn_cast<SCEVAddRecExpr>(S);
|
||||
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(dbgs() << " Testing subscript: " << *A << ", " << *B << "\n");
|
||||
|
||||
if (A == B) {
|
||||
DEBUG(dbgs() << " -> [D] same SCEV\n");
|
||||
return Dependent;
|
||||
}
|
||||
|
||||
if (!isAffine(A) || !isAffine(B)) {
|
||||
DEBUG(dbgs() << " -> [?] not affine\n");
|
||||
return Unknown;
|
||||
}
|
||||
|
||||
if (isZIVPair(A, B))
|
||||
return analyseZIV(A, B, S);
|
||||
|
||||
if (isSIVPair(A, B))
|
||||
return analyseSIV(A, B, S);
|
||||
|
||||
return analyseMIV(A, B, S);
|
||||
}
|
||||
|
||||
LoopDependenceAnalysis::DependenceResult
|
||||
LoopDependenceAnalysis::analysePair(DependencePair *P) const {
|
||||
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(dbgs() << "--> [?] no load/store\n");
|
||||
return Unknown;
|
||||
}
|
||||
|
||||
Value *aPtr = GetPointerOperand(P->A);
|
||||
Value *bPtr = GetPointerOperand(P->B);
|
||||
|
||||
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(dbgs() << "---> [?] may alias\n");
|
||||
return Unknown;
|
||||
|
||||
case AliasAnalysis::NoAlias:
|
||||
// If the objects noalias, they are distinct, accesses are independent.
|
||||
DEBUG(dbgs() << "---> [I] no alias\n");
|
||||
return Independent;
|
||||
|
||||
case AliasAnalysis::MustAlias:
|
||||
break; // The underlying objects alias, test accesses for dependence.
|
||||
}
|
||||
|
||||
const GEPOperator *aGEP = dyn_cast<GEPOperator>(aPtr);
|
||||
const GEPOperator *bGEP = dyn_cast<GEPOperator>(bPtr);
|
||||
|
||||
if (!aGEP || !bGEP)
|
||||
return Unknown;
|
||||
|
||||
// FIXME: Is filtering coupled subscripts necessary?
|
||||
|
||||
// Collect GEP operand pairs (FIXME: use GetGEPOperands from BasicAA), adding
|
||||
// trailing zeroes to the smaller GEP, if needed.
|
||||
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(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);
|
||||
}
|
||||
// We successfully analysed all subscripts but failed to prove independence.
|
||||
return Dependent;
|
||||
}
|
||||
|
||||
bool LoopDependenceAnalysis::depends(Value *A, Value *B) {
|
||||
assert(isDependencePair(A, B) && "Values form no dependence pair!");
|
||||
++NumAnswered;
|
||||
|
||||
DependencePair *p;
|
||||
if (!findOrInsertDependencePair(A, B, p)) {
|
||||
// The pair is not cached, so analyse it.
|
||||
++NumAnalysed;
|
||||
switch (p->Result = analysePair(p)) {
|
||||
case Dependent: ++NumDependent; break;
|
||||
case Independent: ++NumIndependent; break;
|
||||
case Unknown: ++NumUnknown; break;
|
||||
}
|
||||
}
|
||||
return p->Result != Independent;
|
||||
}
|
||||
|
||||
//===----------------------------------------------------------------------===//
|
||||
// LoopDependenceAnalysis Implementation
|
||||
//===----------------------------------------------------------------------===//
|
||||
|
||||
bool LoopDependenceAnalysis::runOnLoop(Loop *L, LPPassManager &) {
|
||||
this->L = L;
|
||||
AA = &getAnalysis<AliasAnalysis>();
|
||||
SE = &getAnalysis<ScalarEvolution>();
|
||||
return false;
|
||||
}
|
||||
|
||||
void LoopDependenceAnalysis::releaseMemory() {
|
||||
Pairs.clear();
|
||||
PairAllocator.Reset();
|
||||
}
|
||||
|
||||
void LoopDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
|
||||
AU.setPreservesAll();
|
||||
AU.addRequiredTransitive<AliasAnalysis>();
|
||||
AU.addRequiredTransitive<ScalarEvolution>();
|
||||
}
|
||||
|
||||
static void PrintLoopInfo(raw_ostream &OS,
|
||||
LoopDependenceAnalysis *LDA, const Loop *L) {
|
||||
if (!L->empty()) return; // ignore non-innermost loops
|
||||
|
||||
SmallVector<Instruction*, 8> memrefs;
|
||||
GetMemRefInstrs(L, memrefs);
|
||||
|
||||
OS << "Loop at depth " << L->getLoopDepth() << ", header block: ";
|
||||
WriteAsOperand(OS, L->getHeader(), false);
|
||||
OS << "\n";
|
||||
|
||||
OS << " Load/store instructions: " << memrefs.size() << "\n";
|
||||
for (SmallVector<Instruction*, 8>::const_iterator x = memrefs.begin(),
|
||||
end = memrefs.end(); x != end; ++x)
|
||||
OS << "\t" << (x - memrefs.begin()) << ": " << **x << "\n";
|
||||
|
||||
OS << " Pairwise dependence results:\n";
|
||||
for (SmallVector<Instruction*, 8>::const_iterator x = memrefs.begin(),
|
||||
end = memrefs.end(); x != end; ++x)
|
||||
for (SmallVector<Instruction*, 8>::const_iterator y = x + 1;
|
||||
y != end; ++y)
|
||||
if (LDA->isDependencePair(*x, *y))
|
||||
OS << "\t" << (x - memrefs.begin()) << "," << (y - memrefs.begin())
|
||||
<< ": " << (LDA->depends(*x, *y) ? "dependent" : "independent")
|
||||
<< "\n";
|
||||
}
|
||||
|
||||
void LoopDependenceAnalysis::print(raw_ostream &OS, const Module*) const {
|
||||
// TODO: doc why const_cast is safe
|
||||
PrintLoopInfo(OS, const_cast<LoopDependenceAnalysis*>(this), this->L);
|
||||
}
|
||||
@ -1,44 +0,0 @@
|
||||
; RUN: opt < %s -analyze -basicaa -lda | FileCheck %s
|
||||
|
||||
;; x[5] = x[6] // with x being a pointer passed as argument
|
||||
|
||||
define void @f1(i32* nocapture %xptr) nounwind {
|
||||
entry:
|
||||
%x.ld.addr = getelementptr i32* %xptr, i64 6
|
||||
%x.st.addr = getelementptr i32* %xptr, i64 5
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%x = load i32* %x.ld.addr
|
||||
store i32 %x, i32* %x.st.addr
|
||||
; CHECK: 0,1: dep
|
||||
%i.next = add i64 %i, 1
|
||||
%exitcond = icmp eq i64 %i.next, 256
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
|
||||
;; x[5] = x[6] // with x being an array on the stack
|
||||
|
||||
define void @foo(...) nounwind {
|
||||
entry:
|
||||
%xptr = alloca [256 x i32], align 4
|
||||
%x.ld.addr = getelementptr [256 x i32]* %xptr, i64 0, i64 6
|
||||
%x.st.addr = getelementptr [256 x i32]* %xptr, i64 0, i64 5
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%x = load i32* %x.ld.addr
|
||||
store i32 %x, i32* %x.st.addr
|
||||
; CHECK: 0,1: ind
|
||||
%i.next = add i64 %i, 1
|
||||
%exitcond = icmp eq i64 %i.next, 256
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
@ -1 +0,0 @@
|
||||
config.suffixes = ['.ll', '.c', '.cpp']
|
||||
@ -1,110 +0,0 @@
|
||||
; RUN: opt < %s -analyze -basicaa -lda | FileCheck %s
|
||||
|
||||
@x = common global [256 x i32] zeroinitializer, align 4
|
||||
@y = common global [256 x i32] zeroinitializer, align 4
|
||||
|
||||
;; for (i = 0; i < 256; i++)
|
||||
;; x[i] = x[i] + y[i]
|
||||
|
||||
define void @f1(...) nounwind {
|
||||
entry:
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%y.addr = getelementptr [256 x i32]* @y, i64 0, i64 %i
|
||||
%x.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i
|
||||
%x = load i32* %x.addr ; 0
|
||||
%y = load i32* %y.addr ; 1
|
||||
%r = add i32 %y, %x
|
||||
store i32 %r, i32* %x.addr ; 2
|
||||
; CHECK: 0,2: dep
|
||||
; CHECK: 1,2: ind
|
||||
%i.next = add i64 %i, 1
|
||||
%exitcond = icmp eq i64 %i.next, 256
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
|
||||
;; for (i = 0; i < 256; i++)
|
||||
;; x[i+1] = x[i] + y[i]
|
||||
|
||||
define void @f2(...) nounwind {
|
||||
entry:
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%y.ld.addr = getelementptr [256 x i32]* @y, i64 0, i64 %i
|
||||
%x.ld.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i
|
||||
%i.next = add i64 %i, 1
|
||||
%x.st.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i.next
|
||||
%x = load i32* %x.ld.addr ; 0
|
||||
%y = load i32* %y.ld.addr ; 1
|
||||
%r = add i32 %y, %x
|
||||
store i32 %r, i32* %x.st.addr ; 2
|
||||
; CHECK: 0,2: dep
|
||||
; CHECK: 1,2: ind
|
||||
%exitcond = icmp eq i64 %i.next, 256
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
|
||||
;; for (i = 0; i < 10; i++)
|
||||
;; x[i+20] = x[i] + y[i]
|
||||
|
||||
define void @f3(...) nounwind {
|
||||
entry:
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%y.ld.addr = getelementptr [256 x i32]* @y, i64 0, i64 %i
|
||||
%x.ld.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i
|
||||
%i.20 = add i64 %i, 20
|
||||
%x.st.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i.20
|
||||
%x = load i32* %x.ld.addr ; 0
|
||||
%y = load i32* %y.ld.addr ; 1
|
||||
%r = add i32 %y, %x
|
||||
store i32 %r, i32* %x.st.addr ; 2
|
||||
; CHECK: 0,2: dep
|
||||
; CHECK: 1,2: ind
|
||||
%i.next = add i64 %i, 1
|
||||
%exitcond = icmp eq i64 %i.next, 10
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
|
||||
;; for (i = 0; i < 10; i++)
|
||||
;; x[10*i+1] = x[10*i] + y[i]
|
||||
|
||||
define void @f4(...) nounwind {
|
||||
entry:
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%i.10 = mul i64 %i, 10
|
||||
%y.ld.addr = getelementptr [256 x i32]* @y, i64 0, i64 %i.10
|
||||
%x.ld.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i.10
|
||||
%i.10.1 = add i64 %i.10, 1
|
||||
%x.st.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i.10.1
|
||||
%x = load i32* %x.ld.addr ; 0
|
||||
%y = load i32* %y.ld.addr ; 1
|
||||
%r = add i32 %y, %x
|
||||
store i32 %r, i32* %x.st.addr ; 2
|
||||
; CHECK: 0,2: dep
|
||||
; CHECK: 1,2: ind
|
||||
%i.next = add i64 %i, 1
|
||||
%exitcond = icmp eq i64 %i.next, 10
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
@ -1,118 +0,0 @@
|
||||
; RUN: opt < %s -analyze -basicaa -lda | FileCheck %s
|
||||
|
||||
@x = common global [256 x i32] zeroinitializer, align 4
|
||||
@y = common global [256 x i32] zeroinitializer, align 4
|
||||
|
||||
;; for (i = 0; i < 256; i++)
|
||||
;; x[i] = x[255 - i] + y[i]
|
||||
|
||||
define void @f1(...) nounwind {
|
||||
entry:
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%i.255 = sub i64 255, %i
|
||||
%y.ld.addr = getelementptr [256 x i32]* @y, i64 0, i64 %i
|
||||
%x.ld.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i.255
|
||||
%x.st.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i
|
||||
%x = load i32* %x.ld.addr ; 0
|
||||
%y = load i32* %y.ld.addr ; 1
|
||||
%r = add i32 %y, %x
|
||||
store i32 %r, i32* %x.st.addr ; 2
|
||||
; CHECK: 0,2: dep
|
||||
; CHECK: 1,2: ind
|
||||
%i.next = add i64 %i, 1
|
||||
%exitcond = icmp eq i64 %i.next, 256
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
|
||||
;; for (i = 0; i < 100; i++)
|
||||
;; x[i] = x[255 - i] + y[i]
|
||||
|
||||
define void @f2(...) nounwind {
|
||||
entry:
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%i.255 = sub i64 255, %i
|
||||
%y.ld.addr = getelementptr [256 x i32]* @y, i64 0, i64 %i
|
||||
%x.ld.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i.255
|
||||
%x.st.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i
|
||||
%x = load i32* %x.ld.addr ; 0
|
||||
%y = load i32* %y.ld.addr ; 1
|
||||
%r = add i32 %y, %x
|
||||
store i32 %r, i32* %x.st.addr ; 2
|
||||
; CHECK: 0,2: dep
|
||||
; CHECK: 1,2: ind
|
||||
%i.next = add i64 %i, 1
|
||||
%exitcond = icmp eq i64 %i.next, 100
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
|
||||
;; // the first iteration (i=0) leads to an out-of-bounds access of x. as the
|
||||
;; // result of this access is undefined, _any_ dependence result is safe.
|
||||
;; for (i = 0; i < 256; i++)
|
||||
;; x[i] = x[256 - i] + y[i]
|
||||
|
||||
define void @f3(...) nounwind {
|
||||
entry:
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%i.256 = sub i64 0, %i
|
||||
%y.ld.addr = getelementptr [256 x i32]* @y, i64 0, i64 %i
|
||||
%x.ld.addr = getelementptr [256 x i32]* @x, i64 1, i64 %i.256
|
||||
%x.st.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i
|
||||
%x = load i32* %x.ld.addr ; 0
|
||||
%y = load i32* %y.ld.addr ; 1
|
||||
%r = add i32 %y, %x
|
||||
store i32 %r, i32* %x.st.addr ; 2
|
||||
; CHECK: 0,2: dep
|
||||
; CHECK: 1,2:
|
||||
%i.next = add i64 %i, 1
|
||||
%exitcond = icmp eq i64 %i.next, 256
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
|
||||
;; // slightly contrived but valid IR for the following loop, where all
|
||||
;; // accesses in all iterations are within bounds. while this example's first
|
||||
;; // (ZIV-)subscript is (0, 1), accesses are dependent.
|
||||
;; for (i = 1; i < 256; i++)
|
||||
;; x[i] = x[256 - i] + y[i]
|
||||
|
||||
define void @f4(...) nounwind {
|
||||
entry:
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%i.1 = add i64 1, %i
|
||||
%i.256 = sub i64 -1, %i
|
||||
%y.ld.addr = getelementptr [256 x i32]* @y, i64 0, i64 %i.1
|
||||
%x.ld.addr = getelementptr [256 x i32]* @x, i64 1, i64 %i.256
|
||||
%x.st.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i.1
|
||||
%x = load i32* %x.ld.addr ; 0
|
||||
%y = load i32* %y.ld.addr ; 1
|
||||
%r = add i32 %y, %x
|
||||
store i32 %r, i32* %x.st.addr ; 2
|
||||
; CHECK: 0,2: dep
|
||||
; CHECK: 1,2: ind
|
||||
%i.next = add i64 %i, 1
|
||||
%exitcond = icmp eq i64 %i.next, 256
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
@ -1,56 +0,0 @@
|
||||
; RUN: opt < %s -analyze -basicaa -lda | FileCheck %s
|
||||
|
||||
@x = common global [256 x i32] zeroinitializer, align 4
|
||||
@y = common global [256 x i32] zeroinitializer, align 4
|
||||
|
||||
;; for (i = 0; i < 256; i++)
|
||||
;; x[i] = x[42] + y[i]
|
||||
|
||||
define void @f1(...) nounwind {
|
||||
entry:
|
||||
%x.ld.addr = getelementptr [256 x i32]* @x, i64 0, i64 42
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%x.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i
|
||||
%y.addr = getelementptr [256 x i32]* @y, i64 0, i64 %i
|
||||
%x = load i32* %x.ld.addr ; 0
|
||||
%y = load i32* %y.addr ; 1
|
||||
%r = add i32 %y, %x
|
||||
store i32 %r, i32* %x.addr ; 2
|
||||
; CHECK: 0,2: dep
|
||||
; CHECK: 1,2: ind
|
||||
%i.next = add i64 %i, 1
|
||||
%exitcond = icmp eq i64 %i.next, 256
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
|
||||
;; for (i = 0; i < 250; i++)
|
||||
;; x[i] = x[255] + y[i]
|
||||
|
||||
define void @f2(...) nounwind {
|
||||
entry:
|
||||
%x.ld.addr = getelementptr [256 x i32]* @x, i64 0, i64 255
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%x.addr = getelementptr [256 x i32]* @x, i64 0, i64 %i
|
||||
%y.addr = getelementptr [256 x i32]* @y, i64 0, i64 %i
|
||||
%x = load i32* %x.ld.addr ; 0
|
||||
%y = load i32* %y.addr ; 1
|
||||
%r = add i32 %y, %x
|
||||
store i32 %r, i32* %x.addr ; 2
|
||||
; CHECK: 0,2: dep
|
||||
; CHECK: 1,2: ind
|
||||
%i.next = add i64 %i, 1
|
||||
%exitcond = icmp eq i64 %i.next, 250
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
@ -1,63 +0,0 @@
|
||||
; RUN: opt < %s -analyze -basicaa -lda | FileCheck %s
|
||||
|
||||
@x = common global [256 x i32] zeroinitializer, align 4
|
||||
|
||||
;; x[5] = x[6]
|
||||
|
||||
define void @f1(...) nounwind {
|
||||
entry:
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%x = load i32* getelementptr ([256 x i32]* @x, i32 0, i64 6)
|
||||
store i32 %x, i32* getelementptr ([256 x i32]* @x, i32 0, i64 5)
|
||||
; CHECK: 0,1: ind
|
||||
%i.next = add i64 %i, 1
|
||||
%exitcond = icmp eq i64 %i.next, 256
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
|
||||
;; x[c] = x[c+1] // with c being a loop-invariant constant
|
||||
|
||||
define void @f2(i64 %c0) nounwind {
|
||||
entry:
|
||||
%c1 = add i64 %c0, 1
|
||||
%x.ld.addr = getelementptr [256 x i32]* @x, i64 0, i64 %c0
|
||||
%x.st.addr = getelementptr [256 x i32]* @x, i64 0, i64 %c1
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%x = load i32* %x.ld.addr
|
||||
store i32 %x, i32* %x.st.addr
|
||||
; CHECK: 0,1: ind
|
||||
%i.next = add i64 %i, 1
|
||||
%exitcond = icmp eq i64 %i.next, 256
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
|
||||
;; x[6] = x[6]
|
||||
|
||||
define void @f3(...) nounwind {
|
||||
entry:
|
||||
br label %for.body
|
||||
|
||||
for.body:
|
||||
%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
|
||||
%x = load i32* getelementptr ([256 x i32]* @x, i32 0, i64 6)
|
||||
store i32 %x, i32* getelementptr ([256 x i32]* @x, i32 0, i64 6)
|
||||
; CHECK: 0,1: dep
|
||||
%i.next = add i64 %i, 1
|
||||
%exitcond = icmp eq i64 %i.next, 256
|
||||
br i1 %exitcond, label %for.end, label %for.body
|
||||
|
||||
for.end:
|
||||
ret void
|
||||
}
|
||||
Loading…
Reference in New Issue
Block a user