//
// The LLVM Compiler Infrastructure
//
-// This file was developed by the Owen Anderson and is distributed under
-// the University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
#include "llvm/Function.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Support/CFG.h"
+#include "llvm/Support/CommandLine.h"
#include "llvm/Target/TargetData.h"
+#include "llvm/ADT/Statistic.h"
+
+#define DEBUG_TYPE "memdep"
using namespace llvm;
+// Control the calculation of non-local dependencies by only examining the
+// predecessors if the basic block has less than X amount (50 by default).
+static cl::opt<int>
+PredLimit("nonlocaldep-threshold", cl::Hidden, cl::init(50),
+ cl::desc("Control the calculation of non-local"
+ "dependencies (default = 50)"));
+
+STATISTIC(NumCacheNonlocal, "Number of cached non-local responses");
+STATISTIC(NumUncacheNonlocal, "Number of uncached non-local responses");
+
char MemoryDependenceAnalysis::ID = 0;
-const Instruction* MemoryDependenceAnalysis::NonLocal = (Instruction*)-3;
-const Instruction* MemoryDependenceAnalysis::None = (Instruction*)-4;
+Instruction* const MemoryDependenceAnalysis::NonLocal = (Instruction*)-3;
+Instruction* const MemoryDependenceAnalysis::None = (Instruction*)-4;
+Instruction* const MemoryDependenceAnalysis::Dirty = (Instruction*)-5;
// Register this pass...
static RegisterPass<MemoryDependenceAnalysis> X("memdep",
- "Memory Dependence Analysis");
+ "Memory Dependence Analysis", false, true);
+
+void MemoryDependenceAnalysis::ping(Instruction *D) {
+ for (depMapType::iterator I = depGraphLocal.begin(), E = depGraphLocal.end();
+ I != E; ++I) {
+ assert(I->first != D);
+ assert(I->second.first != D);
+ }
+
+ for (nonLocalDepMapType::iterator I = depGraphNonLocal.begin(), E = depGraphNonLocal.end();
+ I != E; ++I) {
+ assert(I->first != D);
+ for (DenseMap<BasicBlock*, Value*>::iterator II = I->second.begin(),
+ EE = I->second.end(); II != EE; ++II)
+ assert(II->second != D);
+ }
+
+ for (reverseDepMapType::iterator I = reverseDep.begin(), E = reverseDep.end();
+ I != E; ++I)
+ for (SmallPtrSet<Instruction*, 4>::iterator II = I->second.begin(), EE = I->second.end();
+ II != EE; ++II)
+ assert(*II != D);
+
+ for (reverseDepMapType::iterator I = reverseDepNonLocal.begin(), E = reverseDepNonLocal.end();
+ I != E; ++I)
+ for (SmallPtrSet<Instruction*, 4>::iterator II = I->second.begin(), EE = I->second.end();
+ II != EE; ++II)
+ assert(*II != D);
+}
/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
///
/// getCallSiteDependency - Private helper for finding the local dependencies
/// of a call site.
-const Instruction* MemoryDependenceAnalysis::getCallSiteDependency(CallSite C,
+Instruction* MemoryDependenceAnalysis::getCallSiteDependency(CallSite C,
Instruction* start,
BasicBlock* block) {
+ std::pair<Instruction*, bool>& cachedResult =
+ depGraphLocal[C.getInstruction()];
AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
TargetData& TD = getAnalysis<TargetData>();
BasicBlock::iterator blockBegin = C.getInstruction()->getParent()->begin();
// If the starting point was specifiy, use it
if (start) {
QI = start;
- blockBegin = start->getParent()->end();
+ blockBegin = start->getParent()->begin();
// If the starting point wasn't specified, but the block was, use it
} else if (!start && block) {
QI = block->end();
- blockBegin = block->end();
+ blockBegin = block->begin();
}
// Walk backwards through the block, looking for dependencies
uint64_t pointerSize = 0;
if (StoreInst* S = dyn_cast<StoreInst>(QI)) {
pointer = S->getPointerOperand();
- pointerSize = TD.getTypeSize(S->getOperand(0)->getType());
- } else if (LoadInst* L = dyn_cast<LoadInst>(QI)) {
- pointer = L->getPointerOperand();
- pointerSize = TD.getTypeSize(L->getType());
+ pointerSize = TD.getTypeStoreSize(S->getOperand(0)->getType());
} else if (AllocationInst* AI = dyn_cast<AllocationInst>(QI)) {
pointer = AI;
if (ConstantInt* C = dyn_cast<ConstantInt>(AI->getArraySize()))
pointerSize = C->getZExtValue() * \
- TD.getTypeSize(AI->getAllocatedType());
+ TD.getABITypeSize(AI->getAllocatedType());
else
pointerSize = ~0UL;
} else if (VAArgInst* V = dyn_cast<VAArgInst>(QI)) {
pointer = V->getOperand(0);
- pointerSize = TD.getTypeSize(V->getType());
+ pointerSize = TD.getTypeStoreSize(V->getType());
} else if (FreeInst* F = dyn_cast<FreeInst>(QI)) {
pointer = F->getPointerOperand();
// FreeInsts erase the entire structure
pointerSize = ~0UL;
} else if (CallSite::get(QI).getInstruction() != 0) {
- if (AA.getModRefInfo(C, CallSite::get(QI)) != AliasAnalysis::NoModRef) {
+ AliasAnalysis::ModRefBehavior result =
+ AA.getModRefBehavior(CallSite::get(QI));
+ if (result != AliasAnalysis::DoesNotAccessMemory) {
if (!start && !block) {
- depGraphLocal.insert(std::make_pair(C.getInstruction(),
- std::make_pair(QI, true)));
+ cachedResult.first = QI;
+ cachedResult.second = true;
reverseDep[QI].insert(C.getInstruction());
}
return QI;
if (AA.getModRefInfo(C, pointer, pointerSize) != AliasAnalysis::NoModRef) {
if (!start && !block) {
- depGraphLocal.insert(std::make_pair(C.getInstruction(),
- std::make_pair(QI, true)));
+ cachedResult.first = QI;
+ cachedResult.second = true;
reverseDep[QI].insert(C.getInstruction());
}
return QI;
}
// No dependence found
- depGraphLocal.insert(std::make_pair(C.getInstruction(),
- std::make_pair(NonLocal, true)));
+ cachedResult.first = NonLocal;
+ cachedResult.second = true;
reverseDep[NonLocal].insert(C.getInstruction());
return NonLocal;
}
DenseMap<BasicBlock*, Value*>& resp) {
// Set of blocks that we've already visited in our DFS
SmallPtrSet<BasicBlock*, 4> visited;
+ // If we're updating a dirtied cache entry, we don't need to reprocess
+ // already computed entries.
+ for (DenseMap<BasicBlock*, Value*>::iterator I = resp.begin(),
+ E = resp.end(); I != E; ++I)
+ if (I->second != Dirty)
+ visited.insert(I->first);
+
// Current stack of the DFS
SmallVector<BasicBlock*, 4> stack;
- stack.push_back(block);
+ for (pred_iterator PI = pred_begin(block), PE = pred_end(block);
+ PI != PE; ++PI)
+ stack.push_back(*PI);
// Do a basic DFS
while (!stack.empty()) {
if (BB != block) {
visited.insert(BB);
- const Instruction* localDep = getDependency(query, 0, BB);
+ Instruction* localDep = getDependency(query, 0, BB);
if (localDep != NonLocal) {
- resp.insert(std::make_pair(BB, const_cast<Instruction*>(localDep)));
+ resp.insert(std::make_pair(BB, localDep));
stack.pop_back();
continue;
// If we re-encounter the starting block, we still need to search it
// because there might be a dependency in the starting block AFTER
// the position of the query. This is necessary to get loops right.
- } else if (BB == block && stack.size() > 1) {
+ } else if (BB == block) {
visited.insert(BB);
- const Instruction* localDep = getDependency(query, 0, BB);
+ Instruction* localDep = getDependency(query, 0, BB);
if (localDep != query)
- resp.insert(std::make_pair(BB, const_cast<Instruction*>(localDep)));
+ resp.insert(std::make_pair(BB, localDep));
stack.pop_back();
}
// If we didn't find anything, recurse on the precessors of this block
+ // Only do this for blocks with a small number of predecessors.
bool predOnStack = false;
bool inserted = false;
- for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
- PI != PE; ++PI)
- if (!visited.count(*PI)) {
- stack.push_back(*PI);
- inserted = true;
- } else
- predOnStack = true;
+ if (std::distance(pred_begin(BB), pred_end(BB)) <= PredLimit) {
+ for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
+ PI != PE; ++PI)
+ if (!visited.count(*PI)) {
+ stack.push_back(*PI);
+ inserted = true;
+ } else
+ predOnStack = true;
+ }
// If we inserted a new predecessor, then we'll come back to this block
if (inserted)
// If we didn't insert because we have no predecessors, then this
// query has no dependency at all.
else if (!inserted && !predOnStack) {
- resp.insert(std::make_pair(BB, const_cast<Instruction*>(None)));
+ resp.insert(std::make_pair(BB, None));
// If we didn't insert because our predecessors are already on the stack,
// then we might still have a dependency, but it will be discovered during
// backtracking.
} else if (!inserted && predOnStack){
- resp.insert(std::make_pair(BB, const_cast<Instruction*>(NonLocal)));
+ resp.insert(std::make_pair(BB, NonLocal));
}
stack.pop_back();
/// blocks between the query and its dependencies.
void MemoryDependenceAnalysis::getNonLocalDependency(Instruction* query,
DenseMap<BasicBlock*, Value*>& resp) {
- // First check that we don't actually have a local dependency.
- const Instruction* localDep = getDependency(query);
- if (localDep != NonLocal) {
- resp.insert(std::make_pair(query->getParent(),
- const_cast<Instruction*>(localDep)));
+ if (depGraphNonLocal.count(query)) {
+ DenseMap<BasicBlock*, Value*>& cached = depGraphNonLocal[query];
+ NumCacheNonlocal++;
+
+ SmallVector<BasicBlock*, 4> dirtied;
+ for (DenseMap<BasicBlock*, Value*>::iterator I = cached.begin(),
+ E = cached.end(); I != E; ++I)
+ if (I->second == Dirty)
+ dirtied.push_back(I->first);
+
+ for (SmallVector<BasicBlock*, 4>::iterator I = dirtied.begin(),
+ E = dirtied.end(); I != E; ++I) {
+ Instruction* localDep = getDependency(query, 0, *I);
+ if (localDep != NonLocal)
+ cached[*I] = localDep;
+ else {
+ cached.erase(*I);
+ nonLocalHelper(query, *I, cached);
+ }
+ }
+
+ resp = cached;
+
+ // Update the reverse non-local dependency cache
+ for (DenseMap<BasicBlock*, Value*>::iterator I = resp.begin(), E = resp.end();
+ I != E; ++I)
+ reverseDepNonLocal[I->second].insert(query);
+
return;
- }
+ } else
+ NumUncacheNonlocal++;
- // If not, go ahead and search for non-local ones.
+ // If not, go ahead and search for non-local deps.
nonLocalHelper(query, query->getParent(), resp);
+
+ // Update the non-local dependency cache
+ for (DenseMap<BasicBlock*, Value*>::iterator I = resp.begin(), E = resp.end();
+ I != E; ++I) {
+ depGraphNonLocal[query].insert(*I);
+ reverseDepNonLocal[I->second].insert(query);
+ }
}
/// getDependency - Return the instruction on which a memory operation
-/// depends. The local paramter indicates if the query should only
+/// depends. The local parameter indicates if the query should only
/// evaluate dependencies within the same basic block.
-const Instruction* MemoryDependenceAnalysis::getDependency(Instruction* query,
+Instruction* MemoryDependenceAnalysis::getDependency(Instruction* query,
Instruction* start,
BasicBlock* block) {
// Start looking for dependencies with the queried inst
BasicBlock::iterator QI = query;
// Check for a cached result
- std::pair<const Instruction*, bool> cachedResult = depGraphLocal[query];
+ std::pair<Instruction*, bool>& cachedResult = depGraphLocal[query];
// If we have a _confirmed_ cached entry, return it
- if (cachedResult.second)
- return cachedResult.first;
- else if (cachedResult.first && cachedResult.first != NonLocal)
- // If we have an unconfirmed cached entry, we can start our search from there
- QI = const_cast<Instruction*>(cachedResult.first);
+ if (!block && !start) {
+ if (cachedResult.second)
+ return cachedResult.first;
+ else if (cachedResult.first && cachedResult.first != NonLocal)
+ // If we have an unconfirmed cached entry, we can start our search from there
+ QI = cachedResult.first;
+ }
if (start)
QI = start;
bool queryIsVolatile = false;
if (StoreInst* S = dyn_cast<StoreInst>(query)) {
dependee = S->getPointerOperand();
- dependeeSize = TD.getTypeSize(S->getOperand(0)->getType());
+ dependeeSize = TD.getTypeStoreSize(S->getOperand(0)->getType());
queryIsVolatile = S->isVolatile();
} else if (LoadInst* L = dyn_cast<LoadInst>(query)) {
dependee = L->getPointerOperand();
- dependeeSize = TD.getTypeSize(L->getType());
+ dependeeSize = TD.getTypeStoreSize(L->getType());
queryIsVolatile = L->isVolatile();
} else if (VAArgInst* V = dyn_cast<VAArgInst>(query)) {
dependee = V->getOperand(0);
- dependeeSize = TD.getTypeSize(V->getType());
+ dependeeSize = TD.getTypeStoreSize(V->getType());
} else if (FreeInst* F = dyn_cast<FreeInst>(query)) {
dependee = F->getPointerOperand();
// All volatile loads/stores depend on each other
if (queryIsVolatile && S->isVolatile()) {
if (!start && !block) {
- depGraphLocal.insert(std::make_pair(query, std::make_pair(S, true)));
+ cachedResult.first = S;
+ cachedResult.second = true;
reverseDep[S].insert(query);
}
}
pointer = S->getPointerOperand();
- pointerSize = TD.getTypeSize(S->getOperand(0)->getType());
+ pointerSize = TD.getTypeStoreSize(S->getOperand(0)->getType());
} else if (LoadInst* L = dyn_cast<LoadInst>(QI)) {
// All volatile loads/stores depend on each other
if (queryIsVolatile && L->isVolatile()) {
if (!start && !block) {
- depGraphLocal.insert(std::make_pair(query, std::make_pair(L, true)));
+ cachedResult.first = L;
+ cachedResult.second = true;
reverseDep[L].insert(query);
}
}
pointer = L->getPointerOperand();
- pointerSize = TD.getTypeSize(L->getType());
+ pointerSize = TD.getTypeStoreSize(L->getType());
} else if (AllocationInst* AI = dyn_cast<AllocationInst>(QI)) {
pointer = AI;
if (ConstantInt* C = dyn_cast<ConstantInt>(AI->getArraySize()))
pointerSize = C->getZExtValue() * \
- TD.getTypeSize(AI->getAllocatedType());
+ TD.getABITypeSize(AI->getAllocatedType());
else
pointerSize = ~0UL;
} else if (VAArgInst* V = dyn_cast<VAArgInst>(QI)) {
pointer = V->getOperand(0);
- pointerSize = TD.getTypeSize(V->getType());
+ pointerSize = TD.getTypeStoreSize(V->getType());
} else if (FreeInst* F = dyn_cast<FreeInst>(QI)) {
pointer = F->getPointerOperand();
continue;
if (!start && !block) {
- depGraphLocal.insert(std::make_pair(query,
- std::make_pair(QI, true)));
+ cachedResult.first = QI;
+ cachedResult.second = true;
reverseDep[QI].insert(query);
}
continue;
if (!start && !block) {
- depGraphLocal.insert(std::make_pair(query,
- std::make_pair(QI, true)));
+ cachedResult.first = QI;
+ cachedResult.second = true;
reverseDep[QI].insert(query);
}
// If we found nothing, return the non-local flag
if (!start && !block) {
- depGraphLocal.insert(std::make_pair(query,
- std::make_pair(NonLocal, true)));
+ cachedResult.first = NonLocal;
+ cachedResult.second = true;
reverseDep[NonLocal].insert(query);
}
return NonLocal;
}
+/// dropInstruction - Remove an instruction from the analysis, making
+/// absolutely conservative assumptions when updating the cache. This is
+/// useful, for example when an instruction is changed rather than removed.
+void MemoryDependenceAnalysis::dropInstruction(Instruction* drop) {
+ depMapType::iterator depGraphEntry = depGraphLocal.find(drop);
+ if (depGraphEntry != depGraphLocal.end())
+ reverseDep[depGraphEntry->second.first].erase(drop);
+
+ // Drop dependency information for things that depended on this instr
+ SmallPtrSet<Instruction*, 4>& set = reverseDep[drop];
+ for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
+ I != E; ++I)
+ depGraphLocal.erase(*I);
+
+ depGraphLocal.erase(drop);
+ reverseDep.erase(drop);
+
+ for (DenseMap<BasicBlock*, Value*>::iterator DI =
+ depGraphNonLocal[drop].begin(), DE = depGraphNonLocal[drop].end();
+ DI != DE; ++DI)
+ if (DI->second != None)
+ reverseDepNonLocal[DI->second].erase(drop);
+
+ if (reverseDepNonLocal.count(drop)) {
+ SmallPtrSet<Instruction*, 4>& set = reverseDepNonLocal[drop];
+ for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
+ I != E; ++I)
+ for (DenseMap<BasicBlock*, Value*>::iterator DI =
+ depGraphNonLocal[*I].begin(), DE = depGraphNonLocal[*I].end();
+ DI != DE; ++DI)
+ if (DI->second == drop)
+ DI->second = Dirty;
+ }
+
+ reverseDepNonLocal.erase(drop);
+ nonLocalDepMapType::iterator I = depGraphNonLocal.find(drop);
+ if (I != depGraphNonLocal.end())
+ depGraphNonLocal.erase(I);
+}
+
/// removeInstruction - Remove an instruction from the dependence analysis,
/// updating the dependence of instructions that previously depended on it.
/// This method attempts to keep the cache coherent using the reverse map.
void MemoryDependenceAnalysis::removeInstruction(Instruction* rem) {
// Figure out the new dep for things that currently depend on rem
- const Instruction* newDep = NonLocal;
+ Instruction* newDep = NonLocal;
+
+ for (DenseMap<BasicBlock*, Value*>::iterator DI =
+ depGraphNonLocal[rem].begin(), DE = depGraphNonLocal[rem].end();
+ DI != DE; ++DI)
+ if (DI->second != None)
+ reverseDepNonLocal[DI->second].erase(rem);
depMapType::iterator depGraphEntry = depGraphLocal.find(rem);
- // We assume here that it's not in the reverse map if it's not in
- // the dep map. Checking it could be expensive, so don't do it.
if (depGraphEntry != depGraphLocal.end()) {
+ reverseDep[depGraphEntry->second.first].erase(rem);
+
if (depGraphEntry->second.first != NonLocal &&
+ depGraphEntry->second.first != None &&
depGraphEntry->second.second) {
// If we have dep info for rem, set them to it
- BasicBlock::iterator RI =
- const_cast<Instruction*>(depGraphEntry->second.first);
+ BasicBlock::iterator RI = depGraphEntry->second.first;
RI++;
newDep = RI;
- } else if (depGraphEntry->second.first == NonLocal &&
+ } else if ( (depGraphEntry->second.first == NonLocal ||
+ depGraphEntry->second.first == None ) &&
depGraphEntry->second.second ) {
// If we have a confirmed non-local flag, use it
- newDep = NonLocal;
+ newDep = depGraphEntry->second.first;
} else {
// Otherwise, use the immediate successor of rem
// NOTE: This is because, when getDependence is called, it will first
RI++;
newDep = RI;
}
-
- SmallPtrSet<Instruction*, 4>& set = reverseDep[rem];
+ } else {
+ // Otherwise, use the immediate successor of rem
+ // NOTE: This is because, when getDependence is called, it will first
+ // check the immediate predecessor of what is in the cache.
+ BasicBlock::iterator RI = rem;
+ RI++;
+ newDep = RI;
+ }
+
+ SmallPtrSet<Instruction*, 4>& set = reverseDep[rem];
+ for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
+ I != E; ++I) {
+ // Insert the new dependencies
+ // Mark it as unconfirmed as long as it is not the non-local flag
+ depGraphLocal[*I] = std::make_pair(newDep, (newDep == NonLocal ||
+ newDep == None));
+ }
+
+ depGraphLocal.erase(rem);
+ reverseDep.erase(rem);
+
+ if (reverseDepNonLocal.count(rem)) {
+ SmallPtrSet<Instruction*, 4>& set = reverseDepNonLocal[rem];
for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
- I != E; ++I) {
- // Insert the new dependencies
- // Mark it as unconfirmed as long as it is not the non-local flag
- depGraphLocal[*I] = std::make_pair(newDep, !newDep);
- }
- reverseDep.erase(rem);
+ I != E; ++I)
+ for (DenseMap<BasicBlock*, Value*>::iterator DI =
+ depGraphNonLocal[*I].begin(), DE = depGraphNonLocal[*I].end();
+ DI != DE; ++DI)
+ if (DI->second == rem)
+ DI->second = Dirty;
+
}
+
+ reverseDepNonLocal.erase(rem);
+ nonLocalDepMapType::iterator I = depGraphNonLocal.find(rem);
+ if (I != depGraphNonLocal.end())
+ depGraphNonLocal.erase(I);
getAnalysis<AliasAnalysis>().deleteValue(rem);
}
projects / oota-llvm.git / blobdiff