void DataDependenceAnalysis::analyze(ir::IRKernel& kernel)
{
auto dfg = static_cast<DataflowGraph*>(
getAnalysis("DataflowGraphAnalysis"));
report("Running data dependence analysis on kernel " << kernel.name);
for(auto block = dfg->begin(); block != dfg->end(); ++block)
{
analyzeBlock(block, dfg, _nodes, _instructionToNodes);
}
}
void SCCP::analyze() {
// Queue the first block to start iteration.
cfgwork.push(ir->begin);
while (!cfgwork.empty()) {
while (!cfgwork.empty())
analyzeBlock(cfgwork.pop());
while (!ssawork.empty())
analyzeSSA(ssawork.pop());
}
// TODO own cmdline flag? something less intrusive
// if ((enable_verbose || enable_printir) && ir->size() > 1000)
// printCounts(ir, eval_counts);
}
bool AllocaMerging::runOnFunction(Function& F)
{
cheerp::PointerAnalyzer & PA = getAnalysis<cheerp::PointerAnalyzer>();
cheerp::Registerize & registerize = getAnalysis<cheerp::Registerize>();
cheerp::TypeSupport types(*F.getParent());
AllocaInfos allocaInfos;
// Gather all the allocas
for(BasicBlock& BB: F)
analyzeBlock(registerize, BB, allocaInfos);
if (allocaInfos.size() < 2)
return false;
bool Changed = false;
BasicBlock& entryBlock=F.getEntryBlock();
// Look if we can merge allocas of the same type
for(auto targetCandidate=allocaInfos.begin();targetCandidate!=allocaInfos.end();++targetCandidate)
{
AllocaInst* targetAlloca = targetCandidate->first;
Type* targetType = targetAlloca->getAllocatedType();
// The range storing the sum of all ranges merged into target
cheerp::Registerize::LiveRange targetRange(targetCandidate->second);
// If the range is empty, we have an alloca that we can't analyze
if (targetRange.empty())
continue;
std::vector<AllocaInfos::iterator> mergeSet;
auto sourceCandidate=targetCandidate;
++sourceCandidate;
for(;sourceCandidate!=allocaInfos.end();++sourceCandidate)
{
AllocaInst* sourceAlloca = sourceCandidate->first;
Type* sourceType = sourceAlloca->getAllocatedType();
// Bail out for non compatible types
if(!areTypesEquivalent(types, PA, targetType, sourceType))
continue;
const cheerp::Registerize::LiveRange& sourceRange = sourceCandidate->second;
// Bail out if this source candidate is not analyzable
if(sourceRange.empty())
continue;
// Bail out if the allocas interfere
if(targetRange.doesInterfere(sourceRange))
continue;
// Add the range to the target range and the source alloca to the mergeSet
mergeSet.push_back(sourceCandidate);
PA.invalidate(sourceAlloca);
targetRange.merge(sourceRange);
}
// If the merge set is empty try another target
if(mergeSet.empty())
continue;
PA.invalidate(targetAlloca);
if(!Changed)
registerize.invalidateLiveRangeForAllocas(F);
// Make sure that this alloca is in the entry block
if(targetAlloca->getParent()!=&entryBlock)
targetAlloca->moveBefore(entryBlock.begin());
// We can merge the allocas
for(const AllocaInfos::iterator& it: mergeSet)
{
AllocaInst* allocaToMerge = it->first;
Instruction* targetVal=targetAlloca;
if(targetVal->getType()!=allocaToMerge->getType())
{
targetVal=new BitCastInst(targetVal, allocaToMerge->getType());
targetVal->insertAfter(targetAlloca);
}
allocaToMerge->replaceAllUsesWith(targetVal);
allocaToMerge->eraseFromParent();
if(targetVal != targetAlloca)
PA.getPointerKind(targetVal);
allocaInfos.erase(it);
NumAllocaMerged++;
}
PA.getPointerKind(targetAlloca);
Changed = true;
}
if(Changed)
registerize.computeLiveRangeForAllocas(F);
return Changed;
}
bool AllocaArraysMerging::runOnFunction(Function& F)
{
class ArraysToMerge
{
private:
std::map<AllocaInst*, uint32_t> arraysToMerge;
uint32_t currentOffset;
public:
ArraysToMerge():currentOffset(0)
{
}
bool empty() const
{
return arraysToMerge.empty();
}
std::map<AllocaInst*, uint32_t>::iterator begin()
{
return arraysToMerge.begin();
}
std::map<AllocaInst*, uint32_t>::iterator end()
{
return arraysToMerge.end();
}
void add(AllocaInst* a)
{
arraysToMerge.insert(std::make_pair(a, currentOffset));
currentOffset+=cast<ArrayType>(a->getAllocatedType())->getNumElements();
}
uint32_t getNewSize() const
{
return currentOffset;
}
};
cheerp::PointerAnalyzer & PA = getAnalysis<cheerp::PointerAnalyzer>();
cheerp::Registerize & registerize = getAnalysis<cheerp::Registerize>();
cheerp::GlobalDepsAnalyzer & GDA = getAnalysis<cheerp::GlobalDepsAnalyzer>();
std::list<std::pair<AllocaInst*, cheerp::Registerize::LiveRange>> allocaInfos;
// Gather all the allocas
for(BasicBlock& BB: F)
analyzeBlock(registerize, BB, allocaInfos);
if (allocaInfos.size() < 2)
return false;
bool Changed = false;
// We can also try to merge arrays of the same type, if only pointers to values are passed around
while(!allocaInfos.empty())
{
// Build a map of array to be merged and their offseet into the new array
ArraysToMerge arraysToMerge;
auto targetCandidate = allocaInfos.begin();
AllocaInst* targetAlloca = targetCandidate->first;
if(!targetAlloca->getAllocatedType()->isArrayTy() ||
// Check target uses
!checkUsesForArrayMerging(targetAlloca))
{
allocaInfos.erase(targetCandidate);
continue;
}
Type* targetElementType = targetAlloca->getAllocatedType()->getSequentialElementType();
auto sourceCandidate=targetCandidate;
++sourceCandidate;
// Now that we have computed the sourceCandidate we can invalidate the targetCandidate
allocaInfos.erase(targetCandidate);
while(sourceCandidate!=allocaInfos.end())
{
AllocaInst* sourceAlloca = sourceCandidate->first;
// Check that allocas are arrays of the same type
if(!sourceAlloca->getAllocatedType()->isArrayTy())
{
++sourceCandidate;
continue;
}
// Both are arrays, check the types
if(targetElementType != sourceAlloca->getAllocatedType()->getSequentialElementType())
{
++sourceCandidate;
continue;
}
// Verify that the source candidate has supported uses
if(!checkUsesForArrayMerging(sourceAlloca))
{
++sourceCandidate;
continue;
}
// We can merge the source and the target
// If the set is empty add the target as well
if(arraysToMerge.empty())
arraysToMerge.add(targetAlloca);
arraysToMerge.add(sourceAlloca);
auto oldCandidate = sourceCandidate;
++sourceCandidate;
// Now that we have moved to the next candidate, we can invalidate the old one
allocaInfos.erase(oldCandidate);
}
// If we have a non-empty set of alloca merge them
if (arraysToMerge.empty())
continue;
if(!Changed)
registerize.invalidateLiveRangeForAllocas(F);
// Build new alloca
Type* newAllocaType = ArrayType::get(targetElementType, arraysToMerge.getNewSize());
// Add the new struct type to the GlobalDepsAnalyzer, it may need the createArray helper
GDA.visitType(newAllocaType, /*forceTypedArray*/ false);
AllocaInst* newAlloca = new AllocaInst(newAllocaType, "mergedArray", &(*F.getEntryBlock().begin()));
Type* indexType = IntegerType::get(newAllocaType->getContext(), 32);
// Change every use of every merged array with an appropiate GEP
for(auto it: arraysToMerge)
{
AllocaInst* allocaToMerge = it.first;
uint32_t baseOffset = it.second;
SmallVector<User*, 4> users(allocaToMerge->users());
for(User* u: users)
{
if(GetElementPtrInst* oldGep = dyn_cast<GetElementPtrInst>(u))
{
// Build 2 GEPs, one to reach the first element in the merged array
// and the other for the rest of the offsets
SmallVector<Value*, 4> indices;
// Dereference array
indices.push_back(ConstantInt::get(indexType, 0));
// Reach offset
indices.push_back(ConstantInt::get(indexType, baseOffset));
Value* gep1 = GetElementPtrInst::Create(newAlloca, indices, "", oldGep);
// Apply all the old offsets but the first one using a new GEP
indices.clear();
indices.insert(indices.begin(), oldGep->idx_begin()+1, oldGep->idx_end());
Value* gep2 = GetElementPtrInst::Create(gep1, indices, "", oldGep);
// Replace all uses with gep2
oldGep->replaceAllUsesWith(gep2);
PA.invalidate(oldGep);
oldGep->eraseFromParent();
}
else if(BitCastInst* BI=dyn_cast<BitCastInst>(u))
{
//Only used for lifetime intrinsics
Value* newBitCast=new BitCastInst(newAlloca, BI->getType(), "", BI);
BI->replaceAllUsesWith(newBitCast);
PA.invalidate(BI);
BI->eraseFromParent();
}
else
assert(false && "Unexpected use while merging arrays");
}
// Kill the alloca itself now
PA.invalidate(allocaToMerge);
allocaToMerge->eraseFromParent();
Changed = true;
}
}
if(Changed)
registerize.computeLiveRangeForAllocas(F);
return Changed;
}