void DSGraphStats::visitLoad(LoadInst &LI) {
if (isNodeForValueUntyped(LI.getOperand(0), 0,LI.getParent()->getParent())) {
NumUntypedMemAccesses++;
} else {
NumTypedMemAccesses++;
}
}
bool SFVInstVisitor::visitLoadInst(LoadInst &LI) {
if ((Callbacks & Visit::Loads) == 0) return false;
if (LatticeValue *LV = getLatticeValueForField(LI.getOperand(0)))
if (LV->visitLoad(LI))
return RemoveLatticeValueAtBottom(LV);
return false;
}
/// Move cond_fail down if it can potentially help register promotion later.
static bool sinkCondFail(SILLoop *Loop) {
// Only handle innermost loops for now.
if (!Loop->getSubLoops().empty())
return false;
bool Changed = false;
for (auto *BB : Loop->getBlocks()) {
// A list of CondFails that can be moved down.
SmallVector<CondFailInst*, 4> CFs;
// A pair of load and store that are independent of the CondFails and
// can potentially access the same memory.
LoadInst *LIOfPair = nullptr;
bool foundPair = false;
for (auto &Inst : *BB) {
if (foundPair) {
// Move CFs to right before Inst.
for (unsigned I = 0, E = CFs.size(); I < E; I++) {
DEBUG(llvm::dbgs() << "sinking cond_fail down ");
DEBUG(CFs[I]->dump());
DEBUG(llvm::dbgs() << " before ");
DEBUG(Inst.dump());
CFs[I]->moveBefore(&Inst);
}
Changed = true;
foundPair = false;
LIOfPair = nullptr;
}
if (auto CF = dyn_cast<CondFailInst>(&Inst)) {
CFs.push_back(CF);
} else if (auto LI = dyn_cast<LoadInst>(&Inst)) {
if (addressIndependent(LI->getOperand())) {
LIOfPair = LI;
} else {
CFs.clear();
LIOfPair = nullptr;
}
} else if (auto SI = dyn_cast<StoreInst>(&Inst)) {
if (addressIndependent(SI->getDest())) {
if (LIOfPair &&
addressCanPairUp(SI->getDest(), LIOfPair->getOperand()))
foundPair = true;
} else {
CFs.clear();
LIOfPair = nullptr;
}
} else if (Inst.mayHaveSideEffects()) {
CFs.clear();
LIOfPair = nullptr;
}
}
}
return Changed;
}
string instructionToString(Instruction* curr_Ins)
{
Instruction *InsOp1, *InsOp2;
Value *v;
string op1="'",op2="'",exprsn,opcodeName,opc;
LoadInst *LD;
InsOp1 = cast<Instruction>(curr_Ins->getOperand(0));
InsOp2 = cast<Instruction>(curr_Ins->getOperand(1));
opcodeName=InsOp1->getOpcodeName();
if(opcodeName=="load")
{
LD = cast<LoadInst>(curr_Ins->getOperand(0));
v=LD->getOperand(0);
op1=v->getName().str();
}
else
{
op1=instructionToString(InsOp1);
}
opcodeName=InsOp2->getOpcodeName();
if(opcodeName=="load")
{
LD = cast<LoadInst>(curr_Ins->getOperand(1));
v=LD->getOperand(0);
op2=v->getName().str();
}
else
{
op2=instructionToString(InsOp2);
}
switch(curr_Ins->getOpcode()){
case Instruction::Add: opc="+";
break;
case Instruction::Sub: opc="-";
break;
case 21: opc="/";
break;
case Instruction::Mul: opc="*";
break;}
exprsn=op1+opc+op2;
return exprsn;
}
bool CallAnalyzer::visitLoad(LoadInst &I) {
Value *SROAArg;
DenseMap<Value *, int>::iterator CostIt;
if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
if (I.isSimple()) {
accumulateSROACost(CostIt, InlineConstants::InstrCost);
return true;
}
disableSROA(CostIt);
}
return false;
}
void FuncTransform::visitLoadInst(LoadInst &LI) {
//
// Record the use of the pool handle for the pointer being dereferenced.
//
if (Value *PH = getPoolHandle(LI.getOperand(0)))
AddPoolUse(LI, PH, PoolUses);
//
// If this is a volatile load, then record a use of the pool handle for the
// loaded value, even if it is never used.
//
if (LI.isVolatile()) {
if (Value *PH = getPoolHandle(&LI))
AddPoolUse (LI, PH, PoolUses);
}
visitInstruction(LI);
}
PointerAnalysisFlow* PointerAnalysis::operation_X_Y(PointerAnalysisFlow* in, Instruction* instruction) {
LoadInst* load = static_cast<LoadInst*>(instruction);
PointerAnalysisFlow* f = new PointerAnalysisFlow(in);
Value* Y = load->getOperand(0); //RO
Value* X = load->getNextNode()->getOperand(1); //LO
// X = Y
// Check if X Y are pointers.
if (isPointer(Y) && isPointer(X)) {
if (isVariable(Y) && isVariable(X)) {
//x points to what y points to
PointerAnalysisFlow* ff = new PointerAnalysisFlow();
map<string, set<string> > value;
value[X->getName()] = in->value[Y->getName()];;
ff->value = value;
PointerAnalysisFlow* tmp = static_cast<PointerAnalysisFlow*>(ff->join(f));
delete ff;
delete f;
f = tmp;
}
}
return f;
}
//------------------------------------------------------------------------------
void SymbolicExecution::visitLoadInst(LoadInst &loadInst) {
Instruction *inst = &loadInst;
Value *pointer = loadInst.getOperand(0);
visitPointer(inst, pointer, loadBankConflicts, loadTransactions);
}
/// DoPromotion - This method actually performs the promotion of the specified
/// arguments, and returns the new function. At this point, we know that it's
/// safe to do so.
CallGraphNode *ArgPromotion::DoPromotion(Function *F,
SmallPtrSet<Argument*, 8> &ArgsToPromote,
SmallPtrSet<Argument*, 8> &ByValArgsToTransform) {
// Start by computing a new prototype for the function, which is the same as
// the old function, but has modified arguments.
const FunctionType *FTy = F->getFunctionType();
std::vector<const Type*> Params;
typedef std::set<IndicesVector> ScalarizeTable;
// ScalarizedElements - If we are promoting a pointer that has elements
// accessed out of it, keep track of which elements are accessed so that we
// can add one argument for each.
//
// Arguments that are directly loaded will have a zero element value here, to
// handle cases where there are both a direct load and GEP accesses.
//
std::map<Argument*, ScalarizeTable> ScalarizedElements;
// OriginalLoads - Keep track of a representative load instruction from the
// original function so that we can tell the alias analysis implementation
// what the new GEP/Load instructions we are inserting look like.
std::map<IndicesVector, LoadInst*> OriginalLoads;
// Attributes - Keep track of the parameter attributes for the arguments
// that we are *not* promoting. For the ones that we do promote, the parameter
// attributes are lost
SmallVector<AttributeWithIndex, 8> AttributesVec;
const AttrListPtr &PAL = F->getAttributes();
// Add any return attributes.
if (Attributes attrs = PAL.getRetAttributes())
AttributesVec.push_back(AttributeWithIndex::get(0, attrs));
// First, determine the new argument list
unsigned ArgIndex = 1;
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
++I, ++ArgIndex) {
if (ByValArgsToTransform.count(I)) {
// Simple byval argument? Just add all the struct element types.
const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
const StructType *STy = cast<StructType>(AgTy);
for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
Params.push_back(STy->getElementType(i));
++NumByValArgsPromoted;
} else if (!ArgsToPromote.count(I)) {
// Unchanged argument
Params.push_back(I->getType());
if (Attributes attrs = PAL.getParamAttributes(ArgIndex))
AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs));
} else if (I->use_empty()) {
// Dead argument (which are always marked as promotable)
++NumArgumentsDead;
} else {
// Okay, this is being promoted. This means that the only uses are loads
// or GEPs which are only used by loads
// In this table, we will track which indices are loaded from the argument
// (where direct loads are tracked as no indices).
ScalarizeTable &ArgIndices = ScalarizedElements[I];
for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
++UI) {
Instruction *User = cast<Instruction>(*UI);
assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User));
IndicesVector Indices;
Indices.reserve(User->getNumOperands() - 1);
// Since loads will only have a single operand, and GEPs only a single
// non-index operand, this will record direct loads without any indices,
// and gep+loads with the GEP indices.
for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end();
II != IE; ++II)
Indices.push_back(cast<ConstantInt>(*II)->getSExtValue());
// GEPs with a single 0 index can be merged with direct loads
if (Indices.size() == 1 && Indices.front() == 0)
Indices.clear();
ArgIndices.insert(Indices);
LoadInst *OrigLoad;
if (LoadInst *L = dyn_cast<LoadInst>(User))
OrigLoad = L;
else
// Take any load, we will use it only to update Alias Analysis
OrigLoad = cast<LoadInst>(User->use_back());
OriginalLoads[Indices] = OrigLoad;
}
// Add a parameter to the function for each element passed in.
for (ScalarizeTable::iterator SI = ArgIndices.begin(),
E = ArgIndices.end(); SI != E; ++SI) {
// not allowed to dereference ->begin() if size() is 0
Params.push_back(GetElementPtrInst::getIndexedType(I->getType(),
SI->begin(),
SI->end()));
assert(Params.back());
}
if (ArgIndices.size() == 1 && ArgIndices.begin()->empty())
++NumArgumentsPromoted;
else
++NumAggregatesPromoted;
}
}
// Add any function attributes.
if (Attributes attrs = PAL.getFnAttributes())
AttributesVec.push_back(AttributeWithIndex::get(~0, attrs));
const Type *RetTy = FTy->getReturnType();
// Work around LLVM bug PR56: the CWriter cannot emit varargs functions which
// have zero fixed arguments.
bool ExtraArgHack = false;
if (Params.empty() && FTy->isVarArg()) {
ExtraArgHack = true;
Params.push_back(Type::getInt32Ty(F->getContext()));
}
// Construct the new function type using the new arguments.
FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
// Create the new function body and insert it into the module.
Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
NF->copyAttributesFrom(F);
DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n"
<< "From: " << *F);
// Recompute the parameter attributes list based on the new arguments for
// the function.
NF->setAttributes(AttrListPtr::get(AttributesVec.begin(),
AttributesVec.end()));
AttributesVec.clear();
F->getParent()->getFunctionList().insert(F, NF);
NF->takeName(F);
// Get the alias analysis information that we need to update to reflect our
// changes.
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
// Get the callgraph information that we need to update to reflect our
// changes.
CallGraph &CG = getAnalysis<CallGraph>();
// Get a new callgraph node for NF.
CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF);
// Loop over all of the callers of the function, transforming the call sites
// to pass in the loaded pointers.
//
SmallVector<Value*, 16> Args;
while (!F->use_empty()) {
CallSite CS = CallSite::get(F->use_back());
assert(CS.getCalledFunction() == F);
Instruction *Call = CS.getInstruction();
const AttrListPtr &CallPAL = CS.getAttributes();
// Add any return attributes.
if (Attributes attrs = CallPAL.getRetAttributes())
AttributesVec.push_back(AttributeWithIndex::get(0, attrs));
// Loop over the operands, inserting GEP and loads in the caller as
// appropriate.
CallSite::arg_iterator AI = CS.arg_begin();
ArgIndex = 1;
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
I != E; ++I, ++AI, ++ArgIndex)
if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
Args.push_back(*AI); // Unmodified argument
if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex))
AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
} else if (ByValArgsToTransform.count(I)) {
// Emit a GEP and load for each element of the struct.
const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
const StructType *STy = cast<StructType>(AgTy);
Value *Idxs[2] = {
ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };
for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
Value *Idx = GetElementPtrInst::Create(*AI, Idxs, Idxs+2,
(*AI)->getName()+"."+utostr(i),
Call);
// TODO: Tell AA about the new values?
Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call));
}
} else if (!I->use_empty()) {
// Non-dead argument: insert GEPs and loads as appropriate.
ScalarizeTable &ArgIndices = ScalarizedElements[I];
// Store the Value* version of the indices in here, but declare it now
// for reuse.
std::vector<Value*> Ops;
for (ScalarizeTable::iterator SI = ArgIndices.begin(),
E = ArgIndices.end(); SI != E; ++SI) {
Value *V = *AI;
LoadInst *OrigLoad = OriginalLoads[*SI];
if (!SI->empty()) {
Ops.reserve(SI->size());
const Type *ElTy = V->getType();
for (IndicesVector::const_iterator II = SI->begin(),
IE = SI->end(); II != IE; ++II) {
// Use i32 to index structs, and i64 for others (pointers/arrays).
// This satisfies GEP constraints.
const Type *IdxTy = (ElTy->isStructTy() ?
Type::getInt32Ty(F->getContext()) :
Type::getInt64Ty(F->getContext()));
Ops.push_back(ConstantInt::get(IdxTy, *II));
// Keep track of the type we're currently indexing.
ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II);
}
// And create a GEP to extract those indices.
V = GetElementPtrInst::Create(V, Ops.begin(), Ops.end(),
V->getName()+".idx", Call);
Ops.clear();
AA.copyValue(OrigLoad->getOperand(0), V);
}
// Since we're replacing a load make sure we take the alignment
// of the previous load.
LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call);
newLoad->setAlignment(OrigLoad->getAlignment());
Args.push_back(newLoad);
AA.copyValue(OrigLoad, Args.back());
}
}
if (ExtraArgHack)
Args.push_back(Constant::getNullValue(Type::getInt32Ty(F->getContext())));
// Push any varargs arguments on the list.
for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
Args.push_back(*AI);
if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex))
AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
}
// Add any function attributes.
if (Attributes attrs = CallPAL.getFnAttributes())
AttributesVec.push_back(AttributeWithIndex::get(~0, attrs));
Instruction *New;
if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
Args.begin(), Args.end(), "", Call);
cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(),
AttributesVec.end()));
} else {
New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call);
cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
cast<CallInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(),
AttributesVec.end()));
if (cast<CallInst>(Call)->isTailCall())
cast<CallInst>(New)->setTailCall();
}
Args.clear();
AttributesVec.clear();
// Update the alias analysis implementation to know that we are replacing
// the old call with a new one.
AA.replaceWithNewValue(Call, New);
// Update the callgraph to know that the callsite has been transformed.
CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
CalleeNode->replaceCallEdge(Call, New, NF_CGN);
if (!Call->use_empty()) {
Call->replaceAllUsesWith(New);
New->takeName(Call);
}
// Finally, remove the old call from the program, reducing the use-count of
// F.
Call->eraseFromParent();
}
// Since we have now created the new function, splice the body of the old
// function right into the new function, leaving the old rotting hulk of the
// function empty.
NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
// Loop over the argument list, transfering uses of the old arguments over to
// the new arguments, also transfering over the names as well.
//
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
I2 = NF->arg_begin(); I != E; ++I) {
if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
// If this is an unmodified argument, move the name and users over to the
// new version.
I->replaceAllUsesWith(I2);
I2->takeName(I);
AA.replaceWithNewValue(I, I2);
++I2;
continue;
}
if (ByValArgsToTransform.count(I)) {
// In the callee, we create an alloca, and store each of the new incoming
// arguments into the alloca.
Instruction *InsertPt = NF->begin()->begin();
// Just add all the struct element types.
const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt);
const StructType *STy = cast<StructType>(AgTy);
Value *Idxs[2] = {
ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };
for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
Value *Idx =
GetElementPtrInst::Create(TheAlloca, Idxs, Idxs+2,
TheAlloca->getName()+"."+Twine(i),
InsertPt);
I2->setName(I->getName()+"."+Twine(i));
new StoreInst(I2++, Idx, InsertPt);
}
// Anything that used the arg should now use the alloca.
I->replaceAllUsesWith(TheAlloca);
TheAlloca->takeName(I);
AA.replaceWithNewValue(I, TheAlloca);
continue;
}
if (I->use_empty()) {
AA.deleteValue(I);
continue;
}
// Otherwise, if we promoted this argument, then all users are load
// instructions (or GEPs with only load users), and all loads should be
// using the new argument that we added.
ScalarizeTable &ArgIndices = ScalarizedElements[I];
while (!I->use_empty()) {
if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) {
assert(ArgIndices.begin()->empty() &&
"Load element should sort to front!");
I2->setName(I->getName()+".val");
LI->replaceAllUsesWith(I2);
AA.replaceWithNewValue(LI, I2);
LI->eraseFromParent();
DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
<< "' in function '" << F->getName() << "'\n");
} else {
GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back());
IndicesVector Operands;
Operands.reserve(GEP->getNumIndices());
for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
II != IE; ++II)
Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());
// GEPs with a single 0 index can be merged with direct loads
if (Operands.size() == 1 && Operands.front() == 0)
Operands.clear();
Function::arg_iterator TheArg = I2;
for (ScalarizeTable::iterator It = ArgIndices.begin();
*It != Operands; ++It, ++TheArg) {
assert(It != ArgIndices.end() && "GEP not handled??");
}
std::string NewName = I->getName();
for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
NewName += "." + utostr(Operands[i]);
}
NewName += ".val";
TheArg->setName(NewName);
DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName()
<< "' of function '" << NF->getName() << "'\n");
// All of the uses must be load instructions. Replace them all with
// the argument specified by ArgNo.
while (!GEP->use_empty()) {
LoadInst *L = cast<LoadInst>(GEP->use_back());
L->replaceAllUsesWith(TheArg);
AA.replaceWithNewValue(L, TheArg);
L->eraseFromParent();
}
AA.deleteValue(GEP);
GEP->eraseFromParent();
}
}
// Increment I2 past all of the arguments added for this promoted pointer.
for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i)
++I2;
}
// Notify the alias analysis implementation that we inserted a new argument.
if (ExtraArgHack)
AA.copyValue(Constant::getNullValue(Type::getInt32Ty(F->getContext())),
NF->arg_begin());
// Tell the alias analysis that the old function is about to disappear.
AA.replaceWithNewValue(F, NF);
NF_CGN->stealCalledFunctionsFrom(CG[F]);
// Now that the old function is dead, delete it. If there is a dangling
// reference to the CallgraphNode, just leave the dead function around for
// someone else to nuke.
CallGraphNode *CGN = CG[F];
if (CGN->getNumReferences() == 0)
delete CG.removeFunctionFromModule(CGN);
else
F->setLinkage(Function::ExternalLinkage);
return NF_CGN;
}
bool NVPTXLowerAggrCopies::runOnFunction(Function &F) {
SmallVector<LoadInst *, 4> aggrLoads;
SmallVector<MemTransferInst *, 4> aggrMemcpys;
SmallVector<MemSetInst *, 4> aggrMemsets;
DataLayout *TD = &getAnalysis<DataLayout>();
LLVMContext &Context = F.getParent()->getContext();
//
// Collect all the aggrLoads, aggrMemcpys and addrMemsets.
//
//const BasicBlock *firstBB = &F.front(); // first BB in F
for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
//BasicBlock *bb = BI;
for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;
++II) {
if (LoadInst * load = dyn_cast<LoadInst>(II)) {
if (load->hasOneUse() == false) continue;
if (TD->getTypeStoreSize(load->getType()) < MaxAggrCopySize) continue;
User *use = *(load->use_begin());
if (StoreInst * store = dyn_cast<StoreInst>(use)) {
if (store->getOperand(0) != load) //getValueOperand
continue;
aggrLoads.push_back(load);
}
} else if (MemTransferInst * intr = dyn_cast<MemTransferInst>(II)) {
Value *len = intr->getLength();
// If the number of elements being copied is greater
// than MaxAggrCopySize, lower it to a loop
if (ConstantInt * len_int = dyn_cast < ConstantInt > (len)) {
if (len_int->getZExtValue() >= MaxAggrCopySize) {
aggrMemcpys.push_back(intr);
}
} else {
// turn variable length memcpy/memmov into loop
aggrMemcpys.push_back(intr);
}
} else if (MemSetInst * memsetintr = dyn_cast<MemSetInst>(II)) {
Value *len = memsetintr->getLength();
if (ConstantInt * len_int = dyn_cast<ConstantInt>(len)) {
if (len_int->getZExtValue() >= MaxAggrCopySize) {
aggrMemsets.push_back(memsetintr);
}
} else {
// turn variable length memset into loop
aggrMemsets.push_back(memsetintr);
}
}
}
}
if ((aggrLoads.size() == 0) && (aggrMemcpys.size() == 0)
&& (aggrMemsets.size() == 0)) return false;
//
// Do the transformation of an aggr load/copy/set to a loop
//
for (unsigned i = 0, e = aggrLoads.size(); i != e; ++i) {
LoadInst *load = aggrLoads[i];
StoreInst *store = dyn_cast<StoreInst>(*load->use_begin());
Value *srcAddr = load->getOperand(0);
Value *dstAddr = store->getOperand(1);
unsigned numLoads = TD->getTypeStoreSize(load->getType());
Value *len = ConstantInt::get(Type::getInt32Ty(Context), numLoads);
convertTransferToLoop(store, srcAddr, dstAddr, len, load->isVolatile(),
store->isVolatile(), Context, F);
store->eraseFromParent();
load->eraseFromParent();
}
for (unsigned i = 0, e = aggrMemcpys.size(); i != e; ++i) {
MemTransferInst *cpy = aggrMemcpys[i];
Value *len = cpy->getLength();
// llvm 2.7 version of memcpy does not have volatile
// operand yet. So always making it non-volatile
// optimistically, so that we don't see unnecessary
// st.volatile in ptx
convertTransferToLoop(cpy, cpy->getSource(), cpy->getDest(), len, false,
false, Context, F);
cpy->eraseFromParent();
}
for (unsigned i = 0, e = aggrMemsets.size(); i != e; ++i) {
MemSetInst *memsetinst = aggrMemsets[i];
Value *len = memsetinst->getLength();
Value *val = memsetinst->getValue();
convertMemSetToLoop(memsetinst, memsetinst->getDest(), len, val, Context,
F);
memsetinst->eraseFromParent();
}
return true;
}
int qdp_jit_vec::vectorize_loads( std::vector<std::vector<Instruction*> >& load_instructions )
{
DEBUG(dbgs() << "Vectorize loads, total of " << load_instructions.size() << "\n");
//std::vector<std::pair<Value*,Value*> > scalar_vector_loads;
scalar_vector_pairs.clear();
if (load_instructions.empty())
return 0;
int load_vec_elem = 0;
for( std::vector<Instruction*>& VI : load_instructions ) {
DEBUG(dbgs() << "Processing vector of loads number " << load_vec_elem++ << "\n");
assert( VI.size() == vec_len && "length of vector of loads does not match vec_len" );
int loads_consec = true;
uint64_t lo,hi;
bool first = true;
for( Instruction* I : VI ) {
GetElementPtrInst* GEP;
if ((GEP = dyn_cast<GetElementPtrInst>(I->getOperand(0)))) {
if (first) {
ConstantInt * CI;
if ((CI = dyn_cast<ConstantInt>(GEP->getOperand(1)))) {
lo = CI->getZExtValue();
hi = lo+1;
first=false;
} else {
DEBUG(dbgs() << "First load in the chain: Operand of GEP not a ConstantInt" << *GEP->getOperand(1) << "\n");
assert( 0 && "First load in the chain: Operand of GEP not a ConstantInt\n");
exit(0);
}
} else {
ConstantInt * CI;
if ((CI = dyn_cast<ConstantInt>(GEP->getOperand(1)))) {
if (hi != CI->getZExtValue()) {
DEBUG(dbgs() << "Loads not consecutive lo=" << lo << " hi=" << hi << " this=" << CI->getZExtValue() << "\n");
loads_consec = false;
} else {
hi++;
}
}
}
} else {
DEBUG(dbgs() << "Operand of load not a GEP " << *I->getOperand(0) << "\n");
assert( 0 && "Operand of load not a GEP" );
exit(0);
loads_consec = false;
}
}
if (loads_consec) {
DEBUG(dbgs() << "Loads consecuetive\n");
LoadInst* LI = cast<LoadInst>(VI.at(0));
GetElementPtrInst* GEP = cast<GetElementPtrInst>(LI->getOperand(0));
Instruction* GEPcl = clone_with_operands(GEP);
unsigned AS = LI->getPointerAddressSpace();
VectorType *VecTy = VectorType::get( LI->getType() , vec_len );
unsigned bitwidth = LI->getType()->getPrimitiveSizeInBits();
unsigned bytewidth = bitwidth == 1 ? 1 : bitwidth/8;
DEBUG(dbgs() << "bit/byte width of load instr trype: " << bitwidth << "/" << bytewidth << "\n");
//Builder->SetInsertPoint( GEP );
Value *VecPtr = Builder->CreateBitCast(GEPcl,VecTy->getPointerTo(AS));
//Value *VecLoad = Builder->CreateLoad( VecPtr );
unsigned align = lo % vec_len == 0 ? bytewidth * vec_len : bytewidth;
Value *VecLoad = Builder->CreateAlignedLoad( VecPtr , align );
//DEBUG(dbgs() << "created vector load: " << *VecLoad << "\n");
//function->dump();
// unsigned AS = LI->getPointerAddressSpace();
// VectorType *VecTy = VectorType::get( LI->getType() , vec_len );
// Builder->SetInsertPoint( LI );
// Value *VecPtr = Builder->CreateBitCast(LI->getPointerOperand(),VecTy->getPointerTo(AS));
// Value *VecLoad = Builder->CreateLoad( VecPtr );
scalar_vector_pairs.push_back( std::make_pair( VI.at(0) , VecLoad ) );
} else {
DEBUG(dbgs() << "Loads not consecutive:\n");
for (Value* V: VI) {
DEBUG(dbgs() << *V << "\n");
}
//Instruction* I = dyn_cast<Instruction>(VI.back()->getNextNode());
//DEBUG(dbgs() << *I << "\n");
//Builder->SetInsertPoint( VI.at(0) );
std::vector<Instruction*> VIcl;
for( Instruction* I : VI ) {
VIcl.push_back( clone_with_operands(I) );
}
VectorType *VecTy = VectorType::get( VI.at(0)->getType() , vec_len );
Value *Vec = UndefValue::get(VecTy);
int i=0;
for( Instruction* I : VIcl ) {
Vec = Builder->CreateInsertElement(Vec, I, Builder->getInt32(i++));
}
scalar_vector_pairs.push_back( std::make_pair( VI.at(0) , Vec ) );
}
}
//vectorize_all_uses( scalar_vector_loads );
DEBUG(dbgs() << "Searching for the stores:\n");
//function->dump();
//
// Vectorize all StoreInst reachable by the first load of each vector of loads
//
{
SetVector<Instruction*> to_visit;
SetVector<Instruction*> stores_processed;
for( std::vector<Instruction*>& VI : load_instructions ) {
to_visit.insert(VI.at(0));
}
while (!to_visit.empty()) {
Instruction* I = to_visit.back();
to_visit.pop_back();
DEBUG(dbgs() << "visiting " << *I << "\n");
if (StoreInst* SI = dyn_cast<StoreInst>(I)) {
if (!stores_processed.count(SI)) {
get_vector_version( SI );
stores_processed.insert( SI );
}
} else {
for (Use &U : I->uses()) {
Value* V = U.getUser();
to_visit.insert(cast<Instruction>(V));
}
}
}
}
// DEBUG(dbgs() << "After vectorizing the stores\n");
// function->dump();
//
// Mark all stores as being processed
//
SetVector<Instruction*> to_visit;
for( std::vector<Instruction*>& VI : load_instructions ) {
for( Instruction* I : VI ) {
to_visit.insert(I);
if (GetElementPtrInst* GEP = dyn_cast<GetElementPtrInst>(I->getOperand(0))) {
for_erasure.insert(GEP);
}
}
}
while (!to_visit.empty()) {
Instruction* I = to_visit.back();
to_visit.pop_back();
for_erasure.insert(I);
if (StoreInst* SI = dyn_cast<StoreInst>(I)) {
stores_processed.insert(SI);
if (GetElementPtrInst* GEP = dyn_cast<GetElementPtrInst>(SI->getOperand(1))) {
for_erasure.insert(GEP);
}
} else {
for (Use &U : I->uses()) {
Value* V = U.getUser();
to_visit.insert(cast<Instruction>(V));
}
}
}
DEBUG(dbgs() << "----------------------------------------\n");
DEBUG(dbgs() << "After vectorize_loads\n");
//function->dump();
return 0;
}
virtual bool runOnFunction(Function &F) {
//F.dump();
bool changed = false;
for (inst_iterator inst_it = inst_begin(F), _inst_end = inst_end(F); inst_it != _inst_end; ++inst_it) {
LoadInst *li = dyn_cast<LoadInst>(&*inst_it);
if (!li) continue;
ConstantExpr *ce = dyn_cast<ConstantExpr>(li->getOperand(0));
// Not 100% sure what the isGEPWithNoNotionalOverIndexing() means, but
// at least it checks if it's a gep:
if (ce && ce->isGEPWithNoNotionalOverIndexing() && ce->getOperand(0)->getType() == g.llvm_flavor_type_ptr) {
changed = handleFlavor(li, ce);
}
GlobalVariable *gv = dyn_cast<GlobalVariable>(li->getOperand(0));
if (!gv) continue;
llvm::Type* gv_t = gv->getType();
if (gv_t == g.llvm_bool_type_ptr->getPointerTo()) {
changed = handleBool(li, gv) || changed;
continue;
}
if (gv_t == g.llvm_class_type_ptr->getPointerTo()) {
changed = handleCls(li, gv) || changed;
continue;
}
}
return changed;
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass. Search for insert/extractvalue instructions
// that can be simplified.
//
// Inputs:
// M - A reference to the LLVM module to transform.
//
// Outputs:
// M - The transformed LLVM module.
//
// Return value:
// true - The module was modified.
// false - The module was not modified.
//
bool SimplifyLoad::runOnModule(Module& M) {
// Repeat till no change
bool changed;
do {
changed = false;
for (Module::iterator F = M.begin(); F != M.end(); ++F) {
for (Function::iterator B = F->begin(), FE = F->end(); B != FE; ++B) {
for (BasicBlock::iterator I = B->begin(), BE = B->end(); I != BE;) {
LoadInst *LI = dyn_cast<LoadInst>(I++);
if(!LI)
continue;
if(LI->hasOneUse()) {
if(CastInst *CI = dyn_cast<CastInst>(*(LI->use_begin()))) {
if(LI->getType()->isPointerTy()) {
if(ConstantExpr *CE = dyn_cast<ConstantExpr>(LI->getOperand(0))) {
if(const PointerType *PTy = dyn_cast<PointerType>(CE->getOperand(0)->getType()))
if(PTy->getElementType() == CI->getType()) {
LoadInst *LINew = new LoadInst(CE->getOperand(0), "", LI);
CI->replaceAllUsesWith(LINew);
}
}
}
}
}
}
}
}
} while(changed);
return (numErased > 0);
}