void GraphBuilder::visitLoadInst(LoadInst &LI) {
//
// Create a DSNode for the pointer dereferenced by the load. If the DSNode
// is NULL, do nothing more (this can occur if the load is loading from a
// NULL pointer constant (bugpoint can generate such code).
//
DSNodeHandle Ptr = getValueDest(LI.getPointerOperand());
if (Ptr.isNull()) return; // Load from null
// Make that the node is read from...
Ptr.getNode()->setReadMarker();
// Ensure a typerecord exists...
Ptr.getNode()->growSizeForType(LI.getType(), Ptr.getOffset());
if (isa<PointerType>(LI.getType()))
setDestTo(LI, getLink(Ptr));
// check that it is the inserted value
if(TypeInferenceOptimize)
if(LI.hasOneUse())
if(StoreInst *SI = dyn_cast<StoreInst>(*(LI.use_begin())))
if(SI->getOperand(0) == &LI) {
++NumIgnoredInst;
return;
}
Ptr.getNode()->mergeTypeInfo(LI.getType(), Ptr.getOffset());
}
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;
}
//
// 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);
}
