bool AMDGPUPromoteAlloca::runOnFunction(Function &F) {
const FunctionType *FTy = F.getFunctionType();
LocalMemAvailable = ST.getLocalMemorySize();
// If the function has any arguments in the local address space, then it's
// possible these arguments require the entire local memory space, so
// we cannot use local memory in the pass.
for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i) {
const Type *ParamTy = FTy->getParamType(i);
if (ParamTy->isPointerTy() &&
ParamTy->getPointerAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) {
LocalMemAvailable = 0;
DEBUG(dbgs() << "Function has local memory argument. Promoting to "
"local memory disabled.\n");
break;
}
}
if (LocalMemAvailable > 0) {
// Check how much local memory is being used by global objects
for (Module::global_iterator I = Mod->global_begin(),
E = Mod->global_end(); I != E; ++I) {
GlobalVariable *GV = I;
PointerType *GVTy = GV->getType();
if (GVTy->getAddressSpace() != AMDGPUAS::LOCAL_ADDRESS)
continue;
for (Value::use_iterator U = GV->use_begin(),
UE = GV->use_end(); U != UE; ++U) {
Instruction *Use = dyn_cast<Instruction>(*U);
if (!Use)
continue;
if (Use->getParent()->getParent() == &F)
LocalMemAvailable -=
Mod->getDataLayout()->getTypeAllocSize(GVTy->getElementType());
}
}
}
LocalMemAvailable = std::max(0, LocalMemAvailable);
DEBUG(dbgs() << LocalMemAvailable << "bytes free in local memory.\n");
visit(F);
return false;
}
static void rewriteTlsVars(Module &M, std::vector<VarInfo> *TlsVars,
PointerType *TemplatePtrType) {
// Set up the intrinsic that reads the thread pointer.
Function *ReadTpFunc = Intrinsic::getDeclaration(&M, Intrinsic::nacl_read_tp);
for (std::vector<VarInfo>::iterator VarInfo = TlsVars->begin();
VarInfo != TlsVars->end();
++VarInfo) {
GlobalVariable *Var = VarInfo->TlsVar;
while (!Var->use_empty()) {
Use *U = &Var->use_begin().getUse();
Instruction *InsertPt = PhiSafeInsertPt(U);
Value *RawThreadPtr = CallInst::Create(ReadTpFunc, "tls_raw", InsertPt);
Value *TypedThreadPtr = new BitCastInst(RawThreadPtr, TemplatePtrType,
"tls_struct", InsertPt);
SmallVector<Value*, 3> Indexes;
// We use -1 because we use the x86-style TLS layout in which
// the TLS data is stored at addresses below the thread pointer.
// This is largely because a check in nacl_irt_thread_create()
// in irt/irt_thread.c requires the thread pointer to be a
// self-pointer on x86-32.
// TODO(mseaborn): I intend to remove that check because it is
// non-portable. In the mean time, we want PNaCl pexes to work
// in older Chromium releases when translated to nexes.
Indexes.push_back(ConstantInt::get(
M.getContext(), APInt(32, -1)));
Indexes.push_back(ConstantInt::get(
M.getContext(), APInt(32, VarInfo->IsBss ? 1 : 0)));
Indexes.push_back(ConstantInt::get(
M.getContext(), APInt(32, VarInfo->TemplateIndex)));
Value *TlsField = GetElementPtrInst::Create(TypedThreadPtr, Indexes,
"field", InsertPt);
PhiSafeReplaceUses(U, TlsField);
}
VarInfo->TlsVar->eraseFromParent();
}
}
bool GenericToNVVM::runOnModule(Module &M) {
// Create a clone of each global variable that has the default address space.
// The clone is created with the global address space specifier, and the pair
// of original global variable and its clone is placed in the GVMap for later
// use.
for (Module::global_iterator I = M.global_begin(), E = M.global_end();
I != E;) {
GlobalVariable *GV = I++;
if (GV->getType()->getAddressSpace() == llvm::ADDRESS_SPACE_GENERIC &&
!llvm::isTexture(*GV) && !llvm::isSurface(*GV) &&
!GV->getName().startswith("llvm.")) {
GlobalVariable *NewGV = new GlobalVariable(
M, GV->getType()->getElementType(), GV->isConstant(),
GV->getLinkage(),
GV->hasInitializer() ? GV->getInitializer() : nullptr,
"", GV, GV->getThreadLocalMode(), llvm::ADDRESS_SPACE_GLOBAL);
NewGV->copyAttributesFrom(GV);
GVMap[GV] = NewGV;
}
}
// Return immediately, if every global variable has a specific address space
// specifier.
if (GVMap.empty()) {
return false;
}
// Walk through the instructions in function defitinions, and replace any use
// of original global variables in GVMap with a use of the corresponding
// copies in GVMap. If necessary, promote constants to instructions.
for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
if (I->isDeclaration()) {
continue;
}
IRBuilder<> Builder(I->getEntryBlock().getFirstNonPHIOrDbg());
for (Function::iterator BBI = I->begin(), BBE = I->end(); BBI != BBE;
++BBI) {
for (BasicBlock::iterator II = BBI->begin(), IE = BBI->end(); II != IE;
++II) {
for (unsigned i = 0, e = II->getNumOperands(); i < e; ++i) {
Value *Operand = II->getOperand(i);
if (isa<Constant>(Operand)) {
II->setOperand(
i, remapConstant(&M, I, cast<Constant>(Operand), Builder));
}
}
}
}
ConstantToValueMap.clear();
}
// Walk through the metadata section and update the debug information
// associated with the global variables in the default address space.
for (Module::named_metadata_iterator I = M.named_metadata_begin(),
E = M.named_metadata_end();
I != E; I++) {
remapNamedMDNode(&M, I);
}
// Walk through the global variable initializers, and replace any use of
// original global variables in GVMap with a use of the corresponding copies
// in GVMap. The copies need to be bitcast to the original global variable
// types, as we cannot use cvta in global variable initializers.
for (GVMapTy::iterator I = GVMap.begin(), E = GVMap.end(); I != E;) {
GlobalVariable *GV = I->first;
GlobalVariable *NewGV = I->second;
++I;
Constant *BitCastNewGV = ConstantExpr::getPointerCast(NewGV, GV->getType());
// At this point, the remaining uses of GV should be found only in global
// variable initializers, as other uses have been already been removed
// while walking through the instructions in function definitions.
for (Value::use_iterator UI = GV->use_begin(), UE = GV->use_end();
UI != UE;)
(UI++)->set(BitCastNewGV);
std::string Name = GV->getName();
GV->removeDeadConstantUsers();
GV->eraseFromParent();
NewGV->setName(Name);
}
GVMap.clear();
return true;
}