/// RewriteSingleStoreAlloca - If there is only a single store to this value,
/// replace any loads of it that are directly dominated by the definition with
/// the value stored.
void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
AllocaInfo &Info,
LargeBlockInfo &LBI) {
StoreInst *OnlyStore = Info.OnlyStore;
bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
BasicBlock *StoreBB = OnlyStore->getParent();
int StoreIndex = -1;
// Clear out UsingBlocks. We will reconstruct it here if needed.
Info.UsingBlocks.clear();
for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
Instruction *UserInst = cast<Instruction>(*UI++);
if (!isa<LoadInst>(UserInst)) {
assert(UserInst == OnlyStore && "Should only have load/stores");
continue;
}
LoadInst *LI = cast<LoadInst>(UserInst);
// Okay, if we have a load from the alloca, we want to replace it with the
// only value stored to the alloca. We can do this if the value is
// dominated by the store. If not, we use the rest of the mem2reg machinery
// to insert the phi nodes as needed.
if (!StoringGlobalVal) { // Non-instructions are always dominated.
if (LI->getParent() == StoreBB) {
// If we have a use that is in the same block as the store, compare the
// indices of the two instructions to see which one came first. If the
// load came before the store, we can't handle it.
if (StoreIndex == -1)
StoreIndex = LBI.getInstructionIndex(OnlyStore);
if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
// Can't handle this load, bail out.
Info.UsingBlocks.push_back(StoreBB);
continue;
}
} else if (LI->getParent() != StoreBB &&
!dominates(StoreBB, LI->getParent())) {
// If the load and store are in different blocks, use BB dominance to
// check their relationships. If the store doesn't dom the use, bail
// out.
Info.UsingBlocks.push_back(LI->getParent());
continue;
}
}
// Otherwise, we *can* safely rewrite this load.
Value *ReplVal = OnlyStore->getOperand(0);
// If the replacement value is the load, this must occur in unreachable
// code.
if (ReplVal == LI)
ReplVal = UndefValue::get(LI->getType());
LI->replaceAllUsesWith(ReplVal);
if (AST && LI->getType()->isPointerTy())
AST->deleteValue(LI);
LI->eraseFromParent();
LBI.deleteValue(LI);
}
}
bool AMDGPUCodeGenPrepare::visitLoadInst(LoadInst &I) {
if (!WidenLoads)
return false;
if ((I.getPointerAddressSpace() == AMDGPUASI.CONSTANT_ADDRESS ||
I.getPointerAddressSpace() == AMDGPUASI.CONSTANT_ADDRESS_32BIT) &&
canWidenScalarExtLoad(I)) {
IRBuilder<> Builder(&I);
Builder.SetCurrentDebugLocation(I.getDebugLoc());
Type *I32Ty = Builder.getInt32Ty();
Type *PT = PointerType::get(I32Ty, I.getPointerAddressSpace());
Value *BitCast= Builder.CreateBitCast(I.getPointerOperand(), PT);
LoadInst *WidenLoad = Builder.CreateLoad(BitCast);
WidenLoad->copyMetadata(I);
// If we have range metadata, we need to convert the type, and not make
// assumptions about the high bits.
if (auto *Range = WidenLoad->getMetadata(LLVMContext::MD_range)) {
ConstantInt *Lower =
mdconst::extract<ConstantInt>(Range->getOperand(0));
if (Lower->getValue().isNullValue()) {
WidenLoad->setMetadata(LLVMContext::MD_range, nullptr);
} else {
Metadata *LowAndHigh[] = {
ConstantAsMetadata::get(ConstantInt::get(I32Ty, Lower->getValue().zext(32))),
// Don't make assumptions about the high bits.
ConstantAsMetadata::get(ConstantInt::get(I32Ty, 0))
};
WidenLoad->setMetadata(LLVMContext::MD_range,
MDNode::get(Mod->getContext(), LowAndHigh));
}
}
int TySize = Mod->getDataLayout().getTypeSizeInBits(I.getType());
Type *IntNTy = Builder.getIntNTy(TySize);
Value *ValTrunc = Builder.CreateTrunc(WidenLoad, IntNTy);
Value *ValOrig = Builder.CreateBitCast(ValTrunc, I.getType());
I.replaceAllUsesWith(ValOrig);
I.eraseFromParent();
return true;
}
return false;
}
/// PromoteSingleBlockAlloca - Many allocas are only used within a single basic
/// block. If this is the case, avoid traversing the CFG and inserting a lot of
/// potentially useless PHI nodes by just performing a single linear pass over
/// the basic block using the Alloca.
///
/// If we cannot promote this alloca (because it is read before it is written),
/// return true. This is necessary in cases where, due to control flow, the
/// alloca is potentially undefined on some control flow paths. e.g. code like
/// this is potentially correct:
///
/// for (...) { if (c) { A = undef; undef = B; } }
///
/// ... so long as A is not used before undef is set.
///
void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
LargeBlockInfo &LBI) {
// The trickiest case to handle is when we have large blocks. Because of this,
// this code is optimized assuming that large blocks happen. This does not
// significantly pessimize the small block case. This uses LargeBlockInfo to
// make it efficient to get the index of various operations in the block.
// Clear out UsingBlocks. We will reconstruct it here if needed.
Info.UsingBlocks.clear();
// Walk the use-def list of the alloca, getting the locations of all stores.
typedef SmallVector<std::pair<unsigned, StoreInst*>, 64> StoresByIndexTy;
StoresByIndexTy StoresByIndex;
for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
UI != E; ++UI)
if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
// If there are no stores to the alloca, just replace any loads with undef.
if (StoresByIndex.empty()) {
for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;)
if (LoadInst *LI = dyn_cast<LoadInst>(*UI++)) {
LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
if (AST && LI->getType()->isPointerTy())
AST->deleteValue(LI);
LBI.deleteValue(LI);
LI->eraseFromParent();
}
return;
}
// Sort the stores by their index, making it efficient to do a lookup with a
// binary search.
std::sort(StoresByIndex.begin(), StoresByIndex.end());
// Walk all of the loads from this alloca, replacing them with the nearest
// store above them, if any.
for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
LoadInst *LI = dyn_cast<LoadInst>(*UI++);
if (!LI) continue;
unsigned LoadIdx = LBI.getInstructionIndex(LI);
// Find the nearest store that has a lower than this load.
StoresByIndexTy::iterator I =
std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
std::pair<unsigned, StoreInst*>(LoadIdx, static_cast<StoreInst*>(0)),
StoreIndexSearchPredicate());
// If there is no store before this load, then we can't promote this load.
if (I == StoresByIndex.begin()) {
// Can't handle this load, bail out.
Info.UsingBlocks.push_back(LI->getParent());
continue;
}
// Otherwise, there was a store before this load, the load takes its value.
--I;
LI->replaceAllUsesWith(I->second->getOperand(0));
if (AST && LI->getType()->isPointerTy())
AST->deleteValue(LI);
LI->eraseFromParent();
LBI.deleteValue(LI);
}
}
/// tryAggregating - When scanning forward over instructions, we look for
/// other loads or stores that could be aggregated with this one.
/// Returns the last instruction added (if one was added) since we might have
/// removed some loads or stores and that might invalidate an iterator.
Instruction *AggregateGlobalOpsOpt::tryAggregating(Instruction *StartInst, Value *StartPtr,
bool DebugThis) {
if (TD == 0) return 0;
Module* M = StartInst->getParent()->getParent()->getParent();
LLVMContext& Context = StartInst->getContext();
Type* int8Ty = Type::getInt8Ty(Context);
Type* sizeTy = Type::getInt64Ty(Context);
Type* globalInt8PtrTy = int8Ty->getPointerTo(globalSpace);
bool isLoad = isa<LoadInst>(StartInst);
bool isStore = isa<StoreInst>(StartInst);
Instruction *lastAddedInsn = NULL;
Instruction *LastLoadOrStore = NULL;
SmallVector<Instruction*, 8> toRemove;
// Okay, so we now have a single global load/store. Scan to find
// all subsequent stores of the same value to offset from the same pointer.
// Join these together into ranges, so we can decide whether contiguous blocks
// are stored.
MemOpRanges Ranges(*TD);
// Put the first store in since we want to preserve the order.
Ranges.addInst(0, StartInst);
BasicBlock::iterator BI = StartInst;
for (++BI; !isa<TerminatorInst>(BI); ++BI) {
if( isGlobalLoadOrStore(BI, globalSpace, isLoad, isStore) ) {
// OK!
} else {
// If the instruction is readnone, ignore it, otherwise bail out. We
// don't even allow readonly here because we don't want something like:
// A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
if (BI->mayWriteToMemory())
break;
if (isStore && BI->mayReadFromMemory())
break;
continue;
}
if ( isStore && isa<StoreInst>(BI) ) {
StoreInst *NextStore = cast<StoreInst>(BI);
// If this is a store, see if we can merge it in.
if (!NextStore->isSimple()) break;
// Check to see if this store is to a constant offset from the start ptr.
int64_t Offset;
if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, *TD))
break;
Ranges.addStore(Offset, NextStore);
LastLoadOrStore = NextStore;
} else {
LoadInst *NextLoad = cast<LoadInst>(BI);
if (!NextLoad->isSimple()) break;
// Check to see if this load is to a constant offset from the start ptr.
int64_t Offset;
if (!IsPointerOffset(StartPtr, NextLoad->getPointerOperand(), Offset, *TD))
break;
Ranges.addLoad(Offset, NextLoad);
LastLoadOrStore = NextLoad;
}
}
// If we have no ranges, then we just had a single store with nothing that
// could be merged in. This is a very common case of course.
if (!Ranges.moreThanOneOp())
return 0;
// Divide the instructions between StartInst and LastLoadOrStore into
// addressing, memops, and uses of memops (uses of loads)
reorderAddressingMemopsUses(StartInst, LastLoadOrStore, DebugThis);
Instruction* insertBefore = StartInst;
IRBuilder<> builder(insertBefore);
// Now that we have full information about ranges, loop over the ranges and
// emit memcpy's for anything big enough to be worthwhile.
for (MemOpRanges::const_iterator I = Ranges.begin(), E = Ranges.end();
I != E; ++I) {
const MemOpRange &Range = *I;
Value* oldBaseI = NULL;
Value* newBaseI = NULL;
if (Range.TheStores.size() == 1) continue; // Don't bother if there's only one thing...
builder.SetInsertPoint(insertBefore);
// Otherwise, we do want to transform this! Create a new memcpy.
// Get the starting pointer of the block.
StartPtr = Range.StartPtr;
if( DebugThis ) {
errs() << "base is:";
StartPtr->dump();
}
// Determine alignment
unsigned Alignment = Range.Alignment;
if (Alignment == 0) {
Type *EltType =
cast<PointerType>(StartPtr->getType())->getElementType();
Alignment = TD->getABITypeAlignment(EltType);
}
Instruction *alloc = NULL;
Value *globalPtr = NULL;
// create temporary alloca space to communicate to/from.
alloc = makeAlloca(int8Ty, "agg.tmp", insertBefore,
Range.End-Range.Start, Alignment);
// Generate the old and new base pointers before we output
// anything else.
{
Type* iPtrTy = TD->getIntPtrType(alloc->getType());
Type* iNewBaseTy = TD->getIntPtrType(alloc->getType());
oldBaseI = builder.CreatePtrToInt(StartPtr, iPtrTy, "agg.tmp.oldb.i");
newBaseI = builder.CreatePtrToInt(alloc, iNewBaseTy, "agg.tmp.newb.i");
}
// If storing, do the stores we had into our alloca'd region.
if( isStore ) {
for (SmallVector<Instruction*, 16>::const_iterator
SI = Range.TheStores.begin(),
SE = Range.TheStores.end(); SI != SE; ++SI) {
StoreInst* oldStore = cast<StoreInst>(*SI);
if( DebugThis ) {
errs() << "have store in range:";
oldStore->dump();
}
Value* ptrToAlloc = rebasePointer(oldStore->getPointerOperand(),
StartPtr, alloc, "agg.tmp",
&builder, *TD, oldBaseI, newBaseI);
// Old load must not be volatile or atomic... or we shouldn't have put
// it in ranges
assert(!(oldStore->isVolatile() || oldStore->isAtomic()));
StoreInst* newStore =
builder.CreateStore(oldStore->getValueOperand(), ptrToAlloc);
newStore->setAlignment(oldStore->getAlignment());
newStore->takeName(oldStore);
}
}
// cast the pointer that was load/stored to i8 if necessary.
if( StartPtr->getType()->getPointerElementType() == int8Ty ) {
globalPtr = StartPtr;
} else {
globalPtr = builder.CreatePointerCast(StartPtr, globalInt8PtrTy, "agg.cast");
}
// Get a Constant* for the length.
Constant* len = ConstantInt::get(sizeTy, Range.End-Range.Start, false);
// Now add the memcpy instruction
unsigned addrSpaceDst,addrSpaceSrc;
addrSpaceDst = addrSpaceSrc = 0;
if( isStore ) addrSpaceDst = globalSpace;
if( isLoad ) addrSpaceSrc = globalSpace;
Type *types[3];
types[0] = PointerType::get(int8Ty, addrSpaceDst);
types[1] = PointerType::get(int8Ty, addrSpaceSrc);
types[2] = sizeTy;
Function *func = Intrinsic::getDeclaration(M, Intrinsic::memcpy, types);
Value* args[5]; // dst src len alignment isvolatile
if( isStore ) {
// it's a store (ie put)
args[0] = globalPtr;
args[1] = alloc;
} else {
// it's a load (ie get)
args[0] = alloc;
args[1] = globalPtr;
}
args[2] = len;
// alignment
args[3] = ConstantInt::get(Type::getInt32Ty(Context), 0, false);
// isvolatile
args[4] = ConstantInt::get(Type::getInt1Ty(Context), 0, false);
Instruction* aMemCpy = builder.CreateCall(func, args);
/*
DEBUG(dbgs() << "Replace ops:\n";
for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
dbgs() << *Range.TheStores[i] << '\n';
dbgs() << "With: " << *AMemSet << '\n');
*/
if (!Range.TheStores.empty())
aMemCpy->setDebugLoc(Range.TheStores[0]->getDebugLoc());
lastAddedInsn = aMemCpy;
// If loading, load from the memcpy'd region
if( isLoad ) {
for (SmallVector<Instruction*, 16>::const_iterator
SI = Range.TheStores.begin(),
SE = Range.TheStores.end(); SI != SE; ++SI) {
LoadInst* oldLoad = cast<LoadInst>(*SI);
if( DebugThis ) {
errs() << "have load in range:";
oldLoad->dump();
}
Value* ptrToAlloc = rebasePointer(oldLoad->getPointerOperand(),
StartPtr, alloc, "agg.tmp",
&builder, *TD, oldBaseI, newBaseI);
// Old load must not be volatile or atomic... or we shouldn't have put
// it in ranges
assert(!(oldLoad->isVolatile() || oldLoad->isAtomic()));
LoadInst* newLoad = builder.CreateLoad(ptrToAlloc);
newLoad->setAlignment(oldLoad->getAlignment());
oldLoad->replaceAllUsesWith(newLoad);
newLoad->takeName(oldLoad);
lastAddedInsn = newLoad;
}
}
// Save old loads/stores for removal
for (SmallVector<Instruction*, 16>::const_iterator
SI = Range.TheStores.begin(),
SE = Range.TheStores.end(); SI != SE; ++SI) {
Instruction* insn = *SI;
toRemove.push_back(insn);
}
}
// Zap all the old loads/stores
for (SmallVector<Instruction*, 16>::const_iterator
SI = toRemove.begin(),
SE = toRemove.end(); SI != SE; ++SI) {
(*SI)->eraseFromParent();
}
return lastAddedInsn;
}
/// 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;
}
/// Many allocas are only used within a single basic block. If this is the
/// case, avoid traversing the CFG and inserting a lot of potentially useless
/// PHI nodes by just performing a single linear pass over the basic block
/// using the Alloca.
///
/// If we cannot promote this alloca (because it is read before it is written),
/// return true. This is necessary in cases where, due to control flow, the
/// alloca is potentially undefined on some control flow paths. e.g. code like
/// this is potentially correct:
///
/// for (...) { if (c) { A = undef; undef = B; } }
///
/// ... so long as A is not used before undef is set.
static void promoteSingleBlockAlloca(AllocaInst *AI, const AllocaInfo &Info,
LargeBlockInfo &LBI,
AliasSetTracker *AST) {
// The trickiest case to handle is when we have large blocks. Because of this,
// this code is optimized assuming that large blocks happen. This does not
// significantly pessimize the small block case. This uses LargeBlockInfo to
// make it efficient to get the index of various operations in the block.
// Walk the use-def list of the alloca, getting the locations of all stores.
typedef SmallVector<std::pair<unsigned, StoreInst *>, 64> StoresByIndexTy;
StoresByIndexTy StoresByIndex;
for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;
++UI)
if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
// Sort the stores by their index, making it efficient to do a lookup with a
// binary search.
std::sort(StoresByIndex.begin(), StoresByIndex.end(),
StoreIndexSearchPredicate());
// Walk all of the loads from this alloca, replacing them with the nearest
// store above them, if any.
for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
LoadInst *LI = dyn_cast<LoadInst>(*UI++);
if (!LI)
continue;
unsigned LoadIdx = LBI.getInstructionIndex(LI);
// Find the nearest store that has a lower index than this load.
StoresByIndexTy::iterator I =
std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
std::make_pair(LoadIdx, static_cast<StoreInst *>(0)),
StoreIndexSearchPredicate());
if (I == StoresByIndex.begin())
// If there is no store before this load, the load takes the undef value.
LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
else
// Otherwise, there was a store before this load, the load takes its value.
LI->replaceAllUsesWith(llvm::prior(I)->second->getOperand(0));
if (AST && LI->getType()->isPointerTy())
AST->deleteValue(LI);
LI->eraseFromParent();
LBI.deleteValue(LI);
}
// Remove the (now dead) stores and alloca.
while (!AI->use_empty()) {
StoreInst *SI = cast<StoreInst>(AI->use_back());
// Record debuginfo for the store before removing it.
if (DbgDeclareInst *DDI = Info.DbgDeclare) {
DIBuilder DIB(*AI->getParent()->getParent()->getParent());
ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
}
SI->eraseFromParent();
LBI.deleteValue(SI);
}
if (AST)
AST->deleteValue(AI);
AI->eraseFromParent();
LBI.deleteValue(AI);
// The alloca's debuginfo can be removed as well.
if (DbgDeclareInst *DDI = Info.DbgDeclare)
DDI->eraseFromParent();
++NumLocalPromoted;
}
/// \brief Rewrite as many loads as possible given a single store.
///
/// When there is only a single store, we can use the domtree to trivially
/// replace all of the dominated loads with the stored value. Do so, and return
/// true if this has successfully promoted the alloca entirely. If this returns
/// false there were some loads which were not dominated by the single store
/// and thus must be phi-ed with undef. We fall back to the standard alloca
/// promotion algorithm in that case.
static bool rewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
LargeBlockInfo &LBI,
DominatorTree &DT,
AliasSetTracker *AST) {
StoreInst *OnlyStore = Info.OnlyStore;
bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
BasicBlock *StoreBB = OnlyStore->getParent();
int StoreIndex = -1;
// Clear out UsingBlocks. We will reconstruct it here if needed.
Info.UsingBlocks.clear();
for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
Instruction *UserInst = cast<Instruction>(*UI++);
if (!isa<LoadInst>(UserInst)) {
assert(UserInst == OnlyStore && "Should only have load/stores");
continue;
}
LoadInst *LI = cast<LoadInst>(UserInst);
// Okay, if we have a load from the alloca, we want to replace it with the
// only value stored to the alloca. We can do this if the value is
// dominated by the store. If not, we use the rest of the mem2reg machinery
// to insert the phi nodes as needed.
if (!StoringGlobalVal) { // Non-instructions are always dominated.
if (LI->getParent() == StoreBB) {
// If we have a use that is in the same block as the store, compare the
// indices of the two instructions to see which one came first. If the
// load came before the store, we can't handle it.
if (StoreIndex == -1)
StoreIndex = LBI.getInstructionIndex(OnlyStore);
if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
// Can't handle this load, bail out.
Info.UsingBlocks.push_back(StoreBB);
continue;
}
} else if (LI->getParent() != StoreBB &&
!DT.dominates(StoreBB, LI->getParent())) {
// If the load and store are in different blocks, use BB dominance to
// check their relationships. If the store doesn't dom the use, bail
// out.
Info.UsingBlocks.push_back(LI->getParent());
continue;
}
}
// Otherwise, we *can* safely rewrite this load.
Value *ReplVal = OnlyStore->getOperand(0);
// If the replacement value is the load, this must occur in unreachable
// code.
if (ReplVal == LI)
ReplVal = UndefValue::get(LI->getType());
LI->replaceAllUsesWith(ReplVal);
if (AST && LI->getType()->isPointerTy())
AST->deleteValue(LI);
LI->eraseFromParent();
LBI.deleteValue(LI);
}
// Finally, after the scan, check to see if the store is all that is left.
if (!Info.UsingBlocks.empty())
return false; // If not, we'll have to fall back for the remainder.
// Record debuginfo for the store and remove the declaration's
// debuginfo.
if (DbgDeclareInst *DDI = Info.DbgDeclare) {
DIBuilder DIB(*AI->getParent()->getParent()->getParent());
ConvertDebugDeclareToDebugValue(DDI, Info.OnlyStore, DIB);
DDI->eraseFromParent();
}
// Remove the (now dead) store and alloca.
Info.OnlyStore->eraseFromParent();
LBI.deleteValue(Info.OnlyStore);
if (AST)
AST->deleteValue(AI);
AI->eraseFromParent();
LBI.deleteValue(AI);
return true;
}
/// updateCallSites - Update all sites that call F to use NF.
CallGraphNode *SRETPromotion::updateCallSites(Function *F, Function *NF) {
CallGraph &CG = getAnalysis<CallGraph>();
SmallVector<Value*, 16> Args;
// Attributes - Keep track of the parameter attributes for the arguments.
SmallVector<AttributeWithIndex, 8> ArgAttrsVec;
// Get a new callgraph node for NF.
CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF);
while (!F->use_empty()) {
CallSite CS(*F->use_begin());
Instruction *Call = CS.getInstruction();
const AttrListPtr &PAL = F->getAttributes();
// Add any return attributes.
if (Attributes attrs = PAL.getRetAttributes())
ArgAttrsVec.push_back(AttributeWithIndex::get(0, attrs));
// Copy arguments, however skip first one.
CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
Value *FirstCArg = *AI;
++AI;
// 0th parameter attribute is reserved for return type.
// 1th parameter attribute is for first 1st sret argument.
unsigned ParamIndex = 2;
while (AI != AE) {
Args.push_back(*AI);
if (Attributes Attrs = PAL.getParamAttributes(ParamIndex))
ArgAttrsVec.push_back(AttributeWithIndex::get(ParamIndex - 1, Attrs));
++ParamIndex;
++AI;
}
// Add any function attributes.
if (Attributes attrs = PAL.getFnAttributes())
ArgAttrsVec.push_back(AttributeWithIndex::get(~0, attrs));
AttrListPtr NewPAL = AttrListPtr::get(ArgAttrsVec.begin(), ArgAttrsVec.end());
// Build new call instruction.
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(NewPAL);
} else {
New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call);
cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
cast<CallInst>(New)->setAttributes(NewPAL);
if (cast<CallInst>(Call)->isTailCall())
cast<CallInst>(New)->setTailCall();
}
Args.clear();
ArgAttrsVec.clear();
New->takeName(Call);
// Update the callgraph to know that the callsite has been transformed.
CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
CalleeNode->removeCallEdgeFor(Call);
CalleeNode->addCalledFunction(New, NF_CGN);
// Update all users of sret parameter to extract value using extractvalue.
for (Value::use_iterator UI = FirstCArg->use_begin(),
UE = FirstCArg->use_end(); UI != UE; ) {
User *U2 = *UI++;
CallInst *C2 = dyn_cast<CallInst>(U2);
if (C2 && (C2 == Call))
continue;
GetElementPtrInst *UGEP = cast<GetElementPtrInst>(U2);
ConstantInt *Idx = cast<ConstantInt>(UGEP->getOperand(2));
Value *GR = ExtractValueInst::Create(New, Idx->getZExtValue(),
"evi", UGEP);
while(!UGEP->use_empty()) {
// isSafeToUpdateAllCallers has checked that all GEP uses are
// LoadInsts
LoadInst *L = cast<LoadInst>(*UGEP->use_begin());
L->replaceAllUsesWith(GR);
L->eraseFromParent();
}
UGEP->eraseFromParent();
continue;
}
Call->eraseFromParent();
}
return NF_CGN;
}
void LoadAndStorePromoter::
run(const SmallVectorImpl<Instruction*> &Insts) const {
// First step: bucket up uses of the alloca by the block they occur in.
// This is important because we have to handle multiple defs/uses in a block
// ourselves: SSAUpdater is purely for cross-block references.
// FIXME: Want a TinyVector<Instruction*> since there is often 0/1 element.
DenseMap<BasicBlock*, std::vector<Instruction*> > UsesByBlock;
for (unsigned i = 0, e = Insts.size(); i != e; ++i) {
Instruction *User = Insts[i];
UsesByBlock[User->getParent()].push_back(User);
}
// Okay, now we can iterate over all the blocks in the function with uses,
// processing them. Keep track of which loads are loading a live-in value.
// Walk the uses in the use-list order to be determinstic.
SmallVector<LoadInst*, 32> LiveInLoads;
DenseMap<Value*, Value*> ReplacedLoads;
for (unsigned i = 0, e = Insts.size(); i != e; ++i) {
Instruction *User = Insts[i];
BasicBlock *BB = User->getParent();
std::vector<Instruction*> &BlockUses = UsesByBlock[BB];
// If this block has already been processed, ignore this repeat use.
if (BlockUses.empty()) continue;
// Okay, this is the first use in the block. If this block just has a
// single user in it, we can rewrite it trivially.
if (BlockUses.size() == 1) {
// If it is a store, it is a trivial def of the value in the block.
if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
updateDebugInfo(SI);
SSA.AddAvailableValue(BB, SI->getOperand(0));
} else
// Otherwise it is a load, queue it to rewrite as a live-in load.
LiveInLoads.push_back(cast<LoadInst>(User));
BlockUses.clear();
continue;
}
// Otherwise, check to see if this block is all loads.
bool HasStore = false;
for (unsigned i = 0, e = BlockUses.size(); i != e; ++i) {
if (isa<StoreInst>(BlockUses[i])) {
HasStore = true;
break;
}
}
// If so, we can queue them all as live in loads. We don't have an
// efficient way to tell which on is first in the block and don't want to
// scan large blocks, so just add all loads as live ins.
if (!HasStore) {
for (unsigned i = 0, e = BlockUses.size(); i != e; ++i)
LiveInLoads.push_back(cast<LoadInst>(BlockUses[i]));
BlockUses.clear();
continue;
}
// Otherwise, we have mixed loads and stores (or just a bunch of stores).
// Since SSAUpdater is purely for cross-block values, we need to determine
// the order of these instructions in the block. If the first use in the
// block is a load, then it uses the live in value. The last store defines
// the live out value. We handle this by doing a linear scan of the block.
Value *StoredValue = 0;
for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) {
if (LoadInst *L = dyn_cast<LoadInst>(II)) {
// If this is a load from an unrelated pointer, ignore it.
if (!isInstInList(L, Insts)) continue;
// If we haven't seen a store yet, this is a live in use, otherwise
// use the stored value.
if (StoredValue) {
replaceLoadWithValue(L, StoredValue);
L->replaceAllUsesWith(StoredValue);
ReplacedLoads[L] = StoredValue;
} else {
LiveInLoads.push_back(L);
}
continue;
}
if (StoreInst *SI = dyn_cast<StoreInst>(II)) {
// If this is a store to an unrelated pointer, ignore it.
if (!isInstInList(SI, Insts)) continue;
updateDebugInfo(SI);
// Remember that this is the active value in the block.
StoredValue = SI->getOperand(0);
}
}
// The last stored value that happened is the live-out for the block.
assert(StoredValue && "Already checked that there is a store in block");
SSA.AddAvailableValue(BB, StoredValue);
BlockUses.clear();
}
// Okay, now we rewrite all loads that use live-in values in the loop,
// inserting PHI nodes as necessary.
for (unsigned i = 0, e = LiveInLoads.size(); i != e; ++i) {
LoadInst *ALoad = LiveInLoads[i];
Value *NewVal = SSA.GetValueInMiddleOfBlock(ALoad->getParent());
replaceLoadWithValue(ALoad, NewVal);
// Avoid assertions in unreachable code.
if (NewVal == ALoad) NewVal = UndefValue::get(NewVal->getType());
ALoad->replaceAllUsesWith(NewVal);
ReplacedLoads[ALoad] = NewVal;
}
// Allow the client to do stuff before we start nuking things.
doExtraRewritesBeforeFinalDeletion();
// Now that everything is rewritten, delete the old instructions from the
// function. They should all be dead now.
for (unsigned i = 0, e = Insts.size(); i != e; ++i) {
Instruction *User = Insts[i];
// If this is a load that still has uses, then the load must have been added
// as a live value in the SSAUpdate data structure for a block (e.g. because
// the loaded value was stored later). In this case, we need to recursively
// propagate the updates until we get to the real value.
if (!User->use_empty()) {
Value *NewVal = ReplacedLoads[User];
assert(NewVal && "not a replaced load?");
// Propagate down to the ultimate replacee. The intermediately loads
// could theoretically already have been deleted, so we don't want to
// dereference the Value*'s.
DenseMap<Value*, Value*>::iterator RLI = ReplacedLoads.find(NewVal);
while (RLI != ReplacedLoads.end()) {
NewVal = RLI->second;
RLI = ReplacedLoads.find(NewVal);
}
replaceLoadWithValue(cast<LoadInst>(User), NewVal);
User->replaceAllUsesWith(NewVal);
}
instructionDeleted(User);
User->eraseFromParent();
}
}
/// Many allocas are only used within a single basic block. If this is the
/// case, avoid traversing the CFG and inserting a lot of potentially useless
/// PHI nodes by just performing a single linear pass over the basic block
/// using the Alloca.
///
/// If we cannot promote this alloca (because it is read before it is written),
/// return false. This is necessary in cases where, due to control flow, the
/// alloca is undefined only on some control flow paths. e.g. code like
/// this is correct in LLVM IR:
/// // A is an alloca with no stores so far
/// for (...) {
/// int t = *A;
/// if (!first_iteration)
/// use(t);
/// *A = 42;
/// }
static bool promoteSingleBlockAlloca(AllocaInst *AI, const AllocaInfo &Info,
LargeBlockInfo &LBI,
const DataLayout &DL,
DominatorTree &DT,
AssumptionCache *AC) {
// The trickiest case to handle is when we have large blocks. Because of this,
// this code is optimized assuming that large blocks happen. This does not
// significantly pessimize the small block case. This uses LargeBlockInfo to
// make it efficient to get the index of various operations in the block.
// Walk the use-def list of the alloca, getting the locations of all stores.
using StoresByIndexTy = SmallVector<std::pair<unsigned, StoreInst *>, 64>;
StoresByIndexTy StoresByIndex;
for (User *U : AI->users())
if (StoreInst *SI = dyn_cast<StoreInst>(U))
StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
// Sort the stores by their index, making it efficient to do a lookup with a
// binary search.
llvm::sort(StoresByIndex, less_first());
// Walk all of the loads from this alloca, replacing them with the nearest
// store above them, if any.
for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
LoadInst *LI = dyn_cast<LoadInst>(*UI++);
if (!LI)
continue;
unsigned LoadIdx = LBI.getInstructionIndex(LI);
// Find the nearest store that has a lower index than this load.
StoresByIndexTy::iterator I =
std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
std::make_pair(LoadIdx,
static_cast<StoreInst *>(nullptr)),
less_first());
if (I == StoresByIndex.begin()) {
if (StoresByIndex.empty())
// If there are no stores, the load takes the undef value.
LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
else
// There is no store before this load, bail out (load may be affected
// by the following stores - see main comment).
return false;
} else {
// Otherwise, there was a store before this load, the load takes its value.
// Note, if the load was marked as nonnull we don't want to lose that
// information when we erase it. So we preserve it with an assume.
Value *ReplVal = std::prev(I)->second->getOperand(0);
if (AC && LI->getMetadata(LLVMContext::MD_nonnull) &&
!isKnownNonZero(ReplVal, DL, 0, AC, LI, &DT))
addAssumeNonNull(AC, LI);
// If the replacement value is the load, this must occur in unreachable
// code.
if (ReplVal == LI)
ReplVal = UndefValue::get(LI->getType());
LI->replaceAllUsesWith(ReplVal);
}
LI->eraseFromParent();
LBI.deleteValue(LI);
}
// Remove the (now dead) stores and alloca.
while (!AI->use_empty()) {
StoreInst *SI = cast<StoreInst>(AI->user_back());
// Record debuginfo for the store before removing it.
for (DbgVariableIntrinsic *DII : Info.DbgDeclares) {
DIBuilder DIB(*AI->getModule(), /*AllowUnresolved*/ false);
ConvertDebugDeclareToDebugValue(DII, SI, DIB);
}
SI->eraseFromParent();
LBI.deleteValue(SI);
}
AI->eraseFromParent();
LBI.deleteValue(AI);
// The alloca's debuginfo can be removed as well.
for (DbgVariableIntrinsic *DII : Info.DbgDeclares) {
DII->eraseFromParent();
LBI.deleteValue(DII);
}
++NumLocalPromoted;
return true;
}
/// Rewrite as many loads as possible given a single store.
///
/// When there is only a single store, we can use the domtree to trivially
/// replace all of the dominated loads with the stored value. Do so, and return
/// true if this has successfully promoted the alloca entirely. If this returns
/// false there were some loads which were not dominated by the single store
/// and thus must be phi-ed with undef. We fall back to the standard alloca
/// promotion algorithm in that case.
static bool rewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
LargeBlockInfo &LBI, const DataLayout &DL,
DominatorTree &DT, AssumptionCache *AC) {
StoreInst *OnlyStore = Info.OnlyStore;
bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
BasicBlock *StoreBB = OnlyStore->getParent();
int StoreIndex = -1;
// Clear out UsingBlocks. We will reconstruct it here if needed.
Info.UsingBlocks.clear();
for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
Instruction *UserInst = cast<Instruction>(*UI++);
if (!isa<LoadInst>(UserInst)) {
assert(UserInst == OnlyStore && "Should only have load/stores");
continue;
}
LoadInst *LI = cast<LoadInst>(UserInst);
// Okay, if we have a load from the alloca, we want to replace it with the
// only value stored to the alloca. We can do this if the value is
// dominated by the store. If not, we use the rest of the mem2reg machinery
// to insert the phi nodes as needed.
if (!StoringGlobalVal) { // Non-instructions are always dominated.
if (LI->getParent() == StoreBB) {
// If we have a use that is in the same block as the store, compare the
// indices of the two instructions to see which one came first. If the
// load came before the store, we can't handle it.
if (StoreIndex == -1)
StoreIndex = LBI.getInstructionIndex(OnlyStore);
if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
// Can't handle this load, bail out.
Info.UsingBlocks.push_back(StoreBB);
continue;
}
} else if (LI->getParent() != StoreBB &&
!DT.dominates(StoreBB, LI->getParent())) {
// If the load and store are in different blocks, use BB dominance to
// check their relationships. If the store doesn't dom the use, bail
// out.
Info.UsingBlocks.push_back(LI->getParent());
continue;
}
}
// Otherwise, we *can* safely rewrite this load.
Value *ReplVal = OnlyStore->getOperand(0);
// If the replacement value is the load, this must occur in unreachable
// code.
if (ReplVal == LI)
ReplVal = UndefValue::get(LI->getType());
// If the load was marked as nonnull we don't want to lose
// that information when we erase this Load. So we preserve
// it with an assume.
if (AC && LI->getMetadata(LLVMContext::MD_nonnull) &&
!isKnownNonZero(ReplVal, DL, 0, AC, LI, &DT))
addAssumeNonNull(AC, LI);
LI->replaceAllUsesWith(ReplVal);
LI->eraseFromParent();
LBI.deleteValue(LI);
}
// Finally, after the scan, check to see if the store is all that is left.
if (!Info.UsingBlocks.empty())
return false; // If not, we'll have to fall back for the remainder.
// Record debuginfo for the store and remove the declaration's
// debuginfo.
for (DbgVariableIntrinsic *DII : Info.DbgDeclares) {
DIBuilder DIB(*AI->getModule(), /*AllowUnresolved*/ false);
ConvertDebugDeclareToDebugValue(DII, Info.OnlyStore, DIB);
DII->eraseFromParent();
LBI.deleteValue(DII);
}
// Remove the (now dead) store and alloca.
Info.OnlyStore->eraseFromParent();
LBI.deleteValue(Info.OnlyStore);
AI->eraseFromParent();
LBI.deleteValue(AI);
return true;
}
/// PromoteAliasSet - Try to promote memory values to scalars by sinking
/// stores out of the loop and moving loads to before the loop. We do this by
/// looping over the stores in the loop, looking for stores to Must pointers
/// which are loop invariant.
///
void LICM::PromoteAliasSet(AliasSet &AS) {
// We can promote this alias set if it has a store, if it is a "Must" alias
// set, if the pointer is loop invariant, and if we are not eliminating any
// volatile loads or stores.
if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue()))
return;
assert(!AS.empty() &&
"Must alias set should have at least one pointer element in it!");
Value *SomePtr = AS.begin()->getValue();
// It isn't safe to promote a load/store from the loop if the load/store is
// conditional. For example, turning:
//
// for () { if (c) *P += 1; }
//
// into:
//
// tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
//
// is not safe, because *P may only be valid to access if 'c' is true.
//
// It is safe to promote P if all uses are direct load/stores and if at
// least one is guaranteed to be executed.
bool GuaranteedToExecute = false;
SmallVector<Instruction*, 64> LoopUses;
SmallPtrSet<Value*, 4> PointerMustAliases;
// Check that all of the pointers in the alias set have the same type. We
// cannot (yet) promote a memory location that is loaded and stored in
// different sizes.
for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) {
Value *ASIV = ASI->getValue();
PointerMustAliases.insert(ASIV);
// Check that all of the pointers in the alias set have the same type. We
// cannot (yet) promote a memory location that is loaded and stored in
// different sizes.
if (SomePtr->getType() != ASIV->getType())
return;
for (Value::use_iterator UI = ASIV->use_begin(), UE = ASIV->use_end();
UI != UE; ++UI) {
// Ignore instructions that are outside the loop.
Instruction *Use = dyn_cast<Instruction>(*UI);
if (!Use || !CurLoop->contains(Use))
continue;
// If there is an non-load/store instruction in the loop, we can't promote
// it.
if (isa<LoadInst>(Use))
assert(!cast<LoadInst>(Use)->isVolatile() && "AST broken");
else if (isa<StoreInst>(Use)) {
assert(!cast<StoreInst>(Use)->isVolatile() && "AST broken");
if (Use->getOperand(0) == ASIV) return;
} else
return; // Not a load or store.
if (!GuaranteedToExecute)
GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use);
LoopUses.push_back(Use);
}
}
// If there isn't a guaranteed-to-execute instruction, we can't promote.
if (!GuaranteedToExecute)
return;
// Otherwise, this is safe to promote, lets do it!
DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n');
Changed = true;
++NumPromoted;
// We use the SSAUpdater interface to insert phi nodes as required.
SmallVector<PHINode*, 16> NewPHIs;
SSAUpdater SSA(&NewPHIs);
// It wants to know some value of the same type as what we'll be inserting.
Value *SomeValue;
if (isa<LoadInst>(LoopUses[0]))
SomeValue = LoopUses[0];
else
SomeValue = cast<StoreInst>(LoopUses[0])->getOperand(0);
SSA.Initialize(SomeValue->getType(), SomeValue->getName());
// First step: bucket up uses of the pointers by the block they occur in.
// This is important because we have to handle multiple defs/uses in a block
// ourselves: SSAUpdater is purely for cross-block references.
// FIXME: Want a TinyVector<Instruction*> since there is usually 0/1 element.
DenseMap<BasicBlock*, std::vector<Instruction*> > UsesByBlock;
for (unsigned i = 0, e = LoopUses.size(); i != e; ++i) {
Instruction *User = LoopUses[i];
UsesByBlock[User->getParent()].push_back(User);
}
// Okay, now we can iterate over all the blocks in the loop with uses,
// processing them. Keep track of which loads are loading a live-in value.
SmallVector<LoadInst*, 32> LiveInLoads;
DenseMap<Value*, Value*> ReplacedLoads;
for (unsigned LoopUse = 0, e = LoopUses.size(); LoopUse != e; ++LoopUse) {
Instruction *User = LoopUses[LoopUse];
std::vector<Instruction*> &BlockUses = UsesByBlock[User->getParent()];
// If this block has already been processed, ignore this repeat use.
if (BlockUses.empty()) continue;
// Okay, this is the first use in the block. If this block just has a
// single user in it, we can rewrite it trivially.
if (BlockUses.size() == 1) {
// If it is a store, it is a trivial def of the value in the block.
if (isa<StoreInst>(User)) {
SSA.AddAvailableValue(User->getParent(),
cast<StoreInst>(User)->getOperand(0));
} else {
// Otherwise it is a load, queue it to rewrite as a live-in load.
LiveInLoads.push_back(cast<LoadInst>(User));
}
BlockUses.clear();
continue;
}
// Otherwise, check to see if this block is all loads. If so, we can queue
// them all as live in loads.
bool HasStore = false;
for (unsigned i = 0, e = BlockUses.size(); i != e; ++i) {
if (isa<StoreInst>(BlockUses[i])) {
HasStore = true;
break;
}
}
if (!HasStore) {
for (unsigned i = 0, e = BlockUses.size(); i != e; ++i)
LiveInLoads.push_back(cast<LoadInst>(BlockUses[i]));
BlockUses.clear();
continue;
}
// Otherwise, we have mixed loads and stores (or just a bunch of stores).
// Since SSAUpdater is purely for cross-block values, we need to determine
// the order of these instructions in the block. If the first use in the
// block is a load, then it uses the live in value. The last store defines
// the live out value. We handle this by doing a linear scan of the block.
BasicBlock *BB = User->getParent();
Value *StoredValue = 0;
for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) {
if (LoadInst *L = dyn_cast<LoadInst>(II)) {
// If this is a load from an unrelated pointer, ignore it.
if (!PointerMustAliases.count(L->getOperand(0))) continue;
// If we haven't seen a store yet, this is a live in use, otherwise
// use the stored value.
if (StoredValue) {
L->replaceAllUsesWith(StoredValue);
ReplacedLoads[L] = StoredValue;
} else {
LiveInLoads.push_back(L);
}
continue;
}
if (StoreInst *S = dyn_cast<StoreInst>(II)) {
// If this is a store to an unrelated pointer, ignore it.
if (!PointerMustAliases.count(S->getOperand(1))) continue;
// Remember that this is the active value in the block.
StoredValue = S->getOperand(0);
}
}
// The last stored value that happened is the live-out for the block.
assert(StoredValue && "Already checked that there is a store in block");
SSA.AddAvailableValue(BB, StoredValue);
BlockUses.clear();
}
// Now that all the intra-loop values are classified, set up the preheader.
// It gets a load of the pointer we're promoting, and it is the live-out value
// from the preheader.
LoadInst *PreheaderLoad = new LoadInst(SomePtr,SomePtr->getName()+".promoted",
Preheader->getTerminator());
SSA.AddAvailableValue(Preheader, PreheaderLoad);
// Now that the preheader is good to go, set up the exit blocks. Each exit
// block gets a store of the live-out values that feed them. Since we've
// already told the SSA updater about the defs in the loop and the preheader
// definition, it is all set and we can start using it.
SmallVector<BasicBlock*, 8> ExitBlocks;
CurLoop->getUniqueExitBlocks(ExitBlocks);
for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
BasicBlock *ExitBlock = ExitBlocks[i];
Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
Instruction *InsertPos = ExitBlock->getFirstNonPHI();
new StoreInst(LiveInValue, SomePtr, InsertPos);
}
// Okay, now we rewrite all loads that use live-in values in the loop,
// inserting PHI nodes as necessary.
for (unsigned i = 0, e = LiveInLoads.size(); i != e; ++i) {
LoadInst *ALoad = LiveInLoads[i];
Value *NewVal = SSA.GetValueInMiddleOfBlock(ALoad->getParent());
ALoad->replaceAllUsesWith(NewVal);
CurAST->copyValue(ALoad, NewVal);
ReplacedLoads[ALoad] = NewVal;
}
// If the preheader load is itself a pointer, we need to tell alias analysis
// about the new pointer we created in the preheader block and about any PHI
// nodes that just got inserted.
if (PreheaderLoad->getType()->isPointerTy()) {
// Copy any value stored to or loaded from a must-alias of the pointer.
CurAST->copyValue(SomeValue, PreheaderLoad);
for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
CurAST->copyValue(SomeValue, NewPHIs[i]);
}
// Now that everything is rewritten, delete the old instructions from the body
// of the loop. They should all be dead now.
for (unsigned i = 0, e = LoopUses.size(); i != e; ++i) {
Instruction *User = LoopUses[i];
// If this is a load that still has uses, then the load must have been added
// as a live value in the SSAUpdate data structure for a block (e.g. because
// the loaded value was stored later). In this case, we need to recursively
// propagate the updates until we get to the real value.
if (!User->use_empty()) {
Value *NewVal = ReplacedLoads[User];
assert(NewVal && "not a replaced load?");
// Propagate down to the ultimate replacee. The intermediately loads
// could theoretically already have been deleted, so we don't want to
// dereference the Value*'s.
DenseMap<Value*, Value*>::iterator RLI = ReplacedLoads.find(NewVal);
while (RLI != ReplacedLoads.end()) {
NewVal = RLI->second;
RLI = ReplacedLoads.find(NewVal);
}
User->replaceAllUsesWith(NewVal);
CurAST->copyValue(User, NewVal);
}
CurAST->deleteValue(User);
User->eraseFromParent();
}
// fwew, we're done!
}
/// 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.
static Function *
doPromotion(Function *F, SmallPtrSetImpl<Argument *> &ArgsToPromote,
SmallPtrSetImpl<Argument *> &ByValArgsToTransform,
Optional<function_ref<void(CallSite OldCS, CallSite NewCS)>>
ReplaceCallSite) {
// Start by computing a new prototype for the function, which is the same as
// the old function, but has modified arguments.
FunctionType *FTy = F->getFunctionType();
std::vector<Type *> Params;
using ScalarizeTable = std::set<std::pair<Type *, IndicesVector>>;
// 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.
// We need to keep the original loads for each argument and the elements
// of the argument that are accessed.
std::map<std::pair<Argument *, IndicesVector>, LoadInst *> OriginalLoads;
// Attribute - 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<AttributeSet, 8> ArgAttrVec;
AttributeList PAL = F->getAttributes();
// First, determine the new argument list
unsigned ArgNo = 0;
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
++I, ++ArgNo) {
if (ByValArgsToTransform.count(&*I)) {
// Simple byval argument? Just add all the struct element types.
Type *AgTy = cast<PointerType>(I->getType())->getElementType();
StructType *STy = cast<StructType>(AgTy);
Params.insert(Params.end(), STy->element_begin(), STy->element_end());
ArgAttrVec.insert(ArgAttrVec.end(), STy->getNumElements(),
AttributeSet());
++NumByValArgsPromoted;
} else if (!ArgsToPromote.count(&*I)) {
// Unchanged argument
Params.push_back(I->getType());
ArgAttrVec.push_back(PAL.getParamAttributes(ArgNo));
} else if (I->use_empty()) {
// Dead argument (which are always marked as promotable)
++NumArgumentsDead;
// There may be remaining metadata uses of the argument for things like
// llvm.dbg.value. Replace them with undef.
I->replaceAllUsesWith(UndefValue::get(I->getType()));
} 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 (User *U : I->users()) {
Instruction *UI = cast<Instruction>(U);
Type *SrcTy;
if (LoadInst *L = dyn_cast<LoadInst>(UI))
SrcTy = L->getType();
else
SrcTy = cast<GetElementPtrInst>(UI)->getSourceElementType();
IndicesVector Indices;
Indices.reserve(UI->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 = UI->op_begin() + 1, IE = UI->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(std::make_pair(SrcTy, Indices));
LoadInst *OrigLoad;
if (LoadInst *L = dyn_cast<LoadInst>(UI))
OrigLoad = L;
else
// Take any load, we will use it only to update Alias Analysis
OrigLoad = cast<LoadInst>(UI->user_back());
OriginalLoads[std::make_pair(&*I, Indices)] = OrigLoad;
}
// Add a parameter to the function for each element passed in.
for (const auto &ArgIndex : ArgIndices) {
// not allowed to dereference ->begin() if size() is 0
Params.push_back(GetElementPtrInst::getIndexedType(
cast<PointerType>(I->getType()->getScalarType())->getElementType(),
ArgIndex.second));
ArgAttrVec.push_back(AttributeSet());
assert(Params.back());
}
if (ArgIndices.size() == 1 && ArgIndices.begin()->second.empty())
++NumArgumentsPromoted;
else
++NumAggregatesPromoted;
}
}
Type *RetTy = FTy->getReturnType();
// 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);
// Patch the pointer to LLVM function in debug info descriptor.
NF->setSubprogram(F->getSubprogram());
F->setSubprogram(nullptr);
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(AttributeList::get(F->getContext(), PAL.getFnAttributes(),
PAL.getRetAttributes(), ArgAttrVec));
ArgAttrVec.clear();
F->getParent()->getFunctionList().insert(F->getIterator(), NF);
NF->takeName(F);
// 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(F->user_back());
assert(CS.getCalledFunction() == F);
Instruction *Call = CS.getInstruction();
const AttributeList &CallPAL = CS.getAttributes();
// Loop over the operands, inserting GEP and loads in the caller as
// appropriate.
CallSite::arg_iterator AI = CS.arg_begin();
ArgNo = 0;
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
++I, ++AI, ++ArgNo)
if (!ArgsToPromote.count(&*I) && !ByValArgsToTransform.count(&*I)) {
Args.push_back(*AI); // Unmodified argument
ArgAttrVec.push_back(CallPAL.getParamAttributes(ArgNo));
} else if (ByValArgsToTransform.count(&*I)) {
// Emit a GEP and load for each element of the struct.
Type *AgTy = cast<PointerType>(I->getType())->getElementType();
StructType *STy = cast<StructType>(AgTy);
Value *Idxs[2] = {
ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr};
for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
Value *Idx = GetElementPtrInst::Create(
STy, *AI, Idxs, (*AI)->getName() + "." + Twine(i), Call);
// TODO: Tell AA about the new values?
Args.push_back(new LoadInst(Idx, Idx->getName() + ".val", Call));
ArgAttrVec.push_back(AttributeSet());
}
} 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 (const auto &ArgIndex : ArgIndices) {
Value *V = *AI;
LoadInst *OrigLoad =
OriginalLoads[std::make_pair(&*I, ArgIndex.second)];
if (!ArgIndex.second.empty()) {
Ops.reserve(ArgIndex.second.size());
Type *ElTy = V->getType();
for (auto II : ArgIndex.second) {
// Use i32 to index structs, and i64 for others (pointers/arrays).
// This satisfies GEP constraints.
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.
if (auto *ElPTy = dyn_cast<PointerType>(ElTy))
ElTy = ElPTy->getElementType();
else
ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(II);
}
// And create a GEP to extract those indices.
V = GetElementPtrInst::Create(ArgIndex.first, V, Ops,
V->getName() + ".idx", Call);
Ops.clear();
}
// 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());
// Transfer the AA info too.
AAMDNodes AAInfo;
OrigLoad->getAAMetadata(AAInfo);
newLoad->setAAMetadata(AAInfo);
Args.push_back(newLoad);
ArgAttrVec.push_back(AttributeSet());
}
}
// Push any varargs arguments on the list.
for (; AI != CS.arg_end(); ++AI, ++ArgNo) {
Args.push_back(*AI);
ArgAttrVec.push_back(CallPAL.getParamAttributes(ArgNo));
}
SmallVector<OperandBundleDef, 1> OpBundles;
CS.getOperandBundlesAsDefs(OpBundles);
CallSite NewCS;
if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
NewCS = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
Args, OpBundles, "", Call);
} else {
auto *NewCall = CallInst::Create(NF, Args, OpBundles, "", Call);
NewCall->setTailCallKind(cast<CallInst>(Call)->getTailCallKind());
NewCS = NewCall;
}
NewCS.setCallingConv(CS.getCallingConv());
NewCS.setAttributes(
AttributeList::get(F->getContext(), CallPAL.getFnAttributes(),
CallPAL.getRetAttributes(), ArgAttrVec));
NewCS->setDebugLoc(Call->getDebugLoc());
uint64_t W;
if (Call->extractProfTotalWeight(W))
NewCS->setProfWeight(W);
Args.clear();
ArgAttrVec.clear();
// Update the callgraph to know that the callsite has been transformed.
if (ReplaceCallSite)
(*ReplaceCallSite)(CS, NewCS);
if (!Call->use_empty()) {
Call->replaceAllUsesWith(NewCS.getInstruction());
NewCS->takeName(Call);
}
// Finally, remove the old call from the program, reducing the use-count of
// F.
Call->eraseFromParent();
}
const DataLayout &DL = F->getParent()->getDataLayout();
// 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, transferring uses of the old arguments over to
// the new arguments, also transferring 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);
++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()->front();
// Just add all the struct element types.
Type *AgTy = cast<PointerType>(I->getType())->getElementType();
Value *TheAlloca = new AllocaInst(AgTy, DL.getAllocaAddrSpace(), nullptr,
I->getParamAlignment(), "", InsertPt);
StructType *STy = cast<StructType>(AgTy);
Value *Idxs[2] = {ConstantInt::get(Type::getInt32Ty(F->getContext()), 0),
nullptr};
for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
Value *Idx = GetElementPtrInst::Create(
AgTy, TheAlloca, Idxs, 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);
// If the alloca is used in a call, we must clear the tail flag since
// the callee now uses an alloca from the caller.
for (User *U : TheAlloca->users()) {
CallInst *Call = dyn_cast<CallInst>(U);
if (!Call)
continue;
Call->setTailCall(false);
}
continue;
}
if (I->use_empty())
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->user_back())) {
assert(ArgIndices.begin()->second.empty() &&
"Load element should sort to front!");
I2->setName(I->getName() + ".val");
LI->replaceAllUsesWith(&*I2);
LI->eraseFromParent();
DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
<< "' in function '" << F->getName() << "'\n");
} else {
GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->user_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->second != 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->user_back());
L->replaceAllUsesWith(&*TheArg);
L->eraseFromParent();
}
GEP->eraseFromParent();
}
}
// Increment I2 past all of the arguments added for this promoted pointer.
std::advance(I2, ArgIndices.size());
}
return NF;
}
Function * futamurize( const Function * orig_func, DenseMap<const Value*, Value*> &argmap, std::set<const unsigned char *> &constant_addresses_set )
{
LLVMContext &context = getGlobalContext();
// Make a copy of the function, removing constant arguments
Function * specialized_func = CloneFunction( orig_func, argmap );
specialized_func->setName( orig_func->getNameStr() + "_1" );
// add it to our module
LLVM_Module->getFunctionList().push_back( specialized_func );
printf("\nspecialized_func = %p <%s>\n", specialized_func, specialized_func->getName().data());
//~ specialized_func->dump();
// Optimize it
FunctionPassManager PM( LLVM_Module );
createStandardFunctionPasses( &PM, 3 );
PM.add(createScalarReplAggregatesPass()); // Break up aggregate allocas
PM.add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
PM.add(createJumpThreadingPass()); // Thread jumps.
PM.add(createCFGSimplificationPass()); // Merge & remove BBs
PM.add(createInstructionCombiningPass()); // Combine silly seq's
PM.add(createTailCallEliminationPass()); // Eliminate tail calls
PM.add(createCFGSimplificationPass()); // Merge & remove BBs
PM.add(createReassociatePass()); // Reassociate expressions
PM.add(createLoopRotatePass()); // Rotate Loop
PM.add(createLICMPass()); // Hoist loop invariants
PM.add(createLoopUnswitchPass( false ));
PM.add(createInstructionCombiningPass());
PM.add(createIndVarSimplifyPass()); // Canonicalize indvars
PM.add(createLoopDeletionPass()); // Delete dead loops
PM.add(createLoopUnroll2Pass()); // Unroll small loops
PM.add(createInstructionCombiningPass()); // Clean up after the unroller
PM.add(createGVNPass()); // Remove redundancies
PM.add(createMemCpyOptPass()); // Remove memcpy / form memset
PM.add(createSCCPPass()); // Constant prop with SCCP
PM.add(createPromoteMemoryToRegisterPass());
PM.add(createConstantPropagationPass());
PM.add(createDeadStoreEliminationPass());
PM.add(createAggressiveDCEPass());
PM.add(new MemoryDependenceAnalysis());
//~ PM.add(createAAEvalPass());
const PassInfo * pinfo = Pass::lookupPassInfo( "print-alias-sets" );
if( !pinfo ) { printf( "print-alias-sets not found\n" ); exit(-1); }
PM.add( pinfo->createPass() );
FunctionPassManager PM_Inline( LLVM_Module );
PM_Inline.add(createSingleFunctionInliningPass());
bool Changed = false;
int iterations = 2;
int inline_iterations = 6;
do
{
Changed = false;
// first do some optimizations
PM.doInitialization();
PM.run( *specialized_func );
PM.doFinalization();
// Load from Constant Memory detection
const TargetData *TD = LLVM_EE->getTargetData();
for (inst_iterator I = inst_begin(specialized_func), E = inst_end(specialized_func); I != E; ++I)
{
Instruction * inst = (Instruction *) &*I;
// get all Load instructions
LoadInst * load = dyn_cast<LoadInst>( inst );
if( !load ) continue;
if( load->isVolatile() ) continue;
if (load->use_empty()) continue; // Don't muck with dead instructions...
// get the address loaded by load instruction
Value *ptr_value = load->getPointerOperand();
// we're only interested in constant addresses
ConstantExpr * ptr_constant_expr = dyn_cast<ConstantExpr>( ptr_value );
if( !ptr_constant_expr ) continue;
ptr_constant_expr->dump();
// compute real address of constant pointer expression
Constant * ptr_constant = ConstantFoldConstantExpression( ptr_constant_expr, TD );
if( !ptr_constant ) continue;
ptr_constant->dump();
// convert to int constant
ConstantInt *int_constant = dyn_cast<ConstantInt>( ConstantExpr::getPtrToInt( ptr_constant, Type::getInt64Ty( context )));
if( !int_constant ) continue;
int_constant->dump();
// get data size
int data_length = TD->getTypeAllocSize( load->getType() );
ptr_value->getType()->dump();
// get real address (at last !)
const unsigned char * c_ptr = (const unsigned char *) int_constant->getLimitedValue();
printf( "%ld %d %d\n", c_ptr, constant_addresses_set.count( c_ptr ), data_length );
// check what's in this address
int isconst = 1;
for( int offset=0; offset<data_length; offset++ )
isconst &= constant_addresses_set.count( c_ptr + offset );
if( !isconst ) continue;
printf( "It is constant.\n" );
// make a LLVM const with the data
Constant *new_constant = NULL;
switch( data_length )
{
case 1: new_constant = ConstantInt::get( Type::getInt8Ty( context ), *(uint8_t*)c_ptr, false /* signed */ ); break;
case 2: new_constant = ConstantInt::get( Type::getInt16Ty( context ), *(uint16_t*)c_ptr, false /* signed */ ); break;
case 4: new_constant = ConstantInt::get( Type::getInt32Ty( context ), *(uint32_t*)c_ptr, false /* signed */ ); break;
case 8: new_constant = ConstantInt::get( Type::getInt64Ty( context ), *(uint64_t*)c_ptr, false /* signed */ ); break;
default:
{
StringRef const_data ( (const char *) c_ptr, data_length );
new_constant = ConstantArray::get( context, const_data, false /* dont add terminating null */ );
}
}
if( !new_constant ) continue;
new_constant->dump();
//~ // get the type that is loaded
const Type *Ty = load->getType();
// do we need a cast ?
if( load->getType() != new_constant->getType() )
{
new_constant = ConstantExpr::getBitCast( new_constant, Ty );
new_constant->dump();
}
// zap the load and replace with constant address
load->replaceAllUsesWith( new_constant );
printf( "\nREPLACED :...\n" );
load->dump();
new_constant->dump();
Changed = true;
}
if( Changed )
continue; // re-optimize and do another pass of constant load elimination
// if we can't do anything else, do an inlining pass
if( inline_iterations > 0 )
{
inline_iterations --;
PM_Inline.doInitialization();
Changed |= PM_Inline.run( *specialized_func );
PM_Inline.doFinalization();
//~ for( int i=0; i<3; i++ )
{
PM.doInitialization();
Changed |= PM.run( *specialized_func );
PM.doFinalization();
}
}
if( iterations>0 && !Changed )
iterations--;
} while( Changed || iterations>0 );
return specialized_func;
}