Value *VectorBlockGenerator::generateStrideOneLoad(
ScopStmt &Stmt, const LoadInst *Load, VectorValueMapT &ScalarMaps,
bool NegativeStride = false) {
unsigned VectorWidth = getVectorWidth();
const Value *Pointer = Load->getPointerOperand();
Type *VectorPtrType = getVectorPtrTy(Pointer, VectorWidth);
unsigned Offset = NegativeStride ? VectorWidth - 1 : 0;
Value *NewPointer = nullptr;
NewPointer = generateLocationAccessed(Stmt, Load, Pointer, ScalarMaps[Offset],
GlobalMaps[Offset], VLTS[Offset]);
Value *VectorPtr =
Builder.CreateBitCast(NewPointer, VectorPtrType, "vector_ptr");
LoadInst *VecLoad =
Builder.CreateLoad(VectorPtr, Load->getName() + "_p_vec_full");
if (!Aligned)
VecLoad->setAlignment(8);
if (NegativeStride) {
SmallVector<Constant *, 16> Indices;
for (int i = VectorWidth - 1; i >= 0; i--)
Indices.push_back(ConstantInt::get(Builder.getInt32Ty(), i));
Constant *SV = llvm::ConstantVector::get(Indices);
Value *RevVecLoad = Builder.CreateShuffleVector(
VecLoad, VecLoad, SV, Load->getName() + "_reverse");
return RevVecLoad;
}
return VecLoad;
}
void WorklessInstrument::CreateIfElseBlock(Loop * pLoop, vector<BasicBlock *> & vecAdded)
{
BasicBlock * pPreHeader = pLoop->getLoopPreheader();
BasicBlock * pHeader = pLoop->getHeader();
Function * pInnerFunction = pPreHeader->getParent();
Module * pModule = pPreHeader->getParent()->getParent();
BasicBlock * pElseBody = NULL;
TerminatorInst * pTerminator = NULL;
BranchInst * pBranch = NULL;
LoadInst * pLoad1 = NULL;
LoadInst * pLoad2 = NULL;
LoadInst * pLoadnumGlobalCounter = NULL;
BinaryOperator * pAddOne = NULL;
StoreInst * pStoreNew = NULL;
CmpInst * pCmp = NULL;
CallInst * pCall = NULL;
StoreInst * pStore = NULL;
AttributeSet emptySet;
pTerminator = pPreHeader->getTerminator();
pLoadnumGlobalCounter = new LoadInst(this->numGlobalCounter, "", false, pTerminator);
pLoadnumGlobalCounter->setAlignment(8);
pAddOne = BinaryOperator::Create(Instruction::Add, pLoadnumGlobalCounter, this->ConstantLong1, "add", pTerminator);
pStoreNew = new StoreInst(pAddOne, this->numGlobalCounter, false, pTerminator);
pStoreNew->setAlignment(8);
pElseBody = BasicBlock::Create(pModule->getContext(), ".else.body.CPI", pInnerFunction, 0);
pLoad2 = new LoadInst(this->CURRENT_SAMPLE, "", false, pTerminator);
pLoad2->setAlignment(8);
pCmp = new ICmpInst(pTerminator, ICmpInst::ICMP_SLT, pAddOne, pLoad2, "");
pBranch = BranchInst::Create(pHeader, pElseBody, pCmp );
ReplaceInstWithInst(pTerminator, pBranch);
pLoad1 = new LoadInst(this->SAMPLE_RATE, "", false, pElseBody);
pCall = CallInst::Create(this->geo, pLoad1, "", pElseBody);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
pCall->setAttributes(emptySet);
CastInst * pCast = CastInst::CreateIntegerCast(pCall, this->LongType, true, "", pElseBody);
//pBinary = BinaryOperator::Create(Instruction::Add, pLoad2, pCast, "add", pIfBody);
pStore = new StoreInst(pCast, this->CURRENT_SAMPLE, false, pElseBody);
pStore->setAlignment(8);
pStore = new StoreInst(this->ConstantLong0, this->numGlobalCounter, false, pElseBody);
pStore->setAlignment(8);
pLoad1 = new LoadInst(this->numInstances, "", false, pElseBody);
pLoad1->setAlignment(8);
pAddOne = BinaryOperator::Create(Instruction::Add, pLoad1, this->ConstantLong1, "add", pElseBody);
pStore = new StoreInst(pAddOne, this->numInstances, false, pElseBody);
pStore->setAlignment(8);
vecAdded.push_back(pPreHeader);
vecAdded.push_back(pElseBody);
}
extern "C" LLVMValueRef
LLVMRustBuildAtomicLoad(LLVMBuilderRef B, LLVMValueRef Source, const char *Name,
LLVMAtomicOrdering Order, unsigned Alignment) {
LoadInst *LI = new LoadInst(unwrap(Source), 0);
LI->setAtomic(fromRust(Order));
LI->setAlignment(Alignment);
return wrap(unwrap(B)->Insert(LI, Name));
}
/// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
const DataLayout *DL) {
User *CI = cast<User>(LI.getOperand(0));
Value *CastOp = CI->getOperand(0);
PointerType *DestTy = cast<PointerType>(CI->getType());
Type *DestPTy = DestTy->getElementType();
if (PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
// If the address spaces don't match, don't eliminate the cast.
if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
return 0;
Type *SrcPTy = SrcTy->getElementType();
if (DestPTy->isIntegerTy() || DestPTy->isPointerTy() ||
DestPTy->isVectorTy()) {
// If the source is an array, the code below will not succeed. Check to
// see if a trivial 'gep P, 0, 0' will help matters. Only do this for
// constants.
if (ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
if (Constant *CSrc = dyn_cast<Constant>(CastOp))
if (ASrcTy->getNumElements() != 0) {
Type *IdxTy = DL
? DL->getIntPtrType(SrcTy)
: Type::getInt64Ty(SrcTy->getContext());
Value *Idx = Constant::getNullValue(IdxTy);
Value *Idxs[2] = { Idx, Idx };
CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
SrcTy = cast<PointerType>(CastOp->getType());
SrcPTy = SrcTy->getElementType();
}
if (IC.getDataLayout() &&
(SrcPTy->isIntegerTy() || SrcPTy->isPointerTy() ||
SrcPTy->isVectorTy()) &&
// Do not allow turning this into a load of an integer, which is then
// casted to a pointer, this pessimizes pointer analysis a lot.
(SrcPTy->isPtrOrPtrVectorTy() ==
LI.getType()->isPtrOrPtrVectorTy()) &&
IC.getDataLayout()->getTypeSizeInBits(SrcPTy) ==
IC.getDataLayout()->getTypeSizeInBits(DestPTy)) {
// Okay, we are casting from one integer or pointer type to another of
// the same size. Instead of casting the pointer before the load, cast
// the result of the loaded value.
LoadInst *NewLoad =
IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName());
NewLoad->setAlignment(LI.getAlignment());
NewLoad->setAtomic(LI.getOrdering(), LI.getSynchScope());
// Now cast the result of the load.
return new BitCastInst(NewLoad, LI.getType());
}
}
}
return 0;
}
llvm::Value *StorageSoa::alignedArrayLoad(llvm::Value *val)
{
VectorType *vectorType = VectorType::get(Type::FloatTy, 4);
PointerType *vectorPtr = PointerType::get(vectorType, 0);
CastInst *cast = new BitCastInst(val, vectorPtr, name("toVector"), m_block);
LoadInst *load = new LoadInst(cast, name("alignLoad"), false, m_block);
load->setAlignment(8);
return load;
}
extern "C" LLVMValueRef LLVMBuildAtomicLoad(LLVMBuilderRef B,
LLVMValueRef source,
const char* Name,
AtomicOrdering order) {
LoadInst* li = new LoadInst(unwrap(source),0);
li->setVolatile(true);
li->setAtomic(order);
li->setAlignment(sizeof(intptr_t));
return wrap(unwrap(B)->Insert(li, Name));
}
extern "C" LLVMValueRef LLVMBuildAtomicLoad(LLVMBuilderRef B,
LLVMValueRef source,
const char* Name,
AtomicOrdering order,
unsigned alignment) {
LoadInst* li = new LoadInst(unwrap(source),0);
li->setAtomic(order);
li->setAlignment(alignment);
return wrap(unwrap(B)->Insert(li, Name));
}
LLVMValueRef
mono_llvm_build_aligned_load (LLVMBuilderRef builder, LLVMValueRef PointerVal,
const char *Name, gboolean is_volatile, int alignment)
{
LoadInst *ins;
ins = unwrap(builder)->CreateLoad(unwrap(PointerVal), is_volatile, Name);
ins->setAlignment (alignment);
return wrap(ins);
}
void SanitizerCoverageModule::InjectCoverageAtBlock(Function &F,
BasicBlock &BB) {
BasicBlock::iterator IP = BB.getFirstInsertionPt(), BE = BB.end();
// Skip static allocas at the top of the entry block so they don't become
// dynamic when we split the block. If we used our optimized stack layout,
// then there will only be one alloca and it will come first.
for (; IP != BE; ++IP) {
AllocaInst *AI = dyn_cast<AllocaInst>(IP);
if (!AI || !AI->isStaticAlloca())
break;
}
bool IsEntryBB = &BB == &F.getEntryBlock();
DebugLoc EntryLoc =
IsEntryBB ? IP->getDebugLoc().getFnDebugLoc(*C) : IP->getDebugLoc();
IRBuilder<> IRB(IP);
IRB.SetCurrentDebugLocation(EntryLoc);
SmallVector<Value *, 1> Indices;
Value *GuardP = IRB.CreateAdd(
IRB.CreatePointerCast(GuardArray, IntptrTy),
ConstantInt::get(IntptrTy, (1 + SanCovFunction->getNumUses()) * 4));
Type *Int32PtrTy = PointerType::getUnqual(IRB.getInt32Ty());
GuardP = IRB.CreateIntToPtr(GuardP, Int32PtrTy);
LoadInst *Load = IRB.CreateLoad(GuardP);
Load->setAtomic(Monotonic);
Load->setAlignment(4);
Load->setMetadata(F.getParent()->getMDKindID("nosanitize"),
MDNode::get(*C, None));
Value *Cmp = IRB.CreateICmpSGE(Constant::getNullValue(Load->getType()), Load);
Instruction *Ins = SplitBlockAndInsertIfThen(
Cmp, IP, false, MDBuilder(*C).createBranchWeights(1, 100000));
IRB.SetInsertPoint(Ins);
IRB.SetCurrentDebugLocation(EntryLoc);
// __sanitizer_cov gets the PC of the instruction using GET_CALLER_PC.
IRB.CreateCall(SanCovFunction, GuardP);
IRB.CreateCall(EmptyAsm); // Avoids callback merge.
if (ClExperimentalTracing) {
// Experimental support for tracing.
// Insert a callback with the same guard variable as used for coverage.
IRB.SetInsertPoint(IP);
IRB.CreateCall(IsEntryBB ? SanCovTraceEnter : SanCovTraceBB, GuardP);
}
}
Value *VectorBlockGenerator::generateStrideZeroLoad(const LoadInst *Load,
ValueMapT &BBMap) {
const Value *Pointer = Load->getPointerOperand();
Type *VectorPtrType = getVectorPtrTy(Pointer, 1);
Value *NewPointer =
getNewValue(Pointer, BBMap, GlobalMaps[0], VLTS[0], getLoopForInst(Load));
Value *VectorPtr = Builder.CreateBitCast(NewPointer, VectorPtrType,
Load->getName() + "_p_vec_p");
LoadInst *ScalarLoad =
Builder.CreateLoad(VectorPtr, Load->getName() + "_p_splat_one");
if (!Aligned)
ScalarLoad->setAlignment(8);
Constant *SplatVector = Constant::getNullValue(
VectorType::get(Builder.getInt32Ty(), getVectorWidth()));
Value *VectorLoad = Builder.CreateShuffleVector(
ScalarLoad, ScalarLoad, SplatVector, Load->getName() + "_p_splat");
return VectorLoad;
}
LLVMValueRef
mono_llvm_build_atomic_load (LLVMBuilderRef builder, LLVMValueRef PointerVal,
const char *Name, gboolean is_volatile, int alignment, BarrierKind barrier)
{
LoadInst *ins = unwrap(builder)->CreateLoad(unwrap(PointerVal), is_volatile, Name);
ins->setAlignment (alignment);
switch (barrier) {
case LLVM_BARRIER_NONE:
break;
case LLVM_BARRIER_ACQ:
ins->setOrdering(Acquire);
break;
case LLVM_BARRIER_SEQ:
ins->setOrdering(SequentiallyConsistent);
break;
default:
g_assert_not_reached ();
break;
}
return wrap(ins);
}
void SanitizerCoverageModule::InjectCoverageAtBlock(Function &F, BasicBlock &BB,
size_t Idx, bool UseCalls) {
BasicBlock::iterator IP = BB.getFirstInsertionPt();
bool IsEntryBB = &BB == &F.getEntryBlock();
DebugLoc EntryLoc;
if (IsEntryBB) {
if (auto SP = F.getSubprogram())
EntryLoc = DebugLoc::get(SP->getScopeLine(), 0, SP);
// Keep static allocas and llvm.localescape calls in the entry block. Even
// if we aren't splitting the block, it's nice for allocas to be before
// calls.
IP = PrepareToSplitEntryBlock(BB, IP);
} else {
EntryLoc = IP->getDebugLoc();
}
IRBuilder<> IRB(&*IP);
IRB.SetCurrentDebugLocation(EntryLoc);
if (Options.TracePC) {
IRB.CreateCall(SanCovTracePC); // gets the PC using GET_CALLER_PC.
IRB.CreateCall(EmptyAsm, {}); // Avoids callback merge.
} else if (Options.TracePCGuard) {
//auto GuardVar = new GlobalVariable(
// *F.getParent(), Int64Ty, false, GlobalVariable::LinkOnceODRLinkage,
// Constant::getNullValue(Int64Ty), "__sancov_guard." + F.getName());
// if (auto Comdat = F.getComdat())
// GuardVar->setComdat(Comdat);
// TODO: add debug into to GuardVar.
// GuardVar->setSection(SanCovTracePCGuardSection);
// auto GuardPtr = IRB.CreatePointerCast(GuardVar, IntptrPtrTy);
auto GuardPtr = IRB.CreateIntToPtr(
IRB.CreateAdd(IRB.CreatePointerCast(FunctionGuardArray, IntptrTy),
ConstantInt::get(IntptrTy, Idx * 4)),
Int32PtrTy);
if (!UseCalls) {
auto GuardLoad = IRB.CreateLoad(GuardPtr);
GuardLoad->setAtomic(AtomicOrdering::Monotonic);
GuardLoad->setAlignment(8);
SetNoSanitizeMetadata(GuardLoad); // Don't instrument with e.g. asan.
auto Cmp = IRB.CreateICmpNE(
GuardLoad, Constant::getNullValue(GuardLoad->getType()));
auto Ins = SplitBlockAndInsertIfThen(
Cmp, &*IP, false, MDBuilder(*C).createBranchWeights(1, 100000));
IRB.SetCurrentDebugLocation(EntryLoc);
IRB.SetInsertPoint(Ins);
}
IRB.CreateCall(SanCovTracePCGuard, GuardPtr);
IRB.CreateCall(EmptyAsm, {}); // Avoids callback merge.
} else {
Value *GuardP = IRB.CreateAdd(
IRB.CreatePointerCast(GuardArray, IntptrTy),
ConstantInt::get(IntptrTy, (1 + NumberOfInstrumentedBlocks()) * 4));
GuardP = IRB.CreateIntToPtr(GuardP, Int32PtrTy);
if (Options.TraceBB) {
IRB.CreateCall(IsEntryBB ? SanCovTraceEnter : SanCovTraceBB, GuardP);
} else if (UseCalls) {
IRB.CreateCall(SanCovWithCheckFunction, GuardP);
} else {
LoadInst *Load = IRB.CreateLoad(GuardP);
Load->setAtomic(AtomicOrdering::Monotonic);
Load->setAlignment(4);
SetNoSanitizeMetadata(Load);
Value *Cmp =
IRB.CreateICmpSGE(Constant::getNullValue(Load->getType()), Load);
Instruction *Ins = SplitBlockAndInsertIfThen(
Cmp, &*IP, false, MDBuilder(*C).createBranchWeights(1, 100000));
IRB.SetInsertPoint(Ins);
IRB.SetCurrentDebugLocation(EntryLoc);
// __sanitizer_cov gets the PC of the instruction using GET_CALLER_PC.
IRB.CreateCall(SanCovFunction, GuardP);
IRB.CreateCall(EmptyAsm, {}); // Avoids callback merge.
}
}
if (Options.Use8bitCounters) {
IRB.SetInsertPoint(&*IP);
Value *P = IRB.CreateAdd(
IRB.CreatePointerCast(EightBitCounterArray, IntptrTy),
ConstantInt::get(IntptrTy, NumberOfInstrumentedBlocks() - 1));
P = IRB.CreateIntToPtr(P, IRB.getInt8PtrTy());
LoadInst *LI = IRB.CreateLoad(P);
Value *Inc = IRB.CreateAdd(LI, ConstantInt::get(IRB.getInt8Ty(), 1));
StoreInst *SI = IRB.CreateStore(Inc, P);
SetNoSanitizeMetadata(LI);
SetNoSanitizeMetadata(SI);
}
}
/// 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;
}
/// 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;
}
// UpgradeIntrinsicCall - Upgrade a call to an old intrinsic to be a call the
// upgraded intrinsic. All argument and return casting must be provided in
// order to seamlessly integrate with existing context.
void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
Function *F = CI->getCalledFunction();
LLVMContext &C = CI->getContext();
ImmutableCallSite CS(CI);
assert(F && "CallInst has no function associated with it.");
if (!NewFn) {
if (F->getName() == "llvm.x86.sse.loadu.ps" ||
F->getName() == "llvm.x86.sse2.loadu.dq" ||
F->getName() == "llvm.x86.sse2.loadu.pd") {
// Convert to a native, unaligned load.
const Type *VecTy = CI->getType();
const Type *IntTy = IntegerType::get(C, 128);
IRBuilder<> Builder(C);
Builder.SetInsertPoint(CI->getParent(), CI);
Value *BC = Builder.CreateBitCast(CI->getArgOperand(0),
PointerType::getUnqual(IntTy),
"cast");
LoadInst *LI = Builder.CreateLoad(BC, CI->getName());
LI->setAlignment(1); // Unaligned load.
BC = Builder.CreateBitCast(LI, VecTy, "new.cast");
// Fix up all the uses with our new load.
if (!CI->use_empty())
CI->replaceAllUsesWith(BC);
// Remove intrinsic.
CI->eraseFromParent();
} else if (F->getName() == "llvm.x86.sse.movnt.ps" ||
F->getName() == "llvm.x86.sse2.movnt.dq" ||
F->getName() == "llvm.x86.sse2.movnt.pd" ||
F->getName() == "llvm.x86.sse2.movnt.i") {
IRBuilder<> Builder(C);
Builder.SetInsertPoint(CI->getParent(), CI);
Module *M = F->getParent();
SmallVector<Value *, 1> Elts;
Elts.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
MDNode *Node = MDNode::get(C, Elts);
Value *Arg0 = CI->getArgOperand(0);
Value *Arg1 = CI->getArgOperand(1);
// Convert the type of the pointer to a pointer to the stored type.
Value *BC = Builder.CreateBitCast(Arg0,
PointerType::getUnqual(Arg1->getType()),
"cast");
StoreInst *SI = Builder.CreateStore(Arg1, BC);
SI->setMetadata(M->getMDKindID("nontemporal"), Node);
SI->setAlignment(16);
// Remove intrinsic.
CI->eraseFromParent();
} else {
llvm_unreachable("Unknown function for CallInst upgrade.");
}
return;
}
switch (NewFn->getIntrinsicID()) {
case Intrinsic::prefetch: {
IRBuilder<> Builder(C);
Builder.SetInsertPoint(CI->getParent(), CI);
const llvm::Type *I32Ty = llvm::Type::getInt32Ty(CI->getContext());
// Add the extra "data cache" argument
Value *Operands[4] = { CI->getArgOperand(0), CI->getArgOperand(1),
CI->getArgOperand(2),
llvm::ConstantInt::get(I32Ty, 1) };
CallInst *NewCI = CallInst::Create(NewFn, Operands,
CI->getName(), CI);
NewCI->setTailCall(CI->isTailCall());
NewCI->setCallingConv(CI->getCallingConv());
// Handle any uses of the old CallInst.
if (!CI->use_empty())
// Replace all uses of the old call with the new cast which has the
// correct type.
CI->replaceAllUsesWith(NewCI);
// Clean up the old call now that it has been completely upgraded.
CI->eraseFromParent();
break;
}
}
}
/// 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;
}
void WorklessInstrument::InstrumentWorkless0Star1(Module * pModule, Loop * pLoop)
{
Function * pMain = NULL;
if(strMainName != "" )
{
pMain = pModule->getFunction(strMainName.c_str());
}
else
{
pMain = pModule->getFunction("main");
}
LoadInst * pLoad;
BinaryOperator* pAdd = NULL;
StoreInst * pStore = NULL;
for (Function::iterator BB = pMain->begin(); BB != pMain->end(); ++BB)
{
if(BB->getName().equals("entry"))
{
CallInst * pCall;
StoreInst * pStore;
Instruction * II = BB->begin();
pCall = CallInst::Create(this->InitHooks, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
AttributeSet emptySet;
pCall->setAttributes(emptySet);
pCall = CallInst::Create(this->getenv, this->SAMPLE_RATE_ptr, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
AttributeSet AS;
{
SmallVector<AttributeSet, 4> Attrs;
AttributeSet PAS;
{
AttrBuilder B;
B.addAttribute(Attribute::NoUnwind);
PAS = AttributeSet::get(pModule->getContext(), ~0U, B);
}
Attrs.push_back(PAS);
AS = AttributeSet::get(pModule->getContext(), Attrs);
}
pCall->setAttributes(AS);
pCall = CallInst::Create(this->function_atoi, pCall, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
{
SmallVector<AttributeSet, 4> Attrs;
AttributeSet PAS;
{
AttrBuilder B;
B.addAttribute(Attribute::NoUnwind);
B.addAttribute(Attribute::ReadOnly);
PAS = AttributeSet::get(pModule->getContext(), ~0U, B);
}
Attrs.push_back(PAS);
AS = AttributeSet::get(pModule->getContext(), Attrs);
}
pCall->setAttributes(AS);
pStore = new StoreInst(pCall, this->SAMPLE_RATE, false, II);
pStore->setAlignment(4);
pCall = CallInst::Create(this->geo, pCall, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
pCall->setAttributes(emptySet);
CastInst * pCast = CastInst::CreateIntegerCast(pCall, this->LongType, true, "", II);
pStore = new StoreInst(pCast, this->CURRENT_SAMPLE, false, II);
pStore->setAlignment(8);
vector<Value *> vecParam;
vecParam.push_back(this->Output_Format_String);
vecParam.push_back(pCall);
pCall = CallInst::Create(this->printf, vecParam, "", II);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
pCall->setAttributes(emptySet);
break;
}
}
for (Function::iterator BB = pMain->begin(); BB != pMain->end(); ++BB)
{
for (BasicBlock::iterator Ins = BB->begin(); Ins != BB->end(); ++Ins)
{
if (isa<ReturnInst>(Ins) || isa<ResumeInst>(Ins))
{
vector<Value*> vecParams;
pLoad = new LoadInst(numIterations, "", false, Ins);
pLoad->setAlignment(8);
vecParams.push_back(pLoad);
pLoad = new LoadInst(numInstances, "", false, Ins);
pLoad->setAlignment(8);
vecParams.push_back(pLoad);
CallInst* pCall = CallInst::Create(this->PrintLoopInfo, vecParams, "", Ins);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
AttributeSet aSet;
pCall->setAttributes(aSet);
}
else if(isa<CallInst>(Ins) || isa<InvokeInst>(Ins))
{
CallSite cs(Ins);
Function * pCalled = cs.getCalledFunction();
if(pCalled == NULL)
{
continue;
}
if(pCalled->getName() == "exit" || pCalled->getName() == "_ZL9mysql_endi")
{
vector<Value*> vecParams;
pLoad = new LoadInst(numIterations, "", false, Ins);
pLoad->setAlignment(8);
vecParams.push_back(pLoad);
pLoad = new LoadInst(numInstances, "", false, Ins);
pLoad->setAlignment(8);
vecParams.push_back(pLoad);
CallInst* pCall = CallInst::Create(this->PrintLoopInfo, vecParams, "", Ins);
pCall->setCallingConv(CallingConv::C);
pCall->setTailCall(false);
AttributeSet aSet;
pCall->setAttributes(aSet);
}
}
}
}
BasicBlock * pHeader = pLoop->getHeader();
set<BasicBlock *> setExitBlock;
CollectExitBlock(pLoop, setExitBlock);
vector<BasicBlock *> vecAdded;
CreateIfElseBlock(pLoop, vecAdded);
ValueToValueMapTy VMap;
set<BasicBlock *> setCloned;
CloneInnerLoop(pLoop, vecAdded, VMap, setCloned);
BasicBlock * pPreHeader = vecAdded[1];
pLoad = new LoadInst(this->numIterations, "", false, pPreHeader->getTerminator());
pLoad->setAlignment(8);
BasicBlock * pClonedHeader = cast<BasicBlock>(VMap[pHeader]);
set<BasicBlock *> setPredBlocks;
for(pred_iterator PI = pred_begin(pClonedHeader), E = pred_end(pClonedHeader); PI != E; ++PI)
{
setPredBlocks.insert(*PI);
}
PHINode * pNew = PHINode::Create(pLoad->getType(), setPredBlocks.size(), "numIterations", pClonedHeader->getFirstInsertionPt());
pAdd = BinaryOperator::Create(Instruction::Add, pNew, this->ConstantLong1, "add", pClonedHeader->getFirstInsertionPt());
set<BasicBlock *>::iterator itSetBegin = setPredBlocks.begin();
set<BasicBlock *>::iterator itSetEnd = setPredBlocks.end();
for(; itSetBegin != itSetEnd; itSetBegin ++ )
{
if((*itSetBegin) == pPreHeader)
{
pNew->addIncoming(pLoad, pPreHeader);
}
else
{
pNew->addIncoming(pAdd, *itSetBegin);
}
}
itSetBegin = setExitBlock.begin();
itSetEnd = setExitBlock.end();
for(; itSetBegin != itSetEnd; itSetBegin ++ )
{
SmallVector<BasicBlock*, 8> LoopBlocks;
for(pred_iterator PI = pred_begin(*itSetBegin), E = pred_end(*itSetBegin); PI != E; ++PI)
{
if(setCloned.find(*PI) != setCloned.end())
{
LoopBlocks.push_back(*PI);
}
}
BasicBlock * NewExitBB = SplitBlockPredecessors(*itSetBegin, LoopBlocks, ".WL.loopexit", this);
pStore = new StoreInst(pAdd, this->numIterations, false, NewExitBB->getFirstInsertionPt());
pStore->setAlignment(8);
}
pPreHeader->getParent()->dump();
}
/// 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;
// We start with an alignment of one and try to find instructions that allow
// us to prove better alignment.
unsigned Alignment = 1;
// 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 (LoadInst *load = dyn_cast<LoadInst>(Use)) {
assert(!load->isVolatile() && "AST broken");
if (!load->isSimple())
return;
} else if (StoreInst *store = dyn_cast<StoreInst>(Use)) {
// Stores *of* the pointer are not interesting, only stores *to* the
// pointer.
if (Use->getOperand(1) != ASIV)
continue;
assert(!store->isVolatile() && "AST broken");
if (!store->isSimple())
return;
// Note that we only check GuaranteedToExecute inside the store case
// so that we do not introduce stores where they did not exist before
// (which would break the LLVM concurrency model).
// If the alignment of this instruction allows us to specify a more
// restrictive (and performant) alignment and if we are sure this
// instruction will be executed, update the alignment.
// Larger is better, with the exception of 0 being the best alignment.
unsigned InstAlignment = store->getAlignment();
if ((InstAlignment > Alignment || InstAlignment == 0)
&& (Alignment != 0))
if (isGuaranteedToExecute(*Use)) {
GuaranteedToExecute = true;
Alignment = InstAlignment;
}
if (!GuaranteedToExecute)
GuaranteedToExecute = isGuaranteedToExecute(*Use);
} else
return; // Not a load or store.
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;
// Grab a debug location for the inserted loads/stores; given that the
// inserted loads/stores have little relation to the original loads/stores,
// this code just arbitrarily picks a location from one, since any debug
// location is better than none.
DebugLoc DL = LoopUses[0]->getDebugLoc();
SmallVector<BasicBlock*, 8> ExitBlocks;
CurLoop->getUniqueExitBlocks(ExitBlocks);
// We use the SSAUpdater interface to insert phi nodes as required.
SmallVector<PHINode*, 16> NewPHIs;
SSAUpdater SSA(&NewPHIs);
LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks,
*CurAST, DL, Alignment);
// Set up the preheader to have a definition of the value. It is the live-out
// value from the preheader that uses in the loop will use.
LoadInst *PreheaderLoad =
new LoadInst(SomePtr, SomePtr->getName()+".promoted",
Preheader->getTerminator());
PreheaderLoad->setAlignment(Alignment);
PreheaderLoad->setDebugLoc(DL);
SSA.AddAvailableValue(Preheader, PreheaderLoad);
// Rewrite all the loads in the loop and remember all the definitions from
// stores in the loop.
Promoter.run(LoopUses);
// If the SSAUpdater didn't use the load in the preheader, just zap it now.
if (PreheaderLoad->use_empty())
PreheaderLoad->eraseFromParent();
}
bool llvm::expandAtomicRMWToCmpXchg(AtomicRMWInst *AI,
CreateCmpXchgInstFun CreateCmpXchg) {
assert(AI);
AtomicOrdering MemOpOrder =
AI->getOrdering() == Unordered ? Monotonic : AI->getOrdering();
Value *Addr = AI->getPointerOperand();
BasicBlock *BB = AI->getParent();
Function *F = BB->getParent();
LLVMContext &Ctx = F->getContext();
// Given: atomicrmw some_op iN* %addr, iN %incr ordering
//
// The standard expansion we produce is:
// [...]
// %init_loaded = load atomic iN* %addr
// br label %loop
// loop:
// %loaded = phi iN [ %init_loaded, %entry ], [ %new_loaded, %loop ]
// %new = some_op iN %loaded, %incr
// %pair = cmpxchg iN* %addr, iN %loaded, iN %new
// %new_loaded = extractvalue { iN, i1 } %pair, 0
// %success = extractvalue { iN, i1 } %pair, 1
// br i1 %success, label %atomicrmw.end, label %loop
// atomicrmw.end:
// [...]
BasicBlock *ExitBB = BB->splitBasicBlock(AI->getIterator(), "atomicrmw.end");
BasicBlock *LoopBB = BasicBlock::Create(Ctx, "atomicrmw.start", F, ExitBB);
// This grabs the DebugLoc from AI.
IRBuilder<> Builder(AI);
// The split call above "helpfully" added a branch at the end of BB (to the
// wrong place), but we want a load. It's easiest to just remove
// the branch entirely.
std::prev(BB->end())->eraseFromParent();
Builder.SetInsertPoint(BB);
LoadInst *InitLoaded = Builder.CreateLoad(Addr);
// Atomics require at least natural alignment.
InitLoaded->setAlignment(AI->getType()->getPrimitiveSizeInBits() / 8);
Builder.CreateBr(LoopBB);
// Start the main loop block now that we've taken care of the preliminaries.
Builder.SetInsertPoint(LoopBB);
PHINode *Loaded = Builder.CreatePHI(AI->getType(), 2, "loaded");
Loaded->addIncoming(InitLoaded, BB);
Value *NewVal =
performAtomicOp(AI->getOperation(), Builder, Loaded, AI->getValOperand());
Value *NewLoaded = nullptr;
Value *Success = nullptr;
CreateCmpXchg(Builder, Addr, Loaded, NewVal, MemOpOrder,
Success, NewLoaded);
assert(Success && NewLoaded);
Loaded->addIncoming(NewLoaded, LoopBB);
Builder.CreateCondBr(Success, ExitBB, LoopBB);
Builder.SetInsertPoint(ExitBB, ExitBB->begin());
AI->replaceAllUsesWith(NewLoaded);
AI->eraseFromParent();
return true;
}
Value *BoUpSLP::vectorizeTree(ValueList &VL, int VF) {
Type *ScalarTy = VL[0]->getType();
if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
ScalarTy = SI->getValueOperand()->getType();
VectorType *VecTy = VectorType::get(ScalarTy, VF);
// Check if all of the operands are constants or identical.
bool AllConst = true;
bool AllSameScalar = true;
for (unsigned i = 0, e = VF; i < e; ++i) {
AllConst &= !!dyn_cast<Constant>(VL[i]);
AllSameScalar &= (VL[0] == VL[i]);
// Must have a single use.
Instruction *I = dyn_cast<Instruction>(VL[i]);
if (I && (I->getNumUses() > 1 || I->getParent() != BB))
return Scalarize(VL, VecTy);
}
// Is this a simple vector constant.
if (AllConst || AllSameScalar) return Scalarize(VL, VecTy);
// Scalarize unknown structures.
Instruction *VL0 = dyn_cast<Instruction>(VL[0]);
if (!VL0) return Scalarize(VL, VecTy);
unsigned Opcode = VL0->getOpcode();
for (unsigned i = 0, e = VF; i < e; ++i) {
Instruction *I = dyn_cast<Instruction>(VL[i]);
// If not all of the instructions are identical then we have to scalarize.
if (!I || Opcode != I->getOpcode()) return Scalarize(VL, VecTy);
}
switch (Opcode) {
case Instruction::Add:
case Instruction::FAdd:
case Instruction::Sub:
case Instruction::FSub:
case Instruction::Mul:
case Instruction::FMul:
case Instruction::UDiv:
case Instruction::SDiv:
case Instruction::FDiv:
case Instruction::URem:
case Instruction::SRem:
case Instruction::FRem:
case Instruction::Shl:
case Instruction::LShr:
case Instruction::AShr:
case Instruction::And:
case Instruction::Or:
case Instruction::Xor: {
ValueList LHSVL, RHSVL;
for (int i = 0; i < VF; ++i) {
RHSVL.push_back(cast<Instruction>(VL[i])->getOperand(0));
LHSVL.push_back(cast<Instruction>(VL[i])->getOperand(1));
}
Value *RHS = vectorizeTree(RHSVL, VF);
Value *LHS = vectorizeTree(LHSVL, VF);
IRBuilder<> Builder(GetLastInstr(VL, VF));
BinaryOperator *BinOp = dyn_cast<BinaryOperator>(VL0);
return Builder.CreateBinOp(BinOp->getOpcode(), RHS,LHS);
}
case Instruction::Load: {
LoadInst *LI = dyn_cast<LoadInst>(VL0);
unsigned Alignment = LI->getAlignment();
// Check if all of the loads are consecutive.
for (unsigned i = 1, e = VF; i < e; ++i)
if (!isConsecutiveAccess(VL[i-1], VL[i]))
return Scalarize(VL, VecTy);
IRBuilder<> Builder(GetLastInstr(VL, VF));
Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
VecTy->getPointerTo());
LI = Builder.CreateLoad(VecPtr);
LI->setAlignment(Alignment);
return LI;
}
case Instruction::Store: {
StoreInst *SI = dyn_cast<StoreInst>(VL0);
unsigned Alignment = SI->getAlignment();
ValueList ValueOp;
for (int i = 0; i < VF; ++i)
ValueOp.push_back(cast<StoreInst>(VL[i])->getValueOperand());
Value *VecValue = vectorizeTree(ValueOp, VF);
IRBuilder<> Builder(GetLastInstr(VL, VF));
Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
VecTy->getPointerTo());
Builder.CreateStore(VecValue, VecPtr)->setAlignment(Alignment);
for (int i = 0; i < VF; ++i)
cast<Instruction>(VL[i])->eraseFromParent();
return 0;
}
default:
return Scalarize(VL, VecTy);
}
}
void SanitizerCoverageModule::InjectCoverageAtBlock(Function &F, BasicBlock &BB,
bool UseCalls) {
// Don't insert coverage for unreachable blocks: we will never call
// __sanitizer_cov() for them, so counting them in
// NumberOfInstrumentedBlocks() might complicate calculation of code coverage
// percentage. Also, unreachable instructions frequently have no debug
// locations.
if (isa<UnreachableInst>(BB.getTerminator()))
return;
BasicBlock::iterator IP = BB.getFirstInsertionPt(), BE = BB.end();
// Skip static allocas at the top of the entry block so they don't become
// dynamic when we split the block. If we used our optimized stack layout,
// then there will only be one alloca and it will come first.
for (; IP != BE; ++IP) {
AllocaInst *AI = dyn_cast<AllocaInst>(IP);
if (!AI || !AI->isStaticAlloca())
break;
}
bool IsEntryBB = &BB == &F.getEntryBlock();
DebugLoc EntryLoc;
if (IsEntryBB) {
if (auto SP = getDISubprogram(&F))
EntryLoc = DebugLoc::get(SP->getScopeLine(), 0, SP);
} else {
EntryLoc = IP->getDebugLoc();
}
IRBuilder<> IRB(IP);
IRB.SetCurrentDebugLocation(EntryLoc);
SmallVector<Value *, 1> Indices;
Value *GuardP = IRB.CreateAdd(
IRB.CreatePointerCast(GuardArray, IntptrTy),
ConstantInt::get(IntptrTy, (1 + NumberOfInstrumentedBlocks()) * 4));
Type *Int32PtrTy = PointerType::getUnqual(IRB.getInt32Ty());
GuardP = IRB.CreateIntToPtr(GuardP, Int32PtrTy);
if (UseCalls) {
IRB.CreateCall(SanCovWithCheckFunction, GuardP);
} else {
LoadInst *Load = IRB.CreateLoad(GuardP);
Load->setAtomic(Monotonic);
Load->setAlignment(4);
SetNoSanitizeMetadata(Load);
Value *Cmp = IRB.CreateICmpSGE(Constant::getNullValue(Load->getType()), Load);
Instruction *Ins = SplitBlockAndInsertIfThen(
Cmp, IP, false, MDBuilder(*C).createBranchWeights(1, 100000));
IRB.SetInsertPoint(Ins);
IRB.SetCurrentDebugLocation(EntryLoc);
// __sanitizer_cov gets the PC of the instruction using GET_CALLER_PC.
IRB.CreateCall(SanCovFunction, GuardP);
IRB.CreateCall(EmptyAsm, {}); // Avoids callback merge.
}
if (Options.Use8bitCounters) {
IRB.SetInsertPoint(IP);
Value *P = IRB.CreateAdd(
IRB.CreatePointerCast(EightBitCounterArray, IntptrTy),
ConstantInt::get(IntptrTy, NumberOfInstrumentedBlocks() - 1));
P = IRB.CreateIntToPtr(P, IRB.getInt8PtrTy());
LoadInst *LI = IRB.CreateLoad(P);
Value *Inc = IRB.CreateAdd(LI, ConstantInt::get(IRB.getInt8Ty(), 1));
StoreInst *SI = IRB.CreateStore(Inc, P);
SetNoSanitizeMetadata(LI);
SetNoSanitizeMetadata(SI);
}
if (Options.TraceBB) {
// Experimental support for tracing.
// Insert a callback with the same guard variable as used for coverage.
IRB.SetInsertPoint(IP);
IRB.CreateCall(IsEntryBB ? SanCovTraceEnter : SanCovTraceBB, GuardP);
}
}
void SanitizerCoverageModule::InjectCoverageAtBlock(Function &F, BasicBlock &BB,
bool UseCalls) {
// Don't insert coverage for unreachable blocks: we will never call
// __sanitizer_cov() for them, so counting them in
// NumberOfInstrumentedBlocks() might complicate calculation of code coverage
// percentage. Also, unreachable instructions frequently have no debug
// locations.
if (isa<UnreachableInst>(BB.getTerminator()))
return;
BasicBlock::iterator IP = BB.getFirstInsertionPt();
bool IsEntryBB = &BB == &F.getEntryBlock();
DebugLoc EntryLoc;
if (IsEntryBB) {
if (auto SP = getDISubprogram(&F))
EntryLoc = DebugLoc::get(SP->getScopeLine(), 0, SP);
// Keep static allocas and llvm.localescape calls in the entry block. Even
// if we aren't splitting the block, it's nice for allocas to be before
// calls.
IP = PrepareToSplitEntryBlock(BB, IP);
} else {
EntryLoc = IP->getDebugLoc();
}
IRBuilder<> IRB(&*IP);
IRB.SetCurrentDebugLocation(EntryLoc);
Value *GuardP = IRB.CreateAdd(
IRB.CreatePointerCast(GuardArray, IntptrTy),
ConstantInt::get(IntptrTy, (1 + NumberOfInstrumentedBlocks()) * 4));
Type *Int32PtrTy = PointerType::getUnqual(IRB.getInt32Ty());
GuardP = IRB.CreateIntToPtr(GuardP, Int32PtrTy);
if (Options.TracePC) {
IRB.CreateCall(SanCovTracePC);
} else if (Options.TraceBB) {
IRB.CreateCall(IsEntryBB ? SanCovTraceEnter : SanCovTraceBB, GuardP);
} else if (UseCalls) {
IRB.CreateCall(SanCovWithCheckFunction, GuardP);
} else {
LoadInst *Load = IRB.CreateLoad(GuardP);
Load->setAtomic(Monotonic);
Load->setAlignment(4);
SetNoSanitizeMetadata(Load);
Value *Cmp = IRB.CreateICmpSGE(Constant::getNullValue(Load->getType()), Load);
Instruction *Ins = SplitBlockAndInsertIfThen(
Cmp, &*IP, false, MDBuilder(*C).createBranchWeights(1, 100000));
IRB.SetInsertPoint(Ins);
IRB.SetCurrentDebugLocation(EntryLoc);
// __sanitizer_cov gets the PC of the instruction using GET_CALLER_PC.
IRB.CreateCall(SanCovFunction, GuardP);
IRB.CreateCall(EmptyAsm, {}); // Avoids callback merge.
}
if (Options.Use8bitCounters) {
IRB.SetInsertPoint(&*IP);
Value *P = IRB.CreateAdd(
IRB.CreatePointerCast(EightBitCounterArray, IntptrTy),
ConstantInt::get(IntptrTy, NumberOfInstrumentedBlocks() - 1));
P = IRB.CreateIntToPtr(P, IRB.getInt8PtrTy());
LoadInst *LI = IRB.CreateLoad(P);
Value *Inc = IRB.CreateAdd(LI, ConstantInt::get(IRB.getInt8Ty(), 1));
StoreInst *SI = IRB.CreateStore(Inc, P);
SetNoSanitizeMetadata(LI);
SetNoSanitizeMetadata(SI);
}
}