/// Remove redundant runtime calls for stack allocated buffers.
/// If a buffer is allocated on the stack it's not needed to explicitly set
/// the RC_DEALLOCATING_FLAG flag (except there is code which may depend on it).
/// Also the a call to swift_deallocClassInstance (which stems from an inlined
/// deallocator) is not needed.
///
/// %0 = alloca
/// ...
/// call @swift_setDeallocating(%0) // not needed
/// // code which does not depend on the RC_DEALLOCATING_FLAG flag.
/// call @swift_deallocClassInstance(%0) // not needed
/// call @llvm.lifetime.end(%0)
///
static void removeRedundantRTCalls(CallInst *DeallocCall) {
BasicBlock::iterator Iter(DeallocCall);
BasicBlock::iterator Begin = DeallocCall->getParent()->begin();
Value *Buffer = DeallocCall->getArgOperand(0);
CallInst *RedundantDealloc = nullptr;
CallInst *RedundantSetFlag = nullptr;
SmallVector<Instruction *, 2> ToDelete;
while (Iter != Begin) {
--Iter;
Instruction *I = &*Iter;
if (auto *CI = dyn_cast<CallInst>(I)) {
// Check if we have a runtime function with the buffer as argument.
if (CI->getNumArgOperands() < 1)
break;
if (CI->getArgOperand(0)->stripPointerCasts() != Buffer)
break;
auto *Callee = dyn_cast<Constant>(CI->getCalledValue());
if (!Callee)
break;
// The callee function my be a bitcast constant expression.
if (auto *U = dyn_cast<ConstantExpr>(Callee)) {
if (U->getOpcode() == Instruction::BitCast)
Callee = U->getOperand(0);
}
auto *RTFunc = dyn_cast<Function>(Callee);
if (!RTFunc)
break;
if (RTFunc->getName() == "swift_setDeallocating") {
assert(RedundantDealloc && "dealloc call must follow setDeallocating");
assert(!RedundantSetFlag && "multiple setDeallocating calls");
RedundantSetFlag = CI;
continue;
}
if (RTFunc->getName() == "swift_deallocClassInstance") {
assert(!RedundantSetFlag && "dealloc call must follow setDeallocating");
assert(!RedundantDealloc && "multiple deallocClassInstance calls");
RedundantDealloc = CI;
continue;
}
break;
}
// Bail if we have an instruction which may read the RC_DEALLOCATING_FLAG
// flag.
if (I->mayReadFromMemory())
break;
}
if (RedundantDealloc)
RedundantDealloc->eraseFromParent();
if (RedundantSetFlag)
RedundantSetFlag->eraseFromParent();
}
bool LowerExpectIntrinsic::runOnFunction(Function &F) {
for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
BasicBlock *BB = I++;
// Create "block_weights" metadata.
if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
if (HandleIfExpect(BI))
IfHandled++;
} else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
if (HandleSwitchExpect(SI))
IfHandled++;
}
// remove llvm.expect intrinsics.
for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
BI != BE; ) {
CallInst *CI = dyn_cast<CallInst>(BI++);
if (!CI)
continue;
Function *Fn = CI->getCalledFunction();
if (Fn && Fn->getIntrinsicID() == Intrinsic::expect) {
Value *Exp = CI->getArgOperand(0);
CI->replaceAllUsesWith(Exp);
CI->eraseFromParent();
}
}
}
return false;
}
// visitCallInst - This converts all LLVM call instructions into invoke
// instructions. The except part of the invoke goes to the "LongJmpBlkPre"
// that grabs the exception and proceeds to determine if it's a longjmp
// exception or not.
void LowerSetJmp::visitCallInst(CallInst& CI)
{
if (CI.getCalledFunction())
if (!IsTransformableFunction(CI.getCalledFunction()->getName()) ||
CI.getCalledFunction()->isIntrinsic()) return;
BasicBlock* OldBB = CI.getParent();
// If not reachable from a setjmp call, don't transform.
if (!DFSBlocks.count(OldBB)) return;
BasicBlock* NewBB = OldBB->splitBasicBlock(CI);
assert(NewBB && "Couldn't split BB of \"call\" instruction!!");
DFSBlocks.insert(NewBB);
NewBB->setName("Call2Invoke");
Function* Func = OldBB->getParent();
// Construct the new "invoke" instruction.
TerminatorInst* Term = OldBB->getTerminator();
std::vector<Value*> Params(CI.op_begin() + 1, CI.op_end());
InvokeInst* II =
InvokeInst::Create(CI.getCalledValue(), NewBB, PrelimBBMap[Func],
Params.begin(), Params.end(), CI.getName(), Term);
II->setCallingConv(CI.getCallingConv());
II->setAttributes(CI.getAttributes());
// Replace the old call inst with the invoke inst and remove the call.
CI.replaceAllUsesWith(II);
CI.eraseFromParent();
// The old terminator is useless now that we have the invoke inst.
Term->eraseFromParent();
++CallsTransformed;
}
bool OptimizeFastMemoryChecks::runOnFunction(Function &F) {
ABA = &getAnalysis<ArrayBoundsAnalysis>();
MSCI = &getAnalysis<MSCInfo>();
// Visit all call instructions in the function to find the checks.
visit(F);
for (size_t i = 0, N = FastCheckCalls.size(); i < N; ++i) {
CallInst *CI = FastCheckCalls[i];
CheckInfoType *Info = MSCI->getCheckInfo(CI->getCalledFunction());
assert(Info && Info->isFastMemoryCheck() && "expecting fast memory checks");
Value *AccessPtr = CI->getArgOperand(Info->PtrArgNo);
Value *AccessSize = CI->getArgOperand(Info->SizeArgNo);
Value *ObjectPtr = CI->getArgOperand(Info->ObjArgNo);
Value *ObjectSize = CI->getArgOperand(Info->ObjSizeArgNo);
if (ABA->isAlwaysInBounds(AccessPtr, AccessSize, ObjectPtr, ObjectSize)) {
CI->eraseFromParent();
++FastMemoryChecksRemoved;
}
}
bool ChangedAnything = !FastCheckCalls.empty();
FastCheckCalls.clear();
return ChangedAnything;
}
// StripDebugInfo - Strip debug info in the module if it exists.
// To do this, we remove llvm.dbg.func.start, llvm.dbg.stoppoint, and
// llvm.dbg.region.end calls, and any globals they point to if now dead.
static bool StripDebugInfo(Module &M) {
bool Changed = false;
// Remove all of the calls to the debugger intrinsics, and remove them from
// the module.
if (Function *Declare = M.getFunction("llvm.dbg.declare")) {
while (!Declare->use_empty()) {
CallInst *CI = cast<CallInst>(Declare->use_back());
CI->eraseFromParent();
}
Declare->eraseFromParent();
Changed = true;
}
if (Function *DbgVal = M.getFunction("llvm.dbg.value")) {
while (!DbgVal->use_empty()) {
CallInst *CI = cast<CallInst>(DbgVal->use_back());
CI->eraseFromParent();
}
DbgVal->eraseFromParent();
Changed = true;
}
for (Module::named_metadata_iterator NMI = M.named_metadata_begin(),
NME = M.named_metadata_end(); NMI != NME;) {
NamedMDNode *NMD = NMI;
++NMI;
if (NMD->getName().startswith("llvm.dbg.")) {
NMD->eraseFromParent();
Changed = true;
}
}
for (Module::iterator MI = M.begin(), ME = M.end(); MI != ME; ++MI)
for (Function::iterator FI = MI->begin(), FE = MI->end(); FI != FE;
++FI)
for (BasicBlock::iterator BI = FI->begin(), BE = FI->end(); BI != BE;
++BI) {
if (!BI->getDebugLoc().isUnknown()) {
Changed = true;
BI->setDebugLoc(DebugLoc());
}
}
return Changed;
}
/**
* removeUndefCalls -- remove calls with undef function
*
* These are irrelevant to the code, so may be removed completely.
*/
void FunctionStaticSlicer::removeUndefCalls(ModulePass *MP, Function &F) {
for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E;) {
CallInst *CI = dyn_cast<CallInst>(&*I);
++I;
if (CI && isa<UndefValue>(CI->getCalledValue())) {
CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
CI->eraseFromParent();
}
}
}
static bool ExpandOpForIntSize(Module *M, unsigned Bits, bool Mul) {
IntegerType *IntTy = IntegerType::get(M->getContext(), Bits);
SmallVector<Type *, 1> Types;
Types.push_back(IntTy);
Intrinsic::ID ID = (Mul ? Intrinsic::umul_with_overflow
: Intrinsic::uadd_with_overflow);
std::string Name = Intrinsic::getName(ID, Types);
Function *Intrinsic = M->getFunction(Name);
if (!Intrinsic)
return false;
for (Value::use_iterator CallIter = Intrinsic->use_begin(),
E = Intrinsic->use_end(); CallIter != E; ) {
CallInst *Call = dyn_cast<CallInst>(*CallIter++);
if (!Call) {
report_fatal_error("ExpandArithWithOverflow: Taking the address of a "
"*.with.overflow intrinsic is not allowed");
}
Value *VariableArg;
ConstantInt *ConstantArg;
if (ConstantInt *C = dyn_cast<ConstantInt>(Call->getArgOperand(0))) {
VariableArg = Call->getArgOperand(1);
ConstantArg = C;
} else if (ConstantInt *C = dyn_cast<ConstantInt>(Call->getArgOperand(1))) {
VariableArg = Call->getArgOperand(0);
ConstantArg = C;
} else {
errs() << "Use: " << *Call << "\n";
report_fatal_error("ExpandArithWithOverflow: At least one argument of "
"*.with.overflow must be a constant");
}
Value *ArithResult = BinaryOperator::Create(
(Mul ? Instruction::Mul : Instruction::Add), VariableArg, ConstantArg,
Call->getName() + ".arith", Call);
uint64_t ArgMax;
if (Mul) {
ArgMax = UintTypeMax(Bits) / ConstantArg->getZExtValue();
} else {
ArgMax = UintTypeMax(Bits) - ConstantArg->getZExtValue();
}
Value *OverflowResult = new ICmpInst(
Call, CmpInst::ICMP_UGT, VariableArg, ConstantInt::get(IntTy, ArgMax),
Call->getName() + ".overflow");
// Construct the struct result.
Value *NewStruct = UndefValue::get(Call->getType());
NewStruct = CreateInsertValue(NewStruct, 0, ArithResult, Call);
NewStruct = CreateInsertValue(NewStruct, 1, OverflowResult, Call);
Call->replaceAllUsesWith(NewStruct);
Call->eraseFromParent();
}
Intrinsic->eraseFromParent();
return true;
}
bool llvm::StripDebugInfo(Module &M) {
bool Changed = false;
// Remove all of the calls to the debugger intrinsics, and remove them from
// the module.
if (Function *Declare = M.getFunction("llvm.dbg.declare")) {
while (!Declare->use_empty()) {
CallInst *CI = cast<CallInst>(Declare->user_back());
CI->eraseFromParent();
}
Declare->eraseFromParent();
Changed = true;
}
if (Function *DbgVal = M.getFunction("llvm.dbg.value")) {
while (!DbgVal->use_empty()) {
CallInst *CI = cast<CallInst>(DbgVal->user_back());
CI->eraseFromParent();
}
DbgVal->eraseFromParent();
Changed = true;
}
for (Module::named_metadata_iterator NMI = M.named_metadata_begin(),
NME = M.named_metadata_end(); NMI != NME;) {
NamedMDNode *NMD = NMI;
++NMI;
if (NMD->getName().startswith("llvm.dbg.")) {
NMD->eraseFromParent();
Changed = true;
}
}
for (Function &F : M)
Changed |= stripDebugInfo(F);
if (GVMaterializer *Materializer = M.getMaterializer())
Materializer->setStripDebugInfo();
return Changed;
}
void AsmDirectivesVisitor::visitCallInst(CallInst &CI) {
if (!CI.isInlineAsm() ||
!cast<InlineAsm>(CI.getCalledValue())->isAsmMemory())
return;
// In NaCl ``asm("":::"memory")`` always comes in pairs, straddling a
// sequentially consistent fence. Other passes rewrite this fence to
// an equivalent stable NaCl intrinsic, meaning that this assembly can
// be removed.
CI.eraseFromParent();
ModifiedFunction = true;
}
CallInst* FunctionCalls::changeFunctionCall(Module &module, Change* change) {
FunctionChange *funChange = (FunctionChange*)change;
CallInst *oldCallInst = dyn_cast<CallInst>(funChange->getValue());
CallInst *newCallInst = NULL;
string oldFunction = oldCallInst->getCalledFunction()->getName();
string newFunction = funChange->getSwitch();
// TODO: use the types vector to not assume signature
Function *newCallee = module.getFunction(newFunction);
if (newCallee) {
errs() << "Changing function call from " << oldFunction << " to " << newFunction << "\n";
if (oldFunction != newFunction) {
// retrieving original operand
Value *oldOperand = oldCallInst->getOperand(0);
// downcasting operand
Type *fType = Type::getFloatTy(module.getContext());
FPTruncInst *newOperand = new FPTruncInst(oldOperand, fType, "", oldCallInst);
// populating array of operands
vector<Value*> operands;
operands.push_back(newOperand);
ArrayRef<Value*> *arrayRefOperands = new ArrayRef<Value*>(operands);
// creating the new CallInst
newCallInst = CallInst::Create(newCallee, *arrayRefOperands, "newCall", oldCallInst);
// casting result to double
Type *dType = Type::getDoubleTy(module.getContext());
FPExtInst *result = new FPExtInst(newCallInst, dType, "", oldCallInst);
// replacing all uses of call instruction
oldCallInst->replaceAllUsesWith(result);
// deleting old callInst
oldCallInst->eraseFromParent();
errs() << "\tChange was successful\n";
}
else {
errs() << "\tNo change required\n";
}
}
else {
errs() << "\tDid not find function " << newFunction << "\n";
}
return newCallInst;
}
void RuntimeHelperFixupPass::FixArrayInit(Function *F)
{
Module *M = F->getParent();
LLVMContext &ctx = F->getContext();
for (auto it = F->use_begin(), end = F->use_end(); it != end; ++it)
{
CallInst *CI = dyn_cast<CallInst>(*it);
assert (CI && "Unknown usage for array init helper");
Instruction *field_runtime_handle = dyn_cast<Instruction>(CI->getArgOperand(1));
assert (field_runtime_handle);
MDNode *md = field_runtime_handle->getMetadata("silk_runtime_field_handle");
assert (md);
ConstantInt *handle = dyn_cast<ConstantInt>(md->getOperand(0));
assert(handle);
auto vm_field = reinterpret_cast<VMField*>(handle->getValue().getLimitedValue());
auto field_map = vm_field->field_mapping();
auto vm_named_class = dynamic_cast<VMNamedClass*>(vm_field->type());
auto size = vm_named_class->type_def()->class_size();
raw_istream is(field_map.start(), size);
std::vector<uint8_t> buf;
buf.resize(size);
is.read((char*)buf.data(), size);
Constant *predefined_value = ConstantDataArray::get(ctx, buf);
auto GV = new GlobalVariable(*M, predefined_value->getType(),
true, GlobalValue::InternalLinkage,
predefined_value, ".initdata");
Value *array_ptr = CI->getArgOperand(0);
IRBuilder<> builder(CI);
auto intrinsic_array_base = intrinsic_->array_base_pointer();
auto array_ptr_casted = builder.CreateBitCast(array_ptr, builder.getInt8PtrTy());
auto array_base_ptr = builder.CreateCall(intrinsic_array_base, array_ptr_casted);
builder.CreateMemCpy(array_base_ptr, GV, ConstantInt::get(Type::getInt64Ty(ctx), size), 0);
CI->eraseFromParent();
}
}
bool OptimizeGEPChecks::runOnFunction(Function &F) {
ABC = &getAnalysis<ArrayBoundsCheckLocal>();
MSCI = &getAnalysis<MSCInfo>();
// Visit all call instructions in the function.
visit(F);
for (size_t i = 0, N = ToRemove.size(); i < N; ++i) {
CallInst *CI = ToRemove[i];
CheckInfoType *Info = MSCI->getCheckInfo(CI->getCalledFunction());
CI->replaceAllUsesWith(CI->getArgOperand(Info->DestPtrArgNo));
CI->eraseFromParent();
++SafeGEPs;
}
bool ChangedAnything = !ToRemove.empty();
ToRemove.clear();
return ChangedAnything;
}
void ArgumentRecovery::fixCallSites(Function& base, Function& newTarget, const CallInformation& ci)
{
auto targetInfo = TargetInfo::getTargetInfo(*base.getParent());
// loop over callers and transform call sites.
while (!base.use_empty())
{
CallInst* call = cast<CallInst>(base.user_back());
Function* caller = call->getParent()->getParent();
auto registers = getRegisterPtr(*caller);
auto newCall = createCallSite(*targetInfo, ci, newTarget, *registers, *call);
// replace call
newCall->takeName(call);
call->eraseFromParent();
md::incrementFunctionVersion(*caller);
}
}
/// Replaces
///
/// strong_retain_unowned %x
/// ... // speculatively executable instructions, including loads from %x
/// strong_release %x
///
/// with
///
/// ... // speculatively executable instructions, including loads from %x
/// check_unowned %x
///
static bool performLocalRetainUnownedOpt(CallInst *Retain, BasicBlock &BB,
ARCEntryPointBuilder &B) {
Value *RetainedObject = Retain->getArgOperand(0);
Value *LoadBaseAddr = getBaseAddress(RetainedObject);
auto BBI = Retain->getIterator(), BBE = BB.getTerminator()->getIterator();
// Scan until we get to the end of the block.
for (++BBI; BBI != BBE; ++BBI) {
Instruction &I = *BBI;
if (classifyInstruction(I) == RT_Release) {
CallInst *ThisRelease = cast<CallInst>(&I);
// Is this the trailing release of the unowned-retained reference?
if (ThisRelease->getArgOperand(0) != RetainedObject)
return false;
// Replace the trailing release with a check_unowned.
B.setInsertPoint(ThisRelease);
B.createCheckUnowned(RetainedObject, ThisRelease);
Retain->eraseFromParent();
ThisRelease->eraseFromParent();
++NumRetainReleasePairs;
return true;
}
if (auto *LI = dyn_cast<LoadInst>(&I)) {
// Accept loads from the unowned-referenced object. This may load garbage
// values, but they are not used until the final check_unowned succeeds.
if (getBaseAddress(LI->getPointerOperand()) == LoadBaseAddr)
continue;
}
// Other than loads from the unowned-referenced object we only accept
// speculatively executable instructions.
if (!isSafeToSpeculativelyExecute(&I))
return false;
}
return false;
}
bool StripDebugDeclare::runOnModule(Module &M) {
if (skipModule(M))
return false;
Function *Declare = M.getFunction("llvm.dbg.declare");
std::vector<Constant*> DeadConstants;
if (Declare) {
while (!Declare->use_empty()) {
CallInst *CI = cast<CallInst>(Declare->user_back());
Value *Arg1 = CI->getArgOperand(0);
Value *Arg2 = CI->getArgOperand(1);
assert(CI->use_empty() && "llvm.dbg intrinsic should have void result");
CI->eraseFromParent();
if (Arg1->use_empty()) {
if (Constant *C = dyn_cast<Constant>(Arg1))
DeadConstants.push_back(C);
else
RecursivelyDeleteTriviallyDeadInstructions(Arg1);
}
if (Arg2->use_empty())
if (Constant *C = dyn_cast<Constant>(Arg2))
DeadConstants.push_back(C);
}
Declare->eraseFromParent();
}
while (!DeadConstants.empty()) {
Constant *C = DeadConstants.back();
DeadConstants.pop_back();
if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
if (GV->hasLocalLinkage())
RemoveDeadConstant(GV);
} else
RemoveDeadConstant(C);
}
return true;
}
/// This function strips all debug info intrinsics, except for llvm.dbg.declare.
/// If an llvm.dbg.declare intrinsic is invalid, then this function simply
/// strips that use.
void llvm::CheckDebugInfoIntrinsics(Module *M) {
if (Function *FuncStart = M->getFunction("llvm.dbg.func.start")) {
while (!FuncStart->use_empty())
cast<CallInst>(FuncStart->use_back())->eraseFromParent();
FuncStart->eraseFromParent();
}
if (Function *StopPoint = M->getFunction("llvm.dbg.stoppoint")) {
while (!StopPoint->use_empty())
cast<CallInst>(StopPoint->use_back())->eraseFromParent();
StopPoint->eraseFromParent();
}
if (Function *RegionStart = M->getFunction("llvm.dbg.region.start")) {
while (!RegionStart->use_empty())
cast<CallInst>(RegionStart->use_back())->eraseFromParent();
RegionStart->eraseFromParent();
}
if (Function *RegionEnd = M->getFunction("llvm.dbg.region.end")) {
while (!RegionEnd->use_empty())
cast<CallInst>(RegionEnd->use_back())->eraseFromParent();
RegionEnd->eraseFromParent();
}
if (Function *Declare = M->getFunction("llvm.dbg.declare")) {
if (!Declare->use_empty()) {
DbgDeclareInst *DDI = cast<DbgDeclareInst>(Declare->use_back());
if (!isa<MDNode>(DDI->getArgOperand(0)) ||
!isa<MDNode>(DDI->getArgOperand(1))) {
while (!Declare->use_empty()) {
CallInst *CI = cast<CallInst>(Declare->use_back());
CI->eraseFromParent();
}
Declare->eraseFromParent();
}
}
}
}
bool InlineSpecials::runOnFunction(Function &F) {
std::set<CallInst*> worklist;
for(inst_iterator i = inst_begin(F), e = inst_end(F); i != e; ++i) {
Instruction *I = &*i;
if(CallInst *C = dyn_cast<CallInst>(I)) {
//get the target of the call, if known
Function *invoked = C->getCalledFunction();
if(invoked != NULL) {
worklist.insert(C);
}
}
}
for(std::set<CallInst*>::iterator itr = worklist.begin();
itr != worklist.end();
itr++)
{
CallInst *C = *itr;
Function *invoked = C->getCalledFunction();
string invokedFuncName = invoked->getName();
map<string,ReplaceFunctionPt>::iterator it = specialMap.find(invokedFuncName);
if(it != specialMap.end()) {
ReplaceFunctionPt func = it->second;
Value *newv = func(F.getParent(), C, this);
C->replaceAllUsesWith(newv);
C->eraseFromParent();
}
}
return false;
}
void RuntimeHelperFixupPass::FixStringConstructor(Function *old_construct, Function *new_construct)
{
for (auto it = old_construct->use_begin(), end = old_construct->use_end(); it != end; ++it)
{
CallInst *CI = dyn_cast<CallInst>(*it);
assert (CI);
auto num_ops = CI->getNumArgOperands();
SmallVector<Value*, 4> ops;
for (size_t i = 1; i < num_ops; ++i)
ops.push_back(CI->getArgOperand(i));
IRBuilder<> builder(CI);
auto new_call = builder.CreateCall(new_construct, ops);
BitCastInst *this_ptr = cast<BitCastInst>(CI->getArgOperand(0));
auto alloc = cast<Instruction>(this_ptr->llvm::User::getOperand(0));
this_ptr->replaceAllUsesWith(new_call);
CI->eraseFromParent();
this_ptr->eraseFromParent();
alloc->eraseFromParent();
}
}
void visitCallInst(CallInst &I) {
string intrinsic = I.getCalledFunction()->getName().str();
if(intrinsic.find("modmul") != -1) {
CallInst *enterMontpro1 = enterMontgomery(I.getOperand(0), I.getOperand(2), &I);
CallInst *enterMontpro2 = enterMontgomery(I.getOperand(1), I.getOperand(2), &I);
CallInst *mulMontpro = mulMontgomery(I.getName().str(), enterMontpro1, enterMontpro2, I.getOperand(2), &I);
CallInst *exitMontpro = leaveMontgomery(mulMontpro, I.getOperand(2), &I);
I.replaceAllUsesWith(exitMontpro);
I.removeFromParent();
} else if(intrinsic.find("modexp") != -1) {
CallInst *enterMontpro1 = enterMontgomery(I.getOperand(0), I.getOperand(2), &I);
CallInst *expMontpro = expMontgomery(I.getName().str(), enterMontpro1, I.getOperand(1), I.getOperand(2), &I);
CallInst *exitMontpro = leaveMontgomery(expMontpro, I.getOperand(2), &I);
I.replaceAllUsesWith(exitMontpro);
I.eraseFromParent();
}
}
bool ObjCARCContract::tryToPeepholeInstruction(
Function &F, Instruction *Inst, inst_iterator &Iter,
SmallPtrSetImpl<Instruction *> &DependingInsts,
SmallPtrSetImpl<const BasicBlock *> &Visited,
bool &TailOkForStoreStrongs) {
// Only these library routines return their argument. In particular,
// objc_retainBlock does not necessarily return its argument.
ARCInstKind Class = GetBasicARCInstKind(Inst);
switch (Class) {
case ARCInstKind::FusedRetainAutorelease:
case ARCInstKind::FusedRetainAutoreleaseRV:
return false;
case ARCInstKind::Autorelease:
case ARCInstKind::AutoreleaseRV:
return contractAutorelease(F, Inst, Class, DependingInsts, Visited);
case ARCInstKind::Retain:
// Attempt to convert retains to retainrvs if they are next to function
// calls.
if (!optimizeRetainCall(F, Inst))
return false;
// If we succeed in our optimization, fall through.
// FALLTHROUGH
case ARCInstKind::RetainRV: {
// If we're compiling for a target which needs a special inline-asm
// marker to do the retainAutoreleasedReturnValue optimization,
// insert it now.
if (!RetainRVMarker)
return false;
BasicBlock::iterator BBI = Inst;
BasicBlock *InstParent = Inst->getParent();
// Step up to see if the call immediately precedes the RetainRV call.
// If it's an invoke, we have to cross a block boundary. And we have
// to carefully dodge no-op instructions.
do {
if (&*BBI == InstParent->begin()) {
BasicBlock *Pred = InstParent->getSinglePredecessor();
if (!Pred)
goto decline_rv_optimization;
BBI = Pred->getTerminator();
break;
}
--BBI;
} while (IsNoopInstruction(BBI));
if (&*BBI == GetArgRCIdentityRoot(Inst)) {
DEBUG(dbgs() << "Adding inline asm marker for "
"retainAutoreleasedReturnValue optimization.\n");
Changed = true;
InlineAsm *IA =
InlineAsm::get(FunctionType::get(Type::getVoidTy(Inst->getContext()),
/*isVarArg=*/false),
RetainRVMarker->getString(),
/*Constraints=*/"", /*hasSideEffects=*/true);
CallInst::Create(IA, "", Inst);
}
decline_rv_optimization:
return false;
}
case ARCInstKind::InitWeak: {
// objc_initWeak(p, null) => *p = null
CallInst *CI = cast<CallInst>(Inst);
if (IsNullOrUndef(CI->getArgOperand(1))) {
Value *Null =
ConstantPointerNull::get(cast<PointerType>(CI->getType()));
Changed = true;
new StoreInst(Null, CI->getArgOperand(0), CI);
DEBUG(dbgs() << "OBJCARCContract: Old = " << *CI << "\n"
<< " New = " << *Null << "\n");
CI->replaceAllUsesWith(Null);
CI->eraseFromParent();
}
return true;
}
case ARCInstKind::Release:
// Try to form an objc store strong from our release. If we fail, there is
// nothing further to do below, so continue.
tryToContractReleaseIntoStoreStrong(Inst, Iter);
return true;
case ARCInstKind::User:
// Be conservative if the function has any alloca instructions.
// Technically we only care about escaping alloca instructions,
// but this is sufficient to handle some interesting cases.
if (isa<AllocaInst>(Inst))
TailOkForStoreStrongs = false;
return true;
case ARCInstKind::IntrinsicUser:
// Remove calls to @clang.arc.use(...).
Inst->eraseFromParent();
return true;
default:
return true;
}
}
/// processMemCpy - perform simplication of memcpy's. If we have memcpy A which
/// copies X to Y, and memcpy B which copies Y to Z, then we can rewrite B to be
/// a memcpy from X to Z (or potentially a memmove, depending on circumstances).
/// This allows later passes to remove the first memcpy altogether.
bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>();
// The are two possible optimizations we can do for memcpy:
// a) memcpy-memcpy xform which exposes redundance for DSE.
// b) call-memcpy xform for return slot optimization.
MemDepResult dep = MD.getDependency(M);
if (!dep.isClobber())
return false;
if (!isa<MemCpyInst>(dep.getInst())) {
if (CallInst *C = dyn_cast<CallInst>(dep.getInst()))
return performCallSlotOptzn(M, C);
return false;
}
MemCpyInst *MDep = cast<MemCpyInst>(dep.getInst());
// We can only transforms memcpy's where the dest of one is the source of the
// other
if (M->getSource() != MDep->getDest())
return false;
// Second, the length of the memcpy's must be the same, or the preceeding one
// must be larger than the following one.
ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
ConstantInt *C2 = dyn_cast<ConstantInt>(M->getLength());
if (!C1 || !C2)
return false;
uint64_t DepSize = C1->getValue().getZExtValue();
uint64_t CpySize = C2->getValue().getZExtValue();
if (DepSize < CpySize)
return false;
// Finally, we have to make sure that the dest of the second does not
// alias the source of the first
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
if (AA.alias(M->getRawDest(), CpySize, MDep->getRawSource(), DepSize) !=
AliasAnalysis::NoAlias)
return false;
else if (AA.alias(M->getRawDest(), CpySize, M->getRawSource(), CpySize) !=
AliasAnalysis::NoAlias)
return false;
else if (AA.alias(MDep->getRawDest(), DepSize, MDep->getRawSource(), DepSize)
!= AliasAnalysis::NoAlias)
return false;
// If all checks passed, then we can transform these memcpy's
const Type *ArgTys[3] = { M->getRawDest()->getType(),
MDep->getRawSource()->getType(),
M->getLength()->getType() };
Function *MemCpyFun = Intrinsic::getDeclaration(
M->getParent()->getParent()->getParent(),
M->getIntrinsicID(), ArgTys, 3);
Value *Args[5] = {
M->getRawDest(), MDep->getRawSource(), M->getLength(),
M->getAlignmentCst(), M->getVolatileCst()
};
CallInst *C = CallInst::Create(MemCpyFun, Args, Args+5, "", M);
// If C and M don't interfere, then this is a valid transformation. If they
// did, this would mean that the two sources overlap, which would be bad.
if (MD.getDependency(C) == dep) {
MD.removeInstruction(M);
M->eraseFromParent();
++NumMemCpyInstr;
return true;
}
// Otherwise, there was no point in doing this, so we remove the call we
// inserted and act like nothing happened.
MD.removeInstruction(C);
C->eraseFromParent();
return false;
}
bool ObjCARCContract::runOnFunction(Function &F) {
if (!EnableARCOpts)
return false;
// If nothing in the Module uses ARC, don't do anything.
if (!Run)
return false;
Changed = false;
AA = &getAnalysis<AliasAnalysis>();
DT = &getAnalysis<DominatorTree>();
PA.setAA(&getAnalysis<AliasAnalysis>());
// Track whether it's ok to mark objc_storeStrong calls with the "tail"
// keyword. Be conservative if the function has variadic arguments.
// It seems that functions which "return twice" are also unsafe for the
// "tail" argument, because they are setjmp, which could need to
// return to an earlier stack state.
bool TailOkForStoreStrongs = !F.isVarArg() &&
!F.callsFunctionThatReturnsTwice();
// For ObjC library calls which return their argument, replace uses of the
// argument with uses of the call return value, if it dominates the use. This
// reduces register pressure.
SmallPtrSet<Instruction *, 4> DependingInstructions;
SmallPtrSet<const BasicBlock *, 4> Visited;
for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ) {
Instruction *Inst = &*I++;
DEBUG(dbgs() << "ObjCARCContract: Visiting: " << *Inst << "\n");
// Only these library routines return their argument. In particular,
// objc_retainBlock does not necessarily return its argument.
InstructionClass Class = GetBasicInstructionClass(Inst);
switch (Class) {
case IC_Retain:
case IC_FusedRetainAutorelease:
case IC_FusedRetainAutoreleaseRV:
break;
case IC_Autorelease:
case IC_AutoreleaseRV:
if (ContractAutorelease(F, Inst, Class, DependingInstructions, Visited))
continue;
break;
case IC_RetainRV: {
// If we're compiling for a target which needs a special inline-asm
// marker to do the retainAutoreleasedReturnValue optimization,
// insert it now.
if (!RetainRVMarker)
break;
BasicBlock::iterator BBI = Inst;
BasicBlock *InstParent = Inst->getParent();
// Step up to see if the call immediately precedes the RetainRV call.
// If it's an invoke, we have to cross a block boundary. And we have
// to carefully dodge no-op instructions.
do {
if (&*BBI == InstParent->begin()) {
BasicBlock *Pred = InstParent->getSinglePredecessor();
if (!Pred)
goto decline_rv_optimization;
BBI = Pred->getTerminator();
break;
}
--BBI;
} while (IsNoopInstruction(BBI));
if (&*BBI == GetObjCArg(Inst)) {
DEBUG(dbgs() << "ObjCARCContract: Adding inline asm marker for "
"retainAutoreleasedReturnValue optimization.\n");
Changed = true;
InlineAsm *IA =
InlineAsm::get(FunctionType::get(Type::getVoidTy(Inst->getContext()),
/*isVarArg=*/false),
RetainRVMarker->getString(),
/*Constraints=*/"", /*hasSideEffects=*/true);
CallInst::Create(IA, "", Inst);
}
decline_rv_optimization:
break;
}
case IC_InitWeak: {
// objc_initWeak(p, null) => *p = null
CallInst *CI = cast<CallInst>(Inst);
if (IsNullOrUndef(CI->getArgOperand(1))) {
Value *Null =
ConstantPointerNull::get(cast<PointerType>(CI->getType()));
Changed = true;
new StoreInst(Null, CI->getArgOperand(0), CI);
DEBUG(dbgs() << "OBJCARCContract: Old = " << *CI << "\n"
<< " New = " << *Null << "\n");
CI->replaceAllUsesWith(Null);
CI->eraseFromParent();
}
continue;
}
case IC_Release:
ContractRelease(Inst, I);
continue;
case IC_User:
// Be conservative if the function has any alloca instructions.
// Technically we only care about escaping alloca instructions,
// but this is sufficient to handle some interesting cases.
if (isa<AllocaInst>(Inst))
TailOkForStoreStrongs = false;
continue;
case IC_IntrinsicUser:
// Remove calls to @clang.arc.use(...).
Inst->eraseFromParent();
continue;
default:
continue;
}
DEBUG(dbgs() << "ObjCARCContract: Finished List.\n\n");
// Don't use GetObjCArg because we don't want to look through bitcasts
// and such; to do the replacement, the argument must have type i8*.
const Value *Arg = cast<CallInst>(Inst)->getArgOperand(0);
for (;;) {
// If we're compiling bugpointed code, don't get in trouble.
if (!isa<Instruction>(Arg) && !isa<Argument>(Arg))
break;
// Look through the uses of the pointer.
for (Value::const_use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
UI != UE; ) {
Use &U = UI.getUse();
unsigned OperandNo = UI.getOperandNo();
++UI; // Increment UI now, because we may unlink its element.
// If the call's return value dominates a use of the call's argument
// value, rewrite the use to use the return value. We check for
// reachability here because an unreachable call is considered to
// trivially dominate itself, which would lead us to rewriting its
// argument in terms of its return value, which would lead to
// infinite loops in GetObjCArg.
if (DT->isReachableFromEntry(U) && DT->dominates(Inst, U)) {
Changed = true;
Instruction *Replacement = Inst;
Type *UseTy = U.get()->getType();
if (PHINode *PHI = dyn_cast<PHINode>(U.getUser())) {
// For PHI nodes, insert the bitcast in the predecessor block.
unsigned ValNo = PHINode::getIncomingValueNumForOperand(OperandNo);
BasicBlock *BB = PHI->getIncomingBlock(ValNo);
if (Replacement->getType() != UseTy)
Replacement = new BitCastInst(Replacement, UseTy, "",
&BB->back());
// While we're here, rewrite all edges for this PHI, rather
// than just one use at a time, to minimize the number of
// bitcasts we emit.
for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
if (PHI->getIncomingBlock(i) == BB) {
// Keep the UI iterator valid.
if (&PHI->getOperandUse(
PHINode::getOperandNumForIncomingValue(i)) ==
&UI.getUse())
++UI;
PHI->setIncomingValue(i, Replacement);
}
} else {
if (Replacement->getType() != UseTy)
Replacement = new BitCastInst(Replacement, UseTy, "",
cast<Instruction>(U.getUser()));
U.set(Replacement);
}
}
}
// If Arg is a no-op casted pointer, strip one level of casts and iterate.
if (const BitCastInst *BI = dyn_cast<BitCastInst>(Arg))
Arg = BI->getOperand(0);
else if (isa<GEPOperator>(Arg) &&
cast<GEPOperator>(Arg)->hasAllZeroIndices())
Arg = cast<GEPOperator>(Arg)->getPointerOperand();
else if (isa<GlobalAlias>(Arg) &&
!cast<GlobalAlias>(Arg)->mayBeOverridden())
Arg = cast<GlobalAlias>(Arg)->getAliasee();
else
break;
}
}
// If this function has no escaping allocas or suspicious vararg usage,
// objc_storeStrong calls can be marked with the "tail" keyword.
if (TailOkForStoreStrongs)
for (SmallPtrSet<CallInst *, 8>::iterator I = StoreStrongCalls.begin(),
E = StoreStrongCalls.end(); I != E; ++I)
(*I)->setTailCall();
StoreStrongCalls.clear();
return Changed;
}
int compile(list<string> args, list<string> kgen_args,
string merge, list<string> merge_args,
string input, string output, int arch,
string host_compiler, string fileprefix)
{
//
// The LLVM compiler to emit IR.
//
const char* llvm_compiler = "kernelgen-gfortran";
//
// Interpret kernelgen compile options.
//
for (list<string>::iterator iarg = kgen_args.begin(),
iearg = kgen_args.end(); iarg != iearg; iarg++)
{
const char* arg = (*iarg).c_str();
if (!strncmp(arg, "-Wk,--llvm-compiler=", 20))
llvm_compiler = arg + 20;
}
//
// Generate temporary output file.
// Check if output file is specified in the command line.
// Replace or add output to the temporary file.
//
cfiledesc tmp_output = cfiledesc::mktemp(fileprefix);
bool output_specified = false;
for (list<string>::iterator iarg = args.begin(),
iearg = args.end(); iarg != iearg; iarg++)
{
const char* arg = (*iarg).c_str();
if (!strcmp(arg, "-o"))
{
iarg++;
*iarg = tmp_output.getFilename();
output_specified = true;
break;
}
}
if (!output_specified)
{
args.push_back("-o");
args.push_back(tmp_output.getFilename());
}
//
// 1) Compile source code using regular host compiler.
//
{
if (verbose)
{
cout << host_compiler;
for (list<string>::iterator iarg = args.begin(),
iearg = args.end(); iarg != iearg; iarg++)
cout << " " << *iarg;
cout << endl;
}
int status = execute(host_compiler, args, "", NULL, NULL);
if (status) return status;
}
//
// 2) Emit LLVM IR.
//
string out = "";
{
list<string> emit_ir_args;
for (list<string>::iterator iarg = args.begin(),
iearg = args.end(); iarg != iearg; iarg++)
{
const char* arg = (*iarg).c_str();
if (!strcmp(arg, "-c") || !strcmp(arg, "-o"))
{
iarg++;
continue;
}
if (!strcmp(arg, "-g"))
{
continue;
}
emit_ir_args.push_back(*iarg);
}
emit_ir_args.push_back("-fplugin=/opt/kernelgen/lib/dragonegg.so");
emit_ir_args.push_back("-fplugin-arg-dragonegg-emit-ir");
emit_ir_args.push_back("-S");
emit_ir_args.push_back(input);
emit_ir_args.push_back("-o");
emit_ir_args.push_back("-");
if (verbose)
{
cout << llvm_compiler;
for (list<string>::iterator iarg = emit_ir_args.begin(),
iearg = emit_ir_args.end(); iarg != iearg; iarg++)
cout << " " << *iarg;
cout << endl;
}
int status = execute(llvm_compiler, emit_ir_args, "", &out, NULL);
if (status) return status;
}
//
// 3) Record existing module functions.
//
LLVMContext &context = getGlobalContext();
SMDiagnostic diag;
MemoryBuffer* buffer1 = MemoryBuffer::getMemBuffer(out);
auto_ptr<Module> m1;
m1.reset(ParseIR(buffer1, diag, context));
//m1.get()->dump();
//
// 4) Inline calls and extract loops into new functions.
//
MemoryBuffer* buffer2 = MemoryBuffer::getMemBuffer(out);
auto_ptr<Module> m2;
m2.reset(ParseIR(buffer2, diag, context));
{
PassManager manager;
manager.add(createInstructionCombiningPass());
manager.run(*m2.get());
}
std::vector<CallInst *> LoopFuctionCalls;
{
PassManager manager;
manager.add(createBranchedLoopExtractorPass(LoopFuctionCalls));
manager.run(*m2.get());
}
//m2.get()->dump();
//
// 5) Replace call to loop functions with call to launcher.
// Append "always inline" attribute to all other functions.
//
Type* int32Ty = Type::getInt32Ty(context);
Function* launch = Function::Create(
TypeBuilder<types::i<32>(types::i<8>*, types::i<64>, types::i<32>*), true>::get(context),
GlobalValue::ExternalLinkage, "kernelgen_launch", m2.get());
for (Module::iterator f1 = m2.get()->begin(), fe1 = m2.get()->end(); f1 != fe1; f1++)
{
Function* func = f1;
if (func->isDeclaration()) continue;
// Search for the current function in original module
// functions list.
// If function is not in list of original module, then
// it is generated by the loop extractor.
// Append "always inline" attribute to all other functions.
if (m1.get()->getFunction(func->getName()))
{
const AttrListPtr attr = func->getAttributes();
const AttrListPtr attr_new = attr.addAttr(~0U, Attribute::AlwaysInline);
func->setAttributes(attr_new);
continue;
}
// Each such function must be extracted to the
// standalone module and packed into resulting
// object file data section.
if (verbose)
cout << "Preparing loop function " << func->getName().data() <<
" ..." << endl;
// Reset to default visibility.
func->setVisibility(GlobalValue::DefaultVisibility);
// Reset to default linkage.
func->setLinkage(GlobalValue::ExternalLinkage);
// Replace call to this function in module with call to launcher.
bool found = false;
for (Module::iterator f2 = m2->begin(), fe2 = m2->end(); (f2 != fe2) && !found; f2++)
for (Function::iterator bb = f2->begin(); (bb != f2->end()) && !found; bb++)
for (BasicBlock::iterator i = bb->begin(); i != bb->end(); i++)
{
// Check if instruction in focus is a call.
CallInst* call = dyn_cast<CallInst>(cast<Value>(i));
if (!call) continue;
// Check if function is called (needs -instcombine pass).
Function* callee = call->getCalledFunction();
if (!callee) continue;
if (callee->isDeclaration()) continue;
if (callee->getName() != func->getName()) continue;
// Create a constant array holding original called
// function name.
Constant* name = ConstantArray::get(
context, callee->getName(), true);
// Create and initialize the memory buffer for name.
ArrayType* nameTy = cast<ArrayType>(name->getType());
AllocaInst* nameAlloc = new AllocaInst(nameTy, "", call);
StoreInst* nameInit = new StoreInst(name, nameAlloc, "", call);
Value* Idx[2];
Idx[0] = Constant::getNullValue(Type::getInt32Ty(context));
Idx[1] = ConstantInt::get(Type::getInt32Ty(context), 0);
GetElementPtrInst* namePtr = GetElementPtrInst::Create(nameAlloc, Idx, "", call);
// Add pointer to the original function string name.
SmallVector<Value*, 16> call_args;
call_args.push_back(namePtr);
// Add size of the aggregated arguments structure.
{
BitCastInst* BC = new BitCastInst(
call->getArgOperand(0), Type::getInt64PtrTy(context),
"", call);
LoadInst* LI = new LoadInst(BC, "", call);
call_args.push_back(LI);
}
// Add original aggregated structure argument.
call_args.push_back(call->getArgOperand(0));
// Create new function call with new call arguments
// and copy old call properties.
CallInst* newcall = CallInst::Create(launch, call_args, "", call);
//newcall->takeName(call);
newcall->setCallingConv(call->getCallingConv());
newcall->setAttributes(call->getAttributes());
newcall->setDebugLoc(call->getDebugLoc());
// Replace old call with new one.
call->replaceAllUsesWith(newcall);
call->eraseFromParent();
found = true;
break;
}
}
//m2.get()->dump();
//
// 6) Apply optimization passes to the resulting common
// module.
//
{
PassManager manager;
manager.add(createLowerSetJmpPass());
PassManagerBuilder builder;
builder.Inliner = createFunctionInliningPass();
builder.OptLevel = 3;
builder.DisableSimplifyLibCalls = true;
builder.populateModulePassManager(manager);
manager.run(*m2.get());
}
//m2.get()->dump();
//
// 7) Embed the resulting module into object file.
//
{
string ir_string;
raw_string_ostream ir(ir_string);
ir << (*m2.get());
celf e(tmp_output.getFilename(), output);
e.getSection(".data")->addSymbol(
"__kernelgen_" + string(input),
ir_string.c_str(), ir_string.size() + 1);
}
return 0;
}
/// performLocalReleaseMotion - Scan backwards from the specified release,
/// moving it earlier in the function if possible, over instructions that do not
/// access the released object. If we get to a retain or allocation of the
/// object, zap both.
static bool performLocalReleaseMotion(CallInst &Release, BasicBlock &BB,
SwiftRCIdentity *RC) {
// FIXME: Call classifier should identify the object for us. Too bad C++
// doesn't have nice Swift-style enums.
Value *ReleasedObject = RC->getSwiftRCIdentityRoot(Release.getArgOperand(0));
BasicBlock::iterator BBI = Release.getIterator();
// Scan until we get to the top of the block.
while (BBI != BB.begin()) {
--BBI;
// Don't analyze PHI nodes. We can't move retains before them and they
// aren't "interesting".
if (isa<PHINode>(BBI) ||
// If we found the instruction that defines the value we're releasing,
// don't push the release past it.
&*BBI == Release.getArgOperand(0)) {
++BBI;
goto OutOfLoop;
}
switch (classifyInstruction(*BBI)) {
// These instructions should not reach here based on the pass ordering.
// i.e. LLVMARCOpt -> LLVMContractOpt.
case RT_UnknownRetainN:
case RT_BridgeRetainN:
case RT_RetainN:
case RT_UnknownReleaseN:
case RT_BridgeReleaseN:
case RT_ReleaseN:
llvm_unreachable("These are only created by LLVMARCContract !");
case RT_NoMemoryAccessed:
// Skip over random instructions that don't touch memory. They don't need
// protection by retain/release.
continue;
case RT_UnknownRelease:
case RT_BridgeRelease:
case RT_ObjCRelease:
case RT_Release: {
// If we get to a release, we can generally ignore it and scan past it.
// However, if we get to a release of obviously the same object, we stop
// scanning here because it should have already be moved as early as
// possible, so there is no reason to move its friend to the same place.
//
// NOTE: If this occurs frequently, maybe we can have a release(Obj, N)
// API to drop multiple retain counts at once.
CallInst &ThisRelease = cast<CallInst>(*BBI);
Value *ThisReleasedObject = ThisRelease.getArgOperand(0);
ThisReleasedObject = RC->getSwiftRCIdentityRoot(ThisReleasedObject);
if (ThisReleasedObject == ReleasedObject) {
//Release.dump(); ThisRelease.dump(); BB.getParent()->dump();
++BBI;
goto OutOfLoop;
}
continue;
}
case RT_UnknownRetain:
case RT_BridgeRetain:
case RT_ObjCRetain:
case RT_Retain: { // swift_retain(obj)
CallInst &Retain = cast<CallInst>(*BBI);
Value *RetainedObject = Retain.getArgOperand(0);
RetainedObject = RC->getSwiftRCIdentityRoot(RetainedObject);
// Since we canonicalized earlier, we know that if our retain has any
// uses, they were replaced already. This assertion documents this
// assumption.
assert(Retain.use_empty() && "Retain should have been canonicalized to "
"have no uses.");
// If the retain and release are to obviously pointer-equal objects, then
// we can delete both of them. We have proven that they do not protect
// anything of value.
if (RetainedObject == ReleasedObject) {
Retain.eraseFromParent();
Release.eraseFromParent();
++NumRetainReleasePairs;
return true;
}
// Otherwise, this is a retain of an object that is not statically known
// to be the same object. It may still be dynamically the same object
// though. In this case, we can't move the release past it.
// TODO: Strengthen analysis.
//Release.dump(); ThisRelease.dump(); BB.getParent()->dump();
++BBI;
goto OutOfLoop;
}
case RT_AllocObject: { // %obj = swift_alloc(...)
CallInst &Allocation = cast<CallInst>(*BBI);
// If this is an allocation of an unrelated object, just ignore it.
// TODO: This is not safe without proving the object being released is not
// related to the allocated object. Consider something silly like this:
// A = allocate()
// B = bitcast A to object
// release(B)
if (ReleasedObject != &Allocation) {
// Release.dump(); BB.getParent()->dump();
++BBI;
goto OutOfLoop;
}
// If this is a release right after an allocation of the object, then we
// can zap both.
Allocation.replaceAllUsesWith(UndefValue::get(Allocation.getType()));
Allocation.eraseFromParent();
Release.eraseFromParent();
++NumAllocateReleasePairs;
return true;
}
case RT_FixLifetime:
case RT_RetainUnowned:
case RT_CheckUnowned:
case RT_Unknown:
// Otherwise, we have reached something that we do not understand. Do not
// attempt to shorten the lifetime of this object beyond this point so we
// are conservative.
++BBI;
goto OutOfLoop;
}
}
OutOfLoop:
// If we got to the top of the block, (and if the instruction didn't start
// there) move the release to the top of the block.
// TODO: This is where we'd plug in some global algorithms someday.
if (&*BBI != &Release) {
Release.moveBefore(&*BBI);
return true;
}
return false;
}
bool GambasPass::runOnFunction(Function &F){
IRBuilder<> Builder(F.getContext());
bool changed = false;
for(Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
for(BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ){
ICmpInst* ICI = dyn_cast<ICmpInst>(I);
CallInst* CI = dyn_cast<CallInst>(I++);
if (ICI && ICI->hasMetadata() && ICI->getMetadata("unref_slt") && dyn_cast<LoadInst>(ICI->getOperand(0))){
ICI->replaceAllUsesWith(ConstantInt::get(ICI->getType(), false));
ICI->eraseFromParent();
changed = true;
continue;
}
if (!CI)
continue;
Function* callee = CI->getCalledFunction();
if (callee == NULL || !callee->isDeclaration())
continue;
StringRef name = callee->getName();
if (name == "JR_release_variant" || name == "JR_borrow_variant"){
ConstantInt* vtype_int = dyn_cast<ConstantInt>(CI->getArgOperand(0));
if (!vtype_int)
continue;
uint64_t vtype = vtype_int->getZExtValue();
if (TYPE_is_string(vtype) || TYPE_is_object(vtype))
continue;
CI->eraseFromParent();
changed = true;
} else if (name == FUNCTION_NAME(__finite)){
ConstantFP* op = dyn_cast<ConstantFP>(CI->getArgOperand(0));
if (!op)
continue;
int val = __finite(op->getValueAPF().convertToDouble());
Constant* res = ConstantInt::get(CI->getType(), val);
CI->replaceAllUsesWith(res);
CI->eraseFromParent();
changed = true;
} else if (name == FUNCTION_NAME(__isnan)){
ConstantFP* op = dyn_cast<ConstantFP>(CI->getArgOperand(0));
if (!op)
continue;
int val = __isnan(op->getValueAPF().convertToDouble());
Constant* res = ConstantInt::get(CI->getType(), val);
CI->replaceAllUsesWith(res);
CI->eraseFromParent();
changed = true;
} else if (name == FUNCTION_NAME(__isinf)){
ConstantFP* op = dyn_cast<ConstantFP>(CI->getArgOperand(0));
if (!op)
continue;
int val = __isinf(op->getValueAPF().convertToDouble());
Constant* res = ConstantInt::get(CI->getType(), val);
CI->replaceAllUsesWith(res);
CI->eraseFromParent();
changed = true;
}
}
}
return changed;
}
/// performLocalRetainMotion - Scan forward from the specified retain, moving it
/// later in the function if possible, over instructions that provably can't
/// release the object. If we get to a release of the object, zap both.
///
/// NOTE: this handles both objc_retain and swift_retain.
///
static bool performLocalRetainMotion(CallInst &Retain, BasicBlock &BB,
SwiftRCIdentity *RC) {
// FIXME: Call classifier should identify the object for us. Too bad C++
// doesn't have nice Swift-style enums.
Value *RetainedObject = RC->getSwiftRCIdentityRoot(Retain.getArgOperand(0));
BasicBlock::iterator BBI = Retain.getIterator(),
BBE = BB.getTerminator()->getIterator();
bool isObjCRetain = Retain.getCalledFunction()->getName() == "objc_retain";
bool MadeProgress = false;
// Scan until we get to the end of the block.
for (++BBI; BBI != BBE; ++BBI) {
Instruction &CurInst = *BBI;
// Classify the instruction. This switch does a "break" when the instruction
// can be skipped and is interesting, and a "continue" when it is a retain
// of the same pointer.
switch (classifyInstruction(CurInst)) {
// These instructions should not reach here based on the pass ordering.
// i.e. LLVMARCOpt -> LLVMContractOpt.
case RT_RetainN:
case RT_UnknownRetainN:
case RT_BridgeRetainN:
case RT_ReleaseN:
case RT_UnknownReleaseN:
case RT_BridgeReleaseN:
llvm_unreachable("These are only created by LLVMARCContract !");
case RT_NoMemoryAccessed:
case RT_AllocObject:
case RT_CheckUnowned:
// Skip over random instructions that don't touch memory. They don't need
// protection by retain/release.
break;
case RT_FixLifetime: // This only stops release motion. Retains can move over it.
break;
case RT_Retain:
case RT_UnknownRetain:
case RT_BridgeRetain:
case RT_RetainUnowned:
case RT_ObjCRetain: { // swift_retain(obj)
//CallInst &ThisRetain = cast<CallInst>(CurInst);
//Value *ThisRetainedObject = ThisRetain.getArgOperand(0);
// If we see a retain of the same object, we can skip over it, but we
// can't count it as progress. Just pushing a retain(x) past a retain(y)
// doesn't change the program.
continue;
}
case RT_UnknownRelease:
case RT_BridgeRelease:
case RT_ObjCRelease:
case RT_Release: {
// If we get to a release that is provably to this object, then we can zap
// it and the retain.
CallInst &ThisRelease = cast<CallInst>(CurInst);
Value *ThisReleasedObject = ThisRelease.getArgOperand(0);
ThisReleasedObject = RC->getSwiftRCIdentityRoot(ThisReleasedObject);
if (ThisReleasedObject == RetainedObject) {
Retain.eraseFromParent();
ThisRelease.eraseFromParent();
if (isObjCRetain) {
++NumObjCRetainReleasePairs;
} else {
++NumRetainReleasePairs;
}
return true;
}
// Otherwise, if this is some other pointer, we can only ignore it if we
// can prove that the two objects don't alias.
// Retain.dump(); ThisRelease.dump(); BB.getParent()->dump();
goto OutOfLoop;
}
case RT_Unknown:
// Loads cannot affect the retain.
if (isa<LoadInst>(CurInst))
continue;
// Load, store, memcpy etc can't do a release.
if (isa<LoadInst>(CurInst) || isa<StoreInst>(CurInst) ||
isa<MemIntrinsic>(CurInst))
break;
// CurInst->dump(); BBI->dump();
// Otherwise, we get to something unknown/unhandled. Bail out for now.
goto OutOfLoop;
}
// If the switch did a break, we made some progress moving this retain.
MadeProgress = true;
}
OutOfLoop:
// If we were able to move the retain down, move it now.
// TODO: This is where we'd plug in some global algorithms someday.
if (MadeProgress) {
Retain.moveBefore(&*BBI);
return true;
}
return false;
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass.
// Clone functions that take GEPs as arguments
//
// Inputs:
// M - A reference to the LLVM module to transform
//
// Outputs:
// M - The transformed LLVM module.
//
// Return value:
// true - The module was modified.
// false - The module was not modified.
//
bool GEPExprArgs::runOnModule(Module& M) {
bool changed;
do {
changed = false;
for (Module::iterator F = M.begin(); F != M.end(); ++F){
for (Function::iterator B = F->begin(), FE = F->end(); B != FE; ++B) {
for (BasicBlock::iterator I = B->begin(), BE = B->end(); I != BE;) {
CallInst *CI = dyn_cast<CallInst>(I++);
if(!CI)
continue;
if(CI->hasByValArgument())
continue;
// if the GEP calls a function, that is externally defined,
// or might be changed, ignore this call site.
Function *F = CI->getCalledFunction();
if (!F || (F->isDeclaration() || F->mayBeOverridden()))
continue;
if(F->hasStructRetAttr())
continue;
if(F->isVarArg())
continue;
// find the argument we must replace
Function::arg_iterator ai = F->arg_begin(), ae = F->arg_end();
unsigned argNum = 1;
for(; argNum < CI->getNumOperands();argNum++, ++ai) {
if(ai->use_empty())
continue;
if (isa<GEPOperator>(CI->getOperand(argNum)))
break;
}
// if no argument was a GEP operator to be changed
if(ai == ae)
continue;
GEPOperator *GEP = dyn_cast<GEPOperator>(CI->getOperand(argNum));
if(!GEP->hasAllConstantIndices())
continue;
// Construct the new Type
// Appends the struct Type at the beginning
std::vector<Type*>TP;
TP.push_back(GEP->getPointerOperand()->getType());
for(unsigned c = 1; c < CI->getNumOperands();c++) {
TP.push_back(CI->getOperand(c)->getType());
}
//return type is same as that of original instruction
FunctionType *NewFTy = FunctionType::get(CI->getType(), TP, false);
Function *NewF;
numSimplified++;
if(numSimplified > 800)
return true;
NewF = Function::Create(NewFTy,
GlobalValue::InternalLinkage,
F->getName().str() + ".TEST",
&M);
Function::arg_iterator NI = NewF->arg_begin();
NI->setName("GEParg");
++NI;
ValueToValueMapTy ValueMap;
for (Function::arg_iterator II = F->arg_begin(); NI != NewF->arg_end(); ++II, ++NI) {
ValueMap[II] = NI;
NI->setName(II->getName());
NI->addAttr(F->getAttributes().getParamAttributes(II->getArgNo() + 1));
}
NewF->setAttributes(NewF->getAttributes().addAttr(
0, F->getAttributes().getRetAttributes()));
// Perform the cloning.
SmallVector<ReturnInst*,100> Returns;
CloneFunctionInto(NewF, F, ValueMap, false, Returns);
std::vector<Value*> fargs;
for(Function::arg_iterator ai = NewF->arg_begin(),
ae= NewF->arg_end(); ai != ae; ++ai) {
fargs.push_back(ai);
}
NewF->setAttributes(NewF->getAttributes().addAttr(
~0, F->getAttributes().getFnAttributes()));
//Get the point to insert the GEP instr.
SmallVector<Value*, 8> Ops(CI->op_begin()+1, CI->op_end());
Instruction *InsertPoint;
for (BasicBlock::iterator insrt = NewF->front().begin();
isa<AllocaInst>(InsertPoint = insrt); ++insrt) {;}
NI = NewF->arg_begin();
SmallVector<Value*, 8> Indices;
Indices.append(GEP->op_begin()+1, GEP->op_end());
GetElementPtrInst *GEP_new = GetElementPtrInst::Create(cast<Value>(NI),
Indices,
"", InsertPoint);
fargs.at(argNum)->replaceAllUsesWith(GEP_new);
unsigned j = argNum + 1;
for(; j < CI->getNumOperands();j++) {
if(CI->getOperand(j) == GEP)
fargs.at(j)->replaceAllUsesWith(GEP_new);
}
SmallVector<AttributeWithIndex, 8> AttributesVec;
// Get the initial attributes of the call
AttrListPtr CallPAL = CI->getAttributes();
Attributes RAttrs = CallPAL.getRetAttributes();
Attributes FnAttrs = CallPAL.getFnAttributes();
if (RAttrs)
AttributesVec.push_back(AttributeWithIndex::get(0, RAttrs));
SmallVector<Value*, 8> Args;
Args.push_back(GEP->getPointerOperand());
for(unsigned j =1;j<CI->getNumOperands();j++) {
Args.push_back(CI->getOperand(j));
// position in the AttributesVec
if (Attributes Attrs = CallPAL.getParamAttributes(j))
AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
}
// Create the new attributes vec.
if (FnAttrs != Attribute::None)
AttributesVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
AttrListPtr NewCallPAL = AttrListPtr::get(AttributesVec.begin(),
AttributesVec.end());
CallInst *CallI = CallInst::Create(NewF,Args,"", CI);
CallI->setCallingConv(CI->getCallingConv());
CallI->setAttributes(NewCallPAL);
CI->replaceAllUsesWith(CallI);
CI->eraseFromParent();
changed = true;
}
}
}
} while(changed);
return true;
}
bool LowerExcHandlers::runOnFunction(Function &F) {
if (!except_enter_func)
return false; // No EH frames in this module
/* Step 1: EH Depth Numbering */
std::map<llvm::CallInst *, int> EnterDepth;
std::map<llvm::CallInst *, int> LeaveDepth;
std::map<BasicBlock *, int> ExitDepth;
int MaxDepth = 0;
// Compute EH Depth at each basic block using a DFS traversal.
for (df_iterator<BasicBlock *> I = df_begin(&F.getEntryBlock()),
E = df_end(&F.getEntryBlock()); I != E; ++I) {
auto *BB = *I;
int Depth = 0;
/* Here we use the assumption that all incoming edges have the same
* EH depth.
*/
for (auto *Pred : predecessors(BB)) {
auto it = ExitDepth.find(Pred);
if (it != ExitDepth.end()) {
Depth = it->second;
break;
}
}
/* Compute the depth within the basic block */
for (auto &I : *BB) {
auto *CI = dyn_cast<CallInst>(&I);
if (!CI)
continue;
Function *Callee = CI->getCalledFunction();
if (!Callee)
continue;
if (Callee == except_enter_func)
EnterDepth[CI] = Depth++;
else if (Callee == leave_func) {
LeaveDepth[CI] = Depth;
Depth -= cast<ConstantInt>(CI->getArgOperand(0))->getLimitedValue();
}
assert(Depth >= 0);
if (Depth > MaxDepth)
MaxDepth = Depth;
}
/* Remember the depth at the BB boundary */
ExitDepth[BB] = Depth;
}
/* Step 2: EH Frame lowering */
// Allocate stack space for each handler. We allocate these as separate
// allocas so the optimizer can later merge and reaarange them if it wants
// to.
Value *handler_sz = ConstantInt::get(Type::getInt32Ty(F.getContext()),
sizeof(jl_handler_t));
Value *handler_sz64 = ConstantInt::get(Type::getInt64Ty(F.getContext()),
sizeof(jl_handler_t));
Instruction *firstInst = &F.getEntryBlock().front();
std::vector<AllocaInst *> buffs;
for (int i = 0; i < MaxDepth; ++i) {
auto *buff = new AllocaInst(Type::getInt8Ty(F.getContext()),
0,
handler_sz, "", firstInst);
buff->setAlignment(16);
buffs.push_back(buff);
}
// Lower enter funcs
for (auto it : EnterDepth) {
assert(it.second >= 0);
AllocaInst *buff = buffs[it.second];
CallInst *enter = it.first;
auto new_enter = CallInst::Create(jlenter_func, buff, "", enter);
Value *lifetime_args[] = {
handler_sz64,
buff
};
CallInst::Create(lifetime_start, lifetime_args, "", new_enter);
#ifndef _OS_WINDOWS_
// For LLVM 3.3 compatibility
Value *args[] = {buff,
ConstantInt::get(Type::getInt32Ty(F.getContext()), 0)};
auto sj = CallInst::Create(setjmp_func, args, "", enter);
#else
auto sj = CallInst::Create(setjmp_func, buff, "", enter);
#endif
// We need to mark this on the call site as well. See issue #6757
sj->setCanReturnTwice();
if (auto dbg = enter->getMetadata(LLVMContext::MD_dbg)) {
new_enter->setMetadata(LLVMContext::MD_dbg, dbg);
sj->setMetadata(LLVMContext::MD_dbg, dbg);
}
enter->replaceAllUsesWith(sj);
enter->eraseFromParent();
}
// Insert lifetime end intrinsics after every leave.
for (auto it : LeaveDepth) {
int StartDepth = it.second - 1;
int npops = cast<ConstantInt>(it.first->getArgOperand(0))->getLimitedValue();
for (int i = 0; i < npops; ++i) {
assert(StartDepth-i >= 0);
Value *lifetime_args[] = {
handler_sz64,
buffs[StartDepth-i]
};
auto LifetimeEnd = CallInst::Create(lifetime_end, lifetime_args);
LifetimeEnd->insertAfter(it.first);
}
}
return true;
}
int CGGraph::solve(std::vector<CallInst*> boundsChecks) {
std::vector<CallInst*>::iterator callIt;
std::string varOne;
std::string varTwo;
bool rCheck;
vector<CallInst*> toRemove;
for (callIt = boundsChecks.begin(); callIt != boundsChecks.end(); ++callIt) {
if(std::find(owner->callInstsRemoved.begin(), owner->callInstsRemoved.end(), (*callIt)) != owner->callInstsRemoved.end())
{
continue;
}
else
{
if ((*callIt)->getParent()->getParent()->getName() == this->funcName) {
Value *v0 = (*callIt)->getArgOperand(0);
Value *v1 = (*callIt)->getArgOperand(1);
if (ConstantInt *c0 = dyn_cast<ConstantInt>(v0)) {
if (ConstantInt *c1 = dyn_cast<ConstantInt>(v1)) {
totalChecked++;
if ((c0->getSExtValue() >= 0) && (c0->getSExtValue() < c1->getSExtValue())) {
//trivial constant case
totalRemoved++;
toRemove.push_back(*callIt);
}
}
}
else {
varOne = getNameFromValue((*callIt)->getArgOperand(1), (*callIt)); //length of array
varTwo = getNameFromValue((*callIt)->getArgOperand(0), (*callIt)); //indexing variable
if (hasNode(varOne) && hasNode(varTwo)) {
rCheck = doTraverse(varOne, varTwo);
totalChecked++;
if (rCheck) {
totalRemoved++;
toRemove.push_back(*callIt);
}
}
}
}
}
}
// Remove them outside the traversal loop so you don't screw the iterator.
for(size_t i = 0; i < toRemove.size(); i++)
{
CallInst *deadCall = toRemove[i];
deadCall->eraseFromParent();
owner->callInstsRemoved.push_back(deadCall);
}
//errs() << "Total number of checks performed for function: " << totalChecked << "\n";
//errs() << "Total number of checks removed for function: " << totalRemoved << "\n";
return totalRemoved;
}
void insertCallToAccessFunction(Function *F, Function *cF) {
CallInst *I;
Instruction *bI;
std::vector<Value *> Args;
std::vector<Type *> ArgsTy;
Module *M = F->getParent();
std::string name;
Function *nF, *tF;
FunctionType *FTy;
std::stringstream out;
Value::user_iterator i = F->user_begin(), e = F->user_end();
while (i != e) {
Args.clear();
ArgsTy.clear();
/************* C codes ***********/
if (isa<CallInst>(*i)) {
I = dyn_cast<CallInst>(*i);
// call to the access function F
Args.push_back(I->getArgOperand(0));
ArgsTy.push_back(I->getArgOperand(0)->getType());
// call to the execute function cF
Args.push_back(cF);
ArgsTy.push_back(PointerType::get(cF->getFunctionType(), 0));
unsigned int t;
for (t = 1; t < I->getNumArgOperands(); t++) {
Args.push_back(I->getArgOperand(t));
ArgsTy.push_back(I->getArgOperand(t)->getType());
// errs() << *(I->getArgOperand(t)) << " is or not " <<
// isa<GlobalVariable>(I->getArgOperand(t)) << "\n";
}
tF = dyn_cast<Function>(I->getCalledFunction());
FTy = FunctionType::get(tF->getReturnType(), ArgsTy, 0);
out.str(std::string());
out << "task_DAE_" << I->getNumArgOperands() - 1;
nF = (Function *)M->getOrInsertFunction(out.str(), FTy);
CallInst *ci = CallInst::Create(nF, Args, I->getName(), I);
i++;
I->replaceAllUsesWith(ci);
I->eraseFromParent();
}
/************* C++ codes ***********/
else {
Value::user_iterator bit = (*i)->user_begin(), bite = (*i)->user_end();
Type *iTy = (*i)->getType();
i++;
while (bit != bite) {
Args.clear();
ArgsTy.clear();
I = dyn_cast<CallInst>(*bit);
bit++;
// call to the access function F
Args.push_back(I->getArgOperand(0));
ArgsTy.push_back(I->getArgOperand(0)->getType());
// call to the execute function cF
bI = new BitCastInst(cF, (iTy), "_TPR", I);
Args.push_back(bI);
ArgsTy.push_back(bI->getType());
unsigned int t;
for (t = 1; t < I->getNumArgOperands(); t++) {
Args.push_back(I->getArgOperand(t));
ArgsTy.push_back(I->getArgOperand(t)->getType());
}
tF = dyn_cast<Function>(I->getCalledFunction());
FTy = FunctionType::get(tF->getReturnType(), ArgsTy, 0);
out.str(std::string());
out << "task_DAE_" << I->getNumArgOperands() - 1;
nF = (Function *)M->getOrInsertFunction(out.str(), FTy);
CallInst *ci = CallInst::Create(nF, Args, I->getName(), I);
I->replaceAllUsesWith(ci);
I->eraseFromParent();
}
}
}
}