// 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;
}
/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
/// an invoke, we have to turn all of the calls that can throw into
/// invokes. This function analyze BB to see if there are any calls, and if so,
/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
/// nodes in that block with the values specified in InvokeDestPHIValues.
///
/// Returns true to indicate that the next block should be skipped.
static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
InvokeInliningInfo &Invoke) {
LandingPadInst *LPI = Invoke.getLandingPadInst();
for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
Instruction *I = BBI++;
if (LandingPadInst *L = dyn_cast<LandingPadInst>(I)) {
unsigned NumClauses = LPI->getNumClauses();
L->reserveClauses(NumClauses);
for (unsigned i = 0; i != NumClauses; ++i)
L->addClause(LPI->getClause(i));
}
// We only need to check for function calls: inlined invoke
// instructions require no special handling.
CallInst *CI = dyn_cast<CallInst>(I);
// If this call cannot unwind, don't convert it to an invoke.
// Inline asm calls cannot throw.
if (!CI || CI->doesNotThrow() || isa<InlineAsm>(CI->getCalledValue()))
continue;
// Convert this function call into an invoke instruction. First, split the
// basic block.
BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
// Delete the unconditional branch inserted by splitBasicBlock
BB->getInstList().pop_back();
// Create the new invoke instruction.
ImmutableCallSite CS(CI);
SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split,
Invoke.getOuterResumeDest(),
InvokeArgs, CI->getName(), BB);
II->setCallingConv(CI->getCallingConv());
II->setAttributes(CI->getAttributes());
// Make sure that anything using the call now uses the invoke! This also
// updates the CallGraph if present, because it uses a WeakVH.
CI->replaceAllUsesWith(II);
// Delete the original call
Split->getInstList().pop_front();
// Update any PHI nodes in the exceptional block to indicate that there is
// now a new entry in them.
Invoke.addIncomingPHIValuesFor(BB);
return false;
}
return false;
}
/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
/// an invoke, we have to turn all of the calls that can throw into
/// invokes. This function analyze BB to see if there are any calls, and if so,
/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
/// nodes in that block with the values specified in InvokeDestPHIValues.
///
static void HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
BasicBlock *InvokeDest,
const SmallVectorImpl<Value*> &InvokeDestPHIValues) {
for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
Instruction *I = BBI++;
// We only need to check for function calls: inlined invoke
// instructions require no special handling.
CallInst *CI = dyn_cast<CallInst>(I);
if (CI == 0) continue;
// If this call cannot unwind, don't convert it to an invoke.
if (CI->doesNotThrow())
continue;
// Convert this function call into an invoke instruction.
// First, split the basic block.
BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
// Next, create the new invoke instruction, inserting it at the end
// of the old basic block.
ImmutableCallSite CS(CI);
SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
InvokeInst *II =
InvokeInst::Create(CI->getCalledValue(), Split, InvokeDest,
InvokeArgs.begin(), InvokeArgs.end(),
CI->getName(), BB->getTerminator());
II->setCallingConv(CI->getCallingConv());
II->setAttributes(CI->getAttributes());
// Make sure that anything using the call now uses the invoke! This also
// updates the CallGraph if present, because it uses a WeakVH.
CI->replaceAllUsesWith(II);
// Delete the unconditional branch inserted by splitBasicBlock
BB->getInstList().pop_back();
Split->getInstList().pop_front(); // Delete the original call
// Update any PHI nodes in the exceptional block to indicate that
// there is now a new entry in them.
unsigned i = 0;
for (BasicBlock::iterator I = InvokeDest->begin();
isa<PHINode>(I); ++I, ++i)
cast<PHINode>(I)->addIncoming(InvokeDestPHIValues[i], BB);
// This basic block is now complete, the caller will continue scanning the
// next one.
return;
}
}
/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
/// in the body of the inlined function into invokes and turn unwind
/// instructions into branches to the invoke unwind dest.
///
/// II is the invoke instruction being inlined. FirstNewBlock is the first
/// block of the inlined code (the last block is the end of the function),
/// and InlineCodeInfo is information about the code that got inlined.
static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
ClonedCodeInfo &InlinedCodeInfo) {
BasicBlock *InvokeDest = II->getUnwindDest();
std::vector<Value*> InvokeDestPHIValues;
// If there are PHI nodes in the unwind destination block, we need to
// keep track of which values came into them from this invoke, then remove
// the entry for this block.
BasicBlock *InvokeBlock = II->getParent();
for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) {
PHINode *PN = cast<PHINode>(I);
// Save the value to use for this edge.
InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock));
}
Function *Caller = FirstNewBlock->getParent();
// The inlined code is currently at the end of the function, scan from the
// start of the inlined code to its end, checking for stuff we need to
// rewrite.
if (InlinedCodeInfo.ContainsCalls || InlinedCodeInfo.ContainsUnwinds) {
for (Function::iterator BB = FirstNewBlock, E = Caller->end();
BB != E; ++BB) {
if (InlinedCodeInfo.ContainsCalls) {
for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ){
Instruction *I = BBI++;
// We only need to check for function calls: inlined invoke
// instructions require no special handling.
if (!isa<CallInst>(I)) continue;
CallInst *CI = cast<CallInst>(I);
// If this call cannot unwind, don't convert it to an invoke.
if (CI->doesNotThrow())
continue;
// Convert this function call into an invoke instruction.
// First, split the basic block.
BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
// Next, create the new invoke instruction, inserting it at the end
// of the old basic block.
SmallVector<Value*, 8> InvokeArgs(CI->op_begin()+1, CI->op_end());
InvokeInst *II =
InvokeInst::Create(CI->getCalledValue(), Split, InvokeDest,
InvokeArgs.begin(), InvokeArgs.end(),
CI->getName(), BB->getTerminator());
II->setCallingConv(CI->getCallingConv());
II->setAttributes(CI->getAttributes());
// Make sure that anything using the call now uses the invoke!
CI->replaceAllUsesWith(II);
// Delete the unconditional branch inserted by splitBasicBlock
BB->getInstList().pop_back();
Split->getInstList().pop_front(); // Delete the original call
// Update any PHI nodes in the exceptional block to indicate that
// there is now a new entry in them.
unsigned i = 0;
for (BasicBlock::iterator I = InvokeDest->begin();
isa<PHINode>(I); ++I, ++i) {
PHINode *PN = cast<PHINode>(I);
PN->addIncoming(InvokeDestPHIValues[i], BB);
}
// This basic block is now complete, start scanning the next one.
break;
}
}
if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
// An UnwindInst requires special handling when it gets inlined into an
// invoke site. Once this happens, we know that the unwind would cause
// a control transfer to the invoke exception destination, so we can
// transform it into a direct branch to the exception destination.
BranchInst::Create(InvokeDest, UI);
// Delete the unwind instruction!
UI->eraseFromParent();
// Update any PHI nodes in the exceptional block to indicate that
// there is now a new entry in them.
unsigned i = 0;
for (BasicBlock::iterator I = InvokeDest->begin();
isa<PHINode>(I); ++I, ++i) {
PHINode *PN = cast<PHINode>(I);
PN->addIncoming(InvokeDestPHIValues[i], BB);
}
}
}
}
// Now that everything is happy, we have one final detail. The PHI nodes in
// the exception destination block still have entries due to the original
// invoke instruction. Eliminate these entries (which might even delete the
// PHI node) now.
InvokeDest->removePredecessor(II->getParent());
}
//
// 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;
}
/// Replaces the given call site (Call or Invoke) with a gc.statepoint
/// intrinsic with an empty deoptimization arguments list. This does
/// NOT do explicit relocation for GC support.
static Value *ReplaceWithStatepoint(const CallSite &CS, /* to replace */
Pass *P) {
BasicBlock *BB = CS.getInstruction()->getParent();
Function *F = BB->getParent();
Module *M = F->getParent();
assert(M && "must be set");
// TODO: technically, a pass is not allowed to get functions from within a
// function pass since it might trigger a new function addition. Refactor
// this logic out to the initialization of the pass. Doesn't appear to
// matter in practice.
// Fill in the one generic type'd argument (the function is also vararg)
std::vector<Type *> argTypes;
argTypes.push_back(CS.getCalledValue()->getType());
Function *gc_statepoint_decl = Intrinsic::getDeclaration(
M, Intrinsic::experimental_gc_statepoint, argTypes);
// Then go ahead and use the builder do actually do the inserts. We insert
// immediately before the previous instruction under the assumption that all
// arguments will be available here. We can't insert afterwards since we may
// be replacing a terminator.
Instruction *insertBefore = CS.getInstruction();
IRBuilder<> Builder(insertBefore);
// First, create the statepoint (with all live ptrs as arguments).
std::vector<llvm::Value *> args;
// target, #args, unused, args
Value *Target = CS.getCalledValue();
args.push_back(Target);
int callArgSize = CS.arg_size();
args.push_back(
ConstantInt::get(Type::getInt32Ty(M->getContext()), callArgSize));
// TODO: add a 'Needs GC-rewrite' later flag
args.push_back(ConstantInt::get(Type::getInt32Ty(M->getContext()), 0));
// Copy all the arguments of the original call
args.insert(args.end(), CS.arg_begin(), CS.arg_end());
// Create the statepoint given all the arguments
Instruction *token = nullptr;
AttributeSet return_attributes;
if (CS.isCall()) {
CallInst *toReplace = cast<CallInst>(CS.getInstruction());
CallInst *call =
Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
call->setTailCall(toReplace->isTailCall());
call->setCallingConv(toReplace->getCallingConv());
// Before we have to worry about GC semantics, all attributes are legal
AttributeSet new_attrs = toReplace->getAttributes();
// In case if we can handle this set of sttributes - set up function attrs
// directly on statepoint and return attrs later for gc_result intrinsic.
call->setAttributes(new_attrs.getFnAttributes());
return_attributes = new_attrs.getRetAttributes();
// TODO: handle param attributes
token = call;
// Put the following gc_result and gc_relocate calls immediately after the
// the old call (which we're about to delete)
BasicBlock::iterator next(toReplace);
assert(BB->end() != next && "not a terminator, must have next");
next++;
Instruction *IP = &*(next);
Builder.SetInsertPoint(IP);
Builder.SetCurrentDebugLocation(IP->getDebugLoc());
} else if (CS.isInvoke()) {
InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
// Insert the new invoke into the old block. We'll remove the old one in a
// moment at which point this will become the new terminator for the
// original block.
InvokeInst *invoke = InvokeInst::Create(
gc_statepoint_decl, toReplace->getNormalDest(),
toReplace->getUnwindDest(), args, "", toReplace->getParent());
invoke->setCallingConv(toReplace->getCallingConv());
// Currently we will fail on parameter attributes and on certain
// function attributes.
AttributeSet new_attrs = toReplace->getAttributes();
// In case if we can handle this set of sttributes - set up function attrs
// directly on statepoint and return attrs later for gc_result intrinsic.
invoke->setAttributes(new_attrs.getFnAttributes());
return_attributes = new_attrs.getRetAttributes();
token = invoke;
// We'll insert the gc.result into the normal block
BasicBlock *normalDest = normalizeBBForInvokeSafepoint(
toReplace->getNormalDest(), invoke->getParent());
Instruction *IP = &*(normalDest->getFirstInsertionPt());
Builder.SetInsertPoint(IP);
} else {
llvm_unreachable("unexpect type of CallSite");
}
assert(token);
// Handle the return value of the original call - update all uses to use a
// gc_result hanging off the statepoint node we just inserted
// Only add the gc_result iff there is actually a used result
if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) {
Instruction *gc_result = nullptr;
std::vector<Type *> types; // one per 'any' type
types.push_back(CS.getType()); // result type
auto get_gc_result_id = [&](Type &Ty) {
if (Ty.isIntegerTy()) {
return Intrinsic::experimental_gc_result_int;
} else if (Ty.isFloatingPointTy()) {
return Intrinsic::experimental_gc_result_float;
} else if (Ty.isPointerTy()) {
return Intrinsic::experimental_gc_result_ptr;
} else {
llvm_unreachable("non java type encountered");
}
};
Intrinsic::ID Id = get_gc_result_id(*CS.getType());
Value *gc_result_func = Intrinsic::getDeclaration(M, Id, types);
std::vector<Value *> args;
args.push_back(token);
gc_result = Builder.CreateCall(
gc_result_func, args,
CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : "");
cast<CallInst>(gc_result)->setAttributes(return_attributes);
return gc_result;
} else {
// No return value for the call.
return nullptr;
}
}
/// Replaces the given call site (Call or Invoke) with a gc.statepoint
/// intrinsic with an empty deoptimization arguments list. This does
/// NOT do explicit relocation for GC support.
static Value *ReplaceWithStatepoint(const CallSite &CS, /* to replace */
Pass *P) {
assert(CS.getInstruction()->getParent()->getParent()->getParent() &&
"must be set");
// TODO: technically, a pass is not allowed to get functions from within a
// function pass since it might trigger a new function addition. Refactor
// this logic out to the initialization of the pass. Doesn't appear to
// matter in practice.
// Then go ahead and use the builder do actually do the inserts. We insert
// immediately before the previous instruction under the assumption that all
// arguments will be available here. We can't insert afterwards since we may
// be replacing a terminator.
IRBuilder<> Builder(CS.getInstruction());
// Note: The gc args are not filled in at this time, that's handled by
// RewriteStatepointsForGC (which is currently under review).
// Create the statepoint given all the arguments
Instruction *Token = nullptr;
AttributeSet OriginalAttrs;
if (CS.isCall()) {
CallInst *ToReplace = cast<CallInst>(CS.getInstruction());
CallInst *Call = Builder.CreateGCStatepointCall(
CS.getCalledValue(), makeArrayRef(CS.arg_begin(), CS.arg_end()), None,
None, "safepoint_token");
Call->setTailCall(ToReplace->isTailCall());
Call->setCallingConv(ToReplace->getCallingConv());
// Before we have to worry about GC semantics, all attributes are legal
// TODO: handle param attributes
OriginalAttrs = ToReplace->getAttributes();
// In case if we can handle this set of attributes - set up function
// attributes directly on statepoint and return attributes later for
// gc_result intrinsic.
Call->setAttributes(OriginalAttrs.getFnAttributes());
Token = Call;
// Put the following gc_result and gc_relocate calls immediately after the
// the old call (which we're about to delete).
assert(ToReplace->getNextNode() && "not a terminator, must have next");
Builder.SetInsertPoint(ToReplace->getNextNode());
Builder.SetCurrentDebugLocation(ToReplace->getNextNode()->getDebugLoc());
} else if (CS.isInvoke()) {
InvokeInst *ToReplace = cast<InvokeInst>(CS.getInstruction());
// Insert the new invoke into the old block. We'll remove the old one in a
// moment at which point this will become the new terminator for the
// original block.
Builder.SetInsertPoint(ToReplace->getParent());
InvokeInst *Invoke = Builder.CreateGCStatepointInvoke(
CS.getCalledValue(), ToReplace->getNormalDest(),
ToReplace->getUnwindDest(), makeArrayRef(CS.arg_begin(), CS.arg_end()),
Builder.getInt32(0), None, "safepoint_token");
// Currently we will fail on parameter attributes and on certain
// function attributes.
OriginalAttrs = ToReplace->getAttributes();
// In case if we can handle this set of attributes - set up function
// attributes directly on statepoint and return attributes later for
// gc_result intrinsic.
Invoke->setAttributes(OriginalAttrs.getFnAttributes());
Token = Invoke;
// We'll insert the gc.result into the normal block
BasicBlock *NormalDest = normalizeBBForInvokeSafepoint(
ToReplace->getNormalDest(), Invoke->getParent());
Builder.SetInsertPoint(NormalDest->getFirstInsertionPt());
} else {
llvm_unreachable("unexpect type of CallSite");
}
assert(Token);
// Handle the return value of the original call - update all uses to use a
// gc_result hanging off the statepoint node we just inserted
// Only add the gc_result iff there is actually a used result
if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) {
std::string TakenName =
CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : "";
CallInst *GCResult = Builder.CreateGCResult(Token, CS.getType(), TakenName);
GCResult->setAttributes(OriginalAttrs.getRetAttributes());
return GCResult;
} else {
// No return value for the call.
return nullptr;
}
}
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;
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass.
// Clone functions that take LoadInsts 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 LoadArgs::runOnModule(Module& M) {
std::map<std::pair<Function*, const Type * > , Function* > fnCache;
bool changed;
do {
changed = false;
for (Module::iterator Func = M.begin(); Func != M.end(); ++Func) {
for (Function::iterator B = Func->begin(), FE = Func->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 CallInst 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 = 0;
for(; argNum < CI->getNumArgOperands();argNum++, ++ai) {
// do not care about dead arguments
if(ai->use_empty())
continue;
if(F->getAttributes().getParamAttributes(argNum).hasAttrSomewhere(Attribute::SExt) ||
F->getAttributes().getParamAttributes(argNum).hasAttrSomewhere(Attribute::ZExt))
continue;
if (isa<LoadInst>(CI->getArgOperand(argNum)))
break;
}
// if no argument was a GEP operator to be changed
if(ai == ae)
continue;
LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(argNum));
Instruction * InsertPt = &(Func->getEntryBlock().front());
AllocaInst *NewVal = new AllocaInst(LI->getType(), "",InsertPt);
StoreInst *Copy = new StoreInst(LI, NewVal);
Copy->insertAfter(LI);
/*if(LI->getParent() != CI->getParent())
continue;
// Also check that there is no store after the load.
// TODO: Check if the load/store do not alias.
BasicBlock::iterator bii = LI->getParent()->begin();
Instruction *BII = bii;
while(BII != LI) {
++bii;
BII = bii;
}
while(BII != CI) {
if(isa<StoreInst>(BII))
break;
++bii;
BII = bii;
}
if(isa<StoreInst>(bii)){
continue;
}*/
// Construct the new Type
// Appends the struct Type at the beginning
std::vector<Type*>TP;
for(unsigned c = 0; c < CI->getNumArgOperands();c++) {
if(c == argNum)
TP.push_back(LI->getPointerOperand()->getType());
TP.push_back(CI->getArgOperand(c)->getType());
}
//return type is same as that of original instruction
FunctionType *NewFTy = FunctionType::get(CI->getType(), TP, false);
numSimplified++;
//if(numSimplified > 1000)
//return true;
Function *NewF;
std::map<std::pair<Function*, const Type* > , Function* >::iterator Test;
Test = fnCache.find(std::make_pair(F, NewFTy));
if(Test != fnCache.end()) {
NewF = Test->second;
} else {
NewF = Function::Create(NewFTy,
GlobalValue::InternalLinkage,
F->getName().str() + ".TEST",
&M);
fnCache[std::make_pair(F, NewFTy)] = NewF;
Function::arg_iterator NI = NewF->arg_begin();
ValueToValueMapTy ValueMap;
unsigned count = 0;
for (Function::arg_iterator II = F->arg_begin(); NI != NewF->arg_end(); ++count, ++NI) {
if(count == argNum) {
NI->setName("LDarg");
continue;
}
ValueMap[II] = NI;
NI->setName(II->getName());
NI->addAttr(F->getAttributes().getParamAttributes(II->getArgNo() + 1));
++II;
}
// 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().addAttributes(
F->getContext(), 0, F->getAttributes().getRetAttributes()));
NewF->setAttributes(NewF->getAttributes().addAttributes(
F->getContext(), ~0, F->getAttributes().getFnAttributes()));
//Get the point to insert the GEP instr.
Instruction *InsertPoint;
for (BasicBlock::iterator insrt = NewF->front().begin(); isa<AllocaInst>(InsertPoint = insrt); ++insrt) {;}
LoadInst *LI_new = new LoadInst(fargs.at(argNum), "", InsertPoint);
fargs.at(argNum+1)->replaceAllUsesWith(LI_new);
}
//this does not seem to be a good idea
AttributeSet NewCallPAL=AttributeSet();
// Get the initial attributes of the call
AttributeSet CallPAL = CI->getAttributes();
AttributeSet RAttrs = CallPAL.getRetAttributes();
AttributeSet FnAttrs = CallPAL.getFnAttributes();
if (!RAttrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),0, RAttrs);
SmallVector<Value*, 8> Args;
for(unsigned j =0;j<CI->getNumArgOperands();j++) {
if(j == argNum) {
Args.push_back(NewVal);
}
Args.push_back(CI->getArgOperand(j));
// position in the NewCallPAL
AttributeSet Attrs = CallPAL.getParamAttributes(j+1);
if (!Attrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),Args.size(), Attrs);
}
// Create the new attributes vec.
if (!FnAttrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),~0, FnAttrs);
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;
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass.
// Search for all call sites to casted functions.
// Check if they only differ in an argument type
// Cast the argument, and call the original function
//
// 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 ArgCast::runOnModule(Module& M) {
std::vector<CallInst*> worklist;
for (Module::iterator I = M.begin(); I != M.end(); ++I) {
if (I->mayBeOverridden())
continue;
// Find all uses of this function
for(Value::user_iterator ui = I->user_begin(), ue = I->user_end(); ui != ue; ) {
// check if is ever casted to a different function type
ConstantExpr *CE = dyn_cast<ConstantExpr>(*ui++);
if(!CE)
continue;
if (CE->getOpcode() != Instruction::BitCast)
continue;
if(CE->getOperand(0) != I)
continue;
const PointerType *PTy = dyn_cast<PointerType>(CE->getType());
if (!PTy)
continue;
const Type *ETy = PTy->getElementType();
const FunctionType *FTy = dyn_cast<FunctionType>(ETy);
if(!FTy)
continue;
// casting to a varargs funtion
// or function with same number of arguments
// possibly varying types of arguments
if(FTy->getNumParams() != I->arg_size() && !FTy->isVarArg())
continue;
for(Value::user_iterator uii = CE->user_begin(),
uee = CE->user_end(); uii != uee; ++uii) {
// Find all uses of the casted value, and check if it is
// used in a Call Instruction
if (CallInst* CI = dyn_cast<CallInst>(*uii)) {
// Check that it is the called value, and not an argument
if(CI->getCalledValue() != CE)
continue;
// Check that the number of arguments passed, and expected
// by the function are the same.
if(!I->isVarArg()) {
if(CI->getNumOperands() != I->arg_size() + 1)
continue;
} else {
if(CI->getNumOperands() < I->arg_size() + 1)
continue;
}
// If so, add to worklist
worklist.push_back(CI);
}
}
}
}
// Proces the worklist of potential call sites to transform
while(!worklist.empty()) {
CallInst *CI = worklist.back();
worklist.pop_back();
// Get the called Function
Function *F = cast<Function>(CI->getCalledValue()->stripPointerCasts());
const FunctionType *FTy = F->getFunctionType();
SmallVector<Value*, 8> Args;
unsigned i =0;
for(i =0; i< FTy->getNumParams(); ++i) {
Type *ArgType = CI->getOperand(i+1)->getType();
Type *FormalType = FTy->getParamType(i);
// If the types for this argument match, just add it to the
// parameter list. No cast needs to be inserted.
if(ArgType == FormalType) {
Args.push_back(CI->getOperand(i+1));
}
else if(ArgType->isPointerTy() && FormalType->isPointerTy()) {
CastInst *CastI = CastInst::CreatePointerCast(CI->getOperand(i+1),
FormalType, "", CI);
Args.push_back(CastI);
} else if (ArgType->isIntegerTy() && FormalType->isIntegerTy()) {
unsigned SrcBits = ArgType->getScalarSizeInBits();
unsigned DstBits = FormalType->getScalarSizeInBits();
if(SrcBits > DstBits) {
CastInst *CastI = CastInst::CreateIntegerCast(CI->getOperand(i+1),
FormalType, true, "", CI);
Args.push_back(CastI);
} else {
if (F->getAttributes().hasAttribute(i+1, Attribute::SExt)) {
CastInst *CastI = CastInst::CreateIntegerCast(CI->getOperand(i+1),
FormalType, true, "", CI);
Args.push_back(CastI);
} else if (F->getAttributes().hasAttribute(i+1, Attribute::ZExt)) {
CastInst *CastI = CastInst::CreateIntegerCast(CI->getOperand(i+1),
FormalType, false, "", CI);
Args.push_back(CastI);
} else {
// Use ZExt in default case.
// Derived from InstCombine. Also, the only reason this should happen
// is mismatched prototypes.
// Seen in case of integer constants which get interpreted as i32,
// even if being used as i64.
// TODO: is this correct?
CastInst *CastI = CastInst::CreateIntegerCast(CI->getOperand(i+1),
FormalType, false, "", CI);
Args.push_back(CastI);
}
}
} else {
DEBUG(ArgType->dump());
DEBUG(FormalType->dump());
break;
}
}
// If we found an argument we could not cast, try the next instruction
if(i != FTy->getNumParams()) {
continue;
}
if(FTy->isVarArg()) {
for(; i< CI->getNumOperands() - 1 ;i++) {
Args.push_back(CI->getOperand(i+1));
}
}
// else replace the call instruction
CallInst *CINew = CallInst::Create(F, Args, "", CI);
CINew->setCallingConv(CI->getCallingConv());
CINew->setAttributes(CI->getAttributes());
if(!CI->use_empty()) {
CastInst *RetCast;
if(CI->getType() != CINew->getType()) {
if(CI->getType()->isPointerTy() && CINew->getType()->isPointerTy())
RetCast = CastInst::CreatePointerCast(CINew, CI->getType(), "", CI);
else if(CI->getType()->isIntOrIntVectorTy() && CINew->getType()->isIntOrIntVectorTy())
RetCast = CastInst::CreateIntegerCast(CINew, CI->getType(), false, "", CI);
else if(CI->getType()->isIntOrIntVectorTy() && CINew->getType()->isPointerTy())
RetCast = CastInst::CreatePointerCast(CINew, CI->getType(), "", CI);
else if(CI->getType()->isPointerTy() && CINew->getType()->isIntOrIntVectorTy())
RetCast = new IntToPtrInst(CINew, CI->getType(), "", CI);
else {
// TODO: I'm not sure what right behavior is here, but this case should be handled.
llvm_unreachable("Unexpected type conversion in call!");
abort();
}
CI->replaceAllUsesWith(RetCast);
} else {
CI->replaceAllUsesWith(CINew);
}
}
// Debug printing
DEBUG(errs() << "ARGCAST:");
DEBUG(errs() << "ERASE:");
DEBUG(CI->dump());
DEBUG(errs() << "ARGCAST:");
DEBUG(errs() << "ADDED:");
DEBUG(CINew->dump());
CI->eraseFromParent();
numChanged++;
}
return true;
}
/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
/// an invoke, we have to turn all of the calls that can throw into
/// invokes. This function analyze BB to see if there are any calls, and if so,
/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
/// nodes in that block with the values specified in InvokeDestPHIValues.
///
/// Returns true to indicate that the next block should be skipped.
static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
InvokeInliningInfo &Invoke) {
for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
Instruction *I = BBI++;
// We only need to check for function calls: inlined invoke
// instructions require no special handling.
CallInst *CI = dyn_cast<CallInst>(I);
if (CI == 0) continue;
// LIBUNWIND: merge selector instructions.
if (EHSelectorInst *Inner = dyn_cast<EHSelectorInst>(CI)) {
EHSelectorInst *Outer = Invoke.getOuterSelector();
if (!Outer) continue;
bool innerIsOnlyCleanup = isCleanupOnlySelector(Inner);
bool outerIsOnlyCleanup = isCleanupOnlySelector(Outer);
// If both selectors contain only cleanups, we don't need to do
// anything. TODO: this is really just a very specific instance
// of a much more general optimization.
if (innerIsOnlyCleanup && outerIsOnlyCleanup) continue;
// Otherwise, we just append the outer selector to the inner selector.
SmallVector<Value*, 16> NewSelector;
for (unsigned i = 0, e = Inner->getNumArgOperands(); i != e; ++i)
NewSelector.push_back(Inner->getArgOperand(i));
for (unsigned i = 2, e = Outer->getNumArgOperands(); i != e; ++i)
NewSelector.push_back(Outer->getArgOperand(i));
CallInst *NewInner =
IRBuilder<>(Inner).CreateCall(Inner->getCalledValue(), NewSelector);
// No need to copy attributes, calling convention, etc.
NewInner->takeName(Inner);
Inner->replaceAllUsesWith(NewInner);
Inner->eraseFromParent();
continue;
}
// If this call cannot unwind, don't convert it to an invoke.
if (CI->doesNotThrow())
continue;
// Convert this function call into an invoke instruction.
// First, split the basic block.
BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
// Delete the unconditional branch inserted by splitBasicBlock
BB->getInstList().pop_back();
// LIBUNWIND: If this is a call to @llvm.eh.resume, just branch
// directly to the new landing pad.
if (Invoke.forwardEHResume(CI, BB)) {
// TODO: 'Split' is now unreachable; clean it up.
// We want to leave the original call intact so that the call
// graph and other structures won't get misled. We also have to
// avoid processing the next block, or we'll iterate here forever.
return true;
}
// Otherwise, create the new invoke instruction.
ImmutableCallSite CS(CI);
SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
InvokeInst *II =
InvokeInst::Create(CI->getCalledValue(), Split,
Invoke.getOuterUnwindDest(),
InvokeArgs, CI->getName(), BB);
II->setCallingConv(CI->getCallingConv());
II->setAttributes(CI->getAttributes());
// Make sure that anything using the call now uses the invoke! This also
// updates the CallGraph if present, because it uses a WeakVH.
CI->replaceAllUsesWith(II);
Split->getInstList().pop_front(); // Delete the original call
// Update any PHI nodes in the exceptional block to indicate that
// there is now a new entry in them.
Invoke.addIncomingPHIValuesFor(BB);
return false;
}
return false;
}
bool LowerEmSetjmp::runOnModule(Module &M) {
TheModule = &M;
Function *Setjmp = TheModule->getFunction("setjmp");
Function *Longjmp = TheModule->getFunction("longjmp");
if (!Setjmp && !Longjmp) return false;
Type *i32 = Type::getInt32Ty(M.getContext());
Type *Void = Type::getVoidTy(M.getContext());
// Add functions
Function *EmSetjmp = NULL;
if (Setjmp) {
SmallVector<Type*, 2> EmSetjmpTypes;
EmSetjmpTypes.push_back(Setjmp->getFunctionType()->getParamType(0));
EmSetjmpTypes.push_back(i32); // extra param that says which setjmp in the function it is
FunctionType *EmSetjmpFunc = FunctionType::get(i32, EmSetjmpTypes, false);
EmSetjmp = Function::Create(EmSetjmpFunc, GlobalValue::ExternalLinkage, "emscripten_setjmp", TheModule);
}
Function *EmLongjmp = Longjmp ? Function::Create(Longjmp->getFunctionType(), GlobalValue::ExternalLinkage, "emscripten_longjmp", TheModule) : NULL;
SmallVector<Type*, 1> IntArgTypes;
IntArgTypes.push_back(i32);
FunctionType *IntIntFunc = FunctionType::get(i32, IntArgTypes, false);
Function *CheckLongjmp = Function::Create(IntIntFunc, GlobalValue::ExternalLinkage, "emscripten_check_longjmp", TheModule); // gets control flow
Function *GetLongjmpResult = Function::Create(IntIntFunc, GlobalValue::ExternalLinkage, "emscripten_get_longjmp_result", TheModule); // gets int value longjmp'd
FunctionType *VoidFunc = FunctionType::get(Void, false);
Function *PrepSetjmp = Function::Create(VoidFunc, GlobalValue::ExternalLinkage, "emscripten_prep_setjmp", TheModule);
Function *CleanupSetjmp = Function::Create(VoidFunc, GlobalValue::ExternalLinkage, "emscripten_cleanup_setjmp", TheModule);
Function *PreInvoke = TheModule->getFunction("emscripten_preinvoke");
if (!PreInvoke) PreInvoke = Function::Create(VoidFunc, GlobalValue::ExternalLinkage, "emscripten_preinvoke", TheModule);
FunctionType *IntFunc = FunctionType::get(i32, false);
Function *PostInvoke = TheModule->getFunction("emscripten_postinvoke");
if (!PostInvoke) PostInvoke = Function::Create(IntFunc, GlobalValue::ExternalLinkage, "emscripten_postinvoke", TheModule);
// Process all callers of setjmp and longjmp. Start with setjmp.
typedef std::vector<PHINode*> Phis;
typedef std::map<Function*, Phis> FunctionPhisMap;
FunctionPhisMap SetjmpOutputPhis;
std::vector<Instruction*> ToErase;
if (Setjmp) {
for (Instruction::user_iterator UI = Setjmp->user_begin(), UE = Setjmp->user_end(); UI != UE; ++UI) {
User *U = *UI;
if (CallInst *CI = dyn_cast<CallInst>(U)) {
BasicBlock *SJBB = CI->getParent();
// The tail is everything right after the call, and will be reached once when setjmp is
// called, and later when longjmp returns to the setjmp
BasicBlock *Tail = SplitBlock(SJBB, CI->getNextNode());
// Add a phi to the tail, which will be the output of setjmp, which indicates if this is the
// first call or a longjmp back. The phi directly uses the right value based on where we
// arrive from
PHINode *SetjmpOutput = PHINode::Create(i32, 2, "", Tail->getFirstNonPHI());
SetjmpOutput->addIncoming(ConstantInt::get(i32, 0), SJBB); // setjmp initial call returns 0
CI->replaceAllUsesWith(SetjmpOutput); // The proper output is now this, not the setjmp call itself
// longjmp returns to the setjmp will add themselves to this phi
Phis& P = SetjmpOutputPhis[SJBB->getParent()];
P.push_back(SetjmpOutput);
// fix call target
SmallVector<Value *, 2> Args;
Args.push_back(CI->getArgOperand(0));
Args.push_back(ConstantInt::get(i32, P.size())); // our index in the function is our place in the array + 1
CallInst::Create(EmSetjmp, Args, "", CI);
ToErase.push_back(CI);
} else {
errs() << **UI << "\n";
report_fatal_error("bad use of setjmp, should only call it");
}
}
}
// Update longjmp FIXME: we could avoid throwing in longjmp as an optimization when longjmping back into the current function perhaps?
if (Longjmp) Longjmp->replaceAllUsesWith(EmLongjmp);
// Update all setjmping functions
for (FunctionPhisMap::iterator I = SetjmpOutputPhis.begin(); I != SetjmpOutputPhis.end(); I++) {
Function *F = I->first;
Phis& P = I->second;
CallInst::Create(PrepSetjmp, "", F->begin()->begin());
// Update each call that can longjmp so it can return to a setjmp where relevant
for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ) {
BasicBlock *BB = BBI++;
for (BasicBlock::iterator Iter = BB->begin(), E = BB->end(); Iter != E; ) {
Instruction *I = Iter++;
CallInst *CI;
if ((CI = dyn_cast<CallInst>(I))) {
Value *V = CI->getCalledValue();
if (V == PrepSetjmp || V == EmSetjmp || V == CheckLongjmp || V == GetLongjmpResult || V == PreInvoke || V == PostInvoke) continue;
if (Function *CF = dyn_cast<Function>(V)) if (CF->isIntrinsic()) continue;
// TODO: proper analysis of what can actually longjmp. Currently we assume anything but setjmp can.
// This may longjmp, so we need to check if it did. Split at that point, and
// envelop the call in pre/post invoke, if we need to
CallInst *After;
Instruction *Check = NULL;
if (Iter != E && (After = dyn_cast<CallInst>(Iter)) && After->getCalledValue() == PostInvoke) {
// use the pre|postinvoke that exceptions lowering already made
Check = Iter++;
}
BasicBlock *Tail = SplitBlock(BB, Iter); // Iter already points to the next instruction, as we need
TerminatorInst *TI = BB->getTerminator();
if (!Check) {
// no existing pre|postinvoke, create our own
CallInst::Create(PreInvoke, "", CI);
Check = CallInst::Create(PostInvoke, "", TI); // CI is at the end of the block
// If we are calling a function that is noreturn, we must remove that attribute. The code we
// insert here does expect it to return, after we catch the exception.
if (CI->doesNotReturn()) {
if (Function *F = dyn_cast<Function>(CI->getCalledValue())) {
F->removeFnAttr(Attribute::NoReturn);
}
CI->setAttributes(CI->getAttributes().removeAttribute(TheModule->getContext(), AttributeSet::FunctionIndex, Attribute::NoReturn));
assert(!CI->doesNotReturn());
}
}
// We need to replace the terminator in Tail - SplitBlock makes BB go straight to Tail, we need to check if a longjmp occurred, and
// go to the right setjmp-tail if so
SmallVector<Value *, 1> Args;
Args.push_back(Check);
Instruction *LongjmpCheck = CallInst::Create(CheckLongjmp, Args, "", BB);
Instruction *LongjmpResult = CallInst::Create(GetLongjmpResult, Args, "", BB);
SwitchInst *SI = SwitchInst::Create(LongjmpCheck, Tail, 2, BB);
// -1 means no longjmp happened, continue normally (will hit the default switch case). 0 means a longjmp that is not ours to handle, needs a rethrow. Otherwise
// the index mean is the same as the index in P+1 (to avoid 0).
for (unsigned i = 0; i < P.size(); i++) {
SI->addCase(cast<ConstantInt>(ConstantInt::get(i32, i+1)), P[i]->getParent());
P[i]->addIncoming(LongjmpResult, BB);
}
ToErase.push_back(TI); // new terminator is now the switch
// we are splitting the block here, and must continue to find other calls in the block - which is now split. so continue
// to traverse in the Tail
BB = Tail;
Iter = BB->begin();
E = BB->end();
} else if (InvokeInst *CI = dyn_cast<InvokeInst>(I)) { // XXX check if target is setjmp
(void)CI;
report_fatal_error("TODO: invoke inside setjmping functions");
}
}
}
// add a cleanup before each return
for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ) {
BasicBlock *BB = BBI++;
TerminatorInst *TI = BB->getTerminator();
if (isa<ReturnInst>(TI)) {
CallInst::Create(CleanupSetjmp, "", TI);
}
}
}
for (unsigned i = 0; i < ToErase.size(); i++) {
ToErase[i]->eraseFromParent();
}
// Finally, our modifications to the cfg can break dominance of SSA variables. For example,
// if (x()) { .. setjmp() .. }
// if (y()) { .. longjmp() .. }
// We must split the longjmp block, and it can jump into the setjmp one. But that means that when
// we split the setjmp block, it's first part no longer dominates its second part - there is
// a theoretically possible control flow path where x() is false, then y() is true and we
// reach the second part of the setjmp block, without ever reaching the first part. So,
// we recalculate regs vs. mem
for (FunctionPhisMap::iterator I = SetjmpOutputPhis.begin(); I != SetjmpOutputPhis.end(); I++) {
Function *F = I->first;
doRegToMem(*F);
doMemToReg(*F);
}
return true;
}
//
// Method: runOnModule()
//
// Description:
// Entry point for this LLVM pass.
// If a function returns a struct, make it return
// a pointer to the struct.
//
// 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 StructRet::runOnModule(Module& M) {
const llvm::DataLayout targetData(&M);
std::vector<Function*> worklist;
for (Module::iterator I = M.begin(); I != M.end(); ++I)
if (!I->mayBeOverridden()) {
if(I->hasAddressTaken())
continue;
if(I->getReturnType()->isStructTy()) {
worklist.push_back(I);
}
}
while(!worklist.empty()) {
Function *F = worklist.back();
worklist.pop_back();
Type *NewArgType = F->getReturnType()->getPointerTo();
// Construct the new Type
std::vector<Type*>TP;
TP.push_back(NewArgType);
for (Function::arg_iterator ii = F->arg_begin(), ee = F->arg_end();
ii != ee; ++ii) {
TP.push_back(ii->getType());
}
FunctionType *NFTy = FunctionType::get(F->getReturnType(), TP, F->isVarArg());
// Create the new function body and insert it into the module.
Function *NF = Function::Create(NFTy,
F->getLinkage(),
F->getName(), &M);
ValueToValueMapTy ValueMap;
Function::arg_iterator NI = NF->arg_begin();
NI->setName("ret");
++NI;
for (Function::arg_iterator II = F->arg_begin(); II != F->arg_end(); ++II, ++NI) {
ValueMap[II] = NI;
NI->setName(II->getName());
AttributeSet attrs = F->getAttributes().getParamAttributes(II->getArgNo() + 1);
if (!attrs.isEmpty())
NI->addAttr(attrs);
}
// Perform the cloning.
SmallVector<ReturnInst*,100> Returns;
if (!F->isDeclaration())
CloneFunctionInto(NF, F, ValueMap, false, Returns);
std::vector<Value*> fargs;
for(Function::arg_iterator ai = NF->arg_begin(),
ae= NF->arg_end(); ai != ae; ++ai) {
fargs.push_back(ai);
}
NF->setAttributes(NF->getAttributes().addAttributes(
M.getContext(), 0, F->getAttributes().getRetAttributes()));
NF->setAttributes(NF->getAttributes().addAttributes(
M.getContext(), ~0, F->getAttributes().getFnAttributes()));
for (Function::iterator B = NF->begin(), FE = NF->end(); B != FE; ++B) {
for (BasicBlock::iterator I = B->begin(), BE = B->end(); I != BE;) {
ReturnInst * RI = dyn_cast<ReturnInst>(I++);
if(!RI)
continue;
LoadInst *LI = dyn_cast<LoadInst>(RI->getOperand(0));
assert(LI && "Return should be preceded by a load instruction");
IRBuilder<> Builder(RI);
Builder.CreateMemCpy(fargs.at(0),
LI->getPointerOperand(),
targetData.getTypeStoreSize(LI->getType()),
targetData.getPrefTypeAlignment(LI->getType()));
}
}
for(Value::use_iterator ui = F->use_begin(), ue = F->use_end();
ui != ue; ) {
CallInst *CI = dyn_cast<CallInst>(*ui++);
if(!CI)
continue;
if(CI->getCalledFunction() != F)
continue;
if(CI->hasByValArgument())
continue;
AllocaInst *AllocaNew = new AllocaInst(F->getReturnType(), 0, "", CI);
SmallVector<Value*, 8> Args;
//this should probably be done in a different manner
AttributeSet NewCallPAL=AttributeSet();
// Get the initial attributes of the call
AttributeSet CallPAL = CI->getAttributes();
AttributeSet RAttrs = CallPAL.getRetAttributes();
AttributeSet FnAttrs = CallPAL.getFnAttributes();
if (!RAttrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),0, RAttrs);
Args.push_back(AllocaNew);
for(unsigned j = 0; j < CI->getNumOperands()-1; j++) {
Args.push_back(CI->getOperand(j));
// position in the NewCallPAL
AttributeSet Attrs = CallPAL.getParamAttributes(j);
if (!Attrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),Args.size(), Attrs);
}
// Create the new attributes vec.
if (!FnAttrs.isEmpty())
NewCallPAL=NewCallPAL.addAttributes(F->getContext(),~0, FnAttrs);
CallInst *CallI = CallInst::Create(NF, Args, "", CI);
CallI->setCallingConv(CI->getCallingConv());
CallI->setAttributes(NewCallPAL);
LoadInst *LI = new LoadInst(AllocaNew, "", CI);
CI->replaceAllUsesWith(LI);
CI->eraseFromParent();
}
if(F->use_empty())
F->eraseFromParent();
}
return true;
}
