bool TailCallElim::runTRE(Function &F) {
// If this function is a varargs function, we won't be able to PHI the args
// right, so don't even try to convert it...
if (F.getFunctionType()->isVarArg()) return false;
TTI = &getAnalysis<TargetTransformInfo>();
BasicBlock *OldEntry = nullptr;
bool TailCallsAreMarkedTail = false;
SmallVector<PHINode*, 8> ArgumentPHIs;
bool MadeChange = false;
// CanTRETailMarkedCall - If false, we cannot perform TRE on tail calls
// marked with the 'tail' attribute, because doing so would cause the stack
// size to increase (real TRE would deallocate variable sized allocas, TRE
// doesn't).
bool CanTRETailMarkedCall = CanTRE(F);
// Change any tail recursive calls to loops.
//
// FIXME: The code generator produces really bad code when an 'escaping
// alloca' is changed from being a static alloca to being a dynamic alloca.
// Until this is resolved, disable this transformation if that would ever
// happen. This bug is PR962.
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) {
bool Change = ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail,
ArgumentPHIs, !CanTRETailMarkedCall);
if (!Change && BB->getFirstNonPHIOrDbg() == Ret)
Change = FoldReturnAndProcessPred(BB, Ret, OldEntry,
TailCallsAreMarkedTail, ArgumentPHIs,
!CanTRETailMarkedCall);
MadeChange |= Change;
}
}
// If we eliminated any tail recursions, it's possible that we inserted some
// silly PHI nodes which just merge an initial value (the incoming operand)
// with themselves. Check to see if we did and clean up our mess if so. This
// occurs when a function passes an argument straight through to its tail
// call.
for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) {
PHINode *PN = ArgumentPHIs[i];
// If the PHI Node is a dynamic constant, replace it with the value it is.
if (Value *PNV = SimplifyInstruction(PN)) {
PN->replaceAllUsesWith(PNV);
PN->eraseFromParent();
}
}
return MadeChange;
}
bool TailCallElim::runOnFunction(Function &F) {
// If this function is a varargs function, we won't be able to PHI the args
// right, so don't even try to convert it...
if (F.getFunctionType()->isVarArg()) return false;
TTI = &getAnalysis<TargetTransformInfo>();
BasicBlock *OldEntry = 0;
bool TailCallsAreMarkedTail = false;
SmallVector<PHINode*, 8> ArgumentPHIs;
bool MadeChange = false;
// CanTRETailMarkedCall - If false, we cannot perform TRE on tail calls
// marked with the 'tail' attribute, because doing so would cause the stack
// size to increase (real TRE would deallocate variable sized allocas, TRE
// doesn't).
bool CanTRETailMarkedCall = true;
// Find calls that can be marked tail.
AllocaCaptureTracker ACT;
for (Function::iterator BB = F.begin(), EE = F.end(); BB != EE; ++BB) {
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
CanTRETailMarkedCall &= CanTRE(AI);
PointerMayBeCaptured(AI, &ACT);
// If any allocas are captured, exit.
if (ACT.Captured)
return false;
}
}
}
// Second pass, change any tail recursive calls to loops.
//
// FIXME: The code generator produces really bad code when an 'escaping
// alloca' is changed from being a static alloca to being a dynamic alloca.
// Until this is resolved, disable this transformation if that would ever
// happen. This bug is PR962.
if (ACT.UsesAlloca.empty()) {
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) {
bool Change = ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail,
ArgumentPHIs, !CanTRETailMarkedCall);
if (!Change && BB->getFirstNonPHIOrDbg() == Ret)
Change = FoldReturnAndProcessPred(BB, Ret, OldEntry,
TailCallsAreMarkedTail, ArgumentPHIs,
!CanTRETailMarkedCall);
MadeChange |= Change;
}
}
}
// If we eliminated any tail recursions, it's possible that we inserted some
// silly PHI nodes which just merge an initial value (the incoming operand)
// with themselves. Check to see if we did and clean up our mess if so. This
// occurs when a function passes an argument straight through to its tail
// call.
if (!ArgumentPHIs.empty()) {
for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) {
PHINode *PN = ArgumentPHIs[i];
// If the PHI Node is a dynamic constant, replace it with the value it is.
if (Value *PNV = SimplifyInstruction(PN)) {
PN->replaceAllUsesWith(PNV);
PN->eraseFromParent();
}
}
}
// At this point, we know that the function does not have any captured
// allocas. If additionally the function does not call setjmp, mark all calls
// in the function that do not access stack memory with the tail keyword. This
// implies ensuring that there does not exist any path from a call that takes
// in an alloca but does not capture it and the call which we wish to mark
// with "tail".
if (!F.callsFunctionThatReturnsTwice()) {
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
if (CallInst *CI = dyn_cast<CallInst>(I)) {
if (!ACT.UsesAlloca.count(CI)) {
CI->setTailCall();
MadeChange = true;
}
}
}
}
}
return MadeChange;
}
bool TailCallElim::runOnFunction(Function &F) {
// If this function is a varargs function, we won't be able to PHI the args
// right, so don't even try to convert it...
if (F.getFunctionType()->isVarArg()) return false;
BasicBlock *OldEntry = 0;
bool TailCallsAreMarkedTail = false;
SmallVector<PHINode*, 8> ArgumentPHIs;
bool MadeChange = false;
bool FunctionContainsEscapingAllocas = false;
// CannotTCETailMarkedCall - If true, we cannot perform TCE on tail calls
// marked with the 'tail' attribute, because doing so would cause the stack
// size to increase (real TCE would deallocate variable sized allocas, TCE
// doesn't).
bool CannotTCETailMarkedCall = false;
// Loop over the function, looking for any returning blocks, and keeping track
// of whether this function has any non-trivially used allocas.
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
if (FunctionContainsEscapingAllocas && CannotTCETailMarkedCall)
break;
FunctionContainsEscapingAllocas |=
CheckForEscapingAllocas(BB, CannotTCETailMarkedCall);
}
/// FIXME: The code generator produces really bad code when an 'escaping
/// alloca' is changed from being a static alloca to being a dynamic alloca.
/// Until this is resolved, disable this transformation if that would ever
/// happen. This bug is PR962.
if (FunctionContainsEscapingAllocas)
return false;
// Second pass, change any tail calls to loops.
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) {
bool Change = ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail,
ArgumentPHIs,CannotTCETailMarkedCall);
if (!Change && BB->getFirstNonPHIOrDbg() == Ret)
Change = FoldReturnAndProcessPred(BB, Ret, OldEntry,
TailCallsAreMarkedTail, ArgumentPHIs,
CannotTCETailMarkedCall);
MadeChange |= Change;
}
}
// If we eliminated any tail recursions, it's possible that we inserted some
// silly PHI nodes which just merge an initial value (the incoming operand)
// with themselves. Check to see if we did and clean up our mess if so. This
// occurs when a function passes an argument straight through to its tail
// call.
if (!ArgumentPHIs.empty()) {
for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) {
PHINode *PN = ArgumentPHIs[i];
// If the PHI Node is a dynamic constant, replace it with the value it is.
if (Value *PNV = SimplifyInstruction(PN)) {
PN->replaceAllUsesWith(PNV);
PN->eraseFromParent();
}
}
}
// Finally, if this function contains no non-escaping allocas, or calls
// setjmp, mark all calls in the function as eligible for tail calls
//(there is no stack memory for them to access).
if (!FunctionContainsEscapingAllocas && !F.callsFunctionThatReturnsTwice())
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
if (CallInst *CI = dyn_cast<CallInst>(I)) {
CI->setTailCall();
MadeChange = true;
}
return MadeChange;
}
/// MoveExceptionValueCalls - Ensure that eh.exception is only ever called from
/// landing pads by replacing calls outside of landing pads with direct use of
/// a register holding the appropriate value; this requires adding calls inside
/// all landing pads to initialize the register. Also, move eh.exception calls
/// inside landing pads to the start of the landing pad (optional, but may make
/// things simpler for later passes).
bool DwarfEHPrepare::MoveExceptionValueCalls() {
// If the eh.exception intrinsic is not declared in the module then there is
// nothing to do. Speed up compilation by checking for this common case.
if (!ExceptionValueIntrinsic &&
!F->getParent()->getFunction(Intrinsic::getName(Intrinsic::eh_exception)))
return false;
bool Changed = false;
// Move calls to eh.exception that are inside a landing pad to the start of
// the landing pad.
for (BBSet::const_iterator LI = LandingPads.begin(), LE = LandingPads.end();
LI != LE; ++LI) {
BasicBlock *LP = *LI;
for (BasicBlock::iterator II = LP->getFirstNonPHIOrDbg(), IE = LP->end();
II != IE;)
if (EHExceptionInst *EI = dyn_cast<EHExceptionInst>(II++)) {
// Found a call to eh.exception.
if (!EI->use_empty()) {
// If there is already a call to eh.exception at the start of the
// landing pad, then get hold of it; otherwise create such a call.
Value *CallAtStart = CreateExceptionValueCall(LP);
// If the call was at the start of a landing pad then leave it alone.
if (EI == CallAtStart)
continue;
EI->replaceAllUsesWith(CallAtStart);
}
EI->eraseFromParent();
++NumExceptionValuesMoved;
Changed = true;
}
}
// Look for calls to eh.exception that are not in a landing pad. If one is
// found, then a register that holds the exception value will be created in
// each landing pad, and the SSAUpdater will be used to compute the values
// returned by eh.exception calls outside of landing pads.
SSAUpdater SSA;
// Remember where we found the eh.exception call, to avoid rescanning earlier
// basic blocks which we already know contain no eh.exception calls.
bool FoundCallOutsideLandingPad = false;
Function::iterator BB = F->begin();
for (Function::iterator BE = F->end(); BB != BE; ++BB) {
// Skip over landing pads.
if (LandingPads.count(BB))
continue;
for (BasicBlock::iterator II = BB->getFirstNonPHIOrDbg(), IE = BB->end();
II != IE; ++II)
if (isa<EHExceptionInst>(II)) {
SSA.Initialize(II->getType(), II->getName());
FoundCallOutsideLandingPad = true;
break;
}
if (FoundCallOutsideLandingPad)
break;
}
// If all calls to eh.exception are in landing pads then we are done.
if (!FoundCallOutsideLandingPad)
return Changed;
// Add a call to eh.exception at the start of each landing pad, and tell the
// SSAUpdater that this is the value produced by the landing pad.
for (BBSet::iterator LI = LandingPads.begin(), LE = LandingPads.end();
LI != LE; ++LI)
SSA.AddAvailableValue(*LI, CreateExceptionValueCall(*LI));
// Now turn all calls to eh.exception that are not in a landing pad into a use
// of the appropriate register.
for (Function::iterator BE = F->end(); BB != BE; ++BB) {
// Skip over landing pads.
if (LandingPads.count(BB))
continue;
for (BasicBlock::iterator II = BB->getFirstNonPHIOrDbg(), IE = BB->end();
II != IE;)
if (EHExceptionInst *EI = dyn_cast<EHExceptionInst>(II++)) {
// Found a call to eh.exception, replace it with the value from any
// upstream landing pad(s).
EI->replaceAllUsesWith(SSA.GetValueAtEndOfBlock(BB));
EI->eraseFromParent();
++NumExceptionValuesMoved;
}
}
return true;
}