/// Emit dealloc_stack for all temporaries.
void PartialApplyCombiner::deallocateTemporaries() {
// Insert dealloc_stack instructions at all function exit points.
for (SILBasicBlock &BB : *PAI->getFunction()) {
TermInst *Term = BB.getTerminator();
if (!Term->isFunctionExiting())
continue;
for (auto Op : Tmps) {
Builder.setInsertionPoint(Term);
Builder.createDeallocStack(PAI->getLoc(), Op);
}
}
}
/// Emit code to release/destroy temporaries.
void PartialApplyCombiner::releaseTemporaries() {
// Insert releases and destroy_addrs as early as possible,
// because we don't want to keep objects alive longer than
// its really needed.
for (auto Op : Tmps) {
auto TmpType = Op->getType().getObjectType();
if (TmpType.isTrivial(PAI->getModule()))
continue;
for (auto *EndPoint : PAFrontier) {
Builder.setInsertionPoint(EndPoint);
if (!TmpType.isAddressOnly(PAI->getModule())) {
auto *Load = Builder.createLoad(PAI->getLoc(), Op);
Builder.createReleaseValue(PAI->getLoc(), Load, Atomicity::Atomic);
} else {
Builder.createDestroyAddr(PAI->getLoc(), Op);
}
}
}
}
/// Perform the apply{partial_apply(x,y)}(z) -> apply(z,x,y) peephole
/// by iterating over all uses of the partial_apply and searching
/// for the pattern to transform.
SILInstruction *PartialApplyCombiner::combine() {
if (!canCombinePartialApply(PAI))
return nullptr;
// Only process partial apply if the callee is a known function.
FRI = dyn_cast<FunctionRefInst>(PAI->getCallee());
// Iterate over all uses of the partial_apply
// and look for applies that use it as a callee.
for (auto UI = PAI->use_begin(), UE = PAI->use_end(); UI != UE; ) {
auto Use = *UI;
++UI;
auto User = Use->getUser();
// If this use of a partial_apply is not
// an apply which uses it as a callee, bail.
auto AI = FullApplySite::isa(User);
if (!AI)
continue;
if (AI.getCallee() != PAI)
continue;
// We cannot handle generic apply yet. Bail.
if (AI.hasSubstitutions())
continue;
if (!processSingleApply(AI))
return nullptr;
}
// release/destroy and deallocate introduced temporaries.
if (!Tmps.empty()) {
releaseTemporaries();
deallocateTemporaries();
}
return nullptr;
}
/// Inlines all mandatory inlined functions into the body of a function,
/// first recursively inlining all mandatory apply instructions in those
/// functions into their bodies if necessary.
///
/// \param F the function to be processed
/// \param AI nullptr if this is being called from the top level; the relevant
/// ApplyInst requiring the recursive call when non-null
/// \param FullyInlinedSet the set of all functions already known to be fully
/// processed, to avoid processing them over again
/// \param SetFactory an instance of ImmutableFunctionSet::Factory
/// \param CurrentInliningSet the set of functions currently being inlined in
/// the current call stack of recursive calls
///
/// \returns true if successful, false if failed due to circular inlining.
static bool
runOnFunctionRecursively(SILOptFunctionBuilder &FuncBuilder,
SILFunction *F, FullApplySite AI,
DenseFunctionSet &FullyInlinedSet,
ImmutableFunctionSet::Factory &SetFactory,
ImmutableFunctionSet CurrentInliningSet,
ClassHierarchyAnalysis *CHA) {
// Avoid reprocessing functions needlessly.
if (FullyInlinedSet.count(F))
return true;
// Prevent attempt to circularly inline.
if (CurrentInliningSet.contains(F)) {
// This cannot happen on a top-level call, so AI should be non-null.
assert(AI && "Cannot have circular inline without apply");
SILLocation L = AI.getLoc();
assert(L && "Must have location for transparent inline apply");
diagnose(F->getModule().getASTContext(), L.getStartSourceLoc(),
diag::circular_transparent);
return false;
}
// Add to the current inlining set (immutably, so we only affect the set
// during this call and recursive subcalls).
CurrentInliningSet = SetFactory.add(CurrentInliningSet, F);
SmallVector<std::pair<SILValue, ParameterConvention>, 16> CaptureArgs;
SmallVector<SILValue, 32> FullArgs;
// Visiting blocks in reverse order avoids revisiting instructions after block
// splitting, which would be quadratic.
for (auto BI = F->rbegin(), BE = F->rend(), nextBB = BI; BI != BE;
BI = nextBB) {
// After inlining, the block iterator will be adjusted to point to the last
// block containing inlined instructions. This way, the inlined function
// body will be reprocessed within the caller's context without revisiting
// any original instructions.
nextBB = std::next(BI);
// While iterating over this block, instructions are inserted and deleted.
// To avoid quadratic block splitting, instructions must be processed in
// reverse order (block splitting reassigned the parent pointer of all
// instructions below the split point).
for (auto II = BI->rbegin(); II != BI->rend(); ++II) {
FullApplySite InnerAI = FullApplySite::isa(&*II);
if (!InnerAI)
continue;
// *NOTE* If devirtualization succeeds, devirtInst may not be InnerAI,
// but a casted result of InnerAI or even a block argument due to
// abstraction changes when calling the witness or class method.
auto *devirtInst = tryDevirtualizeApplyHelper(InnerAI, CHA);
// Restore II to the current apply site.
II = devirtInst->getReverseIterator();
// If the devirtualized call result is no longer a invalid FullApplySite,
// then it has succeeded, but the result is not immediately inlinable.
InnerAI = FullApplySite::isa(devirtInst);
if (!InnerAI)
continue;
SILValue CalleeValue = InnerAI.getCallee();
bool IsThick;
PartialApplyInst *PAI;
SILFunction *CalleeFunction = getCalleeFunction(
F, InnerAI, IsThick, CaptureArgs, FullArgs, PAI);
if (!CalleeFunction)
continue;
// Then recursively process it first before trying to inline it.
if (!runOnFunctionRecursively(FuncBuilder, CalleeFunction, InnerAI,
FullyInlinedSet, SetFactory,
CurrentInliningSet, CHA)) {
// If we failed due to circular inlining, then emit some notes to
// trace back the failure if we have more information.
// FIXME: possibly it could be worth recovering and attempting other
// inlines within this same recursive call rather than simply
// propagating the failure.
if (AI) {
SILLocation L = AI.getLoc();
assert(L && "Must have location for transparent inline apply");
diagnose(F->getModule().getASTContext(), L.getStartSourceLoc(),
diag::note_while_inlining);
}
return false;
}
// Get our list of substitutions.
auto Subs = (PAI
? PAI->getSubstitutionMap()
: InnerAI.getSubstitutionMap());
SILOpenedArchetypesTracker OpenedArchetypesTracker(F);
F->getModule().registerDeleteNotificationHandler(
&OpenedArchetypesTracker);
// The callee only needs to know about opened archetypes used in
// the substitution list.
OpenedArchetypesTracker.registerUsedOpenedArchetypes(
InnerAI.getInstruction());
if (PAI) {
OpenedArchetypesTracker.registerUsedOpenedArchetypes(PAI);
}
SILInliner Inliner(FuncBuilder, SILInliner::InlineKind::MandatoryInline,
Subs, OpenedArchetypesTracker);
if (!Inliner.canInlineApplySite(InnerAI))
continue;
// Inline function at I, which also changes I to refer to the first
// instruction inlined in the case that it succeeds. We purposely
// process the inlined body after inlining, because the inlining may
// have exposed new inlining opportunities beyond those present in
// the inlined function when processed independently.
LLVM_DEBUG(llvm::errs() << "Inlining @" << CalleeFunction->getName()
<< " into @" << InnerAI.getFunction()->getName()
<< "\n");
// If we intend to inline a thick function, then we need to balance the
// reference counts for correctness.
if (IsThick) {
bool IsCalleeGuaranteed =
PAI &&
PAI->getType().castTo<SILFunctionType>()->isCalleeGuaranteed();
fixupReferenceCounts(InnerAI.getInstruction(), CalleeValue, CaptureArgs,
IsCalleeGuaranteed);
}
// Register a callback to record potentially unused function values after
// inlining.
ClosureCleanup closureCleanup;
Inliner.setDeletionCallback([&closureCleanup](SILInstruction *I) {
closureCleanup.recordDeadFunction(I);
});
// Inlining deletes the apply, and can introduce multiple new basic
// blocks. After this, CalleeValue and other instructions may be invalid.
// nextBB will point to the last inlined block
auto firstInlinedInstAndLastBB =
Inliner.inlineFunction(CalleeFunction, InnerAI, FullArgs);
nextBB = firstInlinedInstAndLastBB.second->getReverseIterator();
++NumMandatoryInlines;
// The IR is now valid, and trivial dead arguments are removed. However,
// we may be able to remove dead callee computations (e.g. dead
// partial_apply closures).
closureCleanup.cleanupDeadClosures(F);
// Resume inlining within nextBB, which contains only the inlined
// instructions and possibly instructions in the original call block that
// have not yet been visited.
break;
}
}
// Keep track of full inlined functions so we don't waste time recursively
// reprocessing them.
FullyInlinedSet.insert(F);
return true;
}
/// \brief Inlines all mandatory inlined functions into the body of a function,
/// first recursively inlining all mandatory apply instructions in those
/// functions into their bodies if necessary.
///
/// \param F the function to be processed
/// \param AI nullptr if this is being called from the top level; the relevant
/// ApplyInst requiring the recursive call when non-null
/// \param FullyInlinedSet the set of all functions already known to be fully
/// processed, to avoid processing them over again
/// \param SetFactory an instance of ImmutableFunctionSet::Factory
/// \param CurrentInliningSet the set of functions currently being inlined in
/// the current call stack of recursive calls
///
/// \returns true if successful, false if failed due to circular inlining.
static bool
runOnFunctionRecursively(SILFunction *F, FullApplySite AI,
SILModule::LinkingMode Mode,
DenseFunctionSet &FullyInlinedSet,
ImmutableFunctionSet::Factory &SetFactory,
ImmutableFunctionSet CurrentInliningSet,
ClassHierarchyAnalysis *CHA) {
// Avoid reprocessing functions needlessly.
if (FullyInlinedSet.count(F))
return true;
// Prevent attempt to circularly inline.
if (CurrentInliningSet.contains(F)) {
// This cannot happen on a top-level call, so AI should be non-null.
assert(AI && "Cannot have circular inline without apply");
SILLocation L = AI.getLoc();
assert(L && "Must have location for transparent inline apply");
diagnose(F->getModule().getASTContext(), L.getStartSourceLoc(),
diag::circular_transparent);
return false;
}
// Add to the current inlining set (immutably, so we only affect the set
// during this call and recursive subcalls).
CurrentInliningSet = SetFactory.add(CurrentInliningSet, F);
SmallVector<SILValue, 16> CaptureArgs;
SmallVector<SILValue, 32> FullArgs;
for (auto FI = F->begin(), FE = F->end(); FI != FE; ++FI) {
for (auto I = FI->begin(), E = FI->end(); I != E; ++I) {
FullApplySite InnerAI = FullApplySite::isa(&*I);
if (!InnerAI)
continue;
auto *ApplyBlock = InnerAI.getParent();
auto NewInstPair = tryDevirtualizeApply(InnerAI, CHA);
if (auto *NewInst = NewInstPair.first) {
replaceDeadApply(InnerAI, NewInst);
if (auto *II = dyn_cast<SILInstruction>(NewInst))
I = II->getIterator();
else
I = NewInst->getParentBB()->begin();
auto NewAI = FullApplySite::isa(NewInstPair.second.getInstruction());
if (!NewAI)
continue;
InnerAI = NewAI;
}
SILLocation Loc = InnerAI.getLoc();
SILValue CalleeValue = InnerAI.getCallee();
bool IsThick;
PartialApplyInst *PAI;
SILFunction *CalleeFunction = getCalleeFunction(InnerAI, IsThick,
CaptureArgs, FullArgs,
PAI,
Mode);
if (!CalleeFunction ||
CalleeFunction->isTransparent() == IsNotTransparent)
continue;
// Then recursively process it first before trying to inline it.
if (!runOnFunctionRecursively(CalleeFunction, InnerAI, Mode,
FullyInlinedSet, SetFactory,
CurrentInliningSet, CHA)) {
// If we failed due to circular inlining, then emit some notes to
// trace back the failure if we have more information.
// FIXME: possibly it could be worth recovering and attempting other
// inlines within this same recursive call rather than simply
// propagating the failure.
if (AI) {
SILLocation L = AI.getLoc();
assert(L && "Must have location for transparent inline apply");
diagnose(F->getModule().getASTContext(), L.getStartSourceLoc(),
diag::note_while_inlining);
}
return false;
}
// Inline function at I, which also changes I to refer to the first
// instruction inlined in the case that it succeeds. We purposely
// process the inlined body after inlining, because the inlining may
// have exposed new inlining opportunities beyond those present in
// the inlined function when processed independently.
DEBUG(llvm::errs() << "Inlining @" << CalleeFunction->getName()
<< " into @" << InnerAI.getFunction()->getName()
<< "\n");
// Decrement our iterator (carefully, to avoid going off the front) so it
// is valid after inlining is done. Inlining deletes the apply, and can
// introduce multiple new basic blocks.
if (I != ApplyBlock->begin())
--I;
else
I = ApplyBlock->end();
TypeSubstitutionMap ContextSubs;
std::vector<Substitution> ApplySubs(InnerAI.getSubstitutions());
if (PAI) {
auto PAISubs = PAI->getSubstitutions();
ApplySubs.insert(ApplySubs.end(), PAISubs.begin(), PAISubs.end());
}
ContextSubs.copyFrom(CalleeFunction->getContextGenericParams()
->getSubstitutionMap(ApplySubs));
SILInliner Inliner(*F, *CalleeFunction,
SILInliner::InlineKind::MandatoryInline,
ContextSubs, ApplySubs);
if (!Inliner.inlineFunction(InnerAI, FullArgs)) {
I = InnerAI.getInstruction()->getIterator();
continue;
}
// Inlining was successful. Remove the apply.
InnerAI.getInstruction()->eraseFromParent();
// Reestablish our iterator if it wrapped.
if (I == ApplyBlock->end())
I = ApplyBlock->begin();
else
++I;
// If the inlined apply was a thick function, then we need to balance the
// reference counts for correctness.
if (IsThick)
fixupReferenceCounts(I, Loc, CalleeValue, CaptureArgs);
// Now that the IR is correct, see if we can remove dead callee
// computations (e.g. dead partial_apply closures).
cleanupCalleeValue(CalleeValue, CaptureArgs, FullArgs);
// Reposition iterators possibly invalidated by mutation.
FI = SILFunction::iterator(ApplyBlock);
I = ApplyBlock->begin();
E = ApplyBlock->end();
++NumMandatoryInlines;
}
}
// Keep track of full inlined functions so we don't waste time recursively
// reprocessing them.
FullyInlinedSet.insert(F);
return true;
}
/// \brief Inlines all mandatory inlined functions into the body of a function,
/// first recursively inlining all mandatory apply instructions in those
/// functions into their bodies if necessary.
///
/// \param F the function to be processed
/// \param AI nullptr if this is being called from the top level; the relevant
/// ApplyInst requiring the recursive call when non-null
/// \param FullyInlinedSet the set of all functions already known to be fully
/// processed, to avoid processing them over again
/// \param SetFactory an instance of ImmutableFunctionSet::Factory
/// \param CurrentInliningSet the set of functions currently being inlined in
/// the current call stack of recursive calls
///
/// \returns true if successful, false if failed due to circular inlining.
static bool
runOnFunctionRecursively(SILFunction *F, FullApplySite AI,
DenseFunctionSet &FullyInlinedSet,
ImmutableFunctionSet::Factory &SetFactory,
ImmutableFunctionSet CurrentInliningSet,
ClassHierarchyAnalysis *CHA) {
// Avoid reprocessing functions needlessly.
if (FullyInlinedSet.count(F))
return true;
// Prevent attempt to circularly inline.
if (CurrentInliningSet.contains(F)) {
// This cannot happen on a top-level call, so AI should be non-null.
assert(AI && "Cannot have circular inline without apply");
SILLocation L = AI.getLoc();
assert(L && "Must have location for transparent inline apply");
diagnose(F->getModule().getASTContext(), L.getStartSourceLoc(),
diag::circular_transparent);
return false;
}
// Add to the current inlining set (immutably, so we only affect the set
// during this call and recursive subcalls).
CurrentInliningSet = SetFactory.add(CurrentInliningSet, F);
SmallVector<std::pair<SILValue, ParameterConvention>, 16> CaptureArgs;
SmallVector<SILValue, 32> FullArgs;
for (auto BI = F->begin(), BE = F->end(); BI != BE; ++BI) {
for (auto II = BI->begin(), IE = BI->end(); II != IE; ++II) {
FullApplySite InnerAI = FullApplySite::isa(&*II);
if (!InnerAI)
continue;
auto *ApplyBlock = InnerAI.getParent();
// *NOTE* If devirtualization succeeds, sometimes II will not be InnerAI,
// but a casted result of InnerAI or even a block argument due to
// abstraction changes when calling the witness or class method. We still
// know that InnerAI dominates II though.
std::tie(InnerAI, II) = tryDevirtualizeApplyHelper(InnerAI, II, CHA);
if (!InnerAI)
continue;
SILValue CalleeValue = InnerAI.getCallee();
bool IsThick;
PartialApplyInst *PAI;
SILFunction *CalleeFunction = getCalleeFunction(
F, InnerAI, IsThick, CaptureArgs, FullArgs, PAI);
if (!CalleeFunction)
continue;
// Then recursively process it first before trying to inline it.
if (!runOnFunctionRecursively(CalleeFunction, InnerAI,
FullyInlinedSet, SetFactory,
CurrentInliningSet, CHA)) {
// If we failed due to circular inlining, then emit some notes to
// trace back the failure if we have more information.
// FIXME: possibly it could be worth recovering and attempting other
// inlines within this same recursive call rather than simply
// propagating the failure.
if (AI) {
SILLocation L = AI.getLoc();
assert(L && "Must have location for transparent inline apply");
diagnose(F->getModule().getASTContext(), L.getStartSourceLoc(),
diag::note_while_inlining);
}
return false;
}
// Get our list of substitutions.
auto Subs = (PAI
? PAI->getSubstitutionMap()
: InnerAI.getSubstitutionMap());
SILOpenedArchetypesTracker OpenedArchetypesTracker(F);
F->getModule().registerDeleteNotificationHandler(
&OpenedArchetypesTracker);
// The callee only needs to know about opened archetypes used in
// the substitution list.
OpenedArchetypesTracker.registerUsedOpenedArchetypes(
InnerAI.getInstruction());
if (PAI) {
OpenedArchetypesTracker.registerUsedOpenedArchetypes(PAI);
}
SILInliner Inliner(*F, *CalleeFunction,
SILInliner::InlineKind::MandatoryInline, Subs,
OpenedArchetypesTracker);
if (!Inliner.canInlineFunction(InnerAI)) {
// See comment above about casting when devirtualizing and how this
// sometimes causes II and InnerAI to be different and even in different
// blocks.
II = InnerAI.getInstruction()->getIterator();
continue;
}
// Inline function at I, which also changes I to refer to the first
// instruction inlined in the case that it succeeds. We purposely
// process the inlined body after inlining, because the inlining may
// have exposed new inlining opportunities beyond those present in
// the inlined function when processed independently.
LLVM_DEBUG(llvm::errs() << "Inlining @" << CalleeFunction->getName()
<< " into @" << InnerAI.getFunction()->getName()
<< "\n");
// If we intend to inline a thick function, then we need to balance the
// reference counts for correctness.
if (IsThick) {
bool IsCalleeGuaranteed =
PAI &&
PAI->getType().castTo<SILFunctionType>()->isCalleeGuaranteed();
fixupReferenceCounts(II, CalleeValue, CaptureArgs, IsCalleeGuaranteed);
}
// Decrement our iterator (carefully, to avoid going off the front) so it
// is valid after inlining is done. Inlining deletes the apply, and can
// introduce multiple new basic blocks.
II = prev_or_default(II, ApplyBlock->begin(), ApplyBlock->end());
Inliner.inlineFunction(InnerAI, FullArgs);
// We were able to inline successfully. Remove the apply.
InnerAI.getInstruction()->eraseFromParent();
// Reestablish our iterator if it wrapped.
if (II == ApplyBlock->end())
II = ApplyBlock->begin();
// Update the iterator when instructions are removed.
DeleteInstructionsHandler DeletionHandler(II);
// Now that the IR is correct, see if we can remove dead callee
// computations (e.g. dead partial_apply closures).
cleanupCalleeValue(CalleeValue, FullArgs);
// Reposition iterators possibly invalidated by mutation.
BI = SILFunction::iterator(ApplyBlock);
IE = ApplyBlock->end();
assert(BI == SILFunction::iterator(II->getParent()) &&
"Mismatch between the instruction and basic block");
++NumMandatoryInlines;
}
}
// Keep track of full inlined functions so we don't waste time recursively
// reprocessing them.
FullyInlinedSet.insert(F);
return true;
}
/// Process an apply instruction which uses a partial_apply
/// as its callee.
/// Returns true on success.
bool PartialApplyCombiner::processSingleApply(FullApplySite AI) {
Builder.setInsertionPoint(AI.getInstruction());
Builder.setCurrentDebugScope(AI.getDebugScope());
// Prepare the args.
SmallVector<SILValue, 8> Args;
// First the ApplyInst args.
for (auto Op : AI.getArguments())
Args.push_back(Op);
SILInstruction *InsertionPoint = &*Builder.getInsertionPoint();
// Next, the partial apply args.
// Pre-process partial_apply arguments only once, lazily.
if (isFirstTime) {
isFirstTime = false;
if (!allocateTemporaries())
return false;
}
// Now, copy over the partial apply args.
for (auto Op : PAI->getArguments()) {
auto Arg = Op;
// If there is new temporary for this argument, use it instead.
if (isa<AllocStackInst>(Arg)) {
if (ArgToTmp.count(Arg)) {
Op = ArgToTmp.lookup(Arg);
}
}
Args.push_back(Op);
}
Builder.setInsertionPoint(InsertionPoint);
Builder.setCurrentDebugScope(AI.getDebugScope());
// The thunk that implements the partial apply calls the closure function
// that expects all arguments to be consumed by the function. However, the
// captured arguments are not arguments of *this* apply, so they are not
// pre-incremented. When we combine the partial_apply and this apply into
// a new apply we need to retain all of the closure non-address type
// arguments.
auto ParamInfo = PAI->getSubstCalleeType()->getParameters();
auto PartialApplyArgs = PAI->getArguments();
// Set of arguments that need to be released after each invocation.
SmallVector<SILValue, 8> ToBeReleasedArgs;
for (unsigned i = 0, e = PartialApplyArgs.size(); i < e; ++i) {
SILValue Arg = PartialApplyArgs[i];
if (!Arg->getType().isAddress()) {
// Retain the argument as the callee may consume it.
Builder.emitRetainValueOperation(PAI->getLoc(), Arg);
// For non consumed parameters (e.g. guaranteed), we also need to
// insert releases after each apply instruction that we create.
if (!ParamInfo[ParamInfo.size() - PartialApplyArgs.size() + i].
isConsumed())
ToBeReleasedArgs.push_back(Arg);
}
}
auto *F = FRI->getReferencedFunction();
SILType FnType = F->getLoweredType();
SILType ResultTy = F->getLoweredFunctionType()->getSILResult();
ArrayRef<Substitution> Subs = PAI->getSubstitutions();
if (!Subs.empty()) {
FnType = FnType.substGenericArgs(PAI->getModule(), Subs);
ResultTy = FnType.getAs<SILFunctionType>()->getSILResult();
}
FullApplySite NAI;
if (auto *TAI = dyn_cast<TryApplyInst>(AI))
NAI =
Builder.createTryApply(AI.getLoc(), FRI, FnType, Subs, Args,
TAI->getNormalBB(), TAI->getErrorBB());
else
NAI =
Builder.createApply(AI.getLoc(), FRI, FnType, ResultTy, Subs, Args,
cast<ApplyInst>(AI)->isNonThrowing());
// We also need to release the partial_apply instruction itself because it
// is consumed by the apply_instruction.
if (auto *TAI = dyn_cast<TryApplyInst>(AI)) {
Builder.setInsertionPoint(TAI->getNormalBB()->begin());
for (auto Arg : ToBeReleasedArgs) {
Builder.emitReleaseValueOperation(PAI->getLoc(), Arg);
}
Builder.createStrongRelease(AI.getLoc(), PAI, Atomicity::Atomic);
Builder.setInsertionPoint(TAI->getErrorBB()->begin());
// Release the non-consumed parameters.
for (auto Arg : ToBeReleasedArgs) {
Builder.emitReleaseValueOperation(PAI->getLoc(), Arg);
}
Builder.createStrongRelease(AI.getLoc(), PAI, Atomicity::Atomic);
Builder.setInsertionPoint(AI.getInstruction());
} else {
// Release the non-consumed parameters.
for (auto Arg : ToBeReleasedArgs) {
Builder.emitReleaseValueOperation(PAI->getLoc(), Arg);
}
Builder.createStrongRelease(AI.getLoc(), PAI, Atomicity::Atomic);
}
SilCombiner->replaceInstUsesWith(*AI.getInstruction(), NAI.getInstruction());
SilCombiner->eraseInstFromFunction(*AI.getInstruction());
return true;
}
/// Returns true on success.
bool PartialApplyCombiner::allocateTemporaries() {
// Copy the original arguments of the partial_apply into
// newly created temporaries and use these temporaries instead of
// the original arguments afterwards.
// This is done to "extend" the life-time of original partial_apply
// arguments, as they may be destroyed/deallocated before the last
// use by one of the apply instructions.
// TODO:
// Copy arguments of the partial_apply into new temporaries
// only if the lifetime of arguments ends before their uses
// by apply instructions.
bool needsReleases = false;
CanSILFunctionType PAITy =
dyn_cast<SILFunctionType>(PAI->getCallee()->getType().getSwiftType());
// Emit a destroy value for each captured closure argument.
ArrayRef<SILParameterInfo> Params = PAITy->getParameters();
auto Args = PAI->getArguments();
unsigned Delta = Params.size() - Args.size();
llvm::SmallVector<std::pair<SILValue, unsigned>, 8> ArgsToHandle;
for (unsigned AI = 0, AE = Args.size(); AI != AE; ++AI) {
SILValue Arg = Args[AI];
SILParameterInfo Param = Params[AI + Delta];
if (Param.isIndirectMutating())
continue;
// Create a temporary and copy the argument into it, if:
// - the argument stems from an alloc_stack
// - the argument is consumed by the callee and is indirect
// (e.g. it is an @in argument)
if (isa<AllocStackInst>(Arg) ||
(Param.isConsumed() && Param.isIndirect())) {
// If the temporary is non-trivial, we need to release it later.
if (!Arg->getType().isTrivial(PAI->getModule()))
needsReleases = true;
ArgsToHandle.push_back(std::make_pair(Arg, AI));
}
}
if (needsReleases) {
// Compute the set of endpoints, which will be used to insert releases of
// temporaries. This may fail if the frontier is located on a critical edge
// which we may not split (no CFG changes in SILCombine).
ValueLifetimeAnalysis VLA(PAI);
if (!VLA.computeFrontier(PAFrontier, ValueLifetimeAnalysis::DontModifyCFG))
return false;
}
for (auto ArgWithIdx : ArgsToHandle) {
SILValue Arg = ArgWithIdx.first;
Builder.setInsertionPoint(PAI->getFunction()->begin()->begin());
// Create a new temporary at the beginning of a function.
auto *Tmp = Builder.createAllocStack(PAI->getLoc(), Arg->getType(),
{/*Constant*/ true, ArgWithIdx.second});
Builder.setInsertionPoint(PAI);
// Copy argument into this temporary.
Builder.createCopyAddr(PAI->getLoc(), Arg, Tmp,
IsTake_t::IsNotTake,
IsInitialization_t::IsInitialization);
Tmps.push_back(Tmp);
ArgToTmp.insert(std::make_pair(Arg, Tmp));
}
return true;
}
/// \brief Inlines all mandatory inlined functions into the body of a function,
/// first recursively inlining all mandatory apply instructions in those
/// functions into their bodies if necessary.
///
/// \param F the function to be processed
/// \param AI nullptr if this is being called from the top level; the relevant
/// ApplyInst requiring the recursive call when non-null
/// \param FullyInlinedSet the set of all functions already known to be fully
/// processed, to avoid processing them over again
/// \param SetFactory an instance of ImmutableFunctionSet::Factory
/// \param CurrentInliningSet the set of functions currently being inlined in
/// the current call stack of recursive calls
///
/// \returns true if successful, false if failed due to circular inlining.
static bool
runOnFunctionRecursively(SILFunction *F, FullApplySite AI,
SILModule::LinkingMode Mode,
DenseFunctionSet &FullyInlinedSet,
ImmutableFunctionSet::Factory &SetFactory,
ImmutableFunctionSet CurrentInliningSet,
ClassHierarchyAnalysis *CHA) {
// Avoid reprocessing functions needlessly.
if (FullyInlinedSet.count(F))
return true;
// Prevent attempt to circularly inline.
if (CurrentInliningSet.contains(F)) {
// This cannot happen on a top-level call, so AI should be non-null.
assert(AI && "Cannot have circular inline without apply");
SILLocation L = AI.getLoc();
assert(L && "Must have location for transparent inline apply");
diagnose(F->getModule().getASTContext(), L.getStartSourceLoc(),
diag::circular_transparent);
return false;
}
// Add to the current inlining set (immutably, so we only affect the set
// during this call and recursive subcalls).
CurrentInliningSet = SetFactory.add(CurrentInliningSet, F);
SmallVector<SILValue, 16> CaptureArgs;
SmallVector<SILValue, 32> FullArgs;
for (auto FI = F->begin(), FE = F->end(); FI != FE; ++FI) {
for (auto I = FI->begin(), E = FI->end(); I != E; ++I) {
FullApplySite InnerAI = FullApplySite::isa(&*I);
if (!InnerAI)
continue;
auto *ApplyBlock = InnerAI.getParent();
auto NewInstPair = tryDevirtualizeApply(InnerAI, CHA);
if (auto *NewInst = NewInstPair.first) {
replaceDeadApply(InnerAI, NewInst);
if (auto *II = dyn_cast<SILInstruction>(NewInst))
I = II->getIterator();
else
I = NewInst->getParentBlock()->begin();
auto NewAI = FullApplySite::isa(NewInstPair.second.getInstruction());
if (!NewAI)
continue;
InnerAI = NewAI;
}
SILLocation Loc = InnerAI.getLoc();
SILValue CalleeValue = InnerAI.getCallee();
bool IsThick;
PartialApplyInst *PAI;
SILFunction *CalleeFunction = getCalleeFunction(InnerAI, IsThick,
CaptureArgs, FullArgs,
PAI,
Mode);
if (!CalleeFunction ||
CalleeFunction->isTransparent() == IsNotTransparent)
continue;
if (F->isFragile() &&
!CalleeFunction->hasValidLinkageForFragileRef()) {
if (!CalleeFunction->hasValidLinkageForFragileInline()) {
llvm::errs() << "caller: " << F->getName() << "\n";
llvm::errs() << "callee: " << CalleeFunction->getName() << "\n";
llvm_unreachable("Should never be inlining a resilient function into "
"a fragile function");
}
continue;
}
// Then recursively process it first before trying to inline it.
if (!runOnFunctionRecursively(CalleeFunction, InnerAI, Mode,
FullyInlinedSet, SetFactory,
CurrentInliningSet, CHA)) {
// If we failed due to circular inlining, then emit some notes to
// trace back the failure if we have more information.
// FIXME: possibly it could be worth recovering and attempting other
// inlines within this same recursive call rather than simply
// propagating the failure.
if (AI) {
SILLocation L = AI.getLoc();
assert(L && "Must have location for transparent inline apply");
diagnose(F->getModule().getASTContext(), L.getStartSourceLoc(),
diag::note_while_inlining);
}
return false;
}
// Inline function at I, which also changes I to refer to the first
// instruction inlined in the case that it succeeds. We purposely
// process the inlined body after inlining, because the inlining may
// have exposed new inlining opportunities beyond those present in
// the inlined function when processed independently.
DEBUG(llvm::errs() << "Inlining @" << CalleeFunction->getName()
<< " into @" << InnerAI.getFunction()->getName()
<< "\n");
// If we intend to inline a thick function, then we need to balance the
// reference counts for correctness.
if (IsThick && I != ApplyBlock->begin()) {
// We need to find an appropriate location for our fix up code
// We used to do this after inlining Without any modifications
// This caused us to add a release in a wrong place:
// It would release a value *before* retaining it!
// It is really problematic to do this after inlining -
// Finding a valid insertion point is tricky:
// Inlining might add new basic blocks and/or remove the apply
// We want to add the fix up *just before* where the current apply is!
// Unfortunately, we *can't* add the fix up code here:
// Inlining might fail for any reason -
// If that occurred we'd need to undo our fix up code.
// Instead, we split the current basic block -
// Making sure we have a basic block that starts with our apply.
SILBuilderWithScope B(I);
ApplyBlock = splitBasicBlockAndBranch(B, &*I, nullptr, nullptr);
I = ApplyBlock->begin();
}
// Decrement our iterator (carefully, to avoid going off the front) so it
// is valid after inlining is done. Inlining deletes the apply, and can
// introduce multiple new basic blocks.
if (I != ApplyBlock->begin())
--I;
else
I = ApplyBlock->end();
std::vector<Substitution> ApplySubs(InnerAI.getSubstitutions());
if (PAI) {
auto PAISubs = PAI->getSubstitutions();
ApplySubs.insert(ApplySubs.end(), PAISubs.begin(), PAISubs.end());
}
SILOpenedArchetypesTracker OpenedArchetypesTracker(*F);
F->getModule().registerDeleteNotificationHandler(
&OpenedArchetypesTracker);
// The callee only needs to know about opened archetypes used in
// the substitution list.
OpenedArchetypesTracker.registerUsedOpenedArchetypes(InnerAI.getInstruction());
if (PAI) {
OpenedArchetypesTracker.registerUsedOpenedArchetypes(PAI);
}
SILInliner Inliner(*F, *CalleeFunction,
SILInliner::InlineKind::MandatoryInline,
ApplySubs, OpenedArchetypesTracker);
if (!Inliner.inlineFunction(InnerAI, FullArgs)) {
I = InnerAI.getInstruction()->getIterator();
continue;
}
// Inlining was successful. Remove the apply.
InnerAI.getInstruction()->eraseFromParent();
// Reestablish our iterator if it wrapped.
if (I == ApplyBlock->end())
I = ApplyBlock->begin();
// Update the iterator when instructions are removed.
DeleteInstructionsHandler DeletionHandler(I);
// If the inlined apply was a thick function, then we need to balance the
// reference counts for correctness.
if (IsThick)
fixupReferenceCounts(I, Loc, CalleeValue, CaptureArgs);
// Now that the IR is correct, see if we can remove dead callee
// computations (e.g. dead partial_apply closures).
cleanupCalleeValue(CalleeValue, CaptureArgs, FullArgs);
// Reposition iterators possibly invalidated by mutation.
FI = SILFunction::iterator(ApplyBlock);
E = ApplyBlock->end();
assert(FI == SILFunction::iterator(I->getParent()) &&
"Mismatch between the instruction and basic block");
++NumMandatoryInlines;
}
}
// Keep track of full inlined functions so we don't waste time recursively
// reprocessing them.
FullyInlinedSet.insert(F);
return true;
}
static void checkNoEscapePartialApplyUse(Operand *oper, FollowUse followUses) {
SILInstruction *user = oper->getUser();
// Ignore uses that are totally uninteresting.
if (isIncidentalUse(user) || onlyAffectsRefCount(user))
return;
// Before checking conversions in general below (getSingleValueCopyOrCast),
// check for convert_function to [without_actually_escaping]. Assume such
// conversion are not actually escaping without following their uses.
if (auto *CFI = dyn_cast<ConvertFunctionInst>(user)) {
if (CFI->withoutActuallyEscaping())
return;
}
// Look through copies, borrows, and conversions.
//
// Note: This handles ConversionInst, which already includes everything in
// swift::stripConvertFunctions.
if (SingleValueInstruction *copy = getSingleValueCopyOrCast(user)) {
// Only follow the copied operand. Other operands are incidental,
// as in the second operand of mark_dependence.
if (oper->getOperandNumber() == 0)
followUses(copy);
return;
}
switch (user->getKind()) {
default:
break;
// Look through Optionals.
case SILInstructionKind::EnumInst:
// @noescape block storage can be passed as an Optional (Nullable).
followUses(cast<EnumInst>(user));
return;
// Look through Phis.
case SILInstructionKind::BranchInst: {
const SILPhiArgument *arg = cast<BranchInst>(user)->getArgForOperand(oper);
followUses(arg);
return;
}
case SILInstructionKind::CondBranchInst: {
const SILPhiArgument *arg =
cast<CondBranchInst>(user)->getArgForOperand(oper);
if (arg) // If the use isn't the branch condition, follow it.
followUses(arg);
return;
}
// Look through ObjC closures.
case SILInstructionKind::StoreInst:
if (oper->getOperandNumber() == StoreInst::Src) {
if (auto *PBSI = dyn_cast<ProjectBlockStorageInst>(
cast<StoreInst>(user)->getDest())) {
SILValue storageAddr = PBSI->getOperand();
// The closure is stored to block storage. Recursively visit all
// uses of any initialized block storage values derived from this
// storage address..
for (Operand *oper : storageAddr->getUses()) {
if (auto *IBS = dyn_cast<InitBlockStorageHeaderInst>(oper->getUser()))
followUses(IBS);
}
return;
}
}
break;
case SILInstructionKind::IsEscapingClosureInst:
// May be generated by withoutActuallyEscaping.
return;
case SILInstructionKind::PartialApplyInst: {
// Recurse through partial_apply to handle special cases before handling
// ApplySites in general below.
PartialApplyInst *PAI = cast<PartialApplyInst>(user);
// Use the same logic as checkForViolationAtApply applied to a def-use
// traversal.
//
// checkForViolationAtApply recurses through partial_apply chains.
if (oper->get() == PAI->getCallee()) {
followUses(PAI);
return;
}
// checkForViolationAtApply also uses findClosuresForAppliedArg which in
// turn checks isPartialApplyOfReabstractionThunk.
//
// A closure with @inout_aliasable arguments may be applied to a
// thunk as "escaping", but as long as the thunk is only used as a
// '@noescape" type then it is safe.
if (isPartialApplyOfReabstractionThunk(PAI)) {
// Don't follow thunks that were generated by withoutActuallyEscaping.
SILFunction *thunkDef = PAI->getReferencedFunction();
if (!thunkDef->isWithoutActuallyEscapingThunk())
followUses(PAI);
return;
}
// Handle this use like a normal applied argument.
break;
}
};
// Handle ApplySites in general after checking PartialApply above.
if (isa<ApplySite>(user)) {
SILValue arg = oper->get();
auto argumentFnType = getSILFunctionTypeForValue(arg);
if (argumentFnType && argumentFnType->isNoEscape()) {
// Verify that the inverse operation, finding a partial_apply from a
// @noescape argument, is consistent.
TinyPtrVector<PartialApplyInst *> partialApplies;
findClosuresForFunctionValue(arg, partialApplies);
assert(!partialApplies.empty()
&& "cannot find partial_apply from @noescape function argument");
return;
}
llvm::dbgs() << "Applied argument must be @noescape function type: " << *arg;
}
else
llvm::dbgs() << "Unexpected partial_apply use: " << *user;
llvm_unreachable("A partial_apply with @inout_aliasable may only be "
"used as a @noescape function type argument.");
}