/// \brief Check if it is possible to devirtualize an Apply instruction
/// and a class member obtained using the class_method instruction into
/// a direct call to a specific member of a specific class.
///
/// \p AI is the apply to devirtualize.
/// \p ClassOrMetatypeType is the class type or metatype type we are
/// devirtualizing for.
/// return true if it is possible to devirtualize, false - otherwise.
bool swift::canDevirtualizeClassMethod(FullApplySite AI,
SILType ClassOrMetatypeType,
OptRemark::Emitter *ORE,
bool isEffectivelyFinalMethod) {
LLVM_DEBUG(llvm::dbgs() << " Trying to devirtualize : "
<< *AI.getInstruction());
SILModule &Mod = AI.getModule();
// First attempt to lookup the origin for our class method. The origin should
// either be a metatype or an alloc_ref.
LLVM_DEBUG(llvm::dbgs() << " Origin Type: " << ClassOrMetatypeType);
auto *MI = cast<MethodInst>(AI.getCallee());
// Find the implementation of the member which should be invoked.
auto *F = getTargetClassMethod(Mod, ClassOrMetatypeType, MI);
// If we do not find any such function, we have no function to devirtualize
// to... so bail.
if (!F) {
LLVM_DEBUG(llvm::dbgs() << " FAIL: Could not find matching VTable "
"or vtable method for this class.\n");
return false;
}
// We need to disable the “effectively final” opt if a function is inlinable
if (isEffectivelyFinalMethod && AI.getFunction()->getResilienceExpansion() ==
ResilienceExpansion::Minimal) {
LLVM_DEBUG(llvm::dbgs() << " FAIL: Could not optimize function "
"because it is an effectively-final inlinable: "
<< AI.getFunction()->getName() << "\n");
return false;
}
// Mandatory inlining does class method devirtualization. I'm not sure if this
// is really needed, but some test rely on this.
// So even for Onone functions we have to do it if the SILStage is raw.
if (F->getModule().getStage() != SILStage::Raw && !F->shouldOptimize()) {
// Do not consider functions that should not be optimized.
LLVM_DEBUG(llvm::dbgs() << " FAIL: Could not optimize function "
<< " because it is marked no-opt: " << F->getName()
<< "\n");
return false;
}
if (AI.getFunction()->isSerialized()) {
// function_ref inside fragile function cannot reference a private or
// hidden symbol.
if (!F->hasValidLinkageForFragileRef())
return false;
}
return true;
}
/// FIXME: Total hack to work around issues exposed while ripping call
/// graph maintenance from the inliner.
static void tryLinkCallee(FullApplySite Apply) {
auto *F = Apply.getCalleeFunction();
if (!F || F->isDefinition()) return;
auto &M = Apply.getFunction()->getModule();
M.linkFunction(F, SILModule::LinkingMode::LinkAll);
}
bool SILInliner::canInlineFunction(FullApplySite AI) {
// For now, we cannot inline begin_apply at all.
if (isa<BeginApplyInst>(AI))
return false;
return AI.getFunction() != &Original;
}
/// \brief Check if it is possible to devirtualize an Apply instruction
/// and a class member obtained using the class_method instruction into
/// a direct call to a specific member of a specific class.
///
/// \p AI is the apply to devirtualize.
/// \p ClassOrMetatypeType is the class type or metatype type we are
/// devirtualizing for.
/// return true if it is possible to devirtualize, false - otherwise.
bool swift::canDevirtualizeClassMethod(FullApplySite AI,
SILType ClassOrMetatypeType) {
DEBUG(llvm::dbgs() << " Trying to devirtualize : " << *AI.getInstruction());
SILModule &Mod = AI.getModule();
// First attempt to lookup the origin for our class method. The origin should
// either be a metatype or an alloc_ref.
DEBUG(llvm::dbgs() << " Origin Type: " << ClassOrMetatypeType);
auto *MI = cast<MethodInst>(AI.getCallee());
// Find the implementation of the member which should be invoked.
auto *F = getTargetClassMethod(Mod, ClassOrMetatypeType, MI);
// If we do not find any such function, we have no function to devirtualize
// to... so bail.
if (!F) {
DEBUG(llvm::dbgs() << " FAIL: Could not find matching VTable or "
"vtable method for this class.\n");
return false;
}
if (!F->shouldOptimize()) {
// Do not consider functions that should not be optimized.
DEBUG(llvm::dbgs() << " FAIL: Could not optimize function "
<< " because it is marked no-opt: " << F->getName()
<< "\n");
return false;
}
if (AI.getFunction()->isSerialized()) {
// function_ref inside fragile function cannot reference a private or
// hidden symbol.
if (!F->hasValidLinkageForFragileRef())
return false;
}
if (MI->isVolatile()) {
// dynamic dispatch is semantically required, can't devirtualize
return false;
}
return true;
}
SILValue SILInliner::borrowFunctionArgument(SILValue callArg,
FullApplySite AI) {
if (!AI.getFunction()->hasQualifiedOwnership()
|| callArg.getOwnershipKind() != ValueOwnershipKind::Owned) {
return callArg;
}
auto *borrow = getBuilder().createBeginBorrow(AI.getLoc(), callArg);
if (auto *tryAI = dyn_cast<TryApplyInst>(AI)) {
SILBuilder returnBuilder(tryAI->getNormalBB()->begin());
returnBuilder.createEndBorrow(AI.getLoc(), borrow, callArg);
SILBuilder throwBuilder(tryAI->getErrorBB()->begin());
throwBuilder.createEndBorrow(AI.getLoc(), borrow, callArg);
} else {
SILBuilder returnBuilder(std::next(AI.getInstruction()->getIterator()));
returnBuilder.createEndBorrow(AI.getLoc(), borrow, callArg);
}
return borrow;
}
/// 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 Check if it is possible to devirtualize an Apply instruction
/// and a class member obtained using the class_method instruction into
/// a direct call to a specific member of a specific class.
///
/// \p AI is the apply to devirtualize.
/// \p ClassOrMetatypeType is the class type or metatype type we are
/// devirtualizing for.
/// return true if it is possible to devirtualize, false - otherwise.
bool swift::canDevirtualizeClassMethod(FullApplySite AI,
SILType ClassOrMetatypeType) {
DEBUG(llvm::dbgs() << " Trying to devirtualize : " << *AI.getInstruction());
SILModule &Mod = AI.getModule();
// Bail if any generic types parameters of the class instance type are
// unbound.
// We cannot devirtualize unbound generic calls yet.
if (isClassWithUnboundGenericParameters(ClassOrMetatypeType, Mod))
return false;
// First attempt to lookup the origin for our class method. The origin should
// either be a metatype or an alloc_ref.
DEBUG(llvm::dbgs() << " Origin Type: " << ClassOrMetatypeType);
auto *CMI = cast<ClassMethodInst>(AI.getCallee());
// Find the implementation of the member which should be invoked.
auto *F = getTargetClassMethod(Mod, ClassOrMetatypeType, CMI->getMember());
// If we do not find any such function, we have no function to devirtualize
// to... so bail.
if (!F) {
DEBUG(llvm::dbgs() << " FAIL: Could not find matching VTable or "
"vtable method for this class.\n");
return false;
}
if (AI.getFunction()->isFragile()) {
// function_ref inside fragile function cannot reference a private or
// hidden symbol.
if (!(F->isFragile() || isValidLinkageForFragileRef(F->getLinkage()) ||
F->isExternalDeclaration()))
return false;
}
CanSILFunctionType GenCalleeType = F->getLoweredFunctionType();
auto Subs = getSubstitutionsForCallee(Mod, GenCalleeType,
ClassOrMetatypeType, AI);
// For polymorphic functions, bail if the number of substitutions is
// not the same as the number of expected generic parameters.
if (GenCalleeType->isPolymorphic()) {
auto GenericSig = GenCalleeType->getGenericSignature();
// Get the number of expected generic parameters, which
// is a sum of the number of explicit generic parameters
// and the number of their recursive member types exposed
// through protocol requirements.
auto DepTypes = GenericSig->getAllDependentTypes();
unsigned ExpectedGenParamsNum = 0;
for (auto DT: DepTypes) {
(void)DT;
ExpectedGenParamsNum++;
}
if (ExpectedGenParamsNum != Subs.size())
return false;
}
// Check if the optimizer knows how to cast the return type.
CanSILFunctionType SubstCalleeType = GenCalleeType;
if (GenCalleeType->isPolymorphic())
SubstCalleeType =
GenCalleeType->substGenericArgs(Mod, Mod.getSwiftModule(), Subs);
// If we have a direct return type, make sure we use the subst callee return
// type. If we have an indirect return type, AI's return type of the empty
// tuple should be ok.
SILType ReturnType = AI.getType();
if (!SubstCalleeType->hasIndirectResult()) {
ReturnType = SubstCalleeType->getSILResult();
}
if (!canCastValueToABICompatibleType(Mod, ReturnType, AI.getType()))
return false;
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;
}
/// \brief Inlines the callee of a given ApplyInst (which must be the value of a
/// FunctionRefInst referencing a function with a known body), into the caller
/// containing the ApplyInst, which must be the same function as provided to the
/// constructor of SILInliner. It only performs one step of inlining: it does
/// not recursively inline functions called by the callee.
///
/// It is the responsibility of the caller of this function to delete
/// the given ApplyInst when inlining is successful.
///
/// \returns true on success or false if it is unable to inline the function
/// (for any reason).
bool SILInliner::inlineFunction(FullApplySite AI, ArrayRef<SILValue> Args) {
SILFunction *CalleeFunction = &Original;
this->CalleeFunction = CalleeFunction;
// Do not attempt to inline an apply into its parent function.
if (AI.getFunction() == CalleeFunction)
return false;
SILFunction &F = getBuilder().getFunction();
if (CalleeFunction->getName() == "_TTSg5Vs4Int8___TFVs12_ArrayBufferg9_isNativeSb"
&& F.getName() == "_TTSg5Vs4Int8___TFVs12_ArrayBufferg8endIndexSi")
llvm::errs();
assert(AI.getFunction() && AI.getFunction() == &F &&
"Inliner called on apply instruction in wrong function?");
assert(((CalleeFunction->getRepresentation()
!= SILFunctionTypeRepresentation::ObjCMethod &&
CalleeFunction->getRepresentation()
!= SILFunctionTypeRepresentation::CFunctionPointer) ||
IKind == InlineKind::PerformanceInline) &&
"Cannot inline Objective-C methods or C functions in mandatory "
"inlining");
CalleeEntryBB = &*CalleeFunction->begin();
// Compute the SILLocation which should be used by all the inlined
// instructions.
if (IKind == InlineKind::PerformanceInline) {
Loc = InlinedLocation::getInlinedLocation(AI.getLoc());
} else {
assert(IKind == InlineKind::MandatoryInline && "Unknown InlineKind.");
Loc = MandatoryInlinedLocation::getMandatoryInlinedLocation(AI.getLoc());
}
auto AIScope = AI.getDebugScope();
// FIXME: Turn this into an assertion instead.
if (!AIScope)
AIScope = AI.getFunction()->getDebugScope();
if (IKind == InlineKind::MandatoryInline) {
// Mandatory inlining: every instruction inherits scope/location
// from the call site.
CallSiteScope = AIScope;
} else {
// Performance inlining. Construct a proper inline scope pointing
// back to the call site.
CallSiteScope = new (F.getModule())
SILDebugScope(AI.getLoc(), &F, AIScope);
assert(CallSiteScope->getParentFunction() == &F);
}
assert(CallSiteScope && "call site has no scope");
// Increment the ref count for the inlined function, so it doesn't
// get deleted before we can emit abstract debug info for it.
CalleeFunction->setInlined();
// If the caller's BB is not the last BB in the calling function, then keep
// track of the next BB so we always insert new BBs before it; otherwise,
// we just leave the new BBs at the end as they are by default.
auto IBI = std::next(SILFunction::iterator(AI.getParent()));
InsertBeforeBB = IBI != F.end() ? &*IBI : nullptr;
// Clear argument map and map ApplyInst arguments to the arguments of the
// callee's entry block.
ValueMap.clear();
assert(CalleeEntryBB->bbarg_size() == Args.size() &&
"Unexpected number of arguments to entry block of function?");
auto BAI = CalleeEntryBB->bbarg_begin();
for (auto AI = Args.begin(), AE = Args.end(); AI != AE; ++AI, ++BAI)
ValueMap.insert(std::make_pair(*BAI, *AI));
InstructionMap.clear();
BBMap.clear();
// Do not allow the entry block to be cloned again
SILBasicBlock::iterator InsertPoint =
SILBasicBlock::iterator(AI.getInstruction());
BBMap.insert(std::make_pair(CalleeEntryBB, AI.getParent()));
getBuilder().setInsertionPoint(InsertPoint);
// Recursively visit callee's BB in depth-first preorder, starting with the
// entry block, cloning all instructions other than terminators.
visitSILBasicBlock(CalleeEntryBB);
// If we're inlining into a normal apply and the callee's entry
// block ends in a return, then we can avoid a split.
if (auto nonTryAI = dyn_cast<ApplyInst>(AI)) {
if (ReturnInst *RI = dyn_cast<ReturnInst>(CalleeEntryBB->getTerminator())) {
// Replace all uses of the apply instruction with the operands of the
// return instruction, appropriately mapped.
nonTryAI->replaceAllUsesWith(remapValue(RI->getOperand()));
return true;
}
}
// If we're inlining into a try_apply, we already have a return-to BB.
SILBasicBlock *ReturnToBB;
if (auto tryAI = dyn_cast<TryApplyInst>(AI)) {
ReturnToBB = tryAI->getNormalBB();
// Otherwise, split the caller's basic block to create a return-to BB.
} else {
SILBasicBlock *CallerBB = AI.getParent();
// Split the BB and do NOT create a branch between the old and new
// BBs; we will create the appropriate terminator manually later.
ReturnToBB = CallerBB->splitBasicBlock(InsertPoint);
// Place the return-to BB after all the other mapped BBs.
if (InsertBeforeBB)
F.getBlocks().splice(SILFunction::iterator(InsertBeforeBB), F.getBlocks(),
SILFunction::iterator(ReturnToBB));
else
F.getBlocks().splice(F.getBlocks().end(), F.getBlocks(),
SILFunction::iterator(ReturnToBB));
// Create an argument on the return-to BB representing the returned value.
auto *RetArg = new (F.getModule()) SILArgument(ReturnToBB,
AI.getInstruction()->getType());
// Replace all uses of the ApplyInst with the new argument.
AI.getInstruction()->replaceAllUsesWith(RetArg);
}
// Now iterate over the callee BBs and fix up the terminators.
for (auto BI = BBMap.begin(), BE = BBMap.end(); BI != BE; ++BI) {
getBuilder().setInsertionPoint(BI->second);
// Modify return terminators to branch to the return-to BB, rather than
// trying to clone the ReturnInst.
if (ReturnInst *RI = dyn_cast<ReturnInst>(BI->first->getTerminator())) {
auto thrownValue = remapValue(RI->getOperand());
getBuilder().createBranch(Loc.getValue(), ReturnToBB,
thrownValue);
continue;
}
// Modify throw terminators to branch to the error-return BB, rather than
// trying to clone the ThrowInst.
if (ThrowInst *TI = dyn_cast<ThrowInst>(BI->first->getTerminator())) {
if (auto *A = dyn_cast<ApplyInst>(AI)) {
(void)A;
assert(A->isNonThrowing() &&
"apply of a function with error result must be non-throwing");
getBuilder().createUnreachable(Loc.getValue());
continue;
}
auto tryAI = cast<TryApplyInst>(AI);
auto returnedValue = remapValue(TI->getOperand());
getBuilder().createBranch(Loc.getValue(), tryAI->getErrorBB(),
returnedValue);
continue;
}
// Otherwise use normal visitor, which clones the existing instruction
// but remaps basic blocks and values.
visit(BI->first->getTerminator());
}
return true;
}
/// Return true if inlining this call site is profitable.
bool SILPerformanceInliner::isProfitableToInline(FullApplySite AI,
unsigned loopDepthOfAI,
DominanceAnalysis *DA,
SILLoopAnalysis *LA,
ConstantTracker &callerTracker,
unsigned &NumCallerBlocks) {
SILFunction *Callee = AI.getCalleeFunction();
if (Callee->getInlineStrategy() == AlwaysInline)
return true;
ConstantTracker constTracker(Callee, &callerTracker, AI);
DominanceInfo *DT = DA->get(Callee);
SILLoopInfo *LI = LA->get(Callee);
DominanceOrder domOrder(&Callee->front(), DT, Callee->size());
// Calculate the inlining cost of the callee.
unsigned CalleeCost = 0;
unsigned Benefit = InlineCostThreshold > 0 ? InlineCostThreshold :
RemovedCallBenefit;
Benefit += loopDepthOfAI * LoopBenefitFactor;
int testThreshold = TestThreshold;
while (SILBasicBlock *block = domOrder.getNext()) {
constTracker.beginBlock();
unsigned loopDepth = LI->getLoopDepth(block);
for (SILInstruction &I : *block) {
constTracker.trackInst(&I);
auto ICost = instructionInlineCost(I);
if (testThreshold >= 0) {
// We are in test-mode: use a simplified cost model.
CalleeCost += testCost(&I);
} else {
// Use the regular cost model.
CalleeCost += unsigned(ICost);
}
if (ApplyInst *AI = dyn_cast<ApplyInst>(&I)) {
// Check if the callee is passed as an argument. If so, increase the
// threshold, because inlining will (probably) eliminate the closure.
SILInstruction *def = constTracker.getDefInCaller(AI->getCallee());
if (def && (isa<FunctionRefInst>(def) || isa<PartialApplyInst>(def))) {
DEBUG(llvm::dbgs() << " Boost: apply const function at"
<< *AI);
Benefit += ConstCalleeBenefit + loopDepth * LoopBenefitFactor;
testThreshold *= 2;
}
}
}
// Don't count costs in blocks which are dead after inlining.
SILBasicBlock *takenBlock = getTakenBlock(block->getTerminator(),
constTracker);
if (takenBlock) {
Benefit += ConstTerminatorBenefit + TestOpt;
DEBUG(llvm::dbgs() << " Take bb" << takenBlock->getDebugID() <<
" of" << *block->getTerminator());
domOrder.pushChildrenIf(block, [=] (SILBasicBlock *child) {
return child->getSinglePredecessor() != block || child == takenBlock;
});
} else {
domOrder.pushChildren(block);
}
}
unsigned Threshold = Benefit; // The default.
if (testThreshold >= 0) {
// We are in testing mode.
Threshold = testThreshold;
} else if (AI.getFunction()->isThunk()) {
// Only inline trivial functions into thunks (which will not increase the
// code size).
Threshold = TrivialFunctionThreshold;
} else {
// The default case.
// We reduce the benefit if the caller is too large. For this we use a
// cubic function on the number of caller blocks. This starts to prevent
// inlining at about 800 - 1000 caller blocks.
unsigned blockMinus =
(NumCallerBlocks * NumCallerBlocks) / BlockLimitDenominator *
NumCallerBlocks / BlockLimitDenominator;
if (Threshold > blockMinus + TrivialFunctionThreshold)
Threshold -= blockMinus;
else
Threshold = TrivialFunctionThreshold;
}
if (CalleeCost > Threshold) {
DEBUG(llvm::dbgs() << " NO: Function too big to inline, "
"cost: " << CalleeCost << ", threshold: " << Threshold << "\n");
return false;
}
DEBUG(llvm::dbgs() << " YES: ready to inline, "
"cost: " << CalleeCost << ", threshold: " << Threshold << "\n");
NumCallerBlocks += Callee->size();
return true;
}
/// Insert monomorphic inline caches for a specific class or metatype
/// type \p SubClassTy.
static FullApplySite speculateMonomorphicTarget(FullApplySite AI,
SILType SubType,
CheckedCastBranchInst *&CCBI) {
CCBI = nullptr;
// Bail if this class_method cannot be devirtualized.
if (!canDevirtualizeClassMethod(AI, SubType))
return FullApplySite();
// Create a diamond shaped control flow and a checked_cast_branch
// instruction that checks the exact type of the object.
// This cast selects between two paths: one that calls the slow dynamic
// dispatch and one that calls the specific method.
auto It = AI.getInstruction()->getIterator();
SILFunction *F = AI.getFunction();
SILBasicBlock *Entry = AI.getParent();
// Iden is the basic block containing the direct call.
SILBasicBlock *Iden = F->createBasicBlock();
// Virt is the block containing the slow virtual call.
SILBasicBlock *Virt = F->createBasicBlock();
Iden->createBBArg(SubType);
SILBasicBlock *Continue = Entry->splitBasicBlock(It);
SILBuilderWithScope Builder(Entry, AI.getInstruction());
// Create the checked_cast_branch instruction that checks at runtime if the
// class instance is identical to the SILType.
ClassMethodInst *CMI = cast<ClassMethodInst>(AI.getCallee());
CCBI = Builder.createCheckedCastBranch(AI.getLoc(), /*exact*/ true,
CMI->getOperand(), SubType, Iden,
Virt);
It = CCBI->getIterator();
SILBuilderWithScope VirtBuilder(Virt, AI.getInstruction());
SILBuilderWithScope IdenBuilder(Iden, AI.getInstruction());
// This is the class reference downcasted into subclass SubType.
SILValue DownCastedClassInstance = Iden->getBBArg(0);
// Copy the two apply instructions into the two blocks.
FullApplySite IdenAI = CloneApply(AI, IdenBuilder);
FullApplySite VirtAI = CloneApply(AI, VirtBuilder);
// See if Continue has a release on self as the instruction right after the
// apply. If it exists, move it into position in the diamond.
if (auto *Release =
dyn_cast<StrongReleaseInst>(std::next(Continue->begin()))) {
if (Release->getOperand() == CMI->getOperand()) {
VirtBuilder.createStrongRelease(Release->getLoc(), CMI->getOperand());
IdenBuilder.createStrongRelease(Release->getLoc(),
DownCastedClassInstance);
Release->eraseFromParent();
}
}
// Create a PHInode for returning the return value from both apply
// instructions.
SILArgument *Arg = Continue->createBBArg(AI.getType());
if (!isa<TryApplyInst>(AI)) {
IdenBuilder.createBranch(AI.getLoc(), Continue,
ArrayRef<SILValue>(IdenAI.getInstruction()));
VirtBuilder.createBranch(AI.getLoc(), Continue,
ArrayRef<SILValue>(VirtAI.getInstruction()));
}
// Remove the old Apply instruction.
if (!isa<TryApplyInst>(AI))
AI.getInstruction()->replaceAllUsesWith(Arg);
auto *OriginalBB = AI.getParent();
AI.getInstruction()->eraseFromParent();
if (OriginalBB->empty())
OriginalBB->removeFromParent();
// Update the stats.
NumTargetsPredicted++;
// Devirtualize the apply instruction on the identical path.
auto NewInstPair = devirtualizeClassMethod(IdenAI, DownCastedClassInstance);
assert(NewInstPair.first && "Expected to be able to devirtualize apply!");
replaceDeadApply(IdenAI, NewInstPair.first);
// Split critical edges resulting from VirtAI.
if (auto *TAI = dyn_cast<TryApplyInst>(VirtAI)) {
auto *ErrorBB = TAI->getFunction()->createBasicBlock();
ErrorBB->createBBArg(TAI->getErrorBB()->getBBArg(0)->getType());
Builder.setInsertionPoint(ErrorBB);
Builder.createBranch(TAI->getLoc(), TAI->getErrorBB(),
{ErrorBB->getBBArg(0)});
auto *NormalBB = TAI->getFunction()->createBasicBlock();
NormalBB->createBBArg(TAI->getNormalBB()->getBBArg(0)->getType());
Builder.setInsertionPoint(NormalBB);
Builder.createBranch(TAI->getLoc(), TAI->getNormalBB(),
{NormalBB->getBBArg(0) });
Builder.setInsertionPoint(VirtAI.getInstruction());
SmallVector<SILValue, 4> Args;
for (auto Arg : VirtAI.getArguments()) {
Args.push_back(Arg);
}
FullApplySite NewVirtAI = Builder.createTryApply(VirtAI.getLoc(), VirtAI.getCallee(),
VirtAI.getSubstCalleeSILType(), VirtAI.getSubstitutions(),
Args, NormalBB, ErrorBB);
VirtAI.getInstruction()->eraseFromParent();
VirtAI = NewVirtAI;
}
return VirtAI;
}
bool SILPerformanceInliner::isProfitableToInline(FullApplySite AI,
Weight CallerWeight,
ConstantTracker &callerTracker,
int &NumCallerBlocks,
bool IsGeneric) {
SILFunction *Callee = AI.getReferencedFunction();
SILLoopInfo *LI = LA->get(Callee);
ShortestPathAnalysis *SPA = getSPA(Callee, LI);
assert(SPA->isValid());
ConstantTracker constTracker(Callee, &callerTracker, AI);
DominanceInfo *DT = DA->get(Callee);
SILBasicBlock *CalleeEntry = &Callee->front();
DominanceOrder domOrder(CalleeEntry, DT, Callee->size());
// Calculate the inlining cost of the callee.
int CalleeCost = 0;
int Benefit = 0;
// Start with a base benefit.
int BaseBenefit = RemovedCallBenefit;
const SILOptions &Opts = Callee->getModule().getOptions();
// For some reason -Ounchecked can accept a higher base benefit without
// increasing the code size too much.
if (Opts.Optimization == SILOptions::SILOptMode::OptimizeUnchecked)
BaseBenefit *= 2;
CallerWeight.updateBenefit(Benefit, BaseBenefit);
// Go through all blocks of the function, accumulate the cost and find
// benefits.
while (SILBasicBlock *block = domOrder.getNext()) {
constTracker.beginBlock();
Weight BlockW = SPA->getWeight(block, CallerWeight);
for (SILInstruction &I : *block) {
constTracker.trackInst(&I);
CalleeCost += (int)instructionInlineCost(I);
if (FullApplySite AI = FullApplySite::isa(&I)) {
// Check if the callee is passed as an argument. If so, increase the
// threshold, because inlining will (probably) eliminate the closure.
SILInstruction *def = constTracker.getDefInCaller(AI.getCallee());
if (def && (isa<FunctionRefInst>(def) || isa<PartialApplyInst>(def)))
BlockW.updateBenefit(Benefit, RemovedClosureBenefit);
} else if (auto *LI = dyn_cast<LoadInst>(&I)) {
// Check if it's a load from a stack location in the caller. Such a load
// might be optimized away if inlined.
if (constTracker.isStackAddrInCaller(LI->getOperand()))
BlockW.updateBenefit(Benefit, RemovedLoadBenefit);
} else if (auto *SI = dyn_cast<StoreInst>(&I)) {
// Check if it's a store to a stack location in the caller. Such a load
// might be optimized away if inlined.
if (constTracker.isStackAddrInCaller(SI->getDest()))
BlockW.updateBenefit(Benefit, RemovedStoreBenefit);
} else if (isa<StrongReleaseInst>(&I) || isa<ReleaseValueInst>(&I)) {
SILValue Op = stripCasts(I.getOperand(0));
if (SILArgument *Arg = dyn_cast<SILArgument>(Op)) {
if (Arg->isFunctionArg() && Arg->getArgumentConvention() ==
SILArgumentConvention::Direct_Guaranteed) {
BlockW.updateBenefit(Benefit, RefCountBenefit);
}
}
} else if (auto *BI = dyn_cast<BuiltinInst>(&I)) {
if (BI->getBuiltinInfo().ID == BuiltinValueKind::OnFastPath)
BlockW.updateBenefit(Benefit, FastPathBuiltinBenefit);
}
}
// Don't count costs in blocks which are dead after inlining.
SILBasicBlock *takenBlock = constTracker.getTakenBlock(block->getTerminator());
if (takenBlock) {
BlockW.updateBenefit(Benefit, RemovedTerminatorBenefit);
domOrder.pushChildrenIf(block, [=] (SILBasicBlock *child) {
return child->getSinglePredecessor() != block || child == takenBlock;
});
} else {
domOrder.pushChildren(block);
}
}
if (AI.getFunction()->isThunk()) {
// Only inline trivial functions into thunks (which will not increase the
// code size).
if (CalleeCost > TrivialFunctionThreshold)
return false;
DEBUG(
dumpCaller(AI.getFunction());
llvm::dbgs() << " decision {" << CalleeCost << " into thunk} " <<
Callee->getName() << '\n';
);
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;
}
/// \brief Check if it is possible to devirtualize an Apply instruction
/// and a class member obtained using the class_method instruction into
/// a direct call to a specific member of a specific class.
///
/// \p AI is the apply to devirtualize.
/// \p ClassOrMetatypeType is the class type or metatype type we are
/// devirtualizing for.
/// return true if it is possible to devirtualize, false - otherwise.
bool swift::canDevirtualizeClassMethod(FullApplySite AI,
SILType ClassOrMetatypeType) {
DEBUG(llvm::dbgs() << " Trying to devirtualize : " << *AI.getInstruction());
SILModule &Mod = AI.getModule();
// First attempt to lookup the origin for our class method. The origin should
// either be a metatype or an alloc_ref.
DEBUG(llvm::dbgs() << " Origin Type: " << ClassOrMetatypeType);
auto *MI = cast<MethodInst>(AI.getCallee());
// Find the implementation of the member which should be invoked.
auto *F = getTargetClassMethod(Mod, ClassOrMetatypeType, MI);
// If we do not find any such function, we have no function to devirtualize
// to... so bail.
if (!F) {
DEBUG(llvm::dbgs() << " FAIL: Could not find matching VTable or "
"vtable method for this class.\n");
return false;
}
if (!F->shouldOptimize()) {
// Do not consider functions that should not be optimized.
DEBUG(llvm::dbgs() << " FAIL: Could not optimize function "
<< " because it is marked no-opt: " << F->getName()
<< "\n");
return false;
}
if (AI.getFunction()->isFragile()) {
// function_ref inside fragile function cannot reference a private or
// hidden symbol.
if (!F->hasValidLinkageForFragileRef())
return false;
}
// Type of the actual function to be called.
CanSILFunctionType GenCalleeType = F->getLoweredFunctionType();
// Type of the actual function to be called with substitutions applied.
CanSILFunctionType SubstCalleeType = GenCalleeType;
// For polymorphic functions, bail if the number of substitutions is
// not the same as the number of expected generic parameters.
if (GenCalleeType->isPolymorphic()) {
// First, find proper list of substitutions for the concrete
// method to be called.
SmallVector<Substitution, 4> Subs;
getSubstitutionsForCallee(Mod, GenCalleeType,
ClassOrMetatypeType.getSwiftRValueType(),
AI, Subs);
SubstCalleeType = GenCalleeType->substGenericArgs(Mod, Subs);
}
// Check if the optimizer knows how to cast the return type.
SILType ReturnType = SubstCalleeType->getSILResult();
if (!canCastValueToABICompatibleType(Mod, ReturnType, AI.getType()))
return false;
return true;
}