/// Checks whether any of the arguments to the apply are closures and diagnoses
/// if any of the @inout_aliasable captures passed to those closures have
/// in-progress accesses that would conflict with any access the summary
/// says the closure would perform.
//
/// TODO: We currently fail to statically diagnose non-escaping closures pased
/// via @block_storage convention. To enforce this case, we should statically
/// recognize when the apply takes a block argument that has been initialized to
/// a non-escaping closure.
static void checkForViolationsInNoEscapeClosures(
const StorageMap &Accesses, FullApplySite FAS, AccessSummaryAnalysis *ASA,
llvm::SmallVectorImpl<ConflictingAccess> &ConflictingAccesses) {
SILFunction *Callee = FAS.getCalleeFunction();
if (Callee && !Callee->empty()) {
// Check for violation with directly called closure
checkForViolationWithCall(Accesses, Callee, 0, FAS.getArguments(), ASA,
ConflictingAccesses);
}
// Check for violation with closures passed as arguments
for (SILValue Argument : FAS.getArguments()) {
auto *PAI = lookThroughForPartialApply(Argument);
if (!PAI)
continue;
SILFunction *Closure = PAI->getCalleeFunction();
if (!Closure || Closure->empty())
continue;
// Check the closure's captures, which are a suffix of the closure's
// parameters.
unsigned StartIndex =
Closure->getArguments().size() - PAI->getNumCallArguments();
checkForViolationWithCall(Accesses, Closure, StartIndex,
PAI->getArguments(), ASA, ConflictingAccesses);
}
}
/// Attempt to devirtualize the given apply if possible, and return a
/// new instruction in that case, or nullptr otherwise.
DevirtualizationResult
swift::tryDevirtualizeApply(FullApplySite AI, ClassHierarchyAnalysis *CHA) {
DEBUG(llvm::dbgs() << " Trying to devirtualize: " << *AI.getInstruction());
// Devirtualize apply instructions that call witness_method instructions:
//
// %8 = witness_method $Optional<UInt16>, #LogicValue.boolValue!getter.1
// %9 = apply %8<Self = CodeUnit?>(%6#1) : ...
//
if (isa<WitnessMethodInst>(AI.getCallee()))
return tryDevirtualizeWitnessMethod(AI);
/// Optimize a class_method and alloc_ref pair into a direct function
/// reference:
///
/// \code
/// %XX = alloc_ref $Foo
/// %YY = class_method %XX : $Foo, #Foo.get!1 : $@convention(method)...
/// \endcode
///
/// or
///
/// %XX = metatype $...
/// %YY = class_method %XX : ...
///
/// into
///
/// %YY = function_ref @...
if (auto *CMI = dyn_cast<ClassMethodInst>(AI.getCallee())) {
auto &M = AI.getModule();
auto Instance = stripUpCasts(CMI->getOperand());
auto ClassType = Instance->getType();
if (ClassType.is<MetatypeType>())
ClassType = ClassType.getMetatypeInstanceType(M);
auto *CD = ClassType.getClassOrBoundGenericClass();
if (isEffectivelyFinalMethod(AI, ClassType, CD, CHA))
return tryDevirtualizeClassMethod(AI, Instance);
// Try to check if the exact dynamic type of the instance is statically
// known.
if (auto Instance = getInstanceWithExactDynamicType(CMI->getOperand(),
CMI->getModule(),
CHA))
return tryDevirtualizeClassMethod(AI, Instance);
}
if (isa<SuperMethodInst>(AI.getCallee())) {
if (AI.hasSelfArgument()) {
return tryDevirtualizeClassMethod(AI, AI.getSelfArgument());
}
// It is an invocation of a class method.
// Last operand is the metatype that should be used for dispatching.
return tryDevirtualizeClassMethod(AI, AI.getArguments().back());
}
return std::make_pair(nullptr, FullApplySite());
}
// A utility function for cloning the apply instruction.
static FullApplySite CloneApply(FullApplySite AI, SILBuilder &Builder) {
// Clone the Apply.
Builder.setCurrentDebugScope(AI.getDebugScope());
Builder.addOpenedArchetypeOperands(AI.getInstruction());
auto Args = AI.getArguments();
SmallVector<SILValue, 8> Ret(Args.size());
for (unsigned i = 0, e = Args.size(); i != e; ++i)
Ret[i] = Args[i];
FullApplySite NAI;
switch (AI.getInstruction()->getKind()) {
case SILInstructionKind::ApplyInst:
NAI = Builder.createApply(AI.getLoc(), AI.getCallee(),
AI.getSubstitutions(),
Ret,
cast<ApplyInst>(AI)->isNonThrowing());
break;
case SILInstructionKind::TryApplyInst: {
auto *TryApplyI = cast<TryApplyInst>(AI.getInstruction());
NAI = Builder.createTryApply(AI.getLoc(), AI.getCallee(),
AI.getSubstitutions(),
Ret,
TryApplyI->getNormalBB(),
TryApplyI->getErrorBB());
}
break;
default:
llvm_unreachable("Trying to clone an unsupported apply instruction");
}
NAI.getInstruction();
return NAI;
}
/// \brief Devirtualize an apply of a class method.
///
/// \p AI is the apply to devirtualize.
/// \p ClassOrMetatype is a class value or metatype value that is the
/// self argument of the apply we will devirtualize.
/// return the result value of the new ApplyInst if created one or null.
DevirtualizationResult swift::devirtualizeClassMethod(FullApplySite AI,
SILValue ClassOrMetatype) {
DEBUG(llvm::dbgs() << " Trying to devirtualize : " << *AI.getInstruction());
SILModule &Mod = AI.getModule();
auto *CMI = cast<ClassMethodInst>(AI.getCallee());
auto ClassOrMetatypeType = ClassOrMetatype.getType();
auto *F = getTargetClassMethod(Mod, ClassOrMetatypeType, CMI->getMember());
CanSILFunctionType GenCalleeType = F->getLoweredFunctionType();
auto Subs = getSubstitutionsForCallee(Mod, GenCalleeType,
ClassOrMetatypeType, AI);
CanSILFunctionType SubstCalleeType = GenCalleeType;
if (GenCalleeType->isPolymorphic())
SubstCalleeType = GenCalleeType->substGenericArgs(Mod, Mod.getSwiftModule(), Subs);
SILBuilderWithScope B(AI.getInstruction());
FunctionRefInst *FRI = B.createFunctionRef(AI.getLoc(), F);
// Create the argument list for the new apply, casting when needed
// in order to handle covariant indirect return types and
// contravariant argument types.
llvm::SmallVector<SILValue, 8> NewArgs;
auto Args = AI.getArguments();
auto ParamTypes = SubstCalleeType->getParameterSILTypes();
for (unsigned i = 0, e = Args.size() - 1; i != e; ++i)
NewArgs.push_back(castValueToABICompatibleType(&B, AI.getLoc(), Args[i],
Args[i].getType(),
ParamTypes[i]).getValue());
// Add the self argument, upcasting if required because we're
// calling a base class's method.
auto SelfParamTy = SubstCalleeType->getSelfParameter().getSILType();
NewArgs.push_back(castValueToABICompatibleType(&B, AI.getLoc(),
ClassOrMetatype,
ClassOrMetatypeType,
SelfParamTy).getValue());
// 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 ResultTy = AI.getType();
if (!SubstCalleeType->hasIndirectResult()) {
ResultTy = SubstCalleeType->getSILResult();
}
SILType SubstCalleeSILType =
SILType::getPrimitiveObjectType(SubstCalleeType);
FullApplySite NewAI;
SILBasicBlock *ResultBB = nullptr;
SILBasicBlock *NormalBB = nullptr;
SILValue ResultValue;
bool ResultCastRequired = false;
SmallVector<Operand *, 4> OriginalResultUses;
if (!isa<TryApplyInst>(AI)) {
NewAI = B.createApply(AI.getLoc(), FRI, SubstCalleeSILType, ResultTy,
Subs, NewArgs, cast<ApplyInst>(AI)->isNonThrowing());
ResultValue = SILValue(NewAI.getInstruction(), 0);
} else {
auto *TAI = cast<TryApplyInst>(AI);
// Create new normal and error BBs only if:
// - re-using a BB would create a critical edge
// - or, the result of the new apply would be of different
// type than the argument of the original normal BB.
if (TAI->getNormalBB()->getSinglePredecessor())
ResultBB = TAI->getNormalBB();
else {
ResultBB = B.getFunction().createBasicBlock();
ResultBB->createBBArg(ResultTy);
}
NormalBB = TAI->getNormalBB();
SILBasicBlock *ErrorBB = nullptr;
if (TAI->getErrorBB()->getSinglePredecessor())
ErrorBB = TAI->getErrorBB();
else {
ErrorBB = B.getFunction().createBasicBlock();
ErrorBB->createBBArg(TAI->getErrorBB()->getBBArg(0)->getType());
}
NewAI = B.createTryApply(AI.getLoc(), FRI, SubstCalleeSILType,
Subs, NewArgs,
ResultBB, ErrorBB);
if (ErrorBB != TAI->getErrorBB()) {
B.setInsertionPoint(ErrorBB);
B.createBranch(TAI->getLoc(), TAI->getErrorBB(),
{ErrorBB->getBBArg(0)});
}
// Does the result value need to be casted?
ResultCastRequired = ResultTy != NormalBB->getBBArg(0)->getType();
if (ResultBB != NormalBB)
B.setInsertionPoint(ResultBB);
else if (ResultCastRequired) {
B.setInsertionPoint(NormalBB->begin());
// Collect all uses, before casting.
for (auto *Use : NormalBB->getBBArg(0)->getUses()) {
OriginalResultUses.push_back(Use);
}
NormalBB->getBBArg(0)->replaceAllUsesWith(SILUndef::get(AI.getType(), Mod));
NormalBB->replaceBBArg(0, ResultTy, nullptr);
}
// The result value is passed as a parameter to the normal block.
ResultValue = ResultBB->getBBArg(0);
}
// Check if any casting is required for the return value.
ResultValue = castValueToABICompatibleType(&B, NewAI.getLoc(), ResultValue,
ResultTy, AI.getType()).getValue();
DEBUG(llvm::dbgs() << " SUCCESS: " << F->getName() << "\n");
NumClassDevirt++;
if (NormalBB) {
if (NormalBB != ResultBB) {
// If artificial normal BB was introduced, branch
// to the original normal BB.
B.createBranch(NewAI.getLoc(), NormalBB, { ResultValue });
} else if (ResultCastRequired) {
// Update all original uses by the new value.
for(auto *Use: OriginalResultUses) {
Use->set(ResultValue);
}
}
return std::make_pair(NewAI.getInstruction(), NewAI);
}
// We need to return a pair of values here:
// - the first one is the actual result of the devirtualized call, possibly
// casted into an appropriate type. This SILValue may be a BB arg, if it
// was a cast between optional types.
// - the second one is the new apply site.
return std::make_pair(ResultValue.getDef(), NewAI);
}
/// Update the callsite to pass in the correct arguments.
static void rewriteApplyInst(const CallSiteDescriptor &CSDesc,
SILFunction *NewF) {
FullApplySite AI = CSDesc.getApplyInst();
SingleValueInstruction *Closure = CSDesc.getClosure();
SILBuilderWithScope Builder(Closure);
FunctionRefInst *FRI = Builder.createFunctionRef(AI.getLoc(), NewF);
// Create the args for the new apply by removing the closure argument...
llvm::SmallVector<SILValue, 8> NewArgs;
unsigned Index = 0;
for (auto Arg : AI.getArguments()) {
if (Index != CSDesc.getClosureIndex())
NewArgs.push_back(Arg);
Index++;
}
// ... and appending the captured arguments. We also insert retains here at
// the location of the original closure. This is needed to balance the
// implicit release of all captured arguments that occurs when the partial
// apply is destroyed.
SILModule &M = NewF->getModule();
auto ClosureCalleeConv = CSDesc.getClosureCallee()->getConventions();
unsigned ClosureArgIdx =
ClosureCalleeConv.getNumSILArguments() - CSDesc.getNumArguments();
for (auto Arg : CSDesc.getArguments()) {
SILType ArgTy = Arg->getType();
// If our argument is of trivial type, continue...
if (ArgTy.isTrivial(M)) {
NewArgs.push_back(Arg);
++ClosureArgIdx;
continue;
}
auto ArgConvention =
ClosureCalleeConv.getSILArgumentConvention(ClosureArgIdx);
// Non-inout indirect arguments are not supported yet.
assert(ArgTy.isObject() ||
!isNonInoutIndirectSILArgument(Arg, ArgConvention));
// If argument is not an object and it is an inout parameter,
// continue...
if (!ArgTy.isObject() &&
!isNonInoutIndirectSILArgument(Arg, ArgConvention)) {
NewArgs.push_back(Arg);
++ClosureArgIdx;
continue;
}
// TODO: When we support address types, this code path will need to be
// updated.
// We need to balance the consumed argument of the new partial_apply in the
// specialized callee by a retain. If both the original partial_apply and
// the apply of the callee are in the same basic block we can assume they
// are executed the same number of times. Therefore it is sufficient to just
// retain the argument at the site of the original partial_apply.
//
// %closure = partial_apply (%arg)
// = apply %callee(%closure)
// =>
// retain %arg
// %closure = partial_apply (%arg)
// apply %specialized_callee(..., %arg)
//
// However, if they are not in the same basic block the callee might be
// executed more frequently than the closure (for example, if the closure is
// created in a loop preheader and the callee taking the closure is executed
// in the loop). In such a case we must keep the argument live across the
// call site of the callee and emit a matching retain for every invocation
// of the callee.
//
// %closure = partial_apply (%arg)
//
// while () {
// = %callee(%closure)
// }
// =>
// retain %arg
// %closure = partial_apply (%arg)
//
// while () {
// retain %arg
// apply %specialized_callee(.., %arg)
// }
// release %arg
//
if (AI.getParent() != Closure->getParent()) {
// Emit the retain and release that keeps the argument life across the
// callee using the closure.
CSDesc.extendArgumentLifetime(Arg, ArgConvention);
// Emit the retain that matches the captured argument by the
// partial_apply
// in the callee that is consumed by the partial_apply.
Builder.setInsertionPoint(AI.getInstruction());
Builder.createRetainValue(Closure->getLoc(), Arg,
Builder.getDefaultAtomicity());
} else {
Builder.createRetainValue(Closure->getLoc(), Arg,
Builder.getDefaultAtomicity());
}
NewArgs.push_back(Arg);
++ClosureArgIdx;
}
Builder.setInsertionPoint(AI.getInstruction());
FullApplySite NewAI;
switch (AI.getKind()) {
case FullApplySiteKind::TryApplyInst: {
auto *TAI = cast<TryApplyInst>(AI);
NewAI = Builder.createTryApply(AI.getLoc(), FRI,
SubstitutionMap(), NewArgs,
TAI->getNormalBB(), TAI->getErrorBB());
// If we passed in the original closure as @owned, then insert a release
// right after NewAI. This is to balance the +1 from being an @owned
// argument to AI.
if (!CSDesc.isClosureConsumed() || !CSDesc.closureHasRefSemanticContext()) {
break;
}
Builder.setInsertionPoint(TAI->getNormalBB()->begin());
Builder.createReleaseValue(Closure->getLoc(), Closure,
Builder.getDefaultAtomicity());
Builder.setInsertionPoint(TAI->getErrorBB()->begin());
Builder.createReleaseValue(Closure->getLoc(), Closure,
Builder.getDefaultAtomicity());
Builder.setInsertionPoint(AI.getInstruction());
break;
}
case FullApplySiteKind::ApplyInst: {
auto oldApply = cast<ApplyInst>(AI);
auto newApply = Builder.createApply(oldApply->getLoc(), FRI,
SubstitutionMap(),
NewArgs, oldApply->isNonThrowing());
// If we passed in the original closure as @owned, then insert a release
// right after NewAI. This is to balance the +1 from being an @owned
// argument to AI.
if (CSDesc.isClosureConsumed() && CSDesc.closureHasRefSemanticContext())
Builder.createReleaseValue(Closure->getLoc(), Closure,
Builder.getDefaultAtomicity());
// Replace all uses of the old apply with the new apply.
oldApply->replaceAllUsesWith(newApply);
break;
}
case FullApplySiteKind::BeginApplyInst:
llvm_unreachable("Unhandled case");
}
// Erase the old apply.
AI.getInstruction()->eraseFromParent();
// TODO: Maybe include invalidation code for CallSiteDescriptor after we erase
// AI from parent?
}
/// 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;
}
/// Update the callsite to pass in the correct arguments.
static void rewriteApplyInst(const CallSiteDescriptor &CSDesc,
SILFunction *NewF) {
FullApplySite AI = CSDesc.getApplyInst();
SILInstruction *Closure = CSDesc.getClosure();
SILBuilderWithScope Builder(Closure);
FunctionRefInst *FRI = Builder.createFunctionRef(AI.getLoc(), NewF);
// Create the args for the new apply by removing the closure argument...
llvm::SmallVector<SILValue, 8> NewArgs;
unsigned Index = 0;
for (auto Arg : AI.getArguments()) {
if (Index != CSDesc.getClosureIndex())
NewArgs.push_back(Arg);
Index++;
}
// ... and appending the captured arguments. We also insert retains here at
// the location of the original closure. This is needed to balance the
// implicit release of all captured arguments that occurs when the partial
// apply is destroyed.
SILModule &M = NewF->getModule();
for (auto Arg : CSDesc.getArguments()) {
NewArgs.push_back(Arg);
SILType ArgTy = Arg->getType();
// If our argument is of trivial type, continue...
if (ArgTy.isTrivial(M))
continue;
// TODO: When we support address types, this code path will need to be
// updated.
// We need to balance the consumed argument of the new partial_apply in the
// specialized callee by a retain. If both the original partial_apply and
// the apply of the callee are in the same basic block we can assume they
// are executed the same number of times. Therefore it is sufficient to just
// retain the argument at the site of the original partial_apply.
//
// %closure = partial_apply (%arg)
// = apply %callee(%closure)
// =>
// retain %arg
// %closure = partial_apply (%arg)
// apply %specialized_callee(..., %arg)
//
// However, if they are not in the same basic block the callee might be
// executed more frequently than the closure (for example, if the closure is
// created in a loop preheader and the callee taking the closure is executed
// in the loop). In such a case we must keep the argument live across the
// call site of the callee and emit a matching retain for every invocation
// of the callee.
//
// %closure = partial_apply (%arg)
//
// while () {
// = %callee(%closure)
// }
// =>
// retain %arg
// %closure = partial_apply (%arg)
//
// while () {
// retain %arg
// apply %specialized_callee(.., %arg)
// }
// release %arg
//
if (AI.getParent() != Closure->getParent()) {
// Emit the retain and release that keeps the argument life across the
// callee using the closure.
CSDesc.extendArgumentLifetime(Arg);
// Emit the retain that matches the captured argument by the partial_apply
// in the callee that is consumed by the partial_apply.
Builder.setInsertionPoint(AI.getInstruction());
Builder.createRetainValue(Closure->getLoc(), Arg, Atomicity::Atomic);
} else {
Builder.createRetainValue(Closure->getLoc(), Arg, Atomicity::Atomic);
}
}
SILType LoweredType = NewF->getLoweredType();
SILType ResultType = LoweredType.castTo<SILFunctionType>()->getSILResult();
Builder.setInsertionPoint(AI.getInstruction());
FullApplySite NewAI;
if (auto *TAI = dyn_cast<TryApplyInst>(AI)) {
NewAI = Builder.createTryApply(AI.getLoc(), FRI, LoweredType,
ArrayRef<Substitution>(),
NewArgs,
TAI->getNormalBB(), TAI->getErrorBB());
// If we passed in the original closure as @owned, then insert a release
// right after NewAI. This is to balance the +1 from being an @owned
// argument to AI.
if (CSDesc.isClosureConsumed() && CSDesc.closureHasRefSemanticContext()) {
Builder.setInsertionPoint(TAI->getNormalBB()->begin());
Builder.createReleaseValue(Closure->getLoc(), Closure, Atomicity::Atomic);
Builder.setInsertionPoint(TAI->getErrorBB()->begin());
Builder.createReleaseValue(Closure->getLoc(), Closure, Atomicity::Atomic);
Builder.setInsertionPoint(AI.getInstruction());
}
} else {
NewAI = Builder.createApply(AI.getLoc(), FRI, LoweredType,
ResultType, ArrayRef<Substitution>(),
NewArgs, cast<ApplyInst>(AI)->isNonThrowing());
// If we passed in the original closure as @owned, then insert a release
// right after NewAI. This is to balance the +1 from being an @owned
// argument to AI.
if (CSDesc.isClosureConsumed() && CSDesc.closureHasRefSemanticContext())
Builder.createReleaseValue(Closure->getLoc(), Closure, Atomicity::Atomic);
}
// Replace all uses of the old apply with the new apply.
if (isa<ApplyInst>(AI))
AI.getInstruction()->replaceAllUsesWith(NewAI.getInstruction());
// Erase the old apply.
AI.getInstruction()->eraseFromParent();
// TODO: Maybe include invalidation code for CallSiteDescriptor after we erase
// AI from parent?
}
/// 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;
}
/// Attempt to devirtualize the given apply if possible, and return a
/// new instruction in that case, or nullptr otherwise.
ApplySite swift::tryDevirtualizeApply(ApplySite AI,
ClassHierarchyAnalysis *CHA,
OptRemark::Emitter *ORE) {
LLVM_DEBUG(llvm::dbgs() << " Trying to devirtualize: "
<< *AI.getInstruction());
// Devirtualize apply instructions that call witness_method instructions:
//
// %8 = witness_method $Optional<UInt16>, #LogicValue.boolValue!getter.1
// %9 = apply %8<Self = CodeUnit?>(%6#1) : ...
//
if (isa<WitnessMethodInst>(AI.getCallee()))
return tryDevirtualizeWitnessMethod(AI, ORE);
// TODO: check if we can also de-virtualize partial applies of class methods.
FullApplySite FAS = FullApplySite::isa(AI.getInstruction());
if (!FAS)
return ApplySite();
/// Optimize a class_method and alloc_ref pair into a direct function
/// reference:
///
/// \code
/// %XX = alloc_ref $Foo
/// %YY = class_method %XX : $Foo, #Foo.get!1 : $@convention(method)...
/// \endcode
///
/// or
///
/// %XX = metatype $...
/// %YY = class_method %XX : ...
///
/// into
///
/// %YY = function_ref @...
if (auto *CMI = dyn_cast<ClassMethodInst>(FAS.getCallee())) {
auto &M = FAS.getModule();
auto Instance = stripUpCasts(CMI->getOperand());
auto ClassType = Instance->getType();
if (ClassType.is<MetatypeType>())
ClassType = ClassType.getMetatypeInstanceType(M);
auto *CD = ClassType.getClassOrBoundGenericClass();
if (isEffectivelyFinalMethod(FAS, ClassType, CD, CHA))
return tryDevirtualizeClassMethod(FAS, Instance, ORE,
true /*isEffectivelyFinalMethod*/);
// Try to check if the exact dynamic type of the instance is statically
// known.
if (auto Instance = getInstanceWithExactDynamicType(CMI->getOperand(),
CMI->getModule(),
CHA))
return tryDevirtualizeClassMethod(FAS, Instance, ORE);
if (auto ExactTy = getExactDynamicType(CMI->getOperand(), CMI->getModule(),
CHA)) {
if (ExactTy == CMI->getOperand()->getType())
return tryDevirtualizeClassMethod(FAS, CMI->getOperand(), ORE);
}
}
if (isa<SuperMethodInst>(FAS.getCallee())) {
if (FAS.hasSelfArgument()) {
return tryDevirtualizeClassMethod(FAS, FAS.getSelfArgument(), ORE);
}
// It is an invocation of a class method.
// Last operand is the metatype that should be used for dispatching.
return tryDevirtualizeClassMethod(FAS, FAS.getArguments().back(), ORE);
}
return ApplySite();
}