std::pair<Optional<SILValue>, SILLocation>
SILGenFunction::emitEpilogBB(SILLocation TopLevel) {
assert(ReturnDest.getBlock() && "no epilog bb prepared?!");
SILBasicBlock *epilogBB = ReturnDest.getBlock();
SILLocation ImplicitReturnFromTopLevel =
ImplicitReturnLocation::getImplicitReturnLoc(TopLevel);
SILValue returnValue;
Optional<SILLocation> returnLoc = None;
// If the current BB isn't terminated, and we require a return, then we
// are not allowed to fall off the end of the function and can't reach here.
if (NeedsReturn && B.hasValidInsertionPoint())
B.createUnreachable(ImplicitReturnFromTopLevel);
if (epilogBB->pred_empty()) {
bool hadArg = !epilogBB->bbarg_empty();
// If the epilog was not branched to at all, kill the BB and
// just emit the epilog into the current BB.
while (!epilogBB->empty())
epilogBB->back().eraseFromParent();
eraseBasicBlock(epilogBB);
// If the current bb is terminated then the epilog is just unreachable.
if (!B.hasValidInsertionPoint())
return { None, TopLevel };
// We emit the epilog at the current insertion point.
assert(!hadArg && "NeedsReturn is false but epilog had argument?!");
(void)hadArg;
returnLoc = ImplicitReturnFromTopLevel;
} else if (std::next(epilogBB->pred_begin()) == epilogBB->pred_end()
&& !B.hasValidInsertionPoint()) {
// If the epilog has a single predecessor and there's no current insertion
// point to fall through from, then we can weld the epilog to that
// predecessor BB.
bool needsArg = false;
if (!epilogBB->bbarg_empty()) {
assert(epilogBB->bbarg_size() == 1 && "epilog should take 0 or 1 args");
needsArg = true;
}
// Steal the branch argument as the return value if present.
SILBasicBlock *pred = *epilogBB->pred_begin();
BranchInst *predBranch = cast<BranchInst>(pred->getTerminator());
assert(predBranch->getArgs().size() == (needsArg ? 1 : 0) &&
"epilog predecessor arguments does not match block params");
if (needsArg) {
returnValue = predBranch->getArgs()[0];
// RAUW the old BB argument (if any) with the new value.
SILValue(*epilogBB->bbarg_begin(),0).replaceAllUsesWith(returnValue);
}
// If we are optimizing, we should use the return location from the single,
// previously processed, return statement if any.
if (predBranch->getLoc().is<ReturnLocation>()) {
returnLoc = predBranch->getLoc();
} else {
returnLoc = ImplicitReturnFromTopLevel;
}
// Kill the branch to the now-dead epilog BB.
pred->erase(predBranch);
// Move any instructions from the EpilogBB to the end of the 'pred' block.
pred->spliceAtEnd(epilogBB);
// Finally we can erase the epilog BB.
eraseBasicBlock(epilogBB);
// Emit the epilog into its former predecessor.
B.setInsertionPoint(pred);
} else {
// Move the epilog block to the end of the ordinary section.
auto endOfOrdinarySection =
(StartOfPostmatter ? SILFunction::iterator(StartOfPostmatter) : F.end());
B.moveBlockTo(epilogBB, endOfOrdinarySection);
// Emit the epilog into the epilog bb. Its argument is the return value.
if (!epilogBB->bbarg_empty()) {
assert(epilogBB->bbarg_size() == 1 && "epilog should take 0 or 1 args");
returnValue = epilogBB->bbarg_begin()[0];
}
// If we are falling through from the current block, the return is implicit.
B.emitBlock(epilogBB, ImplicitReturnFromTopLevel);
}
// Emit top-level cleanups into the epilog block.
assert(!Cleanups.hasAnyActiveCleanups(getCleanupsDepth(),
ReturnDest.getDepth()) &&
"emitting epilog in wrong scope");
auto cleanupLoc = CleanupLocation::get(TopLevel);
Cleanups.emitCleanupsForReturn(cleanupLoc);
// If the return location is known to be that of an already
// processed return, use it. (This will get triggered when the
// epilog logic is simplified.)
//
// Otherwise make the ret instruction part of the cleanups.
if (!returnLoc) returnLoc = cleanupLoc;
return { returnValue, *returnLoc };
}
std::pair<Optional<SILValue>, SILLocation>
SILGenFunction::emitEpilogBB(SILLocation TopLevel) {
assert(ReturnDest.getBlock() && "no epilog bb prepared?!");
SILBasicBlock *epilogBB = ReturnDest.getBlock();
SILLocation ImplicitReturnFromTopLevel =
ImplicitReturnLocation::getImplicitReturnLoc(TopLevel);
SmallVector<SILValue, 4> directResults;
Optional<SILLocation> returnLoc = None;
// If the current BB isn't terminated, and we require a return, then we
// are not allowed to fall off the end of the function and can't reach here.
if (NeedsReturn && B.hasValidInsertionPoint())
B.createUnreachable(ImplicitReturnFromTopLevel);
if (epilogBB->pred_empty()) {
// If the epilog was not branched to at all, kill the BB and
// just emit the epilog into the current BB.
while (!epilogBB->empty())
epilogBB->back().eraseFromParent();
eraseBasicBlock(epilogBB);
// If the current bb is terminated then the epilog is just unreachable.
if (!B.hasValidInsertionPoint())
return { None, TopLevel };
// We emit the epilog at the current insertion point.
returnLoc = ImplicitReturnFromTopLevel;
} else if (std::next(epilogBB->pred_begin()) == epilogBB->pred_end()
&& !B.hasValidInsertionPoint()) {
// If the epilog has a single predecessor and there's no current insertion
// point to fall through from, then we can weld the epilog to that
// predecessor BB.
// Steal the branch argument as the return value if present.
SILBasicBlock *pred = *epilogBB->pred_begin();
BranchInst *predBranch = cast<BranchInst>(pred->getTerminator());
assert(predBranch->getArgs().size() == epilogBB->bbarg_size() &&
"epilog predecessor arguments does not match block params");
for (auto index : indices(predBranch->getArgs())) {
SILValue result = predBranch->getArgs()[index];
directResults.push_back(result);
epilogBB->getBBArg(index)->replaceAllUsesWith(result);
}
// If we are optimizing, we should use the return location from the single,
// previously processed, return statement if any.
if (predBranch->getLoc().is<ReturnLocation>()) {
returnLoc = predBranch->getLoc();
} else {
returnLoc = ImplicitReturnFromTopLevel;
}
// Kill the branch to the now-dead epilog BB.
pred->erase(predBranch);
// Move any instructions from the EpilogBB to the end of the 'pred' block.
pred->spliceAtEnd(epilogBB);
// Finally we can erase the epilog BB.
eraseBasicBlock(epilogBB);
// Emit the epilog into its former predecessor.
B.setInsertionPoint(pred);
} else {
// Move the epilog block to the end of the ordinary section.
auto endOfOrdinarySection = StartOfPostmatter;
B.moveBlockTo(epilogBB, endOfOrdinarySection);
// Emit the epilog into the epilog bb. Its arguments are the
// direct results.
directResults.append(epilogBB->bbarg_begin(), epilogBB->bbarg_end());
// If we are falling through from the current block, the return is implicit.
B.emitBlock(epilogBB, ImplicitReturnFromTopLevel);
}
// Emit top-level cleanups into the epilog block.
assert(!Cleanups.hasAnyActiveCleanups(getCleanupsDepth(),
ReturnDest.getDepth()) &&
"emitting epilog in wrong scope");
auto cleanupLoc = CleanupLocation::get(TopLevel);
Cleanups.emitCleanupsForReturn(cleanupLoc);
// If the return location is known to be that of an already
// processed return, use it. (This will get triggered when the
// epilog logic is simplified.)
//
// Otherwise make the ret instruction part of the cleanups.
if (!returnLoc) returnLoc = cleanupLoc;
// Build the return value. We don't do this if there are no direct
// results; this can happen for void functions, but also happens when
// prepareEpilog was asked to not add result arguments to the epilog
// block.
SILValue returnValue;
if (!directResults.empty()) {
assert(directResults.size()
== F.getLoweredFunctionType()->getNumDirectResults());
returnValue = buildReturnValue(*this, TopLevel, directResults);
}
return { returnValue, *returnLoc };
}
/// \brief Populate the body of the cloned closure, modifying instructions as
/// necessary. This is where we create the actual specialized BB Arguments.
void ClosureSpecCloner::populateCloned() {
SILFunction *Cloned = getCloned();
SILModule &M = Cloned->getModule();
SILFunction *ClosureUser = CallSiteDesc.getApplyCallee();
// Create arguments for the entry block.
SILBasicBlock *ClosureUserEntryBB = &*ClosureUser->begin();
SILBasicBlock *ClonedEntryBB = new (M) SILBasicBlock(Cloned);
// Remove the closure argument.
SILArgument *ClosureArg = nullptr;
for (size_t i = 0, e = ClosureUserEntryBB->bbarg_size(); i != e; ++i) {
SILArgument *Arg = ClosureUserEntryBB->getBBArg(i);
if (i == CallSiteDesc.getClosureIndex()) {
ClosureArg = Arg;
continue;
}
// Otherwise, create a new argument which copies the original argument
SILValue MappedValue =
new (M) SILArgument(ClonedEntryBB, Arg->getType(), Arg->getDecl());
ValueMap.insert(std::make_pair(Arg, MappedValue));
}
// Next we need to add in any arguments that are not captured as arguments to
// the cloned function.
//
// We do not insert the new mapped arguments into the value map since there by
// definition is nothing in the partial apply user function that references
// such arguments. After this pass is done the only thing that will reference
// the arguments is the partial apply that we will create.
SILFunction *ClosedOverFun = CallSiteDesc.getClosureCallee();
CanSILFunctionType ClosedOverFunTy = ClosedOverFun->getLoweredFunctionType();
unsigned NumTotalParams = ClosedOverFunTy->getParameters().size();
unsigned NumNotCaptured = NumTotalParams - CallSiteDesc.getNumArguments();
llvm::SmallVector<SILValue, 4> NewPAIArgs;
for (auto &PInfo : ClosedOverFunTy->getParameters().slice(NumNotCaptured)) {
SILValue MappedValue =
new (M) SILArgument(ClonedEntryBB, PInfo.getSILType());
NewPAIArgs.push_back(MappedValue);
}
SILBuilder &Builder = getBuilder();
Builder.setInsertionPoint(ClonedEntryBB);
// Clone FRI and PAI, and replace usage of the removed closure argument
// with result of cloned PAI.
SILValue FnVal =
Builder.createFunctionRef(CallSiteDesc.getLoc(), ClosedOverFun);
auto *NewClosure = CallSiteDesc.createNewClosure(Builder, FnVal, NewPAIArgs);
ValueMap.insert(std::make_pair(ClosureArg, SILValue(NewClosure)));
BBMap.insert(std::make_pair(ClosureUserEntryBB, ClonedEntryBB));
// Recursively visit original BBs in depth-first preorder, starting with the
// entry block, cloning all instructions other than terminators.
visitSILBasicBlock(ClosureUserEntryBB);
// Now iterate over the BBs and fix up the terminators.
for (auto BI = BBMap.begin(), BE = BBMap.end(); BI != BE; ++BI) {
Builder.setInsertionPoint(BI->second);
visit(BI->first->getTerminator());
}
// Then insert a release in all non failure exit BBs if our partial apply was
// guaranteed. This is b/c it was passed at +0 originally and we need to
// balance the initial increment of the newly created closure.
if (CallSiteDesc.isClosureGuaranteed() &&
CallSiteDesc.closureHasRefSemanticContext()) {
for (SILBasicBlock *BB : CallSiteDesc.getNonFailureExitBBs()) {
SILBasicBlock *OpBB = BBMap[BB];
TermInst *TI = OpBB->getTerminator();
auto Loc = CleanupLocation::get(NewClosure->getLoc());
// If we have a return, we place the release right before it so we know
// that it will be executed at the end of the epilogue.
if (isa<ReturnInst>(TI)) {
Builder.setInsertionPoint(TI);
Builder.createReleaseValue(Loc, SILValue(NewClosure),
Atomicity::Atomic);
continue;
}
// We use casts where findAllNonFailureExitBBs should have made sure that
// this is true. This will ensure that the code is updated when we hit the
// cast failure in debug builds.
auto *Unreachable = cast<UnreachableInst>(TI);
auto PrevIter = std::prev(SILBasicBlock::iterator(Unreachable));
auto NoReturnApply = FullApplySite::isa(&*PrevIter);
// We insert the release value right before the no return apply so that if
// the partial apply is passed into the no-return function as an @owned
// value, we will retain the partial apply before we release it and
// potentially eliminate it.
Builder.setInsertionPoint(NoReturnApply.getInstruction());
Builder.createReleaseValue(Loc, SILValue(NewClosure), Atomicity::Atomic);
}
}
}
