void SILPerformanceInliner::collectAppliesToInline(
SILFunction *Caller, SmallVectorImpl<FullApplySite> &Applies,
DominanceAnalysis *DA, SILLoopAnalysis *LA) {
DominanceInfo *DT = DA->get(Caller);
SILLoopInfo *LI = LA->get(Caller);
ConstantTracker constTracker(Caller);
DominanceOrder domOrder(&Caller->front(), DT, Caller->size());
// Go through all instructions and find candidates for inlining.
// We do this in dominance order for the constTracker.
SmallVector<FullApplySite, 8> InitialCandidates;
while (SILBasicBlock *block = domOrder.getNext()) {
constTracker.beginBlock();
unsigned loopDepth = LI->getLoopDepth(block);
for (auto I = block->begin(), E = block->end(); I != E; ++I) {
constTracker.trackInst(&*I);
if (!FullApplySite::isa(&*I))
continue;
FullApplySite AI = FullApplySite(&*I);
DEBUG(llvm::dbgs() << " Check:" << *I);
auto *Callee = getEligibleFunction(AI);
if (Callee) {
if (isProfitableToInline(AI, loopDepth, DA, LA, constTracker))
InitialCandidates.push_back(AI);
}
}
domOrder.pushChildrenIf(block, [&] (SILBasicBlock *child) {
if (ColdBlockInfo::isSlowPath(block, child)) {
// Handle cold blocks separately.
visitColdBlocks(InitialCandidates, child, DT);
return false;
}
return true;
});
}
// Calculate how many times a callee is called from this caller.
llvm::DenseMap<SILFunction *, unsigned> CalleeCount;
for (auto AI : InitialCandidates) {
SILFunction *Callee = AI.getCalleeFunction();
assert(Callee && "apply_inst does not have a direct callee anymore");
CalleeCount[Callee]++;
}
// Now copy each candidate callee that has a small enough number of
// call sites into the final set of call sites.
for (auto AI : InitialCandidates) {
SILFunction *Callee = AI.getCalleeFunction();
assert(Callee && "apply_inst does not have a direct callee anymore");
const unsigned CallsToCalleeThreshold = 1024;
if (CalleeCount[Callee] <= CallsToCalleeThreshold)
Applies.push_back(AI);
}
}
/// 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;
}
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;
}