/// In this case, the unconditional branch at the end of the first if can be
/// revectored to the false side of the second if.
class JumpThreadingPass : public PassInfoMixin<JumpThreadingPass> {
- TargetLibraryInfo *TLI;
- TargetTransformInfo *TTI;
- LazyValueInfo *LVI;
- AAResults *AA;
- DomTreeUpdater *DTU;
- std::unique_ptr<BlockFrequencyInfo> BFI;
- std::unique_ptr<BranchProbabilityInfo> BPI;
- bool HasProfileData = false;
+ Function *F = nullptr;
+ FunctionAnalysisManager *FAM = nullptr;
+ TargetLibraryInfo *TLI = nullptr;
+ TargetTransformInfo *TTI = nullptr;
+ LazyValueInfo *LVI = nullptr;
+ AAResults *AA = nullptr;
+ DomTreeUpdater *DTU = nullptr;
+ Optional<BlockFrequencyInfo *> BFI = None;
+ Optional<BranchProbabilityInfo *> BPI = None;
bool HasGuards = false;
+ bool HasProfile = false;
#ifndef LLVM_ENABLE_ABI_BREAKING_CHECKS
SmallPtrSet<const BasicBlock *, 16> LoopHeaders;
#else
JumpThreadingPass(int T = -1);
// Glue for old PM.
- bool runImpl(Function &F, TargetLibraryInfo *TLI, TargetTransformInfo *TTI,
+ bool runImpl(Function &F, FunctionAnalysisManager *FAM,
+ TargetLibraryInfo *TLI, TargetTransformInfo *TTI,
LazyValueInfo *LVI, AAResults *AA, DomTreeUpdater *DTU,
- bool HasProfileData, std::unique_ptr<BlockFrequencyInfo> BFI,
- std::unique_ptr<BranchProbabilityInfo> BPI);
+ Optional<BlockFrequencyInfo *> BFI,
+ Optional<BranchProbabilityInfo *> BPI);
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
- void releaseMemory() {
- BFI.reset();
- BPI.reset();
- }
-
void findLoopHeaders(Function &F);
bool processBlock(BasicBlock *BB);
bool maybeMergeBasicBlockIntoOnlyPred(BasicBlock *BB);
BasicBlock *NewBB, BasicBlock *SuccBB);
/// Check if the block has profile metadata for its outgoing edges.
bool doesBlockHaveProfileData(BasicBlock *BB);
+
+ /// Returns an existing instance of BPI if any, otherwise nullptr. By
+ /// "existing" we mean either cached result provided by FunctionAnalysisManger
+ /// or created by preceding call to 'getOrCreateBPI'.
+ BranchProbabilityInfo *getBPI();
+
+ /// Returns an existing instance of BFI if any, otherwise nullptr. By
+ /// "existing" we mean either cached result provided by FunctionAnalysisManger
+ /// or created by preceding call to 'getOrCreateBFI'.
+ BlockFrequencyInfo *getBFI();
+
+ /// Returns an existing instance of BPI if any, otherwise:
+ /// if 'HasProfile' is true creates new instance through
+ /// FunctionAnalysisManager, otherwise nullptr.
+ BranchProbabilityInfo *getOrCreateBPI();
+
+ /// Returns an existing instance of BFI if any, otherwise:
+ /// if 'HasProfile' is true creates new instance through
+ /// FunctionAnalysisManager, otherwise nullptr.
+ BlockFrequencyInfo *getOrCreateBFI();
};
} // end namespace llvm
AU.addRequired<TargetLibraryInfoWrapperPass>();
AU.addRequired<TargetTransformInfoWrapperPass>();
}
-
- void releaseMemory() override { Impl.releaseMemory(); }
};
} // end anonymous namespace
BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
}
- bool Changed = Impl.runImpl(F, TLI, TTI, LVI, AA, &DTU, F.hasProfileData(),
- std::move(BFI), std::move(BPI));
+ bool Changed =
+ Impl.runImpl(F, nullptr, TLI, TTI, LVI, AA, &DTU, BFI.get(), BPI.get());
if (PrintLVIAfterJumpThreading) {
dbgs() << "LVI for function '" << F.getName() << "':\n";
LVI->printLVI(F, DTU.getDomTree(), dbgs());
auto &AA = AM.getResult<AAManager>(F);
DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
- std::unique_ptr<BlockFrequencyInfo> BFI;
- std::unique_ptr<BranchProbabilityInfo> BPI;
- if (F.hasProfileData()) {
- LoopInfo LI{DT};
- BPI.reset(new BranchProbabilityInfo(F, LI, &TLI));
- BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
- }
-
- bool Changed = runImpl(F, &TLI, &TTI, &LVI, &AA, &DTU, F.hasProfileData(),
- std::move(BFI), std::move(BPI));
+ bool Changed =
+ runImpl(F, &AM, &TLI, &TTI, &LVI, &AA, &DTU, None, None);
if (PrintLVIAfterJumpThreading) {
dbgs() << "LVI for function '" << F.getName() << "':\n";
if (!Changed)
return PreservedAnalyses::all();
+
+#ifndef NDEBUG
+ DTU.flush();
+#if defined(EXPENSIVE_CHECKS)
+ assert(DT.verify(DominatorTree::VerificationLevel::Full) &&
+ "DT broken after JumpThreading");
+#else
+ assert(
+ DT.verify(DominatorTree::VerificationLevel::Fast)
+ && "DT broken after JumpThreading");
+#endif
+#endif
+
PreservedAnalyses PA;
PA.preserve<DominatorTreeAnalysis>();
PA.preserve<LazyValueAnalysis>();
+
+ // TODO: We would like to preserve BPI/BFI. Enable once all paths update them.
+ // TODO: Would be nice to verify BPI/BFI consistency as well.
+
return PA;
}
-bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
+bool JumpThreadingPass::runImpl(Function &F_, FunctionAnalysisManager *FAM_, TargetLibraryInfo *TLI_,
TargetTransformInfo *TTI_, LazyValueInfo *LVI_,
AliasAnalysis *AA_, DomTreeUpdater *DTU_,
- bool HasProfileData_,
- std::unique_ptr<BlockFrequencyInfo> BFI_,
- std::unique_ptr<BranchProbabilityInfo> BPI_) {
- LLVM_DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
+ Optional<BlockFrequencyInfo *> BFI_,
+ Optional<BranchProbabilityInfo *> BPI_) {
+ LLVM_DEBUG(dbgs() << "Jump threading on function '" << F_.getName() << "'\n");
+ F = &F_;
+ FAM = FAM_;
TLI = TLI_;
TTI = TTI_;
LVI = LVI_;
AA = AA_;
DTU = DTU_;
- BFI.reset();
- BPI.reset();
- // When profile data is available, we need to update edge weights after
- // successful jump threading, which requires both BPI and BFI being available.
- HasProfileData = HasProfileData_;
- auto *GuardDecl = F.getParent()->getFunction(
+ BFI = BFI_;
+ BPI = BPI_;
+ auto *GuardDecl = F->getParent()->getFunction(
Intrinsic::getName(Intrinsic::experimental_guard));
HasGuards = GuardDecl && !GuardDecl->use_empty();
- if (HasProfileData) {
- BPI = std::move(BPI_);
- BFI = std::move(BFI_);
- }
// Reduce the number of instructions duplicated when optimizing strictly for
// size.
if (BBDuplicateThreshold.getNumOccurrences())
BBDupThreshold = BBDuplicateThreshold;
- else if (F.hasFnAttribute(Attribute::MinSize))
+ else if (F->hasFnAttribute(Attribute::MinSize))
BBDupThreshold = 3;
else
BBDupThreshold = DefaultBBDupThreshold;
assert(DTU && "DTU isn't passed into JumpThreading before using it.");
assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed.");
DominatorTree &DT = DTU->getDomTree();
- for (auto &BB : F)
+ for (auto &BB : *F) {
if (!DT.isReachableFromEntry(&BB))
Unreachable.insert(&BB);
+ }
if (!ThreadAcrossLoopHeaders)
- findLoopHeaders(F);
+ findLoopHeaders(*F);
+
+ HasProfile = llvm::any_of(*F, [&](BasicBlock &BB) {
+ return this->doesBlockHaveProfileData(&BB);
+ });
bool EverChanged = false;
bool Changed;
do {
Changed = false;
- for (auto &BB : F) {
+ for (auto &BB : *F) {
if (Unreachable.count(&BB))
continue;
while (processBlock(&BB)) // Thread all of the branches we can over BB.
// Stop processing BB if it's the entry or is now deleted. The following
// routines attempt to eliminate BB and locating a suitable replacement
// for the entry is non-trivial.
- if (&BB == &F.getEntryBlock() || DTU->isBBPendingDeletion(&BB))
+ if (&BB == &F->getEntryBlock() || DTU->isBBPendingDeletion(&BB))
continue;
if (pred_empty(&BB)) {
<< '\n');
++NumFolds;
ConstantFoldTerminator(BB, true, nullptr, DTU);
- if (HasProfileData)
+ if (auto *BPI = getBPI())
BPI->eraseBlock(BB);
return true;
}
FICond->eraseFromParent();
DTU->applyUpdatesPermissive({{DominatorTree::Delete, BB, RemoveSucc}});
- if (HasProfileData)
+ if (auto *BPI = getBPI())
BPI->eraseBlock(BB);
return true;
}
++NumFolds;
Term->eraseFromParent();
DTU->applyUpdatesPermissive(Updates);
- if (HasProfileData)
+ if (auto *BPI = getBPI())
BPI->eraseBlock(BB);
// If the condition is now dead due to the removal of the old terminator,
LLVM_DEBUG(dbgs() << " Threading through '" << PredBB->getName() << "' and '"
<< BB->getName() << "'\n");
+ // Build BPI/BFI before any changes are made to IR.
+ auto *BPI = getOrCreateBPI();
+ auto *BFI = getOrCreateBFI();
+
BranchInst *CondBr = cast<BranchInst>(BB->getTerminator());
BranchInst *PredBBBranch = cast<BranchInst>(PredBB->getTerminator());
NewBB->moveAfter(PredBB);
// Set the block frequency of NewBB.
- if (HasProfileData) {
+ if (BFI) {
+ assert(BPI && "It's expected BPI to exist along with BFI");
auto NewBBFreq = BFI->getBlockFreq(PredPredBB) *
BPI->getEdgeProbability(PredPredBB, PredBB);
BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
cloneInstructions(PredBB->begin(), PredBB->end(), NewBB, PredPredBB);
// Copy the edge probabilities from PredBB to NewBB.
- if (HasProfileData)
+ if (BPI)
BPI->copyEdgeProbabilities(PredBB, NewBB);
// Update the terminator of PredPredBB to jump to NewBB instead of PredBB.
assert(!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) &&
"Don't thread across loop headers");
+ // Build BPI/BFI before any changes are made to IR.
+ auto *BPI = getOrCreateBPI();
+ auto *BFI = getOrCreateBFI();
+
// And finally, do it! Start by factoring the predecessors if needed.
BasicBlock *PredBB;
if (PredBBs.size() == 1)
NewBB->moveAfter(PredBB);
// Set the block frequency of NewBB.
- if (HasProfileData) {
+ if (BFI) {
+ assert(BPI && "It's expected BPI to exist along with BFI");
auto NewBBFreq =
BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
// Collect the frequencies of all predecessors of BB, which will be used to
// update the edge weight of the result of splitting predecessors.
DenseMap<BasicBlock *, BlockFrequency> FreqMap;
- if (HasProfileData)
+ auto *BFI = getBFI();
+ auto *BPI = getBPI();
+ if (BFI)
for (auto *Pred : Preds)
FreqMap.insert(std::make_pair(
Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
for (auto *Pred : predecessors(NewBB)) {
Updates.push_back({DominatorTree::Delete, Pred, BB});
Updates.push_back({DominatorTree::Insert, Pred, NewBB});
- if (HasProfileData) // Update frequencies between Pred -> NewBB.
+ if (BFI) // Update frequencies between Pred -> NewBB.
NewBBFreq += FreqMap.lookup(Pred);
}
- if (HasProfileData) // Apply the summed frequency to NewBB.
+ if (BFI) // Apply the summed frequency to NewBB.
BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
}
bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
const Instruction *TI = BB->getTerminator();
- assert(TI->getNumSuccessors() > 1 && "not a split");
+ if (!TI || TI->getNumSuccessors() < 2)
+ return false;
MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
if (!WeightsNode)
BasicBlock *BB,
BasicBlock *NewBB,
BasicBlock *SuccBB) {
- if (!HasProfileData)
+ bool DoesBlockHaveProfile = doesBlockHaveProfileData(BB);
+ auto *BFI = getBFI();
+ auto *BPI = getBPI();
+ assert(
+ (!DoesBlockHaveProfile || (BFI && BPI))
+ && "BFI & BPI should have already been created");
+ assert(
+ ((BFI && BPI) || (!BFI && !BFI))
+ && "It's not expected to have only either BPI or BFI");
+
+ if (!BFI)
return;
- assert(BFI && BPI && "BFI & BPI should have been created here");
-
// As the edge from PredBB to BB is deleted, we have to update the block
// frequency of BB.
auto BBOrigFreq = BFI->getBlockFreq(BB);
// edge probabilities.
SmallVector<uint64_t, 4> BBSuccFreq;
for (BasicBlock *Succ : successors(BB)) {
- auto SuccFreq = (Succ == SuccBB)
- ? BB2SuccBBFreq - NewBBFreq
- : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
+ auto SuccFreq =
+ (Succ == SuccBB) ?
+ BB2SuccBBFreq - NewBBFreq :
+ BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
BBSuccFreq.push_back(SuccFreq.getFrequency());
}
- uint64_t MaxBBSuccFreq =
- *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
+ uint64_t MaxBBSuccFreq = *std::max_element(BBSuccFreq.begin(),
+ BBSuccFreq.end());
SmallVector<BranchProbability, 4> BBSuccProbs;
if (MaxBBSuccFreq == 0)
BBSuccProbs.assign(BBSuccFreq.size(),
- {1, static_cast<unsigned>(BBSuccFreq.size())});
+ { 1, static_cast<unsigned>(BBSuccFreq.size()) });
else {
for (uint64_t Freq : BBSuccFreq)
BBSuccProbs.push_back(
BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
// Normalize edge probabilities so that they sum up to one.
BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
- BBSuccProbs.end());
+ BBSuccProbs.end());
}
// Update edge probabilities in BPI.
// FIXME this locally as well so that BPI and BFI are consistent as well. We
// shouldn't make edges extremely likely or unlikely based solely on static
// estimation.
- if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
+ if (BBSuccProbs.size() >= 2 && DoesBlockHaveProfile) {
SmallVector<uint32_t, 4> Weights;
for (auto Prob : BBSuccProbs)
Weights.push_back(Prob.getNumerator());
// Remove the unconditional branch at the end of the PredBB block.
OldPredBranch->eraseFromParent();
- if (HasProfileData)
+ if (auto *BPI = getBPI())
BPI->copyEdgeProbabilities(BB, PredBB);
DTU->applyUpdatesPermissive(Updates);
}
return true;
}
+
+BranchProbabilityInfo *JumpThreadingPass::getBPI() {
+ if (!BPI) {
+ assert(FAM && "Can't create BPI without FunctionAnalysisManager");
+ BPI = FAM->getCachedResult<BranchProbabilityAnalysis>(*F);
+ }
+ return *BPI;
+}
+
+BlockFrequencyInfo *JumpThreadingPass::getBFI() {
+ if (!BFI) {
+ assert(FAM && "Can't create BFI without FunctionAnalysisManager");
+ BFI = FAM->getCachedResult<BlockFrequencyAnalysis>(*F);
+ }
+ return *BFI;
+}
+
+// Important note on validity of BPI/BFI. JumpThreading tries to preserve
+// BPI/BFI as it goes. Thus if cached instance exists it will be updated.
+// Otherwise, new instance of BPI/BFI is created wich is up to date by
+// definition.
+BranchProbabilityInfo *JumpThreadingPass::getOrCreateBPI() {
+ if (HasProfile && (!BPI || *BPI == nullptr)) {
+ assert(FAM && "Can't create BPI without FunctionAnalysisManager");
+ // BranchProbabilityAnalysis depends on DT. Make sure DT is consistent.
+ if (DTU->hasPendingUpdates()) {
+ DTU->flush();
+ assert(DTU->getDomTree().verify(DominatorTree::VerificationLevel::Fast));
+ }
+ BPI = &FAM->getResult<BranchProbabilityAnalysis>(*F);
+ }
+ return getBPI();
+}
+
+BlockFrequencyInfo *JumpThreadingPass::getOrCreateBFI() {
+ if (HasProfile && (!BFI || *BFI == nullptr)) {
+ assert(FAM && "Can't create BFI without FunctionAnalysisManager");
+ // BlockFrequencyAnalysis depends on DT. Make sure DT is consistent.
+ if (DTU->hasPendingUpdates()) {
+ DTU->flush();
+ assert(DTU->getDomTree().verify(DominatorTree::VerificationLevel::Fast));
+ }
+ BFI = &FAM->getResult<BlockFrequencyAnalysis>(*F);
+ }
+ return getBFI();
+}
+
projects / platform / upstream / llvm.git / commitdiff