1 //===- HotColdSplitting.cpp -- Outline Cold Regions -------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Outline cold regions to a separate function.
11 // TODO: Update BFI and BPI
12 // TODO: Add all the outlined functions to a separate section.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/ADT/PostOrderIterator.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/BlockFrequencyInfo.h"
21 #include "llvm/Analysis/BranchProbabilityInfo.h"
22 #include "llvm/Analysis/CFG.h"
23 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
24 #include "llvm/Analysis/PostDominators.h"
25 #include "llvm/Analysis/ProfileSummaryInfo.h"
26 #include "llvm/Analysis/TargetTransformInfo.h"
27 #include "llvm/IR/BasicBlock.h"
28 #include "llvm/IR/CFG.h"
29 #include "llvm/IR/CallSite.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DiagnosticInfo.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/Instruction.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/IR/Metadata.h"
38 #include "llvm/IR/Module.h"
39 #include "llvm/IR/PassManager.h"
40 #include "llvm/IR/Type.h"
41 #include "llvm/IR/Use.h"
42 #include "llvm/IR/User.h"
43 #include "llvm/IR/Value.h"
44 #include "llvm/Pass.h"
45 #include "llvm/Support/BlockFrequency.h"
46 #include "llvm/Support/BranchProbability.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include "llvm/Transforms/IPO.h"
50 #include "llvm/Transforms/IPO/HotColdSplitting.h"
51 #include "llvm/Transforms/Scalar.h"
52 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
53 #include "llvm/Transforms/Utils/Cloning.h"
54 #include "llvm/Transforms/Utils/CodeExtractor.h"
55 #include "llvm/Transforms/Utils/Local.h"
56 #include "llvm/Transforms/Utils/SSAUpdater.h"
57 #include "llvm/Transforms/Utils/ValueMapper.h"
61 #define DEBUG_TYPE "hotcoldsplit"
63 STATISTIC(NumColdRegionsFound, "Number of cold regions found.");
64 STATISTIC(NumColdRegionsOutlined, "Number of cold regions outlined.");
68 static cl::opt<bool> EnableStaticAnalyis("hot-cold-static-analysis",
69 cl::init(true), cl::Hidden);
72 MinOutliningThreshold("min-outlining-thresh", cl::init(3), cl::Hidden,
73 cl::desc("Code size threshold for outlining within a "
74 "single BB (as a multiple of TCC_Basic)"));
78 struct PostDomTree : PostDomTreeBase<BasicBlock> {
79 PostDomTree(Function &F) { recalculate(F); }
82 /// A sequence of basic blocks.
84 /// A 0-sized SmallVector is slightly cheaper to move than a std::vector.
85 using BlockSequence = SmallVector<BasicBlock *, 0>;
87 // Same as blockEndsInUnreachable in CodeGen/BranchFolding.cpp. Do not modify
88 // this function unless you modify the MBB version as well.
90 /// A no successor, non-return block probably ends in unreachable and is cold.
91 /// Also consider a block that ends in an indirect branch to be a return block,
92 /// since many targets use plain indirect branches to return.
93 bool blockEndsInUnreachable(const BasicBlock &BB) {
98 const Instruction *I = BB.getTerminator();
99 return !(isa<ReturnInst>(I) || isa<IndirectBrInst>(I));
102 bool unlikelyExecuted(BasicBlock &BB) {
103 // Exception handling blocks are unlikely executed.
107 // The block is cold if it calls/invokes a cold function.
108 for (Instruction &I : BB)
109 if (auto CS = CallSite(&I))
110 if (CS.hasFnAttr(Attribute::Cold))
113 // The block is cold if it has an unreachable terminator, unless it's
114 // preceded by a call to a (possibly warm) noreturn call (e.g. longjmp).
115 if (blockEndsInUnreachable(BB)) {
117 dyn_cast_or_null<CallInst>(BB.getTerminator()->getPrevNode()))
118 if (CI->hasFnAttr(Attribute::NoReturn))
126 /// Check whether it's safe to outline \p BB.
127 static bool mayExtractBlock(const BasicBlock &BB) {
128 return !BB.hasAddressTaken() && !BB.isEHPad();
131 /// Check whether \p Region is profitable to outline.
132 static bool isProfitableToOutline(const BlockSequence &Region,
133 TargetTransformInfo &TTI) {
134 if (Region.size() > 1)
138 const BasicBlock &BB = *Region[0];
139 for (const Instruction &I : BB) {
140 if (isa<DbgInfoIntrinsic>(&I) || &I == BB.getTerminator())
143 Cost += TTI.getInstructionCost(&I, TargetTransformInfo::TCK_CodeSize);
145 if (Cost >= (MinOutliningThreshold * TargetTransformInfo::TCC_Basic))
151 /// Mark \p F cold. Return true if it's changed.
152 static bool markEntireFunctionCold(Function &F) {
153 assert(!F.hasFnAttribute(Attribute::OptimizeNone) && "Can't mark this cold");
154 bool Changed = false;
155 if (!F.hasFnAttribute(Attribute::MinSize)) {
156 F.addFnAttr(Attribute::MinSize);
159 // TODO: Move this function into a cold section.
163 class HotColdSplitting {
165 HotColdSplitting(ProfileSummaryInfo *ProfSI,
166 function_ref<BlockFrequencyInfo *(Function &)> GBFI,
167 function_ref<TargetTransformInfo &(Function &)> GTTI,
168 std::function<OptimizationRemarkEmitter &(Function &)> *GORE)
169 : PSI(ProfSI), GetBFI(GBFI), GetTTI(GTTI), GetORE(GORE) {}
173 bool shouldOutlineFrom(const Function &F) const;
174 bool outlineColdRegions(Function &F, ProfileSummaryInfo &PSI,
175 BlockFrequencyInfo *BFI, TargetTransformInfo &TTI,
176 DominatorTree &DT, PostDomTree &PDT,
177 OptimizationRemarkEmitter &ORE);
178 Function *extractColdRegion(const BlockSequence &Region, DominatorTree &DT,
179 BlockFrequencyInfo *BFI, TargetTransformInfo &TTI,
180 OptimizationRemarkEmitter &ORE, unsigned Count);
181 SmallPtrSet<const Function *, 2> OutlinedFunctions;
182 ProfileSummaryInfo *PSI;
183 function_ref<BlockFrequencyInfo *(Function &)> GetBFI;
184 function_ref<TargetTransformInfo &(Function &)> GetTTI;
185 std::function<OptimizationRemarkEmitter &(Function &)> *GetORE;
188 class HotColdSplittingLegacyPass : public ModulePass {
191 HotColdSplittingLegacyPass() : ModulePass(ID) {
192 initializeHotColdSplittingLegacyPassPass(*PassRegistry::getPassRegistry());
195 void getAnalysisUsage(AnalysisUsage &AU) const override {
196 AU.addRequired<AssumptionCacheTracker>();
197 AU.addRequired<BlockFrequencyInfoWrapperPass>();
198 AU.addRequired<ProfileSummaryInfoWrapperPass>();
199 AU.addRequired<TargetTransformInfoWrapperPass>();
202 bool runOnModule(Module &M) override;
205 } // end anonymous namespace
207 // Returns false if the function should not be considered for hot-cold split
209 bool HotColdSplitting::shouldOutlineFrom(const Function &F) const {
210 // Do not try to outline again from an already outlined cold function.
211 if (OutlinedFunctions.count(&F))
217 // TODO: Consider only skipping functions marked `optnone` or `cold`.
219 if (F.hasAddressTaken())
222 if (F.hasFnAttribute(Attribute::AlwaysInline))
225 if (F.hasFnAttribute(Attribute::NoInline))
228 if (F.getCallingConv() == CallingConv::Cold)
231 if (PSI->isFunctionEntryCold(&F))
236 Function *HotColdSplitting::extractColdRegion(const BlockSequence &Region,
238 BlockFrequencyInfo *BFI,
239 TargetTransformInfo &TTI,
240 OptimizationRemarkEmitter &ORE,
242 assert(!Region.empty());
244 // TODO: Pass BFI and BPI to update profile information.
245 CodeExtractor CE(Region, &DT, /* AggregateArgs */ false, /* BFI */ nullptr,
246 /* BPI */ nullptr, /* AllowVarArgs */ false,
247 /* AllowAlloca */ false,
248 /* Suffix */ "cold." + std::to_string(Count));
250 SetVector<Value *> Inputs, Outputs, Sinks;
251 CE.findInputsOutputs(Inputs, Outputs, Sinks);
253 // Do not extract regions that have live exit variables.
254 if (Outputs.size() > 0) {
255 LLVM_DEBUG(llvm::dbgs() << "Not outlining; live outputs\n");
259 // TODO: Run MergeBasicBlockIntoOnlyPred on the outlined function.
260 Function *OrigF = Region[0]->getParent();
261 if (Function *OutF = CE.extractCodeRegion()) {
262 User *U = *OutF->user_begin();
263 CallInst *CI = cast<CallInst>(U);
265 NumColdRegionsOutlined++;
266 if (TTI.useColdCCForColdCall(*OutF)) {
267 OutF->setCallingConv(CallingConv::Cold);
268 CS.setCallingConv(CallingConv::Cold);
272 // Try to make the outlined code as small as possible on the assumption
274 markEntireFunctionCold(*OutF);
276 LLVM_DEBUG(llvm::dbgs() << "Outlined Region: " << *OutF);
278 return OptimizationRemark(DEBUG_TYPE, "HotColdSplit",
279 &*Region[0]->begin())
280 << ore::NV("Original", OrigF) << " split cold code into "
281 << ore::NV("Split", OutF);
287 return OptimizationRemarkMissed(DEBUG_TYPE, "ExtractFailed",
288 &*Region[0]->begin())
289 << "Failed to extract region at block "
290 << ore::NV("Block", Region.front());
295 /// A pair of (basic block, score).
296 using BlockTy = std::pair<BasicBlock *, unsigned>;
299 /// A maximal outlining region. This contains all blocks post-dominated by a
300 /// sink block, the sink block itself, and all blocks dominated by the sink.
301 class OutliningRegion {
302 /// A list of (block, score) pairs. A block's score is non-zero iff it's a
303 /// viable sub-region entry point. Blocks with higher scores are better entry
304 /// points (i.e. they are more distant ancestors of the sink block).
305 SmallVector<BlockTy, 0> Blocks = {};
307 /// The suggested entry point into the region. If the region has multiple
308 /// entry points, all blocks within the region may not be reachable from this
310 BasicBlock *SuggestedEntryPoint = nullptr;
312 /// Whether the entire function is cold.
313 bool EntireFunctionCold = false;
315 /// Whether or not \p BB could be the entry point of an extracted region.
316 static bool isViableEntryPoint(BasicBlock &BB) { return !BB.isEHPad(); }
318 /// If \p BB is a viable entry point, return \p Score. Return 0 otherwise.
319 static unsigned getEntryPointScore(BasicBlock &BB, unsigned Score) {
320 return isViableEntryPoint(BB) ? Score : 0;
323 /// These scores should be lower than the score for predecessor blocks,
324 /// because regions starting at predecessor blocks are typically larger.
325 static constexpr unsigned ScoreForSuccBlock = 1;
326 static constexpr unsigned ScoreForSinkBlock = 1;
328 OutliningRegion(const OutliningRegion &) = delete;
329 OutliningRegion &operator=(const OutliningRegion &) = delete;
332 OutliningRegion() = default;
333 OutliningRegion(OutliningRegion &&) = default;
334 OutliningRegion &operator=(OutliningRegion &&) = default;
336 static OutliningRegion create(BasicBlock &SinkBB, const DominatorTree &DT,
337 const PostDomTree &PDT) {
338 OutliningRegion ColdRegion;
340 SmallPtrSet<BasicBlock *, 4> RegionBlocks;
342 auto addBlockToRegion = [&](BasicBlock *BB, unsigned Score) {
343 RegionBlocks.insert(BB);
344 ColdRegion.Blocks.emplace_back(BB, Score);
345 assert(RegionBlocks.size() == ColdRegion.Blocks.size() && "Duplicate BB");
348 // The ancestor farthest-away from SinkBB, and also post-dominated by it.
349 unsigned SinkScore = getEntryPointScore(SinkBB, ScoreForSinkBlock);
350 ColdRegion.SuggestedEntryPoint = (SinkScore > 0) ? &SinkBB : nullptr;
351 unsigned BestScore = SinkScore;
353 // Visit SinkBB's ancestors using inverse DFS.
354 auto PredIt = ++idf_begin(&SinkBB);
355 auto PredEnd = idf_end(&SinkBB);
356 while (PredIt != PredEnd) {
357 BasicBlock &PredBB = **PredIt;
358 bool SinkPostDom = PDT.dominates(&SinkBB, &PredBB);
360 // If the predecessor is cold and has no predecessors, the entire
361 // function must be cold.
362 if (SinkPostDom && pred_empty(&PredBB)) {
363 ColdRegion.EntireFunctionCold = true;
367 // If SinkBB does not post-dominate a predecessor, do not mark the
368 // predecessor (or any of its predecessors) cold.
369 if (!SinkPostDom || !mayExtractBlock(PredBB)) {
370 PredIt.skipChildren();
374 // Keep track of the post-dominated ancestor farthest away from the sink.
375 // The path length is always >= 2, ensuring that predecessor blocks are
376 // considered as entry points before the sink block.
377 unsigned PredScore = getEntryPointScore(PredBB, PredIt.getPathLength());
378 if (PredScore > BestScore) {
379 ColdRegion.SuggestedEntryPoint = &PredBB;
380 BestScore = PredScore;
383 addBlockToRegion(&PredBB, PredScore);
387 // Add SinkBB to the cold region. It's considered as an entry point before
388 // any sink-successor blocks.
389 addBlockToRegion(&SinkBB, SinkScore);
391 // Find all successors of SinkBB dominated by SinkBB using DFS.
392 auto SuccIt = ++df_begin(&SinkBB);
393 auto SuccEnd = df_end(&SinkBB);
394 while (SuccIt != SuccEnd) {
395 BasicBlock &SuccBB = **SuccIt;
396 bool SinkDom = DT.dominates(&SinkBB, &SuccBB);
398 // Don't allow the backwards & forwards DFSes to mark the same block.
399 bool DuplicateBlock = RegionBlocks.count(&SuccBB);
401 // If SinkBB does not dominate a successor, do not mark the successor (or
402 // any of its successors) cold.
403 if (DuplicateBlock || !SinkDom || !mayExtractBlock(SuccBB)) {
404 SuccIt.skipChildren();
408 unsigned SuccScore = getEntryPointScore(SuccBB, ScoreForSuccBlock);
409 if (SuccScore > BestScore) {
410 ColdRegion.SuggestedEntryPoint = &SuccBB;
411 BestScore = SuccScore;
414 addBlockToRegion(&SuccBB, SuccScore);
421 /// Whether this region has nothing to extract.
422 bool empty() const { return !SuggestedEntryPoint; }
424 /// The blocks in this region.
425 ArrayRef<std::pair<BasicBlock *, unsigned>> blocks() const { return Blocks; }
427 /// Whether the entire function containing this region is cold.
428 bool isEntireFunctionCold() const { return EntireFunctionCold; }
430 /// Remove a sub-region from this region and return it as a block sequence.
431 BlockSequence takeSingleEntrySubRegion(DominatorTree &DT) {
432 assert(!empty() && !isEntireFunctionCold() && "Nothing to extract");
434 // Remove blocks dominated by the suggested entry point from this region.
435 // During the removal, identify the next best entry point into the region.
436 // Ensure that the first extracted block is the suggested entry point.
437 BlockSequence SubRegion = {SuggestedEntryPoint};
438 BasicBlock *NextEntryPoint = nullptr;
439 unsigned NextScore = 0;
440 auto RegionEndIt = Blocks.end();
441 auto RegionStartIt = remove_if(Blocks, [&](const BlockTy &Block) {
442 BasicBlock *BB = Block.first;
443 unsigned Score = Block.second;
445 BB == SuggestedEntryPoint || DT.dominates(SuggestedEntryPoint, BB);
446 if (!InSubRegion && Score > NextScore) {
450 if (InSubRegion && BB != SuggestedEntryPoint)
451 SubRegion.push_back(BB);
454 Blocks.erase(RegionStartIt, RegionEndIt);
456 // Update the suggested entry point.
457 SuggestedEntryPoint = NextEntryPoint;
464 bool HotColdSplitting::outlineColdRegions(Function &F, ProfileSummaryInfo &PSI,
465 BlockFrequencyInfo *BFI,
466 TargetTransformInfo &TTI,
467 DominatorTree &DT, PostDomTree &PDT,
468 OptimizationRemarkEmitter &ORE) {
469 bool Changed = false;
471 // The set of cold blocks.
472 SmallPtrSet<BasicBlock *, 4> ColdBlocks;
474 // The worklist of non-intersecting regions left to outline.
475 SmallVector<OutliningRegion, 2> OutliningWorklist;
477 // Set up an RPO traversal. Experimentally, this performs better (outlines
478 // more) than a PO traversal, because we prevent region overlap by keeping
479 // the first region to contain a block.
480 ReversePostOrderTraversal<Function *> RPOT(&F);
482 // Find all cold regions.
483 for (BasicBlock *BB : RPOT) {
484 // Skip blocks which can't be outlined.
485 if (!mayExtractBlock(*BB))
488 // This block is already part of some outlining region.
489 if (ColdBlocks.count(BB))
492 bool Cold = PSI.isColdBlock(BB, BFI) ||
493 (EnableStaticAnalyis && unlikelyExecuted(*BB));
498 dbgs() << "Found a cold block:\n";
502 auto Region = OutliningRegion::create(*BB, DT, PDT);
506 if (Region.isEntireFunctionCold()) {
507 LLVM_DEBUG(dbgs() << "Entire function is cold\n");
508 return markEntireFunctionCold(F);
511 // If this outlining region intersects with another, drop the new region.
513 // TODO: It's theoretically possible to outline more by only keeping the
514 // largest region which contains a block, but the extra bookkeeping to do
515 // this is tricky/expensive.
516 bool RegionsOverlap = any_of(Region.blocks(), [&](const BlockTy &Block) {
517 return !ColdBlocks.insert(Block.first).second;
522 OutliningWorklist.emplace_back(std::move(Region));
523 ++NumColdRegionsFound;
526 // Outline single-entry cold regions, splitting up larger regions as needed.
527 unsigned OutlinedFunctionID = 1;
528 while (!OutliningWorklist.empty()) {
529 OutliningRegion Region = OutliningWorklist.pop_back_val();
530 assert(!Region.empty() && "Empty outlining region in worklist");
532 BlockSequence SubRegion = Region.takeSingleEntrySubRegion(DT);
533 if (!isProfitableToOutline(SubRegion, TTI)) {
535 dbgs() << "Skipping outlining; not profitable to outline\n";
536 SubRegion[0]->dump();
542 dbgs() << "Hot/cold splitting attempting to outline these blocks:\n";
543 for (BasicBlock *BB : SubRegion)
548 extractColdRegion(SubRegion, DT, BFI, TTI, ORE, OutlinedFunctionID);
550 ++OutlinedFunctionID;
551 OutlinedFunctions.insert(Outlined);
554 } while (!Region.empty());
560 bool HotColdSplitting::run(Module &M) {
561 bool Changed = false;
562 OutlinedFunctions.clear();
564 if (!shouldOutlineFrom(F)) {
565 LLVM_DEBUG(llvm::dbgs() << "Skipping " << F.getName() << "\n");
568 LLVM_DEBUG(llvm::dbgs() << "Outlining in " << F.getName() << "\n");
572 BlockFrequencyInfo *BFI = GetBFI(F);
573 TargetTransformInfo &TTI = GetTTI(F);
574 OptimizationRemarkEmitter &ORE = (*GetORE)(F);
575 Changed |= outlineColdRegions(F, *PSI, BFI, TTI, DT, PDT, ORE);
580 bool HotColdSplittingLegacyPass::runOnModule(Module &M) {
583 ProfileSummaryInfo *PSI =
584 &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
585 auto GTTI = [this](Function &F) -> TargetTransformInfo & {
586 return this->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
588 auto GBFI = [this](Function &F) {
589 return &this->getAnalysis<BlockFrequencyInfoWrapperPass>(F).getBFI();
591 std::unique_ptr<OptimizationRemarkEmitter> ORE;
592 std::function<OptimizationRemarkEmitter &(Function &)> GetORE =
593 [&ORE](Function &F) -> OptimizationRemarkEmitter & {
594 ORE.reset(new OptimizationRemarkEmitter(&F));
598 return HotColdSplitting(PSI, GBFI, GTTI, &GetORE).run(M);
602 HotColdSplittingPass::run(Module &M, ModuleAnalysisManager &AM) {
603 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
605 std::function<AssumptionCache &(Function &)> GetAssumptionCache =
606 [&FAM](Function &F) -> AssumptionCache & {
607 return FAM.getResult<AssumptionAnalysis>(F);
610 auto GBFI = [&FAM](Function &F) {
611 return &FAM.getResult<BlockFrequencyAnalysis>(F);
614 std::function<TargetTransformInfo &(Function &)> GTTI =
615 [&FAM](Function &F) -> TargetTransformInfo & {
616 return FAM.getResult<TargetIRAnalysis>(F);
619 std::unique_ptr<OptimizationRemarkEmitter> ORE;
620 std::function<OptimizationRemarkEmitter &(Function &)> GetORE =
621 [&ORE](Function &F) -> OptimizationRemarkEmitter & {
622 ORE.reset(new OptimizationRemarkEmitter(&F));
626 ProfileSummaryInfo *PSI = &AM.getResult<ProfileSummaryAnalysis>(M);
628 if (HotColdSplitting(PSI, GBFI, GTTI, &GetORE).run(M))
629 return PreservedAnalyses::none();
630 return PreservedAnalyses::all();
633 char HotColdSplittingLegacyPass::ID = 0;
634 INITIALIZE_PASS_BEGIN(HotColdSplittingLegacyPass, "hotcoldsplit",
635 "Hot Cold Splitting", false, false)
636 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
637 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
638 INITIALIZE_PASS_END(HotColdSplittingLegacyPass, "hotcoldsplit",
639 "Hot Cold Splitting", false, false)
641 ModulePass *llvm::createHotColdSplittingPass() {
642 return new HotColdSplittingLegacyPass();