1 //===- HotColdSplitting.cpp -- Outline Cold Regions -------------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 ///
9 /// \file
10 /// The goal of hot/cold splitting is to improve the memory locality of code.
11 /// The splitting pass does this by identifying cold blocks and moving them into
12 /// separate functions.
13 ///
14 /// When the splitting pass finds a cold block (referred to as "the sink"), it
15 /// grows a maximal cold region around that block. The maximal region contains
16 /// all blocks (post-)dominated by the sink [*]. In theory, these blocks are as
17 /// cold as the sink. Once a region is found, it's split out of the original
18 /// function provided it's profitable to do so.
19 ///
20 /// [*] In practice, there is some added complexity because some blocks are not
21 /// safe to extract.
22 ///
23 /// TODO: Use the PM to get domtrees, and preserve BFI/BPI.
24 /// TODO: Reorder outlined functions.
25 ///
26 //===----------------------------------------------------------------------===//
27
28 #include "llvm/Transforms/IPO/HotColdSplitting.h"
29 #include "llvm/ADT/PostOrderIterator.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Analysis/AliasAnalysis.h"
33 #include "llvm/Analysis/BlockFrequencyInfo.h"
34 #include "llvm/Analysis/BranchProbabilityInfo.h"
35 #include "llvm/Analysis/CFG.h"
36 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
37 #include "llvm/Analysis/PostDominators.h"
38 #include "llvm/Analysis/ProfileSummaryInfo.h"
39 #include "llvm/Analysis/TargetTransformInfo.h"
40 #include "llvm/IR/BasicBlock.h"
41 #include "llvm/IR/CFG.h"
42 #include "llvm/IR/CallSite.h"
43 #include "llvm/IR/DataLayout.h"
44 #include "llvm/IR/DiagnosticInfo.h"
45 #include "llvm/IR/Dominators.h"
46 #include "llvm/IR/Function.h"
47 #include "llvm/IR/Instruction.h"
48 #include "llvm/IR/Instructions.h"
49 #include "llvm/IR/IntrinsicInst.h"
50 #include "llvm/IR/Metadata.h"
51 #include "llvm/IR/Module.h"
52 #include "llvm/IR/PassManager.h"
53 #include "llvm/IR/Type.h"
54 #include "llvm/IR/Use.h"
55 #include "llvm/IR/User.h"
56 #include "llvm/IR/Value.h"
57 #include "llvm/InitializePasses.h"
58 #include "llvm/Pass.h"
59 #include "llvm/Support/BlockFrequency.h"
60 #include "llvm/Support/BranchProbability.h"
61 #include "llvm/Support/CommandLine.h"
62 #include "llvm/Support/Debug.h"
63 #include "llvm/Support/raw_ostream.h"
64 #include "llvm/Transforms/IPO.h"
65 #include "llvm/Transforms/Scalar.h"
66 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
67 #include "llvm/Transforms/Utils/Cloning.h"
68 #include "llvm/Transforms/Utils/CodeExtractor.h"
69 #include "llvm/Transforms/Utils/Local.h"
70 #include "llvm/Transforms/Utils/ValueMapper.h"
71 #include <algorithm>
72 #include <cassert>
73
74 #define DEBUG_TYPE "hotcoldsplit"
75
76 STATISTIC(NumColdRegionsFound, "Number of cold regions found.");
77 STATISTIC(NumColdRegionsOutlined, "Number of cold regions outlined.");
78
79 using namespace llvm;
80
81 static cl::opt<bool> EnableStaticAnalyis("hot-cold-static-analysis",
82 cl::init(true), cl::Hidden);
83
84 static cl::opt<int>
85 SplittingThreshold("hotcoldsplit-threshold", cl::init(2), cl::Hidden,
86 cl::desc("Base penalty for splitting cold code (as a "
87 "multiple of TCC_Basic)"));
88
89 namespace {
90 // Same as blockEndsInUnreachable in CodeGen/BranchFolding.cpp. Do not modify
91 // this function unless you modify the MBB version as well.
92 //
93 /// A no successor, non-return block probably ends in unreachable and is cold.
94 /// Also consider a block that ends in an indirect branch to be a return block,
95 /// since many targets use plain indirect branches to return.
blockEndsInUnreachable(const BasicBlock & BB)96 bool blockEndsInUnreachable(const BasicBlock &BB) {
97 if (!succ_empty(&BB))
98 return false;
99 if (BB.empty())
100 return true;
101 const Instruction *I = BB.getTerminator();
102 return !(isa<ReturnInst>(I) || isa<IndirectBrInst>(I));
103 }
104
unlikelyExecuted(BasicBlock & BB)105 bool unlikelyExecuted(BasicBlock &BB) {
106 // Exception handling blocks are unlikely executed.
107 if (BB.isEHPad() || isa<ResumeInst>(BB.getTerminator()))
108 return true;
109
110 // The block is cold if it calls/invokes a cold function. However, do not
111 // mark sanitizer traps as cold.
112 for (Instruction &I : BB)
113 if (auto CS = CallSite(&I))
114 if (CS.hasFnAttr(Attribute::Cold) && !CS->getMetadata("nosanitize"))
115 return true;
116
117 // The block is cold if it has an unreachable terminator, unless it's
118 // preceded by a call to a (possibly warm) noreturn call (e.g. longjmp).
119 if (blockEndsInUnreachable(BB)) {
120 if (auto *CI =
121 dyn_cast_or_null<CallInst>(BB.getTerminator()->getPrevNode()))
122 if (CI->hasFnAttr(Attribute::NoReturn))
123 return false;
124 return true;
125 }
126
127 return false;
128 }
129
130 /// Check whether it's safe to outline \p BB.
mayExtractBlock(const BasicBlock & BB)131 static bool mayExtractBlock(const BasicBlock &BB) {
132 // EH pads are unsafe to outline because doing so breaks EH type tables. It
133 // follows that invoke instructions cannot be extracted, because CodeExtractor
134 // requires unwind destinations to be within the extraction region.
135 //
136 // Resumes that are not reachable from a cleanup landing pad are considered to
137 // be unreachable. It’s not safe to split them out either.
138 auto Term = BB.getTerminator();
139 return !BB.hasAddressTaken() && !BB.isEHPad() && !isa<InvokeInst>(Term) &&
140 !isa<ResumeInst>(Term);
141 }
142
143 /// Mark \p F cold. Based on this assumption, also optimize it for minimum size.
144 /// If \p UpdateEntryCount is true (set when this is a new split function and
145 /// module has profile data), set entry count to 0 to ensure treated as cold.
146 /// Return true if the function is changed.
markFunctionCold(Function & F,bool UpdateEntryCount=false)147 static bool markFunctionCold(Function &F, bool UpdateEntryCount = false) {
148 assert(!F.hasOptNone() && "Can't mark this cold");
149 bool Changed = false;
150 if (!F.hasFnAttribute(Attribute::Cold)) {
151 F.addFnAttr(Attribute::Cold);
152 Changed = true;
153 }
154 if (!F.hasFnAttribute(Attribute::MinSize)) {
155 F.addFnAttr(Attribute::MinSize);
156 Changed = true;
157 }
158 if (UpdateEntryCount) {
159 // Set the entry count to 0 to ensure it is placed in the unlikely text
160 // section when function sections are enabled.
161 F.setEntryCount(0);
162 Changed = true;
163 }
164
165 return Changed;
166 }
167
168 class HotColdSplittingLegacyPass : public ModulePass {
169 public:
170 static char ID;
HotColdSplittingLegacyPass()171 HotColdSplittingLegacyPass() : ModulePass(ID) {
172 initializeHotColdSplittingLegacyPassPass(*PassRegistry::getPassRegistry());
173 }
174
getAnalysisUsage(AnalysisUsage & AU) const175 void getAnalysisUsage(AnalysisUsage &AU) const override {
176 AU.addRequired<BlockFrequencyInfoWrapperPass>();
177 AU.addRequired<ProfileSummaryInfoWrapperPass>();
178 AU.addRequired<TargetTransformInfoWrapperPass>();
179 AU.addUsedIfAvailable<AssumptionCacheTracker>();
180 }
181
182 bool runOnModule(Module &M) override;
183 };
184
185 } // end anonymous namespace
186
187 /// Check whether \p F is inherently cold.
isFunctionCold(const Function & F) const188 bool HotColdSplitting::isFunctionCold(const Function &F) const {
189 if (F.hasFnAttribute(Attribute::Cold))
190 return true;
191
192 if (F.getCallingConv() == CallingConv::Cold)
193 return true;
194
195 if (PSI->isFunctionEntryCold(&F))
196 return true;
197
198 return false;
199 }
200
201 // Returns false if the function should not be considered for hot-cold split
202 // optimization.
shouldOutlineFrom(const Function & F) const203 bool HotColdSplitting::shouldOutlineFrom(const Function &F) const {
204 if (F.hasFnAttribute(Attribute::AlwaysInline))
205 return false;
206
207 if (F.hasFnAttribute(Attribute::NoInline))
208 return false;
209
210 // A function marked `noreturn` may contain unreachable terminators: these
211 // should not be considered cold, as the function may be a trampoline.
212 if (F.hasFnAttribute(Attribute::NoReturn))
213 return false;
214
215 if (F.hasFnAttribute(Attribute::SanitizeAddress) ||
216 F.hasFnAttribute(Attribute::SanitizeHWAddress) ||
217 F.hasFnAttribute(Attribute::SanitizeThread) ||
218 F.hasFnAttribute(Attribute::SanitizeMemory))
219 return false;
220
221 return true;
222 }
223
224 /// Get the benefit score of outlining \p Region.
getOutliningBenefit(ArrayRef<BasicBlock * > Region,TargetTransformInfo & TTI)225 static int getOutliningBenefit(ArrayRef<BasicBlock *> Region,
226 TargetTransformInfo &TTI) {
227 // Sum up the code size costs of non-terminator instructions. Tight coupling
228 // with \ref getOutliningPenalty is needed to model the costs of terminators.
229 int Benefit = 0;
230 for (BasicBlock *BB : Region)
231 for (Instruction &I : BB->instructionsWithoutDebug())
232 if (&I != BB->getTerminator())
233 Benefit +=
234 TTI.getInstructionCost(&I, TargetTransformInfo::TCK_CodeSize);
235
236 return Benefit;
237 }
238
239 /// Get the penalty score for outlining \p Region.
getOutliningPenalty(ArrayRef<BasicBlock * > Region,unsigned NumInputs,unsigned NumOutputs)240 static int getOutliningPenalty(ArrayRef<BasicBlock *> Region,
241 unsigned NumInputs, unsigned NumOutputs) {
242 int Penalty = SplittingThreshold;
243 LLVM_DEBUG(dbgs() << "Applying penalty for splitting: " << Penalty << "\n");
244
245 // If the splitting threshold is set at or below zero, skip the usual
246 // profitability check.
247 if (SplittingThreshold <= 0)
248 return Penalty;
249
250 // The typical code size cost for materializing an argument for the outlined
251 // call.
252 LLVM_DEBUG(dbgs() << "Applying penalty for: " << NumInputs << " inputs\n");
253 const int CostForArgMaterialization = TargetTransformInfo::TCC_Basic;
254 Penalty += CostForArgMaterialization * NumInputs;
255
256 // The typical code size cost for an output alloca, its associated store, and
257 // its associated reload.
258 LLVM_DEBUG(dbgs() << "Applying penalty for: " << NumOutputs << " outputs\n");
259 const int CostForRegionOutput = 3 * TargetTransformInfo::TCC_Basic;
260 Penalty += CostForRegionOutput * NumOutputs;
261
262 // Find the number of distinct exit blocks for the region. Use a conservative
263 // check to determine whether control returns from the region.
264 bool NoBlocksReturn = true;
265 SmallPtrSet<BasicBlock *, 2> SuccsOutsideRegion;
266 for (BasicBlock *BB : Region) {
267 // If a block has no successors, only assume it does not return if it's
268 // unreachable.
269 if (succ_empty(BB)) {
270 NoBlocksReturn &= isa<UnreachableInst>(BB->getTerminator());
271 continue;
272 }
273
274 for (BasicBlock *SuccBB : successors(BB)) {
275 if (find(Region, SuccBB) == Region.end()) {
276 NoBlocksReturn = false;
277 SuccsOutsideRegion.insert(SuccBB);
278 }
279 }
280 }
281
282 // Apply a `noreturn` bonus.
283 if (NoBlocksReturn) {
284 LLVM_DEBUG(dbgs() << "Applying bonus for: " << Region.size()
285 << " non-returning terminators\n");
286 Penalty -= Region.size();
287 }
288
289 // Apply a penalty for having more than one successor outside of the region.
290 // This penalty accounts for the switch needed in the caller.
291 if (!SuccsOutsideRegion.empty()) {
292 LLVM_DEBUG(dbgs() << "Applying penalty for: " << SuccsOutsideRegion.size()
293 << " non-region successors\n");
294 Penalty += (SuccsOutsideRegion.size() - 1) * TargetTransformInfo::TCC_Basic;
295 }
296
297 return Penalty;
298 }
299
extractColdRegion(const BlockSequence & Region,const CodeExtractorAnalysisCache & CEAC,DominatorTree & DT,BlockFrequencyInfo * BFI,TargetTransformInfo & TTI,OptimizationRemarkEmitter & ORE,AssumptionCache * AC,unsigned Count)300 Function *HotColdSplitting::extractColdRegion(
301 const BlockSequence &Region, const CodeExtractorAnalysisCache &CEAC,
302 DominatorTree &DT, BlockFrequencyInfo *BFI, TargetTransformInfo &TTI,
303 OptimizationRemarkEmitter &ORE, AssumptionCache *AC, unsigned Count) {
304 assert(!Region.empty());
305
306 // TODO: Pass BFI and BPI to update profile information.
307 CodeExtractor CE(Region, &DT, /* AggregateArgs */ false, /* BFI */ nullptr,
308 /* BPI */ nullptr, AC, /* AllowVarArgs */ false,
309 /* AllowAlloca */ false,
310 /* Suffix */ "cold." + std::to_string(Count));
311
312 // Perform a simple cost/benefit analysis to decide whether or not to permit
313 // splitting.
314 SetVector<Value *> Inputs, Outputs, Sinks;
315 CE.findInputsOutputs(Inputs, Outputs, Sinks);
316 int OutliningBenefit = getOutliningBenefit(Region, TTI);
317 int OutliningPenalty =
318 getOutliningPenalty(Region, Inputs.size(), Outputs.size());
319 LLVM_DEBUG(dbgs() << "Split profitability: benefit = " << OutliningBenefit
320 << ", penalty = " << OutliningPenalty << "\n");
321 if (OutliningBenefit <= OutliningPenalty)
322 return nullptr;
323
324 Function *OrigF = Region[0]->getParent();
325 if (Function *OutF = CE.extractCodeRegion(CEAC)) {
326 User *U = *OutF->user_begin();
327 CallInst *CI = cast<CallInst>(U);
328 CallSite CS(CI);
329 NumColdRegionsOutlined++;
330 if (TTI.useColdCCForColdCall(*OutF)) {
331 OutF->setCallingConv(CallingConv::Cold);
332 CS.setCallingConv(CallingConv::Cold);
333 }
334 CI->setIsNoInline();
335
336 if (OrigF->hasSection())
337 OutF->setSection(OrigF->getSection());
338
339 markFunctionCold(*OutF, BFI != nullptr);
340
341 LLVM_DEBUG(llvm::dbgs() << "Outlined Region: " << *OutF);
342 ORE.emit([&]() {
343 return OptimizationRemark(DEBUG_TYPE, "HotColdSplit",
344 &*Region[0]->begin())
345 << ore::NV("Original", OrigF) << " split cold code into "
346 << ore::NV("Split", OutF);
347 });
348 return OutF;
349 }
350
351 ORE.emit([&]() {
352 return OptimizationRemarkMissed(DEBUG_TYPE, "ExtractFailed",
353 &*Region[0]->begin())
354 << "Failed to extract region at block "
355 << ore::NV("Block", Region.front());
356 });
357 return nullptr;
358 }
359
360 /// A pair of (basic block, score).
361 using BlockTy = std::pair<BasicBlock *, unsigned>;
362
363 namespace {
364 /// A maximal outlining region. This contains all blocks post-dominated by a
365 /// sink block, the sink block itself, and all blocks dominated by the sink.
366 /// If sink-predecessors and sink-successors cannot be extracted in one region,
367 /// the static constructor returns a list of suitable extraction regions.
368 class OutliningRegion {
369 /// A list of (block, score) pairs. A block's score is non-zero iff it's a
370 /// viable sub-region entry point. Blocks with higher scores are better entry
371 /// points (i.e. they are more distant ancestors of the sink block).
372 SmallVector<BlockTy, 0> Blocks = {};
373
374 /// The suggested entry point into the region. If the region has multiple
375 /// entry points, all blocks within the region may not be reachable from this
376 /// entry point.
377 BasicBlock *SuggestedEntryPoint = nullptr;
378
379 /// Whether the entire function is cold.
380 bool EntireFunctionCold = false;
381
382 /// If \p BB is a viable entry point, return \p Score. Return 0 otherwise.
getEntryPointScore(BasicBlock & BB,unsigned Score)383 static unsigned getEntryPointScore(BasicBlock &BB, unsigned Score) {
384 return mayExtractBlock(BB) ? Score : 0;
385 }
386
387 /// These scores should be lower than the score for predecessor blocks,
388 /// because regions starting at predecessor blocks are typically larger.
389 static constexpr unsigned ScoreForSuccBlock = 1;
390 static constexpr unsigned ScoreForSinkBlock = 1;
391
392 OutliningRegion(const OutliningRegion &) = delete;
393 OutliningRegion &operator=(const OutliningRegion &) = delete;
394
395 public:
396 OutliningRegion() = default;
397 OutliningRegion(OutliningRegion &&) = default;
398 OutliningRegion &operator=(OutliningRegion &&) = default;
399
create(BasicBlock & SinkBB,const DominatorTree & DT,const PostDominatorTree & PDT)400 static std::vector<OutliningRegion> create(BasicBlock &SinkBB,
401 const DominatorTree &DT,
402 const PostDominatorTree &PDT) {
403 std::vector<OutliningRegion> Regions;
404 SmallPtrSet<BasicBlock *, 4> RegionBlocks;
405
406 Regions.emplace_back();
407 OutliningRegion *ColdRegion = &Regions.back();
408
409 auto addBlockToRegion = [&](BasicBlock *BB, unsigned Score) {
410 RegionBlocks.insert(BB);
411 ColdRegion->Blocks.emplace_back(BB, Score);
412 };
413
414 // The ancestor farthest-away from SinkBB, and also post-dominated by it.
415 unsigned SinkScore = getEntryPointScore(SinkBB, ScoreForSinkBlock);
416 ColdRegion->SuggestedEntryPoint = (SinkScore > 0) ? &SinkBB : nullptr;
417 unsigned BestScore = SinkScore;
418
419 // Visit SinkBB's ancestors using inverse DFS.
420 auto PredIt = ++idf_begin(&SinkBB);
421 auto PredEnd = idf_end(&SinkBB);
422 while (PredIt != PredEnd) {
423 BasicBlock &PredBB = **PredIt;
424 bool SinkPostDom = PDT.dominates(&SinkBB, &PredBB);
425
426 // If the predecessor is cold and has no predecessors, the entire
427 // function must be cold.
428 if (SinkPostDom && pred_empty(&PredBB)) {
429 ColdRegion->EntireFunctionCold = true;
430 return Regions;
431 }
432
433 // If SinkBB does not post-dominate a predecessor, do not mark the
434 // predecessor (or any of its predecessors) cold.
435 if (!SinkPostDom || !mayExtractBlock(PredBB)) {
436 PredIt.skipChildren();
437 continue;
438 }
439
440 // Keep track of the post-dominated ancestor farthest away from the sink.
441 // The path length is always >= 2, ensuring that predecessor blocks are
442 // considered as entry points before the sink block.
443 unsigned PredScore = getEntryPointScore(PredBB, PredIt.getPathLength());
444 if (PredScore > BestScore) {
445 ColdRegion->SuggestedEntryPoint = &PredBB;
446 BestScore = PredScore;
447 }
448
449 addBlockToRegion(&PredBB, PredScore);
450 ++PredIt;
451 }
452
453 // If the sink can be added to the cold region, do so. It's considered as
454 // an entry point before any sink-successor blocks.
455 //
456 // Otherwise, split cold sink-successor blocks using a separate region.
457 // This satisfies the requirement that all extraction blocks other than the
458 // first have predecessors within the extraction region.
459 if (mayExtractBlock(SinkBB)) {
460 addBlockToRegion(&SinkBB, SinkScore);
461 } else {
462 Regions.emplace_back();
463 ColdRegion = &Regions.back();
464 BestScore = 0;
465 }
466
467 // Find all successors of SinkBB dominated by SinkBB using DFS.
468 auto SuccIt = ++df_begin(&SinkBB);
469 auto SuccEnd = df_end(&SinkBB);
470 while (SuccIt != SuccEnd) {
471 BasicBlock &SuccBB = **SuccIt;
472 bool SinkDom = DT.dominates(&SinkBB, &SuccBB);
473
474 // Don't allow the backwards & forwards DFSes to mark the same block.
475 bool DuplicateBlock = RegionBlocks.count(&SuccBB);
476
477 // If SinkBB does not dominate a successor, do not mark the successor (or
478 // any of its successors) cold.
479 if (DuplicateBlock || !SinkDom || !mayExtractBlock(SuccBB)) {
480 SuccIt.skipChildren();
481 continue;
482 }
483
484 unsigned SuccScore = getEntryPointScore(SuccBB, ScoreForSuccBlock);
485 if (SuccScore > BestScore) {
486 ColdRegion->SuggestedEntryPoint = &SuccBB;
487 BestScore = SuccScore;
488 }
489
490 addBlockToRegion(&SuccBB, SuccScore);
491 ++SuccIt;
492 }
493
494 return Regions;
495 }
496
497 /// Whether this region has nothing to extract.
empty() const498 bool empty() const { return !SuggestedEntryPoint; }
499
500 /// The blocks in this region.
blocks() const501 ArrayRef<std::pair<BasicBlock *, unsigned>> blocks() const { return Blocks; }
502
503 /// Whether the entire function containing this region is cold.
isEntireFunctionCold() const504 bool isEntireFunctionCold() const { return EntireFunctionCold; }
505
506 /// Remove a sub-region from this region and return it as a block sequence.
takeSingleEntrySubRegion(DominatorTree & DT)507 BlockSequence takeSingleEntrySubRegion(DominatorTree &DT) {
508 assert(!empty() && !isEntireFunctionCold() && "Nothing to extract");
509
510 // Remove blocks dominated by the suggested entry point from this region.
511 // During the removal, identify the next best entry point into the region.
512 // Ensure that the first extracted block is the suggested entry point.
513 BlockSequence SubRegion = {SuggestedEntryPoint};
514 BasicBlock *NextEntryPoint = nullptr;
515 unsigned NextScore = 0;
516 auto RegionEndIt = Blocks.end();
517 auto RegionStartIt = remove_if(Blocks, [&](const BlockTy &Block) {
518 BasicBlock *BB = Block.first;
519 unsigned Score = Block.second;
520 bool InSubRegion =
521 BB == SuggestedEntryPoint || DT.dominates(SuggestedEntryPoint, BB);
522 if (!InSubRegion && Score > NextScore) {
523 NextEntryPoint = BB;
524 NextScore = Score;
525 }
526 if (InSubRegion && BB != SuggestedEntryPoint)
527 SubRegion.push_back(BB);
528 return InSubRegion;
529 });
530 Blocks.erase(RegionStartIt, RegionEndIt);
531
532 // Update the suggested entry point.
533 SuggestedEntryPoint = NextEntryPoint;
534
535 return SubRegion;
536 }
537 };
538 } // namespace
539
outlineColdRegions(Function & F,bool HasProfileSummary)540 bool HotColdSplitting::outlineColdRegions(Function &F, bool HasProfileSummary) {
541 bool Changed = false;
542
543 // The set of cold blocks.
544 SmallPtrSet<BasicBlock *, 4> ColdBlocks;
545
546 // The worklist of non-intersecting regions left to outline.
547 SmallVector<OutliningRegion, 2> OutliningWorklist;
548
549 // Set up an RPO traversal. Experimentally, this performs better (outlines
550 // more) than a PO traversal, because we prevent region overlap by keeping
551 // the first region to contain a block.
552 ReversePostOrderTraversal<Function *> RPOT(&F);
553
554 // Calculate domtrees lazily. This reduces compile-time significantly.
555 std::unique_ptr<DominatorTree> DT;
556 std::unique_ptr<PostDominatorTree> PDT;
557
558 // Calculate BFI lazily (it's only used to query ProfileSummaryInfo). This
559 // reduces compile-time significantly. TODO: When we *do* use BFI, we should
560 // be able to salvage its domtrees instead of recomputing them.
561 BlockFrequencyInfo *BFI = nullptr;
562 if (HasProfileSummary)
563 BFI = GetBFI(F);
564
565 TargetTransformInfo &TTI = GetTTI(F);
566 OptimizationRemarkEmitter &ORE = (*GetORE)(F);
567 AssumptionCache *AC = LookupAC(F);
568
569 // Find all cold regions.
570 for (BasicBlock *BB : RPOT) {
571 // This block is already part of some outlining region.
572 if (ColdBlocks.count(BB))
573 continue;
574
575 bool Cold = (BFI && PSI->isColdBlock(BB, BFI)) ||
576 (EnableStaticAnalyis && unlikelyExecuted(*BB));
577 if (!Cold)
578 continue;
579
580 LLVM_DEBUG({
581 dbgs() << "Found a cold block:\n";
582 BB->dump();
583 });
584
585 if (!DT)
586 DT = std::make_unique<DominatorTree>(F);
587 if (!PDT)
588 PDT = std::make_unique<PostDominatorTree>(F);
589
590 auto Regions = OutliningRegion::create(*BB, *DT, *PDT);
591 for (OutliningRegion &Region : Regions) {
592 if (Region.empty())
593 continue;
594
595 if (Region.isEntireFunctionCold()) {
596 LLVM_DEBUG(dbgs() << "Entire function is cold\n");
597 return markFunctionCold(F);
598 }
599
600 // If this outlining region intersects with another, drop the new region.
601 //
602 // TODO: It's theoretically possible to outline more by only keeping the
603 // largest region which contains a block, but the extra bookkeeping to do
604 // this is tricky/expensive.
605 bool RegionsOverlap = any_of(Region.blocks(), [&](const BlockTy &Block) {
606 return !ColdBlocks.insert(Block.first).second;
607 });
608 if (RegionsOverlap)
609 continue;
610
611 OutliningWorklist.emplace_back(std::move(Region));
612 ++NumColdRegionsFound;
613 }
614 }
615
616 if (OutliningWorklist.empty())
617 return Changed;
618
619 // Outline single-entry cold regions, splitting up larger regions as needed.
620 unsigned OutlinedFunctionID = 1;
621 // Cache and recycle the CodeExtractor analysis to avoid O(n^2) compile-time.
622 CodeExtractorAnalysisCache CEAC(F);
623 do {
624 OutliningRegion Region = OutliningWorklist.pop_back_val();
625 assert(!Region.empty() && "Empty outlining region in worklist");
626 do {
627 BlockSequence SubRegion = Region.takeSingleEntrySubRegion(*DT);
628 LLVM_DEBUG({
629 dbgs() << "Hot/cold splitting attempting to outline these blocks:\n";
630 for (BasicBlock *BB : SubRegion)
631 BB->dump();
632 });
633
634 Function *Outlined = extractColdRegion(SubRegion, CEAC, *DT, BFI, TTI,
635 ORE, AC, OutlinedFunctionID);
636 if (Outlined) {
637 ++OutlinedFunctionID;
638 Changed = true;
639 }
640 } while (!Region.empty());
641 } while (!OutliningWorklist.empty());
642
643 return Changed;
644 }
645
run(Module & M)646 bool HotColdSplitting::run(Module &M) {
647 bool Changed = false;
648 bool HasProfileSummary = (M.getProfileSummary(/* IsCS */ false) != nullptr);
649 for (auto It = M.begin(), End = M.end(); It != End; ++It) {
650 Function &F = *It;
651
652 // Do not touch declarations.
653 if (F.isDeclaration())
654 continue;
655
656 // Do not modify `optnone` functions.
657 if (F.hasOptNone())
658 continue;
659
660 // Detect inherently cold functions and mark them as such.
661 if (isFunctionCold(F)) {
662 Changed |= markFunctionCold(F);
663 continue;
664 }
665
666 if (!shouldOutlineFrom(F)) {
667 LLVM_DEBUG(llvm::dbgs() << "Skipping " << F.getName() << "\n");
668 continue;
669 }
670
671 LLVM_DEBUG(llvm::dbgs() << "Outlining in " << F.getName() << "\n");
672 Changed |= outlineColdRegions(F, HasProfileSummary);
673 }
674 return Changed;
675 }
676
runOnModule(Module & M)677 bool HotColdSplittingLegacyPass::runOnModule(Module &M) {
678 if (skipModule(M))
679 return false;
680 ProfileSummaryInfo *PSI =
681 &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
682 auto GTTI = [this](Function &F) -> TargetTransformInfo & {
683 return this->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
684 };
685 auto GBFI = [this](Function &F) {
686 return &this->getAnalysis<BlockFrequencyInfoWrapperPass>(F).getBFI();
687 };
688 std::unique_ptr<OptimizationRemarkEmitter> ORE;
689 std::function<OptimizationRemarkEmitter &(Function &)> GetORE =
690 [&ORE](Function &F) -> OptimizationRemarkEmitter & {
691 ORE.reset(new OptimizationRemarkEmitter(&F));
692 return *ORE.get();
693 };
694 auto LookupAC = [this](Function &F) -> AssumptionCache * {
695 if (auto *ACT = getAnalysisIfAvailable<AssumptionCacheTracker>())
696 return ACT->lookupAssumptionCache(F);
697 return nullptr;
698 };
699
700 return HotColdSplitting(PSI, GBFI, GTTI, &GetORE, LookupAC).run(M);
701 }
702
703 PreservedAnalyses
run(Module & M,ModuleAnalysisManager & AM)704 HotColdSplittingPass::run(Module &M, ModuleAnalysisManager &AM) {
705 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
706
707 auto LookupAC = [&FAM](Function &F) -> AssumptionCache * {
708 return FAM.getCachedResult<AssumptionAnalysis>(F);
709 };
710
711 auto GBFI = [&FAM](Function &F) {
712 return &FAM.getResult<BlockFrequencyAnalysis>(F);
713 };
714
715 std::function<TargetTransformInfo &(Function &)> GTTI =
716 [&FAM](Function &F) -> TargetTransformInfo & {
717 return FAM.getResult<TargetIRAnalysis>(F);
718 };
719
720 std::unique_ptr<OptimizationRemarkEmitter> ORE;
721 std::function<OptimizationRemarkEmitter &(Function &)> GetORE =
722 [&ORE](Function &F) -> OptimizationRemarkEmitter & {
723 ORE.reset(new OptimizationRemarkEmitter(&F));
724 return *ORE.get();
725 };
726
727 ProfileSummaryInfo *PSI = &AM.getResult<ProfileSummaryAnalysis>(M);
728
729 if (HotColdSplitting(PSI, GBFI, GTTI, &GetORE, LookupAC).run(M))
730 return PreservedAnalyses::none();
731 return PreservedAnalyses::all();
732 }
733
734 char HotColdSplittingLegacyPass::ID = 0;
735 INITIALIZE_PASS_BEGIN(HotColdSplittingLegacyPass, "hotcoldsplit",
736 "Hot Cold Splitting", false, false)
INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)737 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
738 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
739 INITIALIZE_PASS_END(HotColdSplittingLegacyPass, "hotcoldsplit",
740 "Hot Cold Splitting", false, false)
741
742 ModulePass *llvm::createHotColdSplittingPass() {
743 return new HotColdSplittingLegacyPass();
744 }
745