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/AssumptionCache.h" |
33 | #include "llvm/Analysis/BlockFrequencyInfo.h" |
34 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
35 | #include "llvm/Analysis/PostDominators.h" |
36 | #include "llvm/Analysis/ProfileSummaryInfo.h" |
37 | #include "llvm/Analysis/TargetTransformInfo.h" |
38 | #include "llvm/IR/BasicBlock.h" |
39 | #include "llvm/IR/CFG.h" |
40 | #include "llvm/IR/DiagnosticInfo.h" |
41 | #include "llvm/IR/Dominators.h" |
42 | #include "llvm/IR/EHPersonalities.h" |
43 | #include "llvm/IR/Function.h" |
44 | #include "llvm/IR/Instruction.h" |
45 | #include "llvm/IR/Instructions.h" |
46 | #include "llvm/IR/Module.h" |
47 | #include "llvm/IR/PassManager.h" |
48 | #include "llvm/IR/ProfDataUtils.h" |
49 | #include "llvm/IR/User.h" |
50 | #include "llvm/IR/Value.h" |
51 | #include "llvm/Support/CommandLine.h" |
52 | #include "llvm/Support/Debug.h" |
53 | #include "llvm/Support/raw_ostream.h" |
54 | #include "llvm/Transforms/IPO.h" |
55 | #include "llvm/Transforms/Utils/CodeExtractor.h" |
56 | #include <algorithm> |
57 | #include <cassert> |
58 | #include <limits> |
59 | #include <string> |
60 | |
61 | #define DEBUG_TYPE "hotcoldsplit" |
62 | |
63 | STATISTIC(NumColdRegionsFound, "Number of cold regions found." ); |
64 | STATISTIC(NumColdRegionsOutlined, "Number of cold regions outlined." ); |
65 | |
66 | using namespace llvm; |
67 | |
68 | static cl::opt<bool> EnableStaticAnalysis("hot-cold-static-analysis" , |
69 | cl::init(Val: true), cl::Hidden); |
70 | |
71 | static cl::opt<int> |
72 | SplittingThreshold("hotcoldsplit-threshold" , cl::init(Val: 2), cl::Hidden, |
73 | cl::desc("Base penalty for splitting cold code (as a " |
74 | "multiple of TCC_Basic)" )); |
75 | |
76 | static cl::opt<bool> EnableColdSection( |
77 | "enable-cold-section" , cl::init(Val: false), cl::Hidden, |
78 | cl::desc("Enable placement of extracted cold functions" |
79 | " into a separate section after hot-cold splitting." )); |
80 | |
81 | static cl::opt<std::string> |
82 | ColdSectionName("hotcoldsplit-cold-section-name" , cl::init(Val: "__llvm_cold" ), |
83 | cl::Hidden, |
84 | cl::desc("Name for the section containing cold functions " |
85 | "extracted by hot-cold splitting." )); |
86 | |
87 | static cl::opt<int> MaxParametersForSplit( |
88 | "hotcoldsplit-max-params" , cl::init(Val: 4), cl::Hidden, |
89 | cl::desc("Maximum number of parameters for a split function" )); |
90 | |
91 | static cl::opt<int> ColdBranchProbDenom( |
92 | "hotcoldsplit-cold-probability-denom" , cl::init(Val: 100), cl::Hidden, |
93 | cl::desc("Divisor of cold branch probability." |
94 | "BranchProbability = 1/ColdBranchProbDenom" )); |
95 | |
96 | namespace { |
97 | // Same as blockEndsInUnreachable in CodeGen/BranchFolding.cpp. Do not modify |
98 | // this function unless you modify the MBB version as well. |
99 | // |
100 | /// A no successor, non-return block probably ends in unreachable and is cold. |
101 | /// Also consider a block that ends in an indirect branch to be a return block, |
102 | /// since many targets use plain indirect branches to return. |
103 | bool blockEndsInUnreachable(const BasicBlock &BB) { |
104 | if (!succ_empty(BB: &BB)) |
105 | return false; |
106 | if (BB.empty()) |
107 | return true; |
108 | const Instruction *I = BB.getTerminator(); |
109 | return !(isa<ReturnInst>(Val: I) || isa<IndirectBrInst>(Val: I)); |
110 | } |
111 | |
112 | void analyzeProfMetadata(BasicBlock *BB, |
113 | BranchProbability ColdProbThresh, |
114 | SmallPtrSetImpl<BasicBlock *> &AnnotatedColdBlocks) { |
115 | // TODO: Handle branches with > 2 successors. |
116 | BranchInst *CondBr = dyn_cast<BranchInst>(Val: BB->getTerminator()); |
117 | if (!CondBr) |
118 | return; |
119 | |
120 | uint64_t TrueWt, FalseWt; |
121 | if (!extractBranchWeights(I: *CondBr, TrueVal&: TrueWt, FalseVal&: FalseWt)) |
122 | return; |
123 | |
124 | auto SumWt = TrueWt + FalseWt; |
125 | if (SumWt == 0) |
126 | return; |
127 | |
128 | auto TrueProb = BranchProbability::getBranchProbability(Numerator: TrueWt, Denominator: SumWt); |
129 | auto FalseProb = BranchProbability::getBranchProbability(Numerator: FalseWt, Denominator: SumWt); |
130 | |
131 | if (TrueProb <= ColdProbThresh) |
132 | AnnotatedColdBlocks.insert(Ptr: CondBr->getSuccessor(i: 0)); |
133 | |
134 | if (FalseProb <= ColdProbThresh) |
135 | AnnotatedColdBlocks.insert(Ptr: CondBr->getSuccessor(i: 1)); |
136 | } |
137 | |
138 | bool unlikelyExecuted(BasicBlock &BB) { |
139 | // Exception handling blocks are unlikely executed. |
140 | if (BB.isEHPad() || isa<ResumeInst>(Val: BB.getTerminator())) |
141 | return true; |
142 | |
143 | // The block is cold if it calls/invokes a cold function. However, do not |
144 | // mark sanitizer traps as cold. |
145 | for (Instruction &I : BB) |
146 | if (auto *CB = dyn_cast<CallBase>(Val: &I)) |
147 | if (CB->hasFnAttr(Kind: Attribute::Cold) && |
148 | !CB->getMetadata(KindID: LLVMContext::MD_nosanitize)) |
149 | return true; |
150 | |
151 | // The block is cold if it has an unreachable terminator, unless it's |
152 | // preceded by a call to a (possibly warm) noreturn call (e.g. longjmp). |
153 | if (blockEndsInUnreachable(BB)) { |
154 | if (auto *CI = |
155 | dyn_cast_or_null<CallInst>(Val: BB.getTerminator()->getPrevNode())) |
156 | if (CI->hasFnAttr(Kind: Attribute::NoReturn)) |
157 | return false; |
158 | return true; |
159 | } |
160 | |
161 | return false; |
162 | } |
163 | |
164 | /// Check whether it's safe to outline \p BB. |
165 | static bool (const BasicBlock &BB) { |
166 | // EH pads are unsafe to outline because doing so breaks EH type tables. It |
167 | // follows that invoke instructions cannot be extracted, because CodeExtractor |
168 | // requires unwind destinations to be within the extraction region. |
169 | // |
170 | // Resumes that are not reachable from a cleanup landing pad are considered to |
171 | // be unreachable. It’s not safe to split them out either. |
172 | |
173 | if (BB.hasAddressTaken() || BB.isEHPad()) |
174 | return false; |
175 | auto Term = BB.getTerminator(); |
176 | if (isa<InvokeInst>(Val: Term) || isa<ResumeInst>(Val: Term)) |
177 | return false; |
178 | |
179 | // Do not outline basic blocks that have token type instructions. e.g., |
180 | // exception: |
181 | // %0 = cleanuppad within none [] |
182 | // call void @"?terminate@@YAXXZ"() [ "funclet"(token %0) ] |
183 | // br label %continue-exception |
184 | if (llvm::any_of( |
185 | Range: BB, P: [](const Instruction &I) { return I.getType()->isTokenTy(); })) { |
186 | return false; |
187 | } |
188 | |
189 | return true; |
190 | } |
191 | |
192 | /// Mark \p F cold. Based on this assumption, also optimize it for minimum size. |
193 | /// If \p UpdateEntryCount is true (set when this is a new split function and |
194 | /// module has profile data), set entry count to 0 to ensure treated as cold. |
195 | /// Return true if the function is changed. |
196 | static bool markFunctionCold(Function &F, bool UpdateEntryCount = false) { |
197 | assert(!F.hasOptNone() && "Can't mark this cold" ); |
198 | bool Changed = false; |
199 | if (!F.hasFnAttribute(Kind: Attribute::Cold)) { |
200 | F.addFnAttr(Kind: Attribute::Cold); |
201 | Changed = true; |
202 | } |
203 | if (!F.hasFnAttribute(Kind: Attribute::MinSize)) { |
204 | F.addFnAttr(Kind: Attribute::MinSize); |
205 | Changed = true; |
206 | } |
207 | if (UpdateEntryCount) { |
208 | // Set the entry count to 0 to ensure it is placed in the unlikely text |
209 | // section when function sections are enabled. |
210 | F.setEntryCount(Count: 0); |
211 | Changed = true; |
212 | } |
213 | |
214 | return Changed; |
215 | } |
216 | |
217 | } // end anonymous namespace |
218 | |
219 | /// Check whether \p F is inherently cold. |
220 | bool HotColdSplitting::isFunctionCold(const Function &F) const { |
221 | if (F.hasFnAttribute(Kind: Attribute::Cold)) |
222 | return true; |
223 | |
224 | if (F.getCallingConv() == CallingConv::Cold) |
225 | return true; |
226 | |
227 | if (PSI->isFunctionEntryCold(F: &F)) |
228 | return true; |
229 | |
230 | return false; |
231 | } |
232 | |
233 | bool HotColdSplitting::isBasicBlockCold( |
234 | BasicBlock *BB, BranchProbability ColdProbThresh, |
235 | SmallPtrSetImpl<BasicBlock *> &AnnotatedColdBlocks, |
236 | BlockFrequencyInfo *BFI) const { |
237 | if (BFI) { |
238 | if (PSI->isColdBlock(BB, BFI)) |
239 | return true; |
240 | } else { |
241 | // Find cold blocks of successors of BB during a reverse postorder traversal. |
242 | analyzeProfMetadata(BB, ColdProbThresh, AnnotatedColdBlocks); |
243 | |
244 | // A statically cold BB would be known before it is visited |
245 | // because the prof-data of incoming edges are 'analyzed' as part of RPOT. |
246 | if (AnnotatedColdBlocks.count(Ptr: BB)) |
247 | return true; |
248 | } |
249 | |
250 | if (EnableStaticAnalysis && unlikelyExecuted(BB&: *BB)) |
251 | return true; |
252 | |
253 | return false; |
254 | } |
255 | |
256 | // Returns false if the function should not be considered for hot-cold split |
257 | // optimization. |
258 | bool HotColdSplitting::shouldOutlineFrom(const Function &F) const { |
259 | if (F.hasFnAttribute(Kind: Attribute::AlwaysInline)) |
260 | return false; |
261 | |
262 | if (F.hasFnAttribute(Kind: Attribute::NoInline)) |
263 | return false; |
264 | |
265 | // A function marked `noreturn` may contain unreachable terminators: these |
266 | // should not be considered cold, as the function may be a trampoline. |
267 | if (F.hasFnAttribute(Kind: Attribute::NoReturn)) |
268 | return false; |
269 | |
270 | if (F.hasFnAttribute(Kind: Attribute::SanitizeAddress) || |
271 | F.hasFnAttribute(Kind: Attribute::SanitizeHWAddress) || |
272 | F.hasFnAttribute(Kind: Attribute::SanitizeThread) || |
273 | F.hasFnAttribute(Kind: Attribute::SanitizeMemory)) |
274 | return false; |
275 | |
276 | // Do not outline scoped EH personality functions. |
277 | if (F.hasPersonalityFn()) |
278 | if (isScopedEHPersonality(Pers: classifyEHPersonality(Pers: F.getPersonalityFn()))) |
279 | return false; |
280 | |
281 | return true; |
282 | } |
283 | |
284 | /// Get the benefit score of outlining \p Region. |
285 | static InstructionCost getOutliningBenefit(ArrayRef<BasicBlock *> Region, |
286 | TargetTransformInfo &TTI) { |
287 | // Sum up the code size costs of non-terminator instructions. Tight coupling |
288 | // with \ref getOutliningPenalty is needed to model the costs of terminators. |
289 | InstructionCost Benefit = 0; |
290 | for (BasicBlock *BB : Region) |
291 | for (Instruction &I : BB->instructionsWithoutDebug()) |
292 | if (&I != BB->getTerminator()) |
293 | Benefit += |
294 | TTI.getInstructionCost(U: &I, CostKind: TargetTransformInfo::TCK_CodeSize); |
295 | |
296 | return Benefit; |
297 | } |
298 | |
299 | /// Get the penalty score for outlining \p Region. |
300 | static int getOutliningPenalty(ArrayRef<BasicBlock *> Region, |
301 | unsigned NumInputs, unsigned NumOutputs) { |
302 | int Penalty = SplittingThreshold; |
303 | LLVM_DEBUG(dbgs() << "Applying penalty for splitting: " << Penalty << "\n" ); |
304 | |
305 | // If the splitting threshold is set at or below zero, skip the usual |
306 | // profitability check. |
307 | if (SplittingThreshold <= 0) |
308 | return Penalty; |
309 | |
310 | // Find the number of distinct exit blocks for the region. Use a conservative |
311 | // check to determine whether control returns from the region. |
312 | bool NoBlocksReturn = true; |
313 | SmallPtrSet<BasicBlock *, 2> SuccsOutsideRegion; |
314 | for (BasicBlock *BB : Region) { |
315 | // If a block has no successors, only assume it does not return if it's |
316 | // unreachable. |
317 | if (succ_empty(BB)) { |
318 | NoBlocksReturn &= isa<UnreachableInst>(Val: BB->getTerminator()); |
319 | continue; |
320 | } |
321 | |
322 | for (BasicBlock *SuccBB : successors(BB)) { |
323 | if (!is_contained(Range&: Region, Element: SuccBB)) { |
324 | NoBlocksReturn = false; |
325 | SuccsOutsideRegion.insert(Ptr: SuccBB); |
326 | } |
327 | } |
328 | } |
329 | |
330 | // Count the number of phis in exit blocks with >= 2 incoming values from the |
331 | // outlining region. These phis are split (\ref severSplitPHINodesOfExits), |
332 | // and new outputs are created to supply the split phis. CodeExtractor can't |
333 | // report these new outputs until extraction begins, but it's important to |
334 | // factor the cost of the outputs into the cost calculation. |
335 | unsigned NumSplitExitPhis = 0; |
336 | for (BasicBlock *ExitBB : SuccsOutsideRegion) { |
337 | for (PHINode &PN : ExitBB->phis()) { |
338 | // Find all incoming values from the outlining region. |
339 | int NumIncomingVals = 0; |
340 | for (unsigned i = 0; i < PN.getNumIncomingValues(); ++i) |
341 | if (llvm::is_contained(Range&: Region, Element: PN.getIncomingBlock(i))) { |
342 | ++NumIncomingVals; |
343 | if (NumIncomingVals > 1) { |
344 | ++NumSplitExitPhis; |
345 | break; |
346 | } |
347 | } |
348 | } |
349 | } |
350 | |
351 | // Apply a penalty for calling the split function. Factor in the cost of |
352 | // materializing all of the parameters. |
353 | int NumOutputsAndSplitPhis = NumOutputs + NumSplitExitPhis; |
354 | int NumParams = NumInputs + NumOutputsAndSplitPhis; |
355 | if (NumParams > MaxParametersForSplit) { |
356 | LLVM_DEBUG(dbgs() << NumInputs << " inputs and " << NumOutputsAndSplitPhis |
357 | << " outputs exceeds parameter limit (" |
358 | << MaxParametersForSplit << ")\n" ); |
359 | return std::numeric_limits<int>::max(); |
360 | } |
361 | const int CostForArgMaterialization = 2 * TargetTransformInfo::TCC_Basic; |
362 | LLVM_DEBUG(dbgs() << "Applying penalty for: " << NumParams << " params\n" ); |
363 | Penalty += CostForArgMaterialization * NumParams; |
364 | |
365 | // Apply the typical code size cost for an output alloca and its associated |
366 | // reload in the caller. Also penalize the associated store in the callee. |
367 | LLVM_DEBUG(dbgs() << "Applying penalty for: " << NumOutputsAndSplitPhis |
368 | << " outputs/split phis\n" ); |
369 | const int CostForRegionOutput = 3 * TargetTransformInfo::TCC_Basic; |
370 | Penalty += CostForRegionOutput * NumOutputsAndSplitPhis; |
371 | |
372 | // Apply a `noreturn` bonus. |
373 | if (NoBlocksReturn) { |
374 | LLVM_DEBUG(dbgs() << "Applying bonus for: " << Region.size() |
375 | << " non-returning terminators\n" ); |
376 | Penalty -= Region.size(); |
377 | } |
378 | |
379 | // Apply a penalty for having more than one successor outside of the region. |
380 | // This penalty accounts for the switch needed in the caller. |
381 | if (SuccsOutsideRegion.size() > 1) { |
382 | LLVM_DEBUG(dbgs() << "Applying penalty for: " << SuccsOutsideRegion.size() |
383 | << " non-region successors\n" ); |
384 | Penalty += (SuccsOutsideRegion.size() - 1) * TargetTransformInfo::TCC_Basic; |
385 | } |
386 | |
387 | return Penalty; |
388 | } |
389 | |
390 | // Determine if it is beneficial to split the \p Region. |
391 | bool HotColdSplitting::(CodeExtractor &CE, |
392 | const BlockSequence &Region, |
393 | TargetTransformInfo &TTI) { |
394 | assert(!Region.empty()); |
395 | |
396 | // Perform a simple cost/benefit analysis to decide whether or not to permit |
397 | // splitting. |
398 | SetVector<Value *> Inputs, Outputs, Sinks; |
399 | CE.findInputsOutputs(Inputs, Outputs, Allocas: Sinks); |
400 | InstructionCost OutliningBenefit = getOutliningBenefit(Region, TTI); |
401 | int OutliningPenalty = |
402 | getOutliningPenalty(Region, NumInputs: Inputs.size(), NumOutputs: Outputs.size()); |
403 | LLVM_DEBUG(dbgs() << "Split profitability: benefit = " << OutliningBenefit |
404 | << ", penalty = " << OutliningPenalty << "\n" ); |
405 | if (!OutliningBenefit.isValid() || OutliningBenefit <= OutliningPenalty) |
406 | return false; |
407 | |
408 | return true; |
409 | } |
410 | |
411 | // Split the single \p EntryPoint cold region. \p CE is the region code |
412 | // extractor. |
413 | Function *HotColdSplitting::( |
414 | BasicBlock &EntryPoint, CodeExtractor &CE, |
415 | const CodeExtractorAnalysisCache &CEAC, BlockFrequencyInfo *BFI, |
416 | TargetTransformInfo &TTI, OptimizationRemarkEmitter &ORE) { |
417 | Function *OrigF = EntryPoint.getParent(); |
418 | if (Function *OutF = CE.extractCodeRegion(CEAC)) { |
419 | User *U = *OutF->user_begin(); |
420 | CallInst *CI = cast<CallInst>(Val: U); |
421 | NumColdRegionsOutlined++; |
422 | if (TTI.useColdCCForColdCall(F&: *OutF)) { |
423 | OutF->setCallingConv(CallingConv::Cold); |
424 | CI->setCallingConv(CallingConv::Cold); |
425 | } |
426 | CI->setIsNoInline(); |
427 | |
428 | if (EnableColdSection) |
429 | OutF->setSection(ColdSectionName); |
430 | else { |
431 | if (OrigF->hasSection()) |
432 | OutF->setSection(OrigF->getSection()); |
433 | } |
434 | |
435 | markFunctionCold(F&: *OutF, UpdateEntryCount: BFI != nullptr); |
436 | |
437 | LLVM_DEBUG(llvm::dbgs() << "Outlined Region: " << *OutF); |
438 | ORE.emit(RemarkBuilder: [&]() { |
439 | return OptimizationRemark(DEBUG_TYPE, "HotColdSplit" , |
440 | &*EntryPoint.begin()) |
441 | << ore::NV("Original" , OrigF) << " split cold code into " |
442 | << ore::NV("Split" , OutF); |
443 | }); |
444 | return OutF; |
445 | } |
446 | |
447 | ORE.emit(RemarkBuilder: [&]() { |
448 | return OptimizationRemarkMissed(DEBUG_TYPE, "ExtractFailed" , |
449 | &*EntryPoint.begin()) |
450 | << "Failed to extract region at block " |
451 | << ore::NV("Block" , &EntryPoint); |
452 | }); |
453 | return nullptr; |
454 | } |
455 | |
456 | /// A pair of (basic block, score). |
457 | using BlockTy = std::pair<BasicBlock *, unsigned>; |
458 | |
459 | namespace { |
460 | /// A maximal outlining region. This contains all blocks post-dominated by a |
461 | /// sink block, the sink block itself, and all blocks dominated by the sink. |
462 | /// If sink-predecessors and sink-successors cannot be extracted in one region, |
463 | /// the static constructor returns a list of suitable extraction regions. |
464 | class OutliningRegion { |
465 | /// A list of (block, score) pairs. A block's score is non-zero iff it's a |
466 | /// viable sub-region entry point. Blocks with higher scores are better entry |
467 | /// points (i.e. they are more distant ancestors of the sink block). |
468 | SmallVector<BlockTy, 0> Blocks = {}; |
469 | |
470 | /// The suggested entry point into the region. If the region has multiple |
471 | /// entry points, all blocks within the region may not be reachable from this |
472 | /// entry point. |
473 | BasicBlock *SuggestedEntryPoint = nullptr; |
474 | |
475 | /// Whether the entire function is cold. |
476 | bool EntireFunctionCold = false; |
477 | |
478 | /// If \p BB is a viable entry point, return \p Score. Return 0 otherwise. |
479 | static unsigned getEntryPointScore(BasicBlock &BB, unsigned Score) { |
480 | return mayExtractBlock(BB) ? Score : 0; |
481 | } |
482 | |
483 | /// These scores should be lower than the score for predecessor blocks, |
484 | /// because regions starting at predecessor blocks are typically larger. |
485 | static constexpr unsigned ScoreForSuccBlock = 1; |
486 | static constexpr unsigned ScoreForSinkBlock = 1; |
487 | |
488 | OutliningRegion(const OutliningRegion &) = delete; |
489 | OutliningRegion &operator=(const OutliningRegion &) = delete; |
490 | |
491 | public: |
492 | OutliningRegion() = default; |
493 | OutliningRegion(OutliningRegion &&) = default; |
494 | OutliningRegion &operator=(OutliningRegion &&) = default; |
495 | |
496 | static std::vector<OutliningRegion> create(BasicBlock &SinkBB, |
497 | const DominatorTree &DT, |
498 | const PostDominatorTree &PDT) { |
499 | std::vector<OutliningRegion> Regions; |
500 | SmallPtrSet<BasicBlock *, 4> RegionBlocks; |
501 | |
502 | Regions.emplace_back(); |
503 | OutliningRegion *ColdRegion = &Regions.back(); |
504 | |
505 | auto addBlockToRegion = [&](BasicBlock *BB, unsigned Score) { |
506 | RegionBlocks.insert(Ptr: BB); |
507 | ColdRegion->Blocks.emplace_back(Args&: BB, Args&: Score); |
508 | }; |
509 | |
510 | // The ancestor farthest-away from SinkBB, and also post-dominated by it. |
511 | unsigned SinkScore = getEntryPointScore(BB&: SinkBB, Score: ScoreForSinkBlock); |
512 | ColdRegion->SuggestedEntryPoint = (SinkScore > 0) ? &SinkBB : nullptr; |
513 | unsigned BestScore = SinkScore; |
514 | |
515 | // Visit SinkBB's ancestors using inverse DFS. |
516 | auto PredIt = ++idf_begin(G: &SinkBB); |
517 | auto PredEnd = idf_end(G: &SinkBB); |
518 | while (PredIt != PredEnd) { |
519 | BasicBlock &PredBB = **PredIt; |
520 | bool SinkPostDom = PDT.dominates(A: &SinkBB, B: &PredBB); |
521 | |
522 | // If the predecessor is cold and has no predecessors, the entire |
523 | // function must be cold. |
524 | if (SinkPostDom && pred_empty(BB: &PredBB)) { |
525 | ColdRegion->EntireFunctionCold = true; |
526 | return Regions; |
527 | } |
528 | |
529 | // If SinkBB does not post-dominate a predecessor, do not mark the |
530 | // predecessor (or any of its predecessors) cold. |
531 | if (!SinkPostDom || !mayExtractBlock(BB: PredBB)) { |
532 | PredIt.skipChildren(); |
533 | continue; |
534 | } |
535 | |
536 | // Keep track of the post-dominated ancestor farthest away from the sink. |
537 | // The path length is always >= 2, ensuring that predecessor blocks are |
538 | // considered as entry points before the sink block. |
539 | unsigned PredScore = getEntryPointScore(BB&: PredBB, Score: PredIt.getPathLength()); |
540 | if (PredScore > BestScore) { |
541 | ColdRegion->SuggestedEntryPoint = &PredBB; |
542 | BestScore = PredScore; |
543 | } |
544 | |
545 | addBlockToRegion(&PredBB, PredScore); |
546 | ++PredIt; |
547 | } |
548 | |
549 | // If the sink can be added to the cold region, do so. It's considered as |
550 | // an entry point before any sink-successor blocks. |
551 | // |
552 | // Otherwise, split cold sink-successor blocks using a separate region. |
553 | // This satisfies the requirement that all extraction blocks other than the |
554 | // first have predecessors within the extraction region. |
555 | if (mayExtractBlock(BB: SinkBB)) { |
556 | addBlockToRegion(&SinkBB, SinkScore); |
557 | if (pred_empty(BB: &SinkBB)) { |
558 | ColdRegion->EntireFunctionCold = true; |
559 | return Regions; |
560 | } |
561 | } else { |
562 | Regions.emplace_back(); |
563 | ColdRegion = &Regions.back(); |
564 | BestScore = 0; |
565 | } |
566 | |
567 | // Find all successors of SinkBB dominated by SinkBB using DFS. |
568 | auto SuccIt = ++df_begin(G: &SinkBB); |
569 | auto SuccEnd = df_end(G: &SinkBB); |
570 | while (SuccIt != SuccEnd) { |
571 | BasicBlock &SuccBB = **SuccIt; |
572 | bool SinkDom = DT.dominates(A: &SinkBB, B: &SuccBB); |
573 | |
574 | // Don't allow the backwards & forwards DFSes to mark the same block. |
575 | bool DuplicateBlock = RegionBlocks.count(Ptr: &SuccBB); |
576 | |
577 | // If SinkBB does not dominate a successor, do not mark the successor (or |
578 | // any of its successors) cold. |
579 | if (DuplicateBlock || !SinkDom || !mayExtractBlock(BB: SuccBB)) { |
580 | SuccIt.skipChildren(); |
581 | continue; |
582 | } |
583 | |
584 | unsigned SuccScore = getEntryPointScore(BB&: SuccBB, Score: ScoreForSuccBlock); |
585 | if (SuccScore > BestScore) { |
586 | ColdRegion->SuggestedEntryPoint = &SuccBB; |
587 | BestScore = SuccScore; |
588 | } |
589 | |
590 | addBlockToRegion(&SuccBB, SuccScore); |
591 | ++SuccIt; |
592 | } |
593 | |
594 | return Regions; |
595 | } |
596 | |
597 | /// Whether this region has nothing to extract. |
598 | bool empty() const { return !SuggestedEntryPoint; } |
599 | |
600 | /// The blocks in this region. |
601 | ArrayRef<std::pair<BasicBlock *, unsigned>> blocks() const { return Blocks; } |
602 | |
603 | /// Whether the entire function containing this region is cold. |
604 | bool isEntireFunctionCold() const { return EntireFunctionCold; } |
605 | |
606 | /// Remove a sub-region from this region and return it as a block sequence. |
607 | BlockSequence takeSingleEntrySubRegion(DominatorTree &DT) { |
608 | assert(!empty() && !isEntireFunctionCold() && "Nothing to extract" ); |
609 | |
610 | // Remove blocks dominated by the suggested entry point from this region. |
611 | // During the removal, identify the next best entry point into the region. |
612 | // Ensure that the first extracted block is the suggested entry point. |
613 | BlockSequence SubRegion = {SuggestedEntryPoint}; |
614 | BasicBlock *NextEntryPoint = nullptr; |
615 | unsigned NextScore = 0; |
616 | auto RegionEndIt = Blocks.end(); |
617 | auto RegionStartIt = remove_if(Range&: Blocks, P: [&](const BlockTy &Block) { |
618 | BasicBlock *BB = Block.first; |
619 | unsigned Score = Block.second; |
620 | bool InSubRegion = |
621 | BB == SuggestedEntryPoint || DT.dominates(A: SuggestedEntryPoint, B: BB); |
622 | if (!InSubRegion && Score > NextScore) { |
623 | NextEntryPoint = BB; |
624 | NextScore = Score; |
625 | } |
626 | if (InSubRegion && BB != SuggestedEntryPoint) |
627 | SubRegion.push_back(Elt: BB); |
628 | return InSubRegion; |
629 | }); |
630 | Blocks.erase(CS: RegionStartIt, CE: RegionEndIt); |
631 | |
632 | // Update the suggested entry point. |
633 | SuggestedEntryPoint = NextEntryPoint; |
634 | |
635 | return SubRegion; |
636 | } |
637 | }; |
638 | } // namespace |
639 | |
640 | bool HotColdSplitting::outlineColdRegions(Function &F, bool HasProfileSummary) { |
641 | // The set of cold blocks outlined. |
642 | SmallPtrSet<BasicBlock *, 4> ColdBlocks; |
643 | |
644 | // The set of cold blocks cannot be outlined. |
645 | SmallPtrSet<BasicBlock *, 4> CannotBeOutlinedColdBlocks; |
646 | |
647 | // Set of cold blocks obtained with RPOT. |
648 | SmallPtrSet<BasicBlock *, 4> AnnotatedColdBlocks; |
649 | |
650 | // The worklist of non-intersecting regions left to outline. The first member |
651 | // of the pair is the entry point into the region to be outlined. |
652 | SmallVector<std::pair<BasicBlock *, CodeExtractor>, 2> OutliningWorklist; |
653 | |
654 | // Set up an RPO traversal. Experimentally, this performs better (outlines |
655 | // more) than a PO traversal, because we prevent region overlap by keeping |
656 | // the first region to contain a block. |
657 | ReversePostOrderTraversal<Function *> RPOT(&F); |
658 | |
659 | // Calculate domtrees lazily. This reduces compile-time significantly. |
660 | std::unique_ptr<DominatorTree> DT; |
661 | std::unique_ptr<PostDominatorTree> PDT; |
662 | |
663 | // Calculate BFI lazily (it's only used to query ProfileSummaryInfo). This |
664 | // reduces compile-time significantly. TODO: When we *do* use BFI, we should |
665 | // be able to salvage its domtrees instead of recomputing them. |
666 | BlockFrequencyInfo *BFI = nullptr; |
667 | if (HasProfileSummary) |
668 | BFI = GetBFI(F); |
669 | |
670 | TargetTransformInfo &TTI = GetTTI(F); |
671 | OptimizationRemarkEmitter &ORE = (*GetORE)(F); |
672 | AssumptionCache *AC = LookupAC(F); |
673 | auto ColdProbThresh = TTI.getPredictableBranchThreshold().getCompl(); |
674 | |
675 | if (ColdBranchProbDenom.getNumOccurrences()) |
676 | ColdProbThresh = BranchProbability(1, ColdBranchProbDenom.getValue()); |
677 | |
678 | unsigned OutlinedFunctionID = 1; |
679 | // Find all cold regions. |
680 | for (BasicBlock *BB : RPOT) { |
681 | // This block is already part of some outlining region. |
682 | if (ColdBlocks.count(Ptr: BB)) |
683 | continue; |
684 | |
685 | // This block is already part of some region cannot be outlined. |
686 | if (CannotBeOutlinedColdBlocks.count(Ptr: BB)) |
687 | continue; |
688 | |
689 | if (!isBasicBlockCold(BB, ColdProbThresh, AnnotatedColdBlocks, BFI)) |
690 | continue; |
691 | |
692 | LLVM_DEBUG({ |
693 | dbgs() << "Found a cold block:\n" ; |
694 | BB->dump(); |
695 | }); |
696 | |
697 | if (!DT) |
698 | DT = std::make_unique<DominatorTree>(args&: F); |
699 | if (!PDT) |
700 | PDT = std::make_unique<PostDominatorTree>(args&: F); |
701 | |
702 | auto Regions = OutliningRegion::create(SinkBB&: *BB, DT: *DT, PDT: *PDT); |
703 | for (OutliningRegion &Region : Regions) { |
704 | if (Region.empty()) |
705 | continue; |
706 | |
707 | if (Region.isEntireFunctionCold()) { |
708 | LLVM_DEBUG(dbgs() << "Entire function is cold\n" ); |
709 | return markFunctionCold(F); |
710 | } |
711 | |
712 | do { |
713 | BlockSequence SubRegion = Region.takeSingleEntrySubRegion(DT&: *DT); |
714 | LLVM_DEBUG({ |
715 | dbgs() << "Hot/cold splitting attempting to outline these blocks:\n" ; |
716 | for (BasicBlock *BB : SubRegion) |
717 | BB->dump(); |
718 | }); |
719 | |
720 | // TODO: Pass BFI and BPI to update profile information. |
721 | CodeExtractor CE( |
722 | SubRegion, &*DT, /* AggregateArgs */ false, /* BFI */ nullptr, |
723 | /* BPI */ nullptr, AC, /* AllowVarArgs */ false, |
724 | /* AllowAlloca */ false, /* AllocaBlock */ nullptr, |
725 | /* Suffix */ "cold." + std::to_string(val: OutlinedFunctionID)); |
726 | |
727 | if (CE.isEligible() && isSplittingBeneficial(CE, Region: SubRegion, TTI) && |
728 | // If this outlining region intersects with another, drop the new |
729 | // region. |
730 | // |
731 | // TODO: It's theoretically possible to outline more by only keeping |
732 | // the largest region which contains a block, but the extra |
733 | // bookkeeping to do this is tricky/expensive. |
734 | none_of(Range&: SubRegion, P: [&](BasicBlock *Block) { |
735 | return ColdBlocks.contains(Ptr: Block); |
736 | })) { |
737 | ColdBlocks.insert(I: SubRegion.begin(), E: SubRegion.end()); |
738 | |
739 | LLVM_DEBUG({ |
740 | for (auto *Block : SubRegion) |
741 | dbgs() << " contains cold block:" << Block->getName() << "\n" ; |
742 | }); |
743 | |
744 | OutliningWorklist.emplace_back( |
745 | Args: std::make_pair(x&: SubRegion[0], y: std::move(CE))); |
746 | ++OutlinedFunctionID; |
747 | } else { |
748 | // The cold block region cannot be outlined. |
749 | for (auto *Block : SubRegion) |
750 | if ((DT->dominates(A: BB, B: Block) && PDT->dominates(A: Block, B: BB)) || |
751 | (PDT->dominates(A: BB, B: Block) && DT->dominates(A: Block, B: BB))) |
752 | // Will skip this cold block in the loop to save the compile time |
753 | CannotBeOutlinedColdBlocks.insert(Ptr: Block); |
754 | } |
755 | } while (!Region.empty()); |
756 | |
757 | ++NumColdRegionsFound; |
758 | } |
759 | } |
760 | |
761 | if (OutliningWorklist.empty()) |
762 | return false; |
763 | |
764 | // Outline single-entry cold regions, splitting up larger regions as needed. |
765 | // Cache and recycle the CodeExtractor analysis to avoid O(n^2) compile-time. |
766 | CodeExtractorAnalysisCache CEAC(F); |
767 | for (auto &BCE : OutliningWorklist) { |
768 | Function *Outlined = |
769 | extractColdRegion(EntryPoint&: *BCE.first, CE&: BCE.second, CEAC, BFI, TTI, ORE); |
770 | assert(Outlined && "Should be outlined" ); |
771 | (void)Outlined; |
772 | } |
773 | |
774 | return true; |
775 | } |
776 | |
777 | bool HotColdSplitting::run(Module &M) { |
778 | bool Changed = false; |
779 | bool HasProfileSummary = (M.getProfileSummary(/* IsCS */ false) != nullptr); |
780 | for (Function &F : M) { |
781 | // Do not touch declarations. |
782 | if (F.isDeclaration()) |
783 | continue; |
784 | |
785 | // Do not modify `optnone` functions. |
786 | if (F.hasOptNone()) |
787 | continue; |
788 | |
789 | // Detect inherently cold functions and mark them as such. |
790 | if (isFunctionCold(F)) { |
791 | Changed |= markFunctionCold(F); |
792 | continue; |
793 | } |
794 | |
795 | if (!shouldOutlineFrom(F)) { |
796 | LLVM_DEBUG(llvm::dbgs() << "Skipping " << F.getName() << "\n" ); |
797 | continue; |
798 | } |
799 | |
800 | LLVM_DEBUG(llvm::dbgs() << "Outlining in " << F.getName() << "\n" ); |
801 | Changed |= outlineColdRegions(F, HasProfileSummary); |
802 | } |
803 | return Changed; |
804 | } |
805 | |
806 | PreservedAnalyses |
807 | HotColdSplittingPass::run(Module &M, ModuleAnalysisManager &AM) { |
808 | auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager(); |
809 | |
810 | auto LookupAC = [&FAM](Function &F) -> AssumptionCache * { |
811 | return FAM.getCachedResult<AssumptionAnalysis>(IR&: F); |
812 | }; |
813 | |
814 | auto GBFI = [&FAM](Function &F) { |
815 | return &FAM.getResult<BlockFrequencyAnalysis>(IR&: F); |
816 | }; |
817 | |
818 | std::function<TargetTransformInfo &(Function &)> GTTI = |
819 | [&FAM](Function &F) -> TargetTransformInfo & { |
820 | return FAM.getResult<TargetIRAnalysis>(IR&: F); |
821 | }; |
822 | |
823 | std::unique_ptr<OptimizationRemarkEmitter> ORE; |
824 | std::function<OptimizationRemarkEmitter &(Function &)> GetORE = |
825 | [&ORE](Function &F) -> OptimizationRemarkEmitter & { |
826 | ORE.reset(p: new OptimizationRemarkEmitter(&F)); |
827 | return *ORE; |
828 | }; |
829 | |
830 | ProfileSummaryInfo *PSI = &AM.getResult<ProfileSummaryAnalysis>(IR&: M); |
831 | |
832 | if (HotColdSplitting(PSI, GBFI, GTTI, &GetORE, LookupAC).run(M)) |
833 | return PreservedAnalyses::none(); |
834 | return PreservedAnalyses::all(); |
835 | } |
836 | |