1 //===- PartialInlining.cpp - Inline parts of functions --------------------===//
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 // This pass performs partial inlining, typically by inlining an if statement
11 // that surrounds the body of the function.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/IPO/PartialInlining.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/DenseSet.h"
18 #include "llvm/ADT/None.h"
19 #include "llvm/ADT/Optional.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/BlockFrequencyInfo.h"
24 #include "llvm/Analysis/BranchProbabilityInfo.h"
25 #include "llvm/Analysis/InlineCost.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
28 #include "llvm/Analysis/ProfileSummaryInfo.h"
29 #include "llvm/Analysis/TargetLibraryInfo.h"
30 #include "llvm/Analysis/TargetTransformInfo.h"
31 #include "llvm/IR/Attributes.h"
32 #include "llvm/IR/BasicBlock.h"
33 #include "llvm/IR/CFG.h"
34 #include "llvm/IR/CallSite.h"
35 #include "llvm/IR/DebugLoc.h"
36 #include "llvm/IR/DiagnosticInfo.h"
37 #include "llvm/IR/Dominators.h"
38 #include "llvm/IR/Function.h"
39 #include "llvm/IR/InstrTypes.h"
40 #include "llvm/IR/Instruction.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/IntrinsicInst.h"
43 #include "llvm/IR/Intrinsics.h"
44 #include "llvm/IR/Module.h"
45 #include "llvm/IR/User.h"
46 #include "llvm/Pass.h"
47 #include "llvm/Support/BlockFrequency.h"
48 #include "llvm/Support/BranchProbability.h"
49 #include "llvm/Support/Casting.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Transforms/IPO.h"
53 #include "llvm/Transforms/Utils/Cloning.h"
54 #include "llvm/Transforms/Utils/CodeExtractor.h"
55 #include "llvm/Transforms/Utils/ValueMapper.h"
67 #define DEBUG_TYPE "partial-inlining"
69 STATISTIC(NumPartialInlined,
70 "Number of callsites functions partially inlined into.");
71 STATISTIC(NumColdOutlinePartialInlined, "Number of times functions with "
72 "cold outlined regions were partially "
73 "inlined into its caller(s).");
74 STATISTIC(NumColdRegionsFound,
75 "Number of cold single entry/exit regions found.");
76 STATISTIC(NumColdRegionsOutlined,
77 "Number of cold single entry/exit regions outlined.");
79 // Command line option to disable partial-inlining. The default is false:
81 DisablePartialInlining("disable-partial-inlining", cl::init(false),
82 cl::Hidden, cl::desc("Disable partial inlining"));
83 // Command line option to disable multi-region partial-inlining. The default is
85 static cl::opt<bool> DisableMultiRegionPartialInline(
86 "disable-mr-partial-inlining", cl::init(false), cl::Hidden,
87 cl::desc("Disable multi-region partial inlining"));
89 // Command line option to force outlining in regions with live exit variables.
90 // The default is false:
92 ForceLiveExit("pi-force-live-exit-outline", cl::init(false), cl::Hidden,
93 cl::desc("Force outline regions with live exits"));
95 // Command line option to enable marking outline functions with Cold Calling
96 // Convention. The default is false:
98 MarkOutlinedColdCC("pi-mark-coldcc", cl::init(false), cl::Hidden,
99 cl::desc("Mark outline function calls with ColdCC"));
102 // Command line option to debug partial-inlining. The default is none:
103 static cl::opt<bool> TracePartialInlining("trace-partial-inlining",
104 cl::init(false), cl::Hidden,
105 cl::desc("Trace partial inlining."));
108 // This is an option used by testing:
109 static cl::opt<bool> SkipCostAnalysis("skip-partial-inlining-cost-analysis",
110 cl::init(false), cl::ZeroOrMore,
112 cl::desc("Skip Cost Analysis"));
113 // Used to determine if a cold region is worth outlining based on
114 // its inlining cost compared to the original function. Default is set at 10%.
115 // ie. if the cold region reduces the inlining cost of the original function by
117 static cl::opt<float> MinRegionSizeRatio(
118 "min-region-size-ratio", cl::init(0.1), cl::Hidden,
119 cl::desc("Minimum ratio comparing relative sizes of each "
120 "outline candidate and original function"));
121 // Used to tune the minimum number of execution counts needed in the predecessor
122 // block to the cold edge. ie. confidence interval.
123 static cl::opt<unsigned>
124 MinBlockCounterExecution("min-block-execution", cl::init(100), cl::Hidden,
125 cl::desc("Minimum block executions to consider "
126 "its BranchProbabilityInfo valid"));
127 // Used to determine when an edge is considered cold. Default is set to 10%. ie.
128 // if the branch probability is 10% or less, then it is deemed as 'cold'.
129 static cl::opt<float> ColdBranchRatio(
130 "cold-branch-ratio", cl::init(0.1), cl::Hidden,
131 cl::desc("Minimum BranchProbability to consider a region cold."));
133 static cl::opt<unsigned> MaxNumInlineBlocks(
134 "max-num-inline-blocks", cl::init(5), cl::Hidden,
135 cl::desc("Max number of blocks to be partially inlined"));
137 // Command line option to set the maximum number of partial inlining allowed
138 // for the module. The default value of -1 means no limit.
139 static cl::opt<int> MaxNumPartialInlining(
140 "max-partial-inlining", cl::init(-1), cl::Hidden, cl::ZeroOrMore,
141 cl::desc("Max number of partial inlining. The default is unlimited"));
143 // Used only when PGO or user annotated branch data is absent. It is
144 // the least value that is used to weigh the outline region. If BFI
145 // produces larger value, the BFI value will be used.
147 OutlineRegionFreqPercent("outline-region-freq-percent", cl::init(75),
148 cl::Hidden, cl::ZeroOrMore,
149 cl::desc("Relative frequency of outline region to "
152 static cl::opt<unsigned> ExtraOutliningPenalty(
153 "partial-inlining-extra-penalty", cl::init(0), cl::Hidden,
154 cl::desc("A debug option to add additional penalty to the computed one."));
158 struct FunctionOutliningInfo {
159 FunctionOutliningInfo() = default;
161 // Returns the number of blocks to be inlined including all blocks
162 // in Entries and one return block.
163 unsigned GetNumInlinedBlocks() const { return Entries.size() + 1; }
165 // A set of blocks including the function entry that guard
166 // the region to be outlined.
167 SmallVector<BasicBlock *, 4> Entries;
169 // The return block that is not included in the outlined region.
170 BasicBlock *ReturnBlock = nullptr;
172 // The dominating block of the region to be outlined.
173 BasicBlock *NonReturnBlock = nullptr;
175 // The set of blocks in Entries that that are predecessors to ReturnBlock
176 SmallVector<BasicBlock *, 4> ReturnBlockPreds;
179 struct FunctionOutliningMultiRegionInfo {
180 FunctionOutliningMultiRegionInfo()
183 // Container for outline regions
184 struct OutlineRegionInfo {
185 OutlineRegionInfo(SmallVector<BasicBlock *, 8> Region,
186 BasicBlock *EntryBlock, BasicBlock *ExitBlock,
187 BasicBlock *ReturnBlock)
188 : Region(Region), EntryBlock(EntryBlock), ExitBlock(ExitBlock),
189 ReturnBlock(ReturnBlock) {}
190 SmallVector<BasicBlock *, 8> Region;
191 BasicBlock *EntryBlock;
192 BasicBlock *ExitBlock;
193 BasicBlock *ReturnBlock;
196 SmallVector<OutlineRegionInfo, 4> ORI;
199 struct PartialInlinerImpl {
202 std::function<AssumptionCache &(Function &)> *GetAC,
203 std::function<TargetTransformInfo &(Function &)> *GTTI,
204 Optional<function_ref<BlockFrequencyInfo &(Function &)>> GBFI,
205 ProfileSummaryInfo *ProfSI)
206 : GetAssumptionCache(GetAC), GetTTI(GTTI), GetBFI(GBFI), PSI(ProfSI) {}
209 // Main part of the transformation that calls helper functions to find
210 // outlining candidates, clone & outline the function, and attempt to
211 // partially inline the resulting function. Returns true if
212 // inlining was successful, false otherwise. Also returns the outline
213 // function (only if we partially inlined early returns) as there is a
214 // possibility to further "peel" early return statements that were left in the
215 // outline function due to code size.
216 std::pair<bool, Function *> unswitchFunction(Function *F);
218 // This class speculatively clones the function to be partial inlined.
219 // At the end of partial inlining, the remaining callsites to the cloned
220 // function that are not partially inlined will be fixed up to reference
221 // the original function, and the cloned function will be erased.
222 struct FunctionCloner {
223 // Two constructors, one for single region outlining, the other for
224 // multi-region outlining.
225 FunctionCloner(Function *F, FunctionOutliningInfo *OI,
226 OptimizationRemarkEmitter &ORE);
227 FunctionCloner(Function *F, FunctionOutliningMultiRegionInfo *OMRI,
228 OptimizationRemarkEmitter &ORE);
231 // Prepare for function outlining: making sure there is only
232 // one incoming edge from the extracted/outlined region to
234 void NormalizeReturnBlock();
236 // Do function outlining for cold regions.
237 bool doMultiRegionFunctionOutlining();
238 // Do function outlining for region after early return block(s).
239 // NOTE: For vararg functions that do the vararg handling in the outlined
240 // function, we temporarily generate IR that does not properly
241 // forward varargs to the outlined function. Calling InlineFunction
242 // will update calls to the outlined functions to properly forward
244 Function *doSingleRegionFunctionOutlining();
246 Function *OrigFunc = nullptr;
247 Function *ClonedFunc = nullptr;
249 typedef std::pair<Function *, BasicBlock *> FuncBodyCallerPair;
250 // Keep track of Outlined Functions and the basic block they're called from.
251 SmallVector<FuncBodyCallerPair, 4> OutlinedFunctions;
253 // ClonedFunc is inlined in one of its callers after function
255 bool IsFunctionInlined = false;
256 // The cost of the region to be outlined.
257 int OutlinedRegionCost = 0;
258 // ClonedOI is specific to outlining non-early return blocks.
259 std::unique_ptr<FunctionOutliningInfo> ClonedOI = nullptr;
260 // ClonedOMRI is specific to outlining cold regions.
261 std::unique_ptr<FunctionOutliningMultiRegionInfo> ClonedOMRI = nullptr;
262 std::unique_ptr<BlockFrequencyInfo> ClonedFuncBFI = nullptr;
263 OptimizationRemarkEmitter &ORE;
267 int NumPartialInlining = 0;
268 std::function<AssumptionCache &(Function &)> *GetAssumptionCache;
269 std::function<TargetTransformInfo &(Function &)> *GetTTI;
270 Optional<function_ref<BlockFrequencyInfo &(Function &)>> GetBFI;
271 ProfileSummaryInfo *PSI;
273 // Return the frequency of the OutlininingBB relative to F's entry point.
274 // The result is no larger than 1 and is represented using BP.
275 // (Note that the outlined region's 'head' block can only have incoming
276 // edges from the guarding entry blocks).
277 BranchProbability getOutliningCallBBRelativeFreq(FunctionCloner &Cloner);
279 // Return true if the callee of CS should be partially inlined with
281 bool shouldPartialInline(CallSite CS, FunctionCloner &Cloner,
282 BlockFrequency WeightedOutliningRcost,
283 OptimizationRemarkEmitter &ORE);
285 // Try to inline DuplicateFunction (cloned from F with call to
286 // the OutlinedFunction into its callers. Return true
287 // if there is any successful inlining.
288 bool tryPartialInline(FunctionCloner &Cloner);
290 // Compute the mapping from use site of DuplicationFunction to the enclosing
291 // BB's profile count.
292 void computeCallsiteToProfCountMap(Function *DuplicateFunction,
293 DenseMap<User *, uint64_t> &SiteCountMap);
295 bool IsLimitReached() {
296 return (MaxNumPartialInlining != -1 &&
297 NumPartialInlining >= MaxNumPartialInlining);
300 static CallSite getCallSite(User *U) {
302 if (CallInst *CI = dyn_cast<CallInst>(U))
304 else if (InvokeInst *II = dyn_cast<InvokeInst>(U))
307 llvm_unreachable("All uses must be calls");
311 static CallSite getOneCallSiteTo(Function *F) {
312 User *User = *F->user_begin();
313 return getCallSite(User);
316 std::tuple<DebugLoc, BasicBlock *> getOneDebugLoc(Function *F) {
317 CallSite CS = getOneCallSiteTo(F);
318 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
319 BasicBlock *Block = CS.getParent();
320 return std::make_tuple(DLoc, Block);
323 // Returns the costs associated with function outlining:
324 // - The first value is the non-weighted runtime cost for making the call
325 // to the outlined function, including the addtional setup cost in the
326 // outlined function itself;
327 // - The second value is the estimated size of the new call sequence in
328 // basic block Cloner.OutliningCallBB;
329 std::tuple<int, int> computeOutliningCosts(FunctionCloner &Cloner);
331 // Compute the 'InlineCost' of block BB. InlineCost is a proxy used to
332 // approximate both the size and runtime cost (Note that in the current
333 // inline cost analysis, there is no clear distinction there either).
334 static int computeBBInlineCost(BasicBlock *BB);
336 std::unique_ptr<FunctionOutliningInfo> computeOutliningInfo(Function *F);
337 std::unique_ptr<FunctionOutliningMultiRegionInfo>
338 computeOutliningColdRegionsInfo(Function *F, OptimizationRemarkEmitter &ORE);
341 struct PartialInlinerLegacyPass : public ModulePass {
342 static char ID; // Pass identification, replacement for typeid
344 PartialInlinerLegacyPass() : ModulePass(ID) {
345 initializePartialInlinerLegacyPassPass(*PassRegistry::getPassRegistry());
348 void getAnalysisUsage(AnalysisUsage &AU) const override {
349 AU.addRequired<AssumptionCacheTracker>();
350 AU.addRequired<ProfileSummaryInfoWrapperPass>();
351 AU.addRequired<TargetTransformInfoWrapperPass>();
354 bool runOnModule(Module &M) override {
358 AssumptionCacheTracker *ACT = &getAnalysis<AssumptionCacheTracker>();
359 TargetTransformInfoWrapperPass *TTIWP =
360 &getAnalysis<TargetTransformInfoWrapperPass>();
361 ProfileSummaryInfo *PSI =
362 &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
364 std::function<AssumptionCache &(Function &)> GetAssumptionCache =
365 [&ACT](Function &F) -> AssumptionCache & {
366 return ACT->getAssumptionCache(F);
369 std::function<TargetTransformInfo &(Function &)> GetTTI =
370 [&TTIWP](Function &F) -> TargetTransformInfo & {
371 return TTIWP->getTTI(F);
374 return PartialInlinerImpl(&GetAssumptionCache, &GetTTI, NoneType::None, PSI)
379 } // end anonymous namespace
381 std::unique_ptr<FunctionOutliningMultiRegionInfo>
382 PartialInlinerImpl::computeOutliningColdRegionsInfo(Function *F,
383 OptimizationRemarkEmitter &ORE) {
384 BasicBlock *EntryBlock = &F->front();
386 DominatorTree DT(*F);
388 BranchProbabilityInfo BPI(*F, LI);
389 std::unique_ptr<BlockFrequencyInfo> ScopedBFI;
390 BlockFrequencyInfo *BFI;
392 ScopedBFI.reset(new BlockFrequencyInfo(*F, BPI, LI));
393 BFI = ScopedBFI.get();
395 BFI = &(*GetBFI)(*F);
397 // Return if we don't have profiling information.
398 if (!PSI->hasInstrumentationProfile())
399 return std::unique_ptr<FunctionOutliningMultiRegionInfo>();
401 std::unique_ptr<FunctionOutliningMultiRegionInfo> OutliningInfo =
402 llvm::make_unique<FunctionOutliningMultiRegionInfo>();
404 auto IsSingleEntry = [](SmallVectorImpl<BasicBlock *> &BlockList) {
405 BasicBlock *Dom = BlockList.front();
406 return BlockList.size() > 1 && Dom->hasNPredecessors(1);
410 [&ORE](SmallVectorImpl<BasicBlock *> &BlockList) -> BasicBlock * {
411 BasicBlock *ExitBlock = nullptr;
412 for (auto *Block : BlockList) {
413 for (auto SI = succ_begin(Block); SI != succ_end(Block); ++SI) {
414 if (!is_contained(BlockList, *SI)) {
417 return OptimizationRemarkMissed(DEBUG_TYPE, "MultiExitRegion",
419 << "Region dominated by "
420 << ore::NV("Block", BlockList.front()->getName())
421 << " has more than one region exit edge.";
432 auto BBProfileCount = [BFI](BasicBlock *BB) {
433 return BFI->getBlockProfileCount(BB)
434 ? BFI->getBlockProfileCount(BB).getValue()
438 // Use the same computeBBInlineCost function to compute the cost savings of
439 // the outlining the candidate region.
440 int OverallFunctionCost = 0;
442 OverallFunctionCost += computeBBInlineCost(&BB);
445 if (TracePartialInlining)
446 dbgs() << "OverallFunctionCost = " << OverallFunctionCost << "\n";
448 int MinOutlineRegionCost =
449 static_cast<int>(OverallFunctionCost * MinRegionSizeRatio);
450 BranchProbability MinBranchProbability(
451 static_cast<int>(ColdBranchRatio * MinBlockCounterExecution),
452 MinBlockCounterExecution);
453 bool ColdCandidateFound = false;
454 BasicBlock *CurrEntry = EntryBlock;
455 std::vector<BasicBlock *> DFS;
456 DenseMap<BasicBlock *, bool> VisitedMap;
457 DFS.push_back(CurrEntry);
458 VisitedMap[CurrEntry] = true;
459 // Use Depth First Search on the basic blocks to find CFG edges that are
461 // Cold regions considered must also have its inline cost compared to the
462 // overall inline cost of the original function. The region is outlined only
463 // if it reduced the inline cost of the function by 'MinOutlineRegionCost' or
465 while (!DFS.empty()) {
466 auto *thisBB = DFS.back();
468 // Only consider regions with predecessor blocks that are considered
469 // not-cold (default: part of the top 99.99% of all block counters)
470 // AND greater than our minimum block execution count (default: 100).
471 if (PSI->isColdBlock(thisBB, BFI) ||
472 BBProfileCount(thisBB) < MinBlockCounterExecution)
474 for (auto SI = succ_begin(thisBB); SI != succ_end(thisBB); ++SI) {
477 VisitedMap[*SI] = true;
479 // If branch isn't cold, we skip to the next one.
480 BranchProbability SuccProb = BPI.getEdgeProbability(thisBB, *SI);
481 if (SuccProb > MinBranchProbability)
484 if (TracePartialInlining) {
485 dbgs() << "Found cold edge: " << thisBB->getName() << "->"
486 << (*SI)->getName() << "\nBranch Probability = " << SuccProb
490 SmallVector<BasicBlock *, 8> DominateVector;
491 DT.getDescendants(*SI, DominateVector);
492 // We can only outline single entry regions (for now).
493 if (!IsSingleEntry(DominateVector))
495 BasicBlock *ExitBlock = nullptr;
496 // We can only outline single exit regions (for now).
497 if (!(ExitBlock = IsSingleExit(DominateVector)))
499 int OutlineRegionCost = 0;
500 for (auto *BB : DominateVector)
501 OutlineRegionCost += computeBBInlineCost(BB);
504 if (TracePartialInlining)
505 dbgs() << "OutlineRegionCost = " << OutlineRegionCost << "\n";
508 if (OutlineRegionCost < MinOutlineRegionCost) {
510 return OptimizationRemarkAnalysis(DEBUG_TYPE, "TooCostly",
512 << ore::NV("Callee", F) << " inline cost-savings smaller than "
513 << ore::NV("Cost", MinOutlineRegionCost);
517 // For now, ignore blocks that belong to a SISE region that is a
518 // candidate for outlining. In the future, we may want to look
519 // at inner regions because the outer region may have live-exit
521 for (auto *BB : DominateVector)
522 VisitedMap[BB] = true;
523 // ReturnBlock here means the block after the outline call
524 BasicBlock *ReturnBlock = ExitBlock->getSingleSuccessor();
525 // assert(ReturnBlock && "ReturnBlock is NULL somehow!");
526 FunctionOutliningMultiRegionInfo::OutlineRegionInfo RegInfo(
527 DominateVector, DominateVector.front(), ExitBlock, ReturnBlock);
528 RegInfo.Region = DominateVector;
529 OutliningInfo->ORI.push_back(RegInfo);
531 if (TracePartialInlining) {
532 dbgs() << "Found Cold Candidate starting at block: "
533 << DominateVector.front()->getName() << "\n";
536 ColdCandidateFound = true;
537 NumColdRegionsFound++;
540 if (ColdCandidateFound)
541 return OutliningInfo;
543 return std::unique_ptr<FunctionOutliningMultiRegionInfo>();
546 std::unique_ptr<FunctionOutliningInfo>
547 PartialInlinerImpl::computeOutliningInfo(Function *F) {
548 BasicBlock *EntryBlock = &F->front();
549 BranchInst *BR = dyn_cast<BranchInst>(EntryBlock->getTerminator());
550 if (!BR || BR->isUnconditional())
551 return std::unique_ptr<FunctionOutliningInfo>();
553 // Returns true if Succ is BB's successor
554 auto IsSuccessor = [](BasicBlock *Succ, BasicBlock *BB) {
555 return is_contained(successors(BB), Succ);
558 auto IsReturnBlock = [](BasicBlock *BB) {
559 Instruction *TI = BB->getTerminator();
560 return isa<ReturnInst>(TI);
563 auto GetReturnBlock = [&](BasicBlock *Succ1, BasicBlock *Succ2) {
564 if (IsReturnBlock(Succ1))
565 return std::make_tuple(Succ1, Succ2);
566 if (IsReturnBlock(Succ2))
567 return std::make_tuple(Succ2, Succ1);
569 return std::make_tuple<BasicBlock *, BasicBlock *>(nullptr, nullptr);
572 // Detect a triangular shape:
573 auto GetCommonSucc = [&](BasicBlock *Succ1, BasicBlock *Succ2) {
574 if (IsSuccessor(Succ1, Succ2))
575 return std::make_tuple(Succ1, Succ2);
576 if (IsSuccessor(Succ2, Succ1))
577 return std::make_tuple(Succ2, Succ1);
579 return std::make_tuple<BasicBlock *, BasicBlock *>(nullptr, nullptr);
582 std::unique_ptr<FunctionOutliningInfo> OutliningInfo =
583 llvm::make_unique<FunctionOutliningInfo>();
585 BasicBlock *CurrEntry = EntryBlock;
586 bool CandidateFound = false;
588 // The number of blocks to be inlined has already reached
589 // the limit. When MaxNumInlineBlocks is set to 0 or 1, this
590 // disables partial inlining for the function.
591 if (OutliningInfo->GetNumInlinedBlocks() >= MaxNumInlineBlocks)
594 if (succ_size(CurrEntry) != 2)
597 BasicBlock *Succ1 = *succ_begin(CurrEntry);
598 BasicBlock *Succ2 = *(succ_begin(CurrEntry) + 1);
600 BasicBlock *ReturnBlock, *NonReturnBlock;
601 std::tie(ReturnBlock, NonReturnBlock) = GetReturnBlock(Succ1, Succ2);
604 OutliningInfo->Entries.push_back(CurrEntry);
605 OutliningInfo->ReturnBlock = ReturnBlock;
606 OutliningInfo->NonReturnBlock = NonReturnBlock;
607 CandidateFound = true;
611 BasicBlock *CommSucc;
612 BasicBlock *OtherSucc;
613 std::tie(CommSucc, OtherSucc) = GetCommonSucc(Succ1, Succ2);
618 OutliningInfo->Entries.push_back(CurrEntry);
619 CurrEntry = OtherSucc;
623 return std::unique_ptr<FunctionOutliningInfo>();
625 // Do sanity check of the entries: threre should not
626 // be any successors (not in the entry set) other than
627 // {ReturnBlock, NonReturnBlock}
628 assert(OutliningInfo->Entries[0] == &F->front() &&
629 "Function Entry must be the first in Entries vector");
630 DenseSet<BasicBlock *> Entries;
631 for (BasicBlock *E : OutliningInfo->Entries)
634 // Returns true of BB has Predecessor which is not
636 auto HasNonEntryPred = [Entries](BasicBlock *BB) {
637 for (auto Pred : predecessors(BB)) {
638 if (!Entries.count(Pred))
643 auto CheckAndNormalizeCandidate =
644 [Entries, HasNonEntryPred](FunctionOutliningInfo *OutliningInfo) {
645 for (BasicBlock *E : OutliningInfo->Entries) {
646 for (auto Succ : successors(E)) {
647 if (Entries.count(Succ))
649 if (Succ == OutliningInfo->ReturnBlock)
650 OutliningInfo->ReturnBlockPreds.push_back(E);
651 else if (Succ != OutliningInfo->NonReturnBlock)
654 // There should not be any outside incoming edges either:
655 if (HasNonEntryPred(E))
661 if (!CheckAndNormalizeCandidate(OutliningInfo.get()))
662 return std::unique_ptr<FunctionOutliningInfo>();
664 // Now further growing the candidate's inlining region by
665 // peeling off dominating blocks from the outlining region:
666 while (OutliningInfo->GetNumInlinedBlocks() < MaxNumInlineBlocks) {
667 BasicBlock *Cand = OutliningInfo->NonReturnBlock;
668 if (succ_size(Cand) != 2)
671 if (HasNonEntryPred(Cand))
674 BasicBlock *Succ1 = *succ_begin(Cand);
675 BasicBlock *Succ2 = *(succ_begin(Cand) + 1);
677 BasicBlock *ReturnBlock, *NonReturnBlock;
678 std::tie(ReturnBlock, NonReturnBlock) = GetReturnBlock(Succ1, Succ2);
679 if (!ReturnBlock || ReturnBlock != OutliningInfo->ReturnBlock)
682 if (NonReturnBlock->getSinglePredecessor() != Cand)
685 // Now grow and update OutlininigInfo:
686 OutliningInfo->Entries.push_back(Cand);
687 OutliningInfo->NonReturnBlock = NonReturnBlock;
688 OutliningInfo->ReturnBlockPreds.push_back(Cand);
689 Entries.insert(Cand);
692 return OutliningInfo;
695 // Check if there is PGO data or user annoated branch data:
696 static bool hasProfileData(Function *F, FunctionOutliningInfo *OI) {
697 if (F->hasProfileData())
699 // Now check if any of the entry block has MD_prof data:
700 for (auto *E : OI->Entries) {
701 BranchInst *BR = dyn_cast<BranchInst>(E->getTerminator());
702 if (!BR || BR->isUnconditional())
705 if (BR->extractProfMetadata(T, F))
712 PartialInlinerImpl::getOutliningCallBBRelativeFreq(FunctionCloner &Cloner) {
713 BasicBlock *OutliningCallBB = Cloner.OutlinedFunctions.back().second;
715 Cloner.ClonedFuncBFI->getBlockFreq(&Cloner.ClonedFunc->getEntryBlock());
716 auto OutliningCallFreq =
717 Cloner.ClonedFuncBFI->getBlockFreq(OutliningCallBB);
718 // FIXME Hackery needed because ClonedFuncBFI is based on the function BEFORE
719 // we outlined any regions, so we may encounter situations where the
720 // OutliningCallFreq is *slightly* bigger than the EntryFreq.
721 if (OutliningCallFreq.getFrequency() > EntryFreq.getFrequency()) {
722 OutliningCallFreq = EntryFreq;
724 auto OutlineRegionRelFreq = BranchProbability::getBranchProbability(
725 OutliningCallFreq.getFrequency(), EntryFreq.getFrequency());
727 if (hasProfileData(Cloner.OrigFunc, Cloner.ClonedOI.get()))
728 return OutlineRegionRelFreq;
730 // When profile data is not available, we need to be conservative in
731 // estimating the overall savings. Static branch prediction can usually
732 // guess the branch direction right (taken/non-taken), but the guessed
733 // branch probability is usually not biased enough. In case when the
734 // outlined region is predicted to be likely, its probability needs
735 // to be made higher (more biased) to not under-estimate the cost of
736 // function outlining. On the other hand, if the outlined region
737 // is predicted to be less likely, the predicted probablity is usually
738 // higher than the actual. For instance, the actual probability of the
739 // less likely target is only 5%, but the guessed probablity can be
740 // 40%. In the latter case, there is no need for further adjustement.
741 // FIXME: add an option for this.
742 if (OutlineRegionRelFreq < BranchProbability(45, 100))
743 return OutlineRegionRelFreq;
745 OutlineRegionRelFreq = std::max(
746 OutlineRegionRelFreq, BranchProbability(OutlineRegionFreqPercent, 100));
748 return OutlineRegionRelFreq;
751 bool PartialInlinerImpl::shouldPartialInline(
752 CallSite CS, FunctionCloner &Cloner,
753 BlockFrequency WeightedOutliningRcost,
754 OptimizationRemarkEmitter &ORE) {
757 Instruction *Call = CS.getInstruction();
758 Function *Callee = CS.getCalledFunction();
759 assert(Callee == Cloner.ClonedFunc);
761 if (SkipCostAnalysis)
762 return isInlineViable(*Callee);
764 Function *Caller = CS.getCaller();
765 auto &CalleeTTI = (*GetTTI)(*Callee);
766 InlineCost IC = getInlineCost(CS, getInlineParams(), CalleeTTI,
767 *GetAssumptionCache, GetBFI, PSI, &ORE);
771 return OptimizationRemarkAnalysis(DEBUG_TYPE, "AlwaysInline", Call)
772 << NV("Callee", Cloner.OrigFunc)
773 << " should always be fully inlined, not partially";
780 return OptimizationRemarkMissed(DEBUG_TYPE, "NeverInline", Call)
781 << NV("Callee", Cloner.OrigFunc) << " not partially inlined into "
782 << NV("Caller", Caller)
783 << " because it should never be inlined (cost=never)";
790 return OptimizationRemarkAnalysis(DEBUG_TYPE, "TooCostly", Call)
791 << NV("Callee", Cloner.OrigFunc) << " not partially inlined into "
792 << NV("Caller", Caller) << " because too costly to inline (cost="
793 << NV("Cost", IC.getCost()) << ", threshold="
794 << NV("Threshold", IC.getCostDelta() + IC.getCost()) << ")";
798 const DataLayout &DL = Caller->getParent()->getDataLayout();
800 // The savings of eliminating the call:
801 int NonWeightedSavings = getCallsiteCost(CS, DL);
802 BlockFrequency NormWeightedSavings(NonWeightedSavings);
804 // Weighted saving is smaller than weighted cost, return false
805 if (NormWeightedSavings < WeightedOutliningRcost) {
807 return OptimizationRemarkAnalysis(DEBUG_TYPE, "OutliningCallcostTooHigh",
809 << NV("Callee", Cloner.OrigFunc) << " not partially inlined into "
810 << NV("Caller", Caller) << " runtime overhead (overhead="
811 << NV("Overhead", (unsigned)WeightedOutliningRcost.getFrequency())
813 << NV("Savings", (unsigned)NormWeightedSavings.getFrequency())
815 << " of making the outlined call is too high";
822 return OptimizationRemarkAnalysis(DEBUG_TYPE, "CanBePartiallyInlined", Call)
823 << NV("Callee", Cloner.OrigFunc) << " can be partially inlined into "
824 << NV("Caller", Caller) << " with cost=" << NV("Cost", IC.getCost())
826 << NV("Threshold", IC.getCostDelta() + IC.getCost()) << ")";
831 // TODO: Ideally we should share Inliner's InlineCost Analysis code.
832 // For now use a simplified version. The returned 'InlineCost' will be used
833 // to esimate the size cost as well as runtime cost of the BB.
834 int PartialInlinerImpl::computeBBInlineCost(BasicBlock *BB) {
836 const DataLayout &DL = BB->getParent()->getParent()->getDataLayout();
837 for (Instruction &I : BB->instructionsWithoutDebug()) {
838 // Skip free instructions.
839 switch (I.getOpcode()) {
840 case Instruction::BitCast:
841 case Instruction::PtrToInt:
842 case Instruction::IntToPtr:
843 case Instruction::Alloca:
844 case Instruction::PHI:
846 case Instruction::GetElementPtr:
847 if (cast<GetElementPtrInst>(&I)->hasAllZeroIndices())
854 if (I.isLifetimeStartOrEnd())
857 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
858 InlineCost += getCallsiteCost(CallSite(CI), DL);
862 if (InvokeInst *II = dyn_cast<InvokeInst>(&I)) {
863 InlineCost += getCallsiteCost(CallSite(II), DL);
867 if (SwitchInst *SI = dyn_cast<SwitchInst>(&I)) {
868 InlineCost += (SI->getNumCases() + 1) * InlineConstants::InstrCost;
871 InlineCost += InlineConstants::InstrCost;
877 PartialInlinerImpl::computeOutliningCosts(FunctionCloner &Cloner) {
878 int OutliningFuncCallCost = 0, OutlinedFunctionCost = 0;
879 for (auto FuncBBPair : Cloner.OutlinedFunctions) {
880 Function *OutlinedFunc = FuncBBPair.first;
881 BasicBlock* OutliningCallBB = FuncBBPair.second;
882 // Now compute the cost of the call sequence to the outlined function
883 // 'OutlinedFunction' in BB 'OutliningCallBB':
884 OutliningFuncCallCost += computeBBInlineCost(OutliningCallBB);
886 // Now compute the cost of the extracted/outlined function itself:
887 for (BasicBlock &BB : *OutlinedFunc)
888 OutlinedFunctionCost += computeBBInlineCost(&BB);
890 assert(OutlinedFunctionCost >= Cloner.OutlinedRegionCost &&
891 "Outlined function cost should be no less than the outlined region");
893 // The code extractor introduces a new root and exit stub blocks with
894 // additional unconditional branches. Those branches will be eliminated
895 // later with bb layout. The cost should be adjusted accordingly:
896 OutlinedFunctionCost -=
897 2 * InlineConstants::InstrCost * Cloner.OutlinedFunctions.size();
899 int OutliningRuntimeOverhead =
900 OutliningFuncCallCost +
901 (OutlinedFunctionCost - Cloner.OutlinedRegionCost) +
902 ExtraOutliningPenalty;
904 return std::make_tuple(OutliningFuncCallCost, OutliningRuntimeOverhead);
907 // Create the callsite to profile count map which is
908 // used to update the original function's entry count,
909 // after the function is partially inlined into the callsite.
910 void PartialInlinerImpl::computeCallsiteToProfCountMap(
911 Function *DuplicateFunction,
912 DenseMap<User *, uint64_t> &CallSiteToProfCountMap) {
913 std::vector<User *> Users(DuplicateFunction->user_begin(),
914 DuplicateFunction->user_end());
915 Function *CurrentCaller = nullptr;
916 std::unique_ptr<BlockFrequencyInfo> TempBFI;
917 BlockFrequencyInfo *CurrentCallerBFI = nullptr;
919 auto ComputeCurrBFI = [&,this](Function *Caller) {
920 // For the old pass manager:
922 DominatorTree DT(*Caller);
924 BranchProbabilityInfo BPI(*Caller, LI);
925 TempBFI.reset(new BlockFrequencyInfo(*Caller, BPI, LI));
926 CurrentCallerBFI = TempBFI.get();
929 CurrentCallerBFI = &(*GetBFI)(*Caller);
933 for (User *User : Users) {
934 CallSite CS = getCallSite(User);
935 Function *Caller = CS.getCaller();
936 if (CurrentCaller != Caller) {
937 CurrentCaller = Caller;
938 ComputeCurrBFI(Caller);
940 assert(CurrentCallerBFI && "CallerBFI is not set");
942 BasicBlock *CallBB = CS.getInstruction()->getParent();
943 auto Count = CurrentCallerBFI->getBlockProfileCount(CallBB);
945 CallSiteToProfCountMap[User] = *Count;
947 CallSiteToProfCountMap[User] = 0;
951 PartialInlinerImpl::FunctionCloner::FunctionCloner(
952 Function *F, FunctionOutliningInfo *OI, OptimizationRemarkEmitter &ORE)
953 : OrigFunc(F), ORE(ORE) {
954 ClonedOI = llvm::make_unique<FunctionOutliningInfo>();
956 // Clone the function, so that we can hack away on it.
957 ValueToValueMapTy VMap;
958 ClonedFunc = CloneFunction(F, VMap);
960 ClonedOI->ReturnBlock = cast<BasicBlock>(VMap[OI->ReturnBlock]);
961 ClonedOI->NonReturnBlock = cast<BasicBlock>(VMap[OI->NonReturnBlock]);
962 for (BasicBlock *BB : OI->Entries) {
963 ClonedOI->Entries.push_back(cast<BasicBlock>(VMap[BB]));
965 for (BasicBlock *E : OI->ReturnBlockPreds) {
966 BasicBlock *NewE = cast<BasicBlock>(VMap[E]);
967 ClonedOI->ReturnBlockPreds.push_back(NewE);
969 // Go ahead and update all uses to the duplicate, so that we can just
970 // use the inliner functionality when we're done hacking.
971 F->replaceAllUsesWith(ClonedFunc);
974 PartialInlinerImpl::FunctionCloner::FunctionCloner(
975 Function *F, FunctionOutliningMultiRegionInfo *OI,
976 OptimizationRemarkEmitter &ORE)
977 : OrigFunc(F), ORE(ORE) {
978 ClonedOMRI = llvm::make_unique<FunctionOutliningMultiRegionInfo>();
980 // Clone the function, so that we can hack away on it.
981 ValueToValueMapTy VMap;
982 ClonedFunc = CloneFunction(F, VMap);
984 // Go through all Outline Candidate Regions and update all BasicBlock
986 for (FunctionOutliningMultiRegionInfo::OutlineRegionInfo RegionInfo :
988 SmallVector<BasicBlock *, 8> Region;
989 for (BasicBlock *BB : RegionInfo.Region) {
990 Region.push_back(cast<BasicBlock>(VMap[BB]));
992 BasicBlock *NewEntryBlock = cast<BasicBlock>(VMap[RegionInfo.EntryBlock]);
993 BasicBlock *NewExitBlock = cast<BasicBlock>(VMap[RegionInfo.ExitBlock]);
994 BasicBlock *NewReturnBlock = nullptr;
995 if (RegionInfo.ReturnBlock)
996 NewReturnBlock = cast<BasicBlock>(VMap[RegionInfo.ReturnBlock]);
997 FunctionOutliningMultiRegionInfo::OutlineRegionInfo MappedRegionInfo(
998 Region, NewEntryBlock, NewExitBlock, NewReturnBlock);
999 ClonedOMRI->ORI.push_back(MappedRegionInfo);
1001 // Go ahead and update all uses to the duplicate, so that we can just
1002 // use the inliner functionality when we're done hacking.
1003 F->replaceAllUsesWith(ClonedFunc);
1006 void PartialInlinerImpl::FunctionCloner::NormalizeReturnBlock() {
1007 auto getFirstPHI = [](BasicBlock *BB) {
1008 BasicBlock::iterator I = BB->begin();
1009 PHINode *FirstPhi = nullptr;
1010 while (I != BB->end()) {
1011 PHINode *Phi = dyn_cast<PHINode>(I);
1022 // Shouldn't need to normalize PHIs if we're not outlining non-early return
1027 // Special hackery is needed with PHI nodes that have inputs from more than
1028 // one extracted block. For simplicity, just split the PHIs into a two-level
1029 // sequence of PHIs, some of which will go in the extracted region, and some
1030 // of which will go outside.
1031 BasicBlock *PreReturn = ClonedOI->ReturnBlock;
1032 // only split block when necessary:
1033 PHINode *FirstPhi = getFirstPHI(PreReturn);
1034 unsigned NumPredsFromEntries = ClonedOI->ReturnBlockPreds.size();
1036 if (!FirstPhi || FirstPhi->getNumIncomingValues() <= NumPredsFromEntries + 1)
1039 auto IsTrivialPhi = [](PHINode *PN) -> Value * {
1040 Value *CommonValue = PN->getIncomingValue(0);
1041 if (all_of(PN->incoming_values(),
1042 [&](Value *V) { return V == CommonValue; }))
1047 ClonedOI->ReturnBlock = ClonedOI->ReturnBlock->splitBasicBlock(
1048 ClonedOI->ReturnBlock->getFirstNonPHI()->getIterator());
1049 BasicBlock::iterator I = PreReturn->begin();
1050 Instruction *Ins = &ClonedOI->ReturnBlock->front();
1051 SmallVector<Instruction *, 4> DeadPhis;
1052 while (I != PreReturn->end()) {
1053 PHINode *OldPhi = dyn_cast<PHINode>(I);
1058 PHINode::Create(OldPhi->getType(), NumPredsFromEntries + 1, "", Ins);
1059 OldPhi->replaceAllUsesWith(RetPhi);
1060 Ins = ClonedOI->ReturnBlock->getFirstNonPHI();
1062 RetPhi->addIncoming(&*I, PreReturn);
1063 for (BasicBlock *E : ClonedOI->ReturnBlockPreds) {
1064 RetPhi->addIncoming(OldPhi->getIncomingValueForBlock(E), E);
1065 OldPhi->removeIncomingValue(E);
1068 // After incoming values splitting, the old phi may become trivial.
1069 // Keeping the trivial phi can introduce definition inside the outline
1070 // region which is live-out, causing necessary overhead (load, store
1071 // arg passing etc).
1072 if (auto *OldPhiVal = IsTrivialPhi(OldPhi)) {
1073 OldPhi->replaceAllUsesWith(OldPhiVal);
1074 DeadPhis.push_back(OldPhi);
1078 for (auto *DP : DeadPhis)
1079 DP->eraseFromParent();
1081 for (auto E : ClonedOI->ReturnBlockPreds) {
1082 E->getTerminator()->replaceUsesOfWith(PreReturn, ClonedOI->ReturnBlock);
1086 bool PartialInlinerImpl::FunctionCloner::doMultiRegionFunctionOutlining() {
1088 auto ComputeRegionCost = [](SmallVectorImpl<BasicBlock *> &Region) {
1090 for (BasicBlock* BB : Region)
1091 Cost += computeBBInlineCost(BB);
1095 assert(ClonedOMRI && "Expecting OutlineInfo for multi region outline");
1097 if (ClonedOMRI->ORI.empty())
1100 // The CodeExtractor needs a dominator tree.
1102 DT.recalculate(*ClonedFunc);
1104 // Manually calculate a BlockFrequencyInfo and BranchProbabilityInfo.
1106 BranchProbabilityInfo BPI(*ClonedFunc, LI);
1107 ClonedFuncBFI.reset(new BlockFrequencyInfo(*ClonedFunc, BPI, LI));
1109 SetVector<Value *> Inputs, Outputs, Sinks;
1110 for (FunctionOutliningMultiRegionInfo::OutlineRegionInfo RegionInfo :
1112 int CurrentOutlinedRegionCost = ComputeRegionCost(RegionInfo.Region);
1114 CodeExtractor CE(RegionInfo.Region, &DT, /*AggregateArgs*/ false,
1115 ClonedFuncBFI.get(), &BPI, /* AllowVarargs */ false);
1117 CE.findInputsOutputs(Inputs, Outputs, Sinks);
1120 if (TracePartialInlining) {
1121 dbgs() << "inputs: " << Inputs.size() << "\n";
1122 dbgs() << "outputs: " << Outputs.size() << "\n";
1123 for (Value *value : Inputs)
1124 dbgs() << "value used in func: " << *value << "\n";
1125 for (Value *output : Outputs)
1126 dbgs() << "instr used in func: " << *output << "\n";
1129 // Do not extract regions that have live exit variables.
1130 if (Outputs.size() > 0 && !ForceLiveExit)
1133 Function *OutlinedFunc = CE.extractCodeRegion();
1136 CallSite OCS = PartialInlinerImpl::getOneCallSiteTo(OutlinedFunc);
1137 BasicBlock *OutliningCallBB = OCS.getInstruction()->getParent();
1138 assert(OutliningCallBB->getParent() == ClonedFunc);
1139 OutlinedFunctions.push_back(std::make_pair(OutlinedFunc,OutliningCallBB));
1140 NumColdRegionsOutlined++;
1141 OutlinedRegionCost += CurrentOutlinedRegionCost;
1143 if (MarkOutlinedColdCC) {
1144 OutlinedFunc->setCallingConv(CallingConv::Cold);
1145 OCS.setCallingConv(CallingConv::Cold);
1149 return OptimizationRemarkMissed(DEBUG_TYPE, "ExtractFailed",
1150 &RegionInfo.Region.front()->front())
1151 << "Failed to extract region at block "
1152 << ore::NV("Block", RegionInfo.Region.front());
1156 return !OutlinedFunctions.empty();
1160 PartialInlinerImpl::FunctionCloner::doSingleRegionFunctionOutlining() {
1161 // Returns true if the block is to be partial inlined into the caller
1162 // (i.e. not to be extracted to the out of line function)
1163 auto ToBeInlined = [&, this](BasicBlock *BB) {
1164 return BB == ClonedOI->ReturnBlock ||
1165 (std::find(ClonedOI->Entries.begin(), ClonedOI->Entries.end(), BB) !=
1166 ClonedOI->Entries.end());
1169 assert(ClonedOI && "Expecting OutlineInfo for single region outline");
1170 // The CodeExtractor needs a dominator tree.
1172 DT.recalculate(*ClonedFunc);
1174 // Manually calculate a BlockFrequencyInfo and BranchProbabilityInfo.
1176 BranchProbabilityInfo BPI(*ClonedFunc, LI);
1177 ClonedFuncBFI.reset(new BlockFrequencyInfo(*ClonedFunc, BPI, LI));
1179 // Gather up the blocks that we're going to extract.
1180 std::vector<BasicBlock *> ToExtract;
1181 ToExtract.push_back(ClonedOI->NonReturnBlock);
1182 OutlinedRegionCost +=
1183 PartialInlinerImpl::computeBBInlineCost(ClonedOI->NonReturnBlock);
1184 for (BasicBlock &BB : *ClonedFunc)
1185 if (!ToBeInlined(&BB) && &BB != ClonedOI->NonReturnBlock) {
1186 ToExtract.push_back(&BB);
1187 // FIXME: the code extractor may hoist/sink more code
1188 // into the outlined function which may make the outlining
1189 // overhead (the difference of the outlined function cost
1190 // and OutliningRegionCost) look larger.
1191 OutlinedRegionCost += computeBBInlineCost(&BB);
1194 // Extract the body of the if.
1195 Function *OutlinedFunc =
1196 CodeExtractor(ToExtract, &DT, /*AggregateArgs*/ false,
1197 ClonedFuncBFI.get(), &BPI,
1198 /* AllowVarargs */ true)
1199 .extractCodeRegion();
1202 BasicBlock *OutliningCallBB =
1203 PartialInlinerImpl::getOneCallSiteTo(OutlinedFunc)
1206 assert(OutliningCallBB->getParent() == ClonedFunc);
1207 OutlinedFunctions.push_back(std::make_pair(OutlinedFunc, OutliningCallBB));
1210 return OptimizationRemarkMissed(DEBUG_TYPE, "ExtractFailed",
1211 &ToExtract.front()->front())
1212 << "Failed to extract region at block "
1213 << ore::NV("Block", ToExtract.front());
1216 return OutlinedFunc;
1219 PartialInlinerImpl::FunctionCloner::~FunctionCloner() {
1220 // Ditch the duplicate, since we're done with it, and rewrite all remaining
1221 // users (function pointers, etc.) back to the original function.
1222 ClonedFunc->replaceAllUsesWith(OrigFunc);
1223 ClonedFunc->eraseFromParent();
1224 if (!IsFunctionInlined) {
1225 // Remove each function that was speculatively created if there is no
1227 for (auto FuncBBPair : OutlinedFunctions) {
1228 Function *Func = FuncBBPair.first;
1229 Func->eraseFromParent();
1234 std::pair<bool, Function *> PartialInlinerImpl::unswitchFunction(Function *F) {
1236 if (F->hasAddressTaken())
1237 return {false, nullptr};
1239 // Let inliner handle it
1240 if (F->hasFnAttribute(Attribute::AlwaysInline))
1241 return {false, nullptr};
1243 if (F->hasFnAttribute(Attribute::NoInline))
1244 return {false, nullptr};
1246 if (PSI->isFunctionEntryCold(F))
1247 return {false, nullptr};
1249 if (empty(F->users()))
1250 return {false, nullptr};
1252 OptimizationRemarkEmitter ORE(F);
1254 // Only try to outline cold regions if we have a profile summary, which
1255 // implies we have profiling information.
1256 if (PSI->hasProfileSummary() && F->hasProfileData() &&
1257 !DisableMultiRegionPartialInline) {
1258 std::unique_ptr<FunctionOutliningMultiRegionInfo> OMRI =
1259 computeOutliningColdRegionsInfo(F, ORE);
1261 FunctionCloner Cloner(F, OMRI.get(), ORE);
1264 if (TracePartialInlining) {
1265 dbgs() << "HotCountThreshold = " << PSI->getHotCountThreshold() << "\n";
1266 dbgs() << "ColdCountThreshold = " << PSI->getColdCountThreshold()
1270 bool DidOutline = Cloner.doMultiRegionFunctionOutlining();
1274 if (TracePartialInlining) {
1275 dbgs() << ">>>>>> Outlined (Cloned) Function >>>>>>\n";
1276 Cloner.ClonedFunc->print(dbgs());
1277 dbgs() << "<<<<<< Outlined (Cloned) Function <<<<<<\n";
1281 if (tryPartialInline(Cloner))
1282 return {true, nullptr};
1287 // Fall-thru to regular partial inlining if we:
1288 // i) can't find any cold regions to outline, or
1289 // ii) can't inline the outlined function anywhere.
1290 std::unique_ptr<FunctionOutliningInfo> OI = computeOutliningInfo(F);
1292 return {false, nullptr};
1294 FunctionCloner Cloner(F, OI.get(), ORE);
1295 Cloner.NormalizeReturnBlock();
1297 Function *OutlinedFunction = Cloner.doSingleRegionFunctionOutlining();
1299 if (!OutlinedFunction)
1300 return {false, nullptr};
1302 bool AnyInline = tryPartialInline(Cloner);
1305 return {true, OutlinedFunction};
1307 return {false, nullptr};
1310 bool PartialInlinerImpl::tryPartialInline(FunctionCloner &Cloner) {
1311 if (Cloner.OutlinedFunctions.empty())
1315 BlockFrequency WeightedRcost;
1316 int NonWeightedRcost;
1317 std::tie(SizeCost, NonWeightedRcost) = computeOutliningCosts(Cloner);
1319 // Only calculate RelativeToEntryFreq when we are doing single region
1321 BranchProbability RelativeToEntryFreq;
1322 if (Cloner.ClonedOI) {
1323 RelativeToEntryFreq = getOutliningCallBBRelativeFreq(Cloner);
1325 // RelativeToEntryFreq doesn't make sense when we have more than one
1326 // outlined call because each call will have a different relative frequency
1327 // to the entry block. We can consider using the average, but the
1328 // usefulness of that information is questionable. For now, assume we never
1329 // execute the calls to outlined functions.
1330 RelativeToEntryFreq = BranchProbability(0, 1);
1332 WeightedRcost = BlockFrequency(NonWeightedRcost) * RelativeToEntryFreq;
1334 // The call sequence(s) to the outlined function(s) are larger than the sum of
1335 // the original outlined region size(s), it does not increase the chances of
1336 // inlining the function with outlining (The inliner uses the size increase to
1337 // model the cost of inlining a callee).
1338 if (!SkipCostAnalysis && Cloner.OutlinedRegionCost < SizeCost) {
1339 OptimizationRemarkEmitter OrigFuncORE(Cloner.OrigFunc);
1342 std::tie(DLoc, Block) = getOneDebugLoc(Cloner.ClonedFunc);
1343 OrigFuncORE.emit([&]() {
1344 return OptimizationRemarkAnalysis(DEBUG_TYPE, "OutlineRegionTooSmall",
1346 << ore::NV("Function", Cloner.OrigFunc)
1347 << " not partially inlined into callers (Original Size = "
1348 << ore::NV("OutlinedRegionOriginalSize", Cloner.OutlinedRegionCost)
1349 << ", Size of call sequence to outlined function = "
1350 << ore::NV("NewSize", SizeCost) << ")";
1355 assert(empty(Cloner.OrigFunc->users()) &&
1356 "F's users should all be replaced!");
1358 std::vector<User *> Users(Cloner.ClonedFunc->user_begin(),
1359 Cloner.ClonedFunc->user_end());
1361 DenseMap<User *, uint64_t> CallSiteToProfCountMap;
1362 auto CalleeEntryCount = Cloner.OrigFunc->getEntryCount();
1363 if (CalleeEntryCount)
1364 computeCallsiteToProfCountMap(Cloner.ClonedFunc, CallSiteToProfCountMap);
1366 uint64_t CalleeEntryCountV =
1367 (CalleeEntryCount ? CalleeEntryCount.getCount() : 0);
1369 bool AnyInline = false;
1370 for (User *User : Users) {
1371 CallSite CS = getCallSite(User);
1373 if (IsLimitReached())
1376 OptimizationRemarkEmitter CallerORE(CS.getCaller());
1377 if (!shouldPartialInline(CS, Cloner, WeightedRcost, CallerORE))
1380 // Construct remark before doing the inlining, as after successful inlining
1381 // the callsite is removed.
1382 OptimizationRemark OR(DEBUG_TYPE, "PartiallyInlined", CS.getInstruction());
1383 OR << ore::NV("Callee", Cloner.OrigFunc) << " partially inlined into "
1384 << ore::NV("Caller", CS.getCaller());
1386 InlineFunctionInfo IFI(nullptr, GetAssumptionCache, PSI);
1387 // We can only forward varargs when we outlined a single region, else we
1388 // bail on vararg functions.
1389 if (!InlineFunction(CS, IFI, nullptr, true,
1390 (Cloner.ClonedOI ? Cloner.OutlinedFunctions.back().first
1396 // Now update the entry count:
1397 if (CalleeEntryCountV && CallSiteToProfCountMap.count(User)) {
1398 uint64_t CallSiteCount = CallSiteToProfCountMap[User];
1399 CalleeEntryCountV -= std::min(CalleeEntryCountV, CallSiteCount);
1403 NumPartialInlining++;
1405 if (Cloner.ClonedOI)
1406 NumPartialInlined++;
1408 NumColdOutlinePartialInlined++;
1413 Cloner.IsFunctionInlined = true;
1414 if (CalleeEntryCount)
1415 Cloner.OrigFunc->setEntryCount(
1416 CalleeEntryCount.setCount(CalleeEntryCountV));
1417 OptimizationRemarkEmitter OrigFuncORE(Cloner.OrigFunc);
1418 OrigFuncORE.emit([&]() {
1419 return OptimizationRemark(DEBUG_TYPE, "PartiallyInlined", Cloner.OrigFunc)
1420 << "Partially inlined into at least one caller";
1428 bool PartialInlinerImpl::run(Module &M) {
1429 if (DisablePartialInlining)
1432 std::vector<Function *> Worklist;
1433 Worklist.reserve(M.size());
1434 for (Function &F : M)
1435 if (!F.use_empty() && !F.isDeclaration())
1436 Worklist.push_back(&F);
1438 bool Changed = false;
1439 while (!Worklist.empty()) {
1440 Function *CurrFunc = Worklist.back();
1441 Worklist.pop_back();
1443 if (CurrFunc->use_empty())
1446 bool Recursive = false;
1447 for (User *U : CurrFunc->users())
1448 if (Instruction *I = dyn_cast<Instruction>(U))
1449 if (I->getParent()->getParent() == CurrFunc) {
1456 std::pair<bool, Function * > Result = unswitchFunction(CurrFunc);
1458 Worklist.push_back(Result.second);
1459 Changed |= Result.first;
1465 char PartialInlinerLegacyPass::ID = 0;
1467 INITIALIZE_PASS_BEGIN(PartialInlinerLegacyPass, "partial-inliner",
1468 "Partial Inliner", false, false)
1469 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1470 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
1471 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1472 INITIALIZE_PASS_END(PartialInlinerLegacyPass, "partial-inliner",
1473 "Partial Inliner", false, false)
1475 ModulePass *llvm::createPartialInliningPass() {
1476 return new PartialInlinerLegacyPass();
1479 PreservedAnalyses PartialInlinerPass::run(Module &M,
1480 ModuleAnalysisManager &AM) {
1481 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
1483 std::function<AssumptionCache &(Function &)> GetAssumptionCache =
1484 [&FAM](Function &F) -> AssumptionCache & {
1485 return FAM.getResult<AssumptionAnalysis>(F);
1488 std::function<BlockFrequencyInfo &(Function &)> GetBFI =
1489 [&FAM](Function &F) -> BlockFrequencyInfo & {
1490 return FAM.getResult<BlockFrequencyAnalysis>(F);
1493 std::function<TargetTransformInfo &(Function &)> GetTTI =
1494 [&FAM](Function &F) -> TargetTransformInfo & {
1495 return FAM.getResult<TargetIRAnalysis>(F);
1498 ProfileSummaryInfo *PSI = &AM.getResult<ProfileSummaryAnalysis>(M);
1500 if (PartialInlinerImpl(&GetAssumptionCache, &GetTTI, {GetBFI}, PSI)
1502 return PreservedAnalyses::none();
1503 return PreservedAnalyses::all();