1 //===- Inliner.cpp - Code common to all inliners --------------------------===// 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 // This file implements the mechanics required to implement inlining without 10 // missing any calls and updating the call graph. The decisions of which calls 11 // are profitable to inline are implemented elsewhere. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/Transforms/IPO/Inliner.h" 16 #include "llvm/ADT/DenseMap.h" 17 #include "llvm/ADT/None.h" 18 #include "llvm/ADT/Optional.h" 19 #include "llvm/ADT/STLExtras.h" 20 #include "llvm/ADT/ScopeExit.h" 21 #include "llvm/ADT/SetVector.h" 22 #include "llvm/ADT/SmallPtrSet.h" 23 #include "llvm/ADT/SmallVector.h" 24 #include "llvm/ADT/Statistic.h" 25 #include "llvm/ADT/StringRef.h" 26 #include "llvm/Analysis/AssumptionCache.h" 27 #include "llvm/Analysis/BasicAliasAnalysis.h" 28 #include "llvm/Analysis/BlockFrequencyInfo.h" 29 #include "llvm/Analysis/CGSCCPassManager.h" 30 #include "llvm/Analysis/CallGraph.h" 31 #include "llvm/Analysis/GlobalsModRef.h" 32 #include "llvm/Analysis/InlineAdvisor.h" 33 #include "llvm/Analysis/InlineCost.h" 34 #include "llvm/Analysis/LazyCallGraph.h" 35 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 36 #include "llvm/Analysis/ProfileSummaryInfo.h" 37 #include "llvm/Analysis/TargetLibraryInfo.h" 38 #include "llvm/Analysis/TargetTransformInfo.h" 39 #include "llvm/Analysis/Utils/ImportedFunctionsInliningStatistics.h" 40 #include "llvm/IR/Attributes.h" 41 #include "llvm/IR/BasicBlock.h" 42 #include "llvm/IR/DataLayout.h" 43 #include "llvm/IR/DebugLoc.h" 44 #include "llvm/IR/DerivedTypes.h" 45 #include "llvm/IR/DiagnosticInfo.h" 46 #include "llvm/IR/Function.h" 47 #include "llvm/IR/InstIterator.h" 48 #include "llvm/IR/Instruction.h" 49 #include "llvm/IR/Instructions.h" 50 #include "llvm/IR/IntrinsicInst.h" 51 #include "llvm/IR/Metadata.h" 52 #include "llvm/IR/Module.h" 53 #include "llvm/IR/PassManager.h" 54 #include "llvm/IR/User.h" 55 #include "llvm/IR/Value.h" 56 #include "llvm/Pass.h" 57 #include "llvm/Support/Casting.h" 58 #include "llvm/Support/CommandLine.h" 59 #include "llvm/Support/Debug.h" 60 #include "llvm/Support/raw_ostream.h" 61 #include "llvm/Transforms/Utils/CallPromotionUtils.h" 62 #include "llvm/Transforms/Utils/Cloning.h" 63 #include "llvm/Transforms/Utils/Local.h" 64 #include "llvm/Transforms/Utils/ModuleUtils.h" 65 #include <algorithm> 66 #include <cassert> 67 #include <functional> 68 #include <sstream> 69 #include <tuple> 70 #include <utility> 71 #include <vector> 72 73 using namespace llvm; 74 75 #define DEBUG_TYPE "inline" 76 77 STATISTIC(NumInlined, "Number of functions inlined"); 78 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined"); 79 STATISTIC(NumDeleted, "Number of functions deleted because all callers found"); 80 STATISTIC(NumMergedAllocas, "Number of allocas merged together"); 81 82 /// Flag to disable manual alloca merging. 83 /// 84 /// Merging of allocas was originally done as a stack-size saving technique 85 /// prior to LLVM's code generator having support for stack coloring based on 86 /// lifetime markers. It is now in the process of being removed. To experiment 87 /// with disabling it and relying fully on lifetime marker based stack 88 /// coloring, you can pass this flag to LLVM. 89 static cl::opt<bool> 90 DisableInlinedAllocaMerging("disable-inlined-alloca-merging", 91 cl::init(false), cl::Hidden); 92 93 extern cl::opt<InlinerFunctionImportStatsOpts> InlinerFunctionImportStats; 94 95 static cl::opt<std::string> CGSCCInlineReplayFile( 96 "cgscc-inline-replay", cl::init(""), cl::value_desc("filename"), 97 cl::desc( 98 "Optimization remarks file containing inline remarks to be replayed " 99 "by inlining from cgscc inline remarks."), 100 cl::Hidden); 101 102 LegacyInlinerBase::LegacyInlinerBase(char &ID) : CallGraphSCCPass(ID) {} 103 104 LegacyInlinerBase::LegacyInlinerBase(char &ID, bool InsertLifetime) 105 : CallGraphSCCPass(ID), InsertLifetime(InsertLifetime) {} 106 107 /// For this class, we declare that we require and preserve the call graph. 108 /// If the derived class implements this method, it should 109 /// always explicitly call the implementation here. 110 void LegacyInlinerBase::getAnalysisUsage(AnalysisUsage &AU) const { 111 AU.addRequired<AssumptionCacheTracker>(); 112 AU.addRequired<ProfileSummaryInfoWrapperPass>(); 113 AU.addRequired<TargetLibraryInfoWrapperPass>(); 114 getAAResultsAnalysisUsage(AU); 115 CallGraphSCCPass::getAnalysisUsage(AU); 116 } 117 118 using InlinedArrayAllocasTy = DenseMap<ArrayType *, std::vector<AllocaInst *>>; 119 120 /// Look at all of the allocas that we inlined through this call site. If we 121 /// have already inlined other allocas through other calls into this function, 122 /// then we know that they have disjoint lifetimes and that we can merge them. 123 /// 124 /// There are many heuristics possible for merging these allocas, and the 125 /// different options have different tradeoffs. One thing that we *really* 126 /// don't want to hurt is SRoA: once inlining happens, often allocas are no 127 /// longer address taken and so they can be promoted. 128 /// 129 /// Our "solution" for that is to only merge allocas whose outermost type is an 130 /// array type. These are usually not promoted because someone is using a 131 /// variable index into them. These are also often the most important ones to 132 /// merge. 133 /// 134 /// A better solution would be to have real memory lifetime markers in the IR 135 /// and not have the inliner do any merging of allocas at all. This would 136 /// allow the backend to do proper stack slot coloring of all allocas that 137 /// *actually make it to the backend*, which is really what we want. 138 /// 139 /// Because we don't have this information, we do this simple and useful hack. 140 static void mergeInlinedArrayAllocas(Function *Caller, InlineFunctionInfo &IFI, 141 InlinedArrayAllocasTy &InlinedArrayAllocas, 142 int InlineHistory) { 143 SmallPtrSet<AllocaInst *, 16> UsedAllocas; 144 145 // When processing our SCC, check to see if the call site was inlined from 146 // some other call site. For example, if we're processing "A" in this code: 147 // A() { B() } 148 // B() { x = alloca ... C() } 149 // C() { y = alloca ... } 150 // Assume that C was not inlined into B initially, and so we're processing A 151 // and decide to inline B into A. Doing this makes an alloca available for 152 // reuse and makes a callsite (C) available for inlining. When we process 153 // the C call site we don't want to do any alloca merging between X and Y 154 // because their scopes are not disjoint. We could make this smarter by 155 // keeping track of the inline history for each alloca in the 156 // InlinedArrayAllocas but this isn't likely to be a significant win. 157 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC. 158 return; 159 160 // Loop over all the allocas we have so far and see if they can be merged with 161 // a previously inlined alloca. If not, remember that we had it. 162 for (unsigned AllocaNo = 0, E = IFI.StaticAllocas.size(); AllocaNo != E; 163 ++AllocaNo) { 164 AllocaInst *AI = IFI.StaticAllocas[AllocaNo]; 165 166 // Don't bother trying to merge array allocations (they will usually be 167 // canonicalized to be an allocation *of* an array), or allocations whose 168 // type is not itself an array (because we're afraid of pessimizing SRoA). 169 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType()); 170 if (!ATy || AI->isArrayAllocation()) 171 continue; 172 173 // Get the list of all available allocas for this array type. 174 std::vector<AllocaInst *> &AllocasForType = InlinedArrayAllocas[ATy]; 175 176 // Loop over the allocas in AllocasForType to see if we can reuse one. Note 177 // that we have to be careful not to reuse the same "available" alloca for 178 // multiple different allocas that we just inlined, we use the 'UsedAllocas' 179 // set to keep track of which "available" allocas are being used by this 180 // function. Also, AllocasForType can be empty of course! 181 bool MergedAwayAlloca = false; 182 for (AllocaInst *AvailableAlloca : AllocasForType) { 183 Align Align1 = AI->getAlign(); 184 Align Align2 = AvailableAlloca->getAlign(); 185 186 // The available alloca has to be in the right function, not in some other 187 // function in this SCC. 188 if (AvailableAlloca->getParent() != AI->getParent()) 189 continue; 190 191 // If the inlined function already uses this alloca then we can't reuse 192 // it. 193 if (!UsedAllocas.insert(AvailableAlloca).second) 194 continue; 195 196 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare 197 // success! 198 LLVM_DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI 199 << "\n\t\tINTO: " << *AvailableAlloca << '\n'); 200 201 // Move affected dbg.declare calls immediately after the new alloca to 202 // avoid the situation when a dbg.declare precedes its alloca. 203 if (auto *L = LocalAsMetadata::getIfExists(AI)) 204 if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L)) 205 for (User *U : MDV->users()) 206 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U)) 207 DDI->moveBefore(AvailableAlloca->getNextNode()); 208 209 AI->replaceAllUsesWith(AvailableAlloca); 210 211 if (Align1 > Align2) 212 AvailableAlloca->setAlignment(AI->getAlign()); 213 214 AI->eraseFromParent(); 215 MergedAwayAlloca = true; 216 ++NumMergedAllocas; 217 IFI.StaticAllocas[AllocaNo] = nullptr; 218 break; 219 } 220 221 // If we already nuked the alloca, we're done with it. 222 if (MergedAwayAlloca) 223 continue; 224 225 // If we were unable to merge away the alloca either because there are no 226 // allocas of the right type available or because we reused them all 227 // already, remember that this alloca came from an inlined function and mark 228 // it used so we don't reuse it for other allocas from this inline 229 // operation. 230 AllocasForType.push_back(AI); 231 UsedAllocas.insert(AI); 232 } 233 } 234 235 /// If it is possible to inline the specified call site, 236 /// do so and update the CallGraph for this operation. 237 /// 238 /// This function also does some basic book-keeping to update the IR. The 239 /// InlinedArrayAllocas map keeps track of any allocas that are already 240 /// available from other functions inlined into the caller. If we are able to 241 /// inline this call site we attempt to reuse already available allocas or add 242 /// any new allocas to the set if not possible. 243 static InlineResult inlineCallIfPossible( 244 CallBase &CB, InlineFunctionInfo &IFI, 245 InlinedArrayAllocasTy &InlinedArrayAllocas, int InlineHistory, 246 bool InsertLifetime, function_ref<AAResults &(Function &)> &AARGetter, 247 ImportedFunctionsInliningStatistics &ImportedFunctionsStats) { 248 Function *Callee = CB.getCalledFunction(); 249 Function *Caller = CB.getCaller(); 250 251 AAResults &AAR = AARGetter(*Callee); 252 253 // Try to inline the function. Get the list of static allocas that were 254 // inlined. 255 InlineResult IR = InlineFunction(CB, IFI, &AAR, InsertLifetime); 256 if (!IR.isSuccess()) 257 return IR; 258 259 if (InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No) 260 ImportedFunctionsStats.recordInline(*Caller, *Callee); 261 262 AttributeFuncs::mergeAttributesForInlining(*Caller, *Callee); 263 264 if (!DisableInlinedAllocaMerging) 265 mergeInlinedArrayAllocas(Caller, IFI, InlinedArrayAllocas, InlineHistory); 266 267 return IR; // success 268 } 269 270 /// Return true if the specified inline history ID 271 /// indicates an inline history that includes the specified function. 272 static bool inlineHistoryIncludes( 273 Function *F, int InlineHistoryID, 274 const SmallVectorImpl<std::pair<Function *, int>> &InlineHistory) { 275 while (InlineHistoryID != -1) { 276 assert(unsigned(InlineHistoryID) < InlineHistory.size() && 277 "Invalid inline history ID"); 278 if (InlineHistory[InlineHistoryID].first == F) 279 return true; 280 InlineHistoryID = InlineHistory[InlineHistoryID].second; 281 } 282 return false; 283 } 284 285 bool LegacyInlinerBase::doInitialization(CallGraph &CG) { 286 if (InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No) 287 ImportedFunctionsStats.setModuleInfo(CG.getModule()); 288 return false; // No changes to CallGraph. 289 } 290 291 bool LegacyInlinerBase::runOnSCC(CallGraphSCC &SCC) { 292 if (skipSCC(SCC)) 293 return false; 294 return inlineCalls(SCC); 295 } 296 297 static bool 298 inlineCallsImpl(CallGraphSCC &SCC, CallGraph &CG, 299 std::function<AssumptionCache &(Function &)> GetAssumptionCache, 300 ProfileSummaryInfo *PSI, 301 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 302 bool InsertLifetime, 303 function_ref<InlineCost(CallBase &CB)> GetInlineCost, 304 function_ref<AAResults &(Function &)> AARGetter, 305 ImportedFunctionsInliningStatistics &ImportedFunctionsStats) { 306 SmallPtrSet<Function *, 8> SCCFunctions; 307 LLVM_DEBUG(dbgs() << "Inliner visiting SCC:"); 308 for (CallGraphNode *Node : SCC) { 309 Function *F = Node->getFunction(); 310 if (F) 311 SCCFunctions.insert(F); 312 LLVM_DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE")); 313 } 314 315 // Scan through and identify all call sites ahead of time so that we only 316 // inline call sites in the original functions, not call sites that result 317 // from inlining other functions. 318 SmallVector<std::pair<CallBase *, int>, 16> CallSites; 319 320 // When inlining a callee produces new call sites, we want to keep track of 321 // the fact that they were inlined from the callee. This allows us to avoid 322 // infinite inlining in some obscure cases. To represent this, we use an 323 // index into the InlineHistory vector. 324 SmallVector<std::pair<Function *, int>, 8> InlineHistory; 325 326 for (CallGraphNode *Node : SCC) { 327 Function *F = Node->getFunction(); 328 if (!F || F->isDeclaration()) 329 continue; 330 331 OptimizationRemarkEmitter ORE(F); 332 for (BasicBlock &BB : *F) 333 for (Instruction &I : BB) { 334 auto *CB = dyn_cast<CallBase>(&I); 335 // If this isn't a call, or it is a call to an intrinsic, it can 336 // never be inlined. 337 if (!CB || isa<IntrinsicInst>(I)) 338 continue; 339 340 // If this is a direct call to an external function, we can never inline 341 // it. If it is an indirect call, inlining may resolve it to be a 342 // direct call, so we keep it. 343 if (Function *Callee = CB->getCalledFunction()) 344 if (Callee->isDeclaration()) { 345 using namespace ore; 346 347 setInlineRemark(*CB, "unavailable definition"); 348 ORE.emit([&]() { 349 return OptimizationRemarkMissed(DEBUG_TYPE, "NoDefinition", &I) 350 << NV("Callee", Callee) << " will not be inlined into " 351 << NV("Caller", CB->getCaller()) 352 << " because its definition is unavailable" 353 << setIsVerbose(); 354 }); 355 continue; 356 } 357 358 CallSites.push_back(std::make_pair(CB, -1)); 359 } 360 } 361 362 LLVM_DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n"); 363 364 // If there are no calls in this function, exit early. 365 if (CallSites.empty()) 366 return false; 367 368 // Now that we have all of the call sites, move the ones to functions in the 369 // current SCC to the end of the list. 370 unsigned FirstCallInSCC = CallSites.size(); 371 for (unsigned I = 0; I < FirstCallInSCC; ++I) 372 if (Function *F = CallSites[I].first->getCalledFunction()) 373 if (SCCFunctions.count(F)) 374 std::swap(CallSites[I--], CallSites[--FirstCallInSCC]); 375 376 InlinedArrayAllocasTy InlinedArrayAllocas; 377 InlineFunctionInfo InlineInfo(&CG, GetAssumptionCache, PSI); 378 379 // Now that we have all of the call sites, loop over them and inline them if 380 // it looks profitable to do so. 381 bool Changed = false; 382 bool LocalChange; 383 do { 384 LocalChange = false; 385 // Iterate over the outer loop because inlining functions can cause indirect 386 // calls to become direct calls. 387 // CallSites may be modified inside so ranged for loop can not be used. 388 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) { 389 auto &P = CallSites[CSi]; 390 CallBase &CB = *P.first; 391 const int InlineHistoryID = P.second; 392 393 Function *Caller = CB.getCaller(); 394 Function *Callee = CB.getCalledFunction(); 395 396 // We can only inline direct calls to non-declarations. 397 if (!Callee || Callee->isDeclaration()) 398 continue; 399 400 bool IsTriviallyDead = isInstructionTriviallyDead(&CB, &GetTLI(*Caller)); 401 402 if (!IsTriviallyDead) { 403 // If this call site was obtained by inlining another function, verify 404 // that the include path for the function did not include the callee 405 // itself. If so, we'd be recursively inlining the same function, 406 // which would provide the same callsites, which would cause us to 407 // infinitely inline. 408 if (InlineHistoryID != -1 && 409 inlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory)) { 410 setInlineRemark(CB, "recursive"); 411 continue; 412 } 413 } 414 415 // FIXME for new PM: because of the old PM we currently generate ORE and 416 // in turn BFI on demand. With the new PM, the ORE dependency should 417 // just become a regular analysis dependency. 418 OptimizationRemarkEmitter ORE(Caller); 419 420 auto OIC = shouldInline(CB, GetInlineCost, ORE); 421 // If the policy determines that we should inline this function, 422 // delete the call instead. 423 if (!OIC) 424 continue; 425 426 // If this call site is dead and it is to a readonly function, we should 427 // just delete the call instead of trying to inline it, regardless of 428 // size. This happens because IPSCCP propagates the result out of the 429 // call and then we're left with the dead call. 430 if (IsTriviallyDead) { 431 LLVM_DEBUG(dbgs() << " -> Deleting dead call: " << CB << "\n"); 432 // Update the call graph by deleting the edge from Callee to Caller. 433 setInlineRemark(CB, "trivially dead"); 434 CG[Caller]->removeCallEdgeFor(CB); 435 CB.eraseFromParent(); 436 ++NumCallsDeleted; 437 } else { 438 // Get DebugLoc to report. CB will be invalid after Inliner. 439 DebugLoc DLoc = CB.getDebugLoc(); 440 BasicBlock *Block = CB.getParent(); 441 442 // Attempt to inline the function. 443 using namespace ore; 444 445 InlineResult IR = inlineCallIfPossible( 446 CB, InlineInfo, InlinedArrayAllocas, InlineHistoryID, 447 InsertLifetime, AARGetter, ImportedFunctionsStats); 448 if (!IR.isSuccess()) { 449 setInlineRemark(CB, std::string(IR.getFailureReason()) + "; " + 450 inlineCostStr(*OIC)); 451 ORE.emit([&]() { 452 return OptimizationRemarkMissed(DEBUG_TYPE, "NotInlined", DLoc, 453 Block) 454 << NV("Callee", Callee) << " will not be inlined into " 455 << NV("Caller", Caller) << ": " 456 << NV("Reason", IR.getFailureReason()); 457 }); 458 continue; 459 } 460 ++NumInlined; 461 462 emitInlinedInto(ORE, DLoc, Block, *Callee, *Caller, *OIC); 463 464 // If inlining this function gave us any new call sites, throw them 465 // onto our worklist to process. They are useful inline candidates. 466 if (!InlineInfo.InlinedCalls.empty()) { 467 // Create a new inline history entry for this, so that we remember 468 // that these new callsites came about due to inlining Callee. 469 int NewHistoryID = InlineHistory.size(); 470 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID)); 471 472 #ifndef NDEBUG 473 // Make sure no dupplicates in the inline candidates. This could 474 // happen when a callsite is simpilfied to reusing the return value 475 // of another callsite during function cloning, thus the other 476 // callsite will be reconsidered here. 477 DenseSet<CallBase *> DbgCallSites; 478 for (auto &II : CallSites) 479 DbgCallSites.insert(II.first); 480 #endif 481 482 for (Value *Ptr : InlineInfo.InlinedCalls) { 483 #ifndef NDEBUG 484 assert(DbgCallSites.count(dyn_cast<CallBase>(Ptr)) == 0); 485 #endif 486 CallSites.push_back( 487 std::make_pair(dyn_cast<CallBase>(Ptr), NewHistoryID)); 488 } 489 } 490 } 491 492 // If we inlined or deleted the last possible call site to the function, 493 // delete the function body now. 494 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() && 495 // TODO: Can remove if in SCC now. 496 !SCCFunctions.count(Callee) && 497 // The function may be apparently dead, but if there are indirect 498 // callgraph references to the node, we cannot delete it yet, this 499 // could invalidate the CGSCC iterator. 500 CG[Callee]->getNumReferences() == 0) { 501 LLVM_DEBUG(dbgs() << " -> Deleting dead function: " 502 << Callee->getName() << "\n"); 503 CallGraphNode *CalleeNode = CG[Callee]; 504 505 // Remove any call graph edges from the callee to its callees. 506 CalleeNode->removeAllCalledFunctions(); 507 508 // Removing the node for callee from the call graph and delete it. 509 delete CG.removeFunctionFromModule(CalleeNode); 510 ++NumDeleted; 511 } 512 513 // Remove this call site from the list. If possible, use 514 // swap/pop_back for efficiency, but do not use it if doing so would 515 // move a call site to a function in this SCC before the 516 // 'FirstCallInSCC' barrier. 517 if (SCC.isSingular()) { 518 CallSites[CSi] = CallSites.back(); 519 CallSites.pop_back(); 520 } else { 521 CallSites.erase(CallSites.begin() + CSi); 522 } 523 --CSi; 524 525 Changed = true; 526 LocalChange = true; 527 } 528 } while (LocalChange); 529 530 return Changed; 531 } 532 533 bool LegacyInlinerBase::inlineCalls(CallGraphSCC &SCC) { 534 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph(); 535 ACT = &getAnalysis<AssumptionCacheTracker>(); 536 PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); 537 GetTLI = [&](Function &F) -> const TargetLibraryInfo & { 538 return getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); 539 }; 540 auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & { 541 return ACT->getAssumptionCache(F); 542 }; 543 return inlineCallsImpl( 544 SCC, CG, GetAssumptionCache, PSI, GetTLI, InsertLifetime, 545 [&](CallBase &CB) { return getInlineCost(CB); }, LegacyAARGetter(*this), 546 ImportedFunctionsStats); 547 } 548 549 /// Remove now-dead linkonce functions at the end of 550 /// processing to avoid breaking the SCC traversal. 551 bool LegacyInlinerBase::doFinalization(CallGraph &CG) { 552 if (InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No) 553 ImportedFunctionsStats.dump(InlinerFunctionImportStats == 554 InlinerFunctionImportStatsOpts::Verbose); 555 return removeDeadFunctions(CG); 556 } 557 558 /// Remove dead functions that are not included in DNR (Do Not Remove) list. 559 bool LegacyInlinerBase::removeDeadFunctions(CallGraph &CG, 560 bool AlwaysInlineOnly) { 561 SmallVector<CallGraphNode *, 16> FunctionsToRemove; 562 SmallVector<Function *, 16> DeadFunctionsInComdats; 563 564 auto RemoveCGN = [&](CallGraphNode *CGN) { 565 // Remove any call graph edges from the function to its callees. 566 CGN->removeAllCalledFunctions(); 567 568 // Remove any edges from the external node to the function's call graph 569 // node. These edges might have been made irrelegant due to 570 // optimization of the program. 571 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN); 572 573 // Removing the node for callee from the call graph and delete it. 574 FunctionsToRemove.push_back(CGN); 575 }; 576 577 // Scan for all of the functions, looking for ones that should now be removed 578 // from the program. Insert the dead ones in the FunctionsToRemove set. 579 for (const auto &I : CG) { 580 CallGraphNode *CGN = I.second.get(); 581 Function *F = CGN->getFunction(); 582 if (!F || F->isDeclaration()) 583 continue; 584 585 // Handle the case when this function is called and we only want to care 586 // about always-inline functions. This is a bit of a hack to share code 587 // between here and the InlineAlways pass. 588 if (AlwaysInlineOnly && !F->hasFnAttribute(Attribute::AlwaysInline)) 589 continue; 590 591 // If the only remaining users of the function are dead constants, remove 592 // them. 593 F->removeDeadConstantUsers(); 594 595 if (!F->isDefTriviallyDead()) 596 continue; 597 598 // It is unsafe to drop a function with discardable linkage from a COMDAT 599 // without also dropping the other members of the COMDAT. 600 // The inliner doesn't visit non-function entities which are in COMDAT 601 // groups so it is unsafe to do so *unless* the linkage is local. 602 if (!F->hasLocalLinkage()) { 603 if (F->hasComdat()) { 604 DeadFunctionsInComdats.push_back(F); 605 continue; 606 } 607 } 608 609 RemoveCGN(CGN); 610 } 611 if (!DeadFunctionsInComdats.empty()) { 612 // Filter out the functions whose comdats remain alive. 613 filterDeadComdatFunctions(CG.getModule(), DeadFunctionsInComdats); 614 // Remove the rest. 615 for (Function *F : DeadFunctionsInComdats) 616 RemoveCGN(CG[F]); 617 } 618 619 if (FunctionsToRemove.empty()) 620 return false; 621 622 // Now that we know which functions to delete, do so. We didn't want to do 623 // this inline, because that would invalidate our CallGraph::iterator 624 // objects. :( 625 // 626 // Note that it doesn't matter that we are iterating over a non-stable order 627 // here to do this, it doesn't matter which order the functions are deleted 628 // in. 629 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end()); 630 FunctionsToRemove.erase( 631 std::unique(FunctionsToRemove.begin(), FunctionsToRemove.end()), 632 FunctionsToRemove.end()); 633 for (CallGraphNode *CGN : FunctionsToRemove) { 634 delete CG.removeFunctionFromModule(CGN); 635 ++NumDeleted; 636 } 637 return true; 638 } 639 640 InlineAdvisor & 641 InlinerPass::getAdvisor(const ModuleAnalysisManagerCGSCCProxy::Result &MAM, 642 FunctionAnalysisManager &FAM, Module &M) { 643 if (OwnedAdvisor) 644 return *OwnedAdvisor; 645 646 auto *IAA = MAM.getCachedResult<InlineAdvisorAnalysis>(M); 647 if (!IAA) { 648 // It should still be possible to run the inliner as a stand-alone SCC pass, 649 // for test scenarios. In that case, we default to the 650 // DefaultInlineAdvisor, which doesn't need to keep state between SCC pass 651 // runs. It also uses just the default InlineParams. 652 // In this case, we need to use the provided FAM, which is valid for the 653 // duration of the inliner pass, and thus the lifetime of the owned advisor. 654 // The one we would get from the MAM can be invalidated as a result of the 655 // inliner's activity. 656 OwnedAdvisor = 657 std::make_unique<DefaultInlineAdvisor>(M, FAM, getInlineParams()); 658 659 if (!CGSCCInlineReplayFile.empty()) 660 OwnedAdvisor = std::make_unique<ReplayInlineAdvisor>( 661 M, FAM, M.getContext(), std::move(OwnedAdvisor), 662 CGSCCInlineReplayFile, 663 /*EmitRemarks=*/true); 664 665 return *OwnedAdvisor; 666 } 667 assert(IAA->getAdvisor() && 668 "Expected a present InlineAdvisorAnalysis also have an " 669 "InlineAdvisor initialized"); 670 return *IAA->getAdvisor(); 671 } 672 673 PreservedAnalyses InlinerPass::run(LazyCallGraph::SCC &InitialC, 674 CGSCCAnalysisManager &AM, LazyCallGraph &CG, 675 CGSCCUpdateResult &UR) { 676 const auto &MAMProxy = 677 AM.getResult<ModuleAnalysisManagerCGSCCProxy>(InitialC, CG); 678 bool Changed = false; 679 680 assert(InitialC.size() > 0 && "Cannot handle an empty SCC!"); 681 Module &M = *InitialC.begin()->getFunction().getParent(); 682 ProfileSummaryInfo *PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(M); 683 684 FunctionAnalysisManager &FAM = 685 AM.getResult<FunctionAnalysisManagerCGSCCProxy>(InitialC, CG) 686 .getManager(); 687 688 InlineAdvisor &Advisor = getAdvisor(MAMProxy, FAM, M); 689 Advisor.onPassEntry(); 690 691 auto AdvisorOnExit = make_scope_exit([&] { Advisor.onPassExit(); }); 692 693 // We use a single common worklist for calls across the entire SCC. We 694 // process these in-order and append new calls introduced during inlining to 695 // the end. 696 // 697 // Note that this particular order of processing is actually critical to 698 // avoid very bad behaviors. Consider *highly connected* call graphs where 699 // each function contains a small amount of code and a couple of calls to 700 // other functions. Because the LLVM inliner is fundamentally a bottom-up 701 // inliner, it can handle gracefully the fact that these all appear to be 702 // reasonable inlining candidates as it will flatten things until they become 703 // too big to inline, and then move on and flatten another batch. 704 // 705 // However, when processing call edges *within* an SCC we cannot rely on this 706 // bottom-up behavior. As a consequence, with heavily connected *SCCs* of 707 // functions we can end up incrementally inlining N calls into each of 708 // N functions because each incremental inlining decision looks good and we 709 // don't have a topological ordering to prevent explosions. 710 // 711 // To compensate for this, we don't process transitive edges made immediate 712 // by inlining until we've done one pass of inlining across the entire SCC. 713 // Large, highly connected SCCs still lead to some amount of code bloat in 714 // this model, but it is uniformly spread across all the functions in the SCC 715 // and eventually they all become too large to inline, rather than 716 // incrementally maknig a single function grow in a super linear fashion. 717 SmallVector<std::pair<CallBase *, int>, 16> Calls; 718 719 // Populate the initial list of calls in this SCC. 720 for (auto &N : InitialC) { 721 auto &ORE = 722 FAM.getResult<OptimizationRemarkEmitterAnalysis>(N.getFunction()); 723 // We want to generally process call sites top-down in order for 724 // simplifications stemming from replacing the call with the returned value 725 // after inlining to be visible to subsequent inlining decisions. 726 // FIXME: Using instructions sequence is a really bad way to do this. 727 // Instead we should do an actual RPO walk of the function body. 728 for (Instruction &I : instructions(N.getFunction())) 729 if (auto *CB = dyn_cast<CallBase>(&I)) 730 if (Function *Callee = CB->getCalledFunction()) { 731 if (!Callee->isDeclaration()) 732 Calls.push_back({CB, -1}); 733 else if (!isa<IntrinsicInst>(I)) { 734 using namespace ore; 735 setInlineRemark(*CB, "unavailable definition"); 736 ORE.emit([&]() { 737 return OptimizationRemarkMissed(DEBUG_TYPE, "NoDefinition", &I) 738 << NV("Callee", Callee) << " will not be inlined into " 739 << NV("Caller", CB->getCaller()) 740 << " because its definition is unavailable" 741 << setIsVerbose(); 742 }); 743 } 744 } 745 } 746 if (Calls.empty()) 747 return PreservedAnalyses::all(); 748 749 // Capture updatable variable for the current SCC. 750 auto *C = &InitialC; 751 752 // When inlining a callee produces new call sites, we want to keep track of 753 // the fact that they were inlined from the callee. This allows us to avoid 754 // infinite inlining in some obscure cases. To represent this, we use an 755 // index into the InlineHistory vector. 756 SmallVector<std::pair<Function *, int>, 16> InlineHistory; 757 758 // Track a set vector of inlined callees so that we can augment the caller 759 // with all of their edges in the call graph before pruning out the ones that 760 // got simplified away. 761 SmallSetVector<Function *, 4> InlinedCallees; 762 763 // Track the dead functions to delete once finished with inlining calls. We 764 // defer deleting these to make it easier to handle the call graph updates. 765 SmallVector<Function *, 4> DeadFunctions; 766 767 // Loop forward over all of the calls. Note that we cannot cache the size as 768 // inlining can introduce new calls that need to be processed. 769 for (int I = 0; I < (int)Calls.size(); ++I) { 770 // We expect the calls to typically be batched with sequences of calls that 771 // have the same caller, so we first set up some shared infrastructure for 772 // this caller. We also do any pruning we can at this layer on the caller 773 // alone. 774 Function &F = *Calls[I].first->getCaller(); 775 LazyCallGraph::Node &N = *CG.lookup(F); 776 if (CG.lookupSCC(N) != C) 777 continue; 778 779 LLVM_DEBUG(dbgs() << "Inlining calls in: " << F.getName() << "\n"); 780 781 auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & { 782 return FAM.getResult<AssumptionAnalysis>(F); 783 }; 784 785 // Now process as many calls as we have within this caller in the sequence. 786 // We bail out as soon as the caller has to change so we can update the 787 // call graph and prepare the context of that new caller. 788 bool DidInline = false; 789 for (; I < (int)Calls.size() && Calls[I].first->getCaller() == &F; ++I) { 790 auto &P = Calls[I]; 791 CallBase *CB = P.first; 792 const int InlineHistoryID = P.second; 793 Function &Callee = *CB->getCalledFunction(); 794 795 if (InlineHistoryID != -1 && 796 inlineHistoryIncludes(&Callee, InlineHistoryID, InlineHistory)) { 797 setInlineRemark(*CB, "recursive"); 798 continue; 799 } 800 801 // Check if this inlining may repeat breaking an SCC apart that has 802 // already been split once before. In that case, inlining here may 803 // trigger infinite inlining, much like is prevented within the inliner 804 // itself by the InlineHistory above, but spread across CGSCC iterations 805 // and thus hidden from the full inline history. 806 if (CG.lookupSCC(*CG.lookup(Callee)) == C && 807 UR.InlinedInternalEdges.count({&N, C})) { 808 LLVM_DEBUG(dbgs() << "Skipping inlining internal SCC edge from a node " 809 "previously split out of this SCC by inlining: " 810 << F.getName() << " -> " << Callee.getName() << "\n"); 811 setInlineRemark(*CB, "recursive SCC split"); 812 continue; 813 } 814 815 auto Advice = Advisor.getAdvice(*CB, OnlyMandatory); 816 // Check whether we want to inline this callsite. 817 if (!Advice->isInliningRecommended()) { 818 Advice->recordUnattemptedInlining(); 819 continue; 820 } 821 822 // Setup the data structure used to plumb customization into the 823 // `InlineFunction` routine. 824 InlineFunctionInfo IFI( 825 /*cg=*/nullptr, GetAssumptionCache, PSI, 826 &FAM.getResult<BlockFrequencyAnalysis>(*(CB->getCaller())), 827 &FAM.getResult<BlockFrequencyAnalysis>(Callee)); 828 829 InlineResult IR = 830 InlineFunction(*CB, IFI, &FAM.getResult<AAManager>(*CB->getCaller())); 831 if (!IR.isSuccess()) { 832 Advice->recordUnsuccessfulInlining(IR); 833 continue; 834 } 835 836 DidInline = true; 837 InlinedCallees.insert(&Callee); 838 ++NumInlined; 839 840 // Add any new callsites to defined functions to the worklist. 841 if (!IFI.InlinedCallSites.empty()) { 842 int NewHistoryID = InlineHistory.size(); 843 InlineHistory.push_back({&Callee, InlineHistoryID}); 844 845 for (CallBase *ICB : reverse(IFI.InlinedCallSites)) { 846 Function *NewCallee = ICB->getCalledFunction(); 847 if (!NewCallee) { 848 // Try to promote an indirect (virtual) call without waiting for 849 // the post-inline cleanup and the next DevirtSCCRepeatedPass 850 // iteration because the next iteration may not happen and we may 851 // miss inlining it. 852 if (tryPromoteCall(*ICB)) 853 NewCallee = ICB->getCalledFunction(); 854 } 855 if (NewCallee) 856 if (!NewCallee->isDeclaration()) 857 Calls.push_back({ICB, NewHistoryID}); 858 } 859 } 860 861 // Merge the attributes based on the inlining. 862 AttributeFuncs::mergeAttributesForInlining(F, Callee); 863 864 // For local functions, check whether this makes the callee trivially 865 // dead. In that case, we can drop the body of the function eagerly 866 // which may reduce the number of callers of other functions to one, 867 // changing inline cost thresholds. 868 bool CalleeWasDeleted = false; 869 if (Callee.hasLocalLinkage()) { 870 // To check this we also need to nuke any dead constant uses (perhaps 871 // made dead by this operation on other functions). 872 Callee.removeDeadConstantUsers(); 873 if (Callee.use_empty() && !CG.isLibFunction(Callee)) { 874 Calls.erase( 875 std::remove_if(Calls.begin() + I + 1, Calls.end(), 876 [&](const std::pair<CallBase *, int> &Call) { 877 return Call.first->getCaller() == &Callee; 878 }), 879 Calls.end()); 880 // Clear the body and queue the function itself for deletion when we 881 // finish inlining and call graph updates. 882 // Note that after this point, it is an error to do anything other 883 // than use the callee's address or delete it. 884 Callee.dropAllReferences(); 885 assert(!is_contained(DeadFunctions, &Callee) && 886 "Cannot put cause a function to become dead twice!"); 887 DeadFunctions.push_back(&Callee); 888 CalleeWasDeleted = true; 889 } 890 } 891 if (CalleeWasDeleted) 892 Advice->recordInliningWithCalleeDeleted(); 893 else 894 Advice->recordInlining(); 895 } 896 897 // Back the call index up by one to put us in a good position to go around 898 // the outer loop. 899 --I; 900 901 if (!DidInline) 902 continue; 903 Changed = true; 904 905 // At this point, since we have made changes we have at least removed 906 // a call instruction. However, in the process we do some incremental 907 // simplification of the surrounding code. This simplification can 908 // essentially do all of the same things as a function pass and we can 909 // re-use the exact same logic for updating the call graph to reflect the 910 // change. 911 912 // Inside the update, we also update the FunctionAnalysisManager in the 913 // proxy for this particular SCC. We do this as the SCC may have changed and 914 // as we're going to mutate this particular function we want to make sure 915 // the proxy is in place to forward any invalidation events. 916 LazyCallGraph::SCC *OldC = C; 917 C = &updateCGAndAnalysisManagerForCGSCCPass(CG, *C, N, AM, UR, FAM); 918 LLVM_DEBUG(dbgs() << "Updated inlining SCC: " << *C << "\n"); 919 920 // If this causes an SCC to split apart into multiple smaller SCCs, there 921 // is a subtle risk we need to prepare for. Other transformations may 922 // expose an "infinite inlining" opportunity later, and because of the SCC 923 // mutation, we will revisit this function and potentially re-inline. If we 924 // do, and that re-inlining also has the potentially to mutate the SCC 925 // structure, the infinite inlining problem can manifest through infinite 926 // SCC splits and merges. To avoid this, we capture the originating caller 927 // node and the SCC containing the call edge. This is a slight over 928 // approximation of the possible inlining decisions that must be avoided, 929 // but is relatively efficient to store. We use C != OldC to know when 930 // a new SCC is generated and the original SCC may be generated via merge 931 // in later iterations. 932 // 933 // It is also possible that even if no new SCC is generated 934 // (i.e., C == OldC), the original SCC could be split and then merged 935 // into the same one as itself. and the original SCC will be added into 936 // UR.CWorklist again, we want to catch such cases too. 937 // 938 // FIXME: This seems like a very heavyweight way of retaining the inline 939 // history, we should look for a more efficient way of tracking it. 940 if ((C != OldC || UR.CWorklist.count(OldC)) && 941 llvm::any_of(InlinedCallees, [&](Function *Callee) { 942 return CG.lookupSCC(*CG.lookup(*Callee)) == OldC; 943 })) { 944 LLVM_DEBUG(dbgs() << "Inlined an internal call edge and split an SCC, " 945 "retaining this to avoid infinite inlining.\n"); 946 UR.InlinedInternalEdges.insert({&N, OldC}); 947 } 948 InlinedCallees.clear(); 949 } 950 951 // Now that we've finished inlining all of the calls across this SCC, delete 952 // all of the trivially dead functions, updating the call graph and the CGSCC 953 // pass manager in the process. 954 // 955 // Note that this walks a pointer set which has non-deterministic order but 956 // that is OK as all we do is delete things and add pointers to unordered 957 // sets. 958 for (Function *DeadF : DeadFunctions) { 959 // Get the necessary information out of the call graph and nuke the 960 // function there. Also, clear out any cached analyses. 961 auto &DeadC = *CG.lookupSCC(*CG.lookup(*DeadF)); 962 FAM.clear(*DeadF, DeadF->getName()); 963 AM.clear(DeadC, DeadC.getName()); 964 auto &DeadRC = DeadC.getOuterRefSCC(); 965 CG.removeDeadFunction(*DeadF); 966 967 // Mark the relevant parts of the call graph as invalid so we don't visit 968 // them. 969 UR.InvalidatedSCCs.insert(&DeadC); 970 UR.InvalidatedRefSCCs.insert(&DeadRC); 971 972 // And delete the actual function from the module. 973 // The Advisor may use Function pointers to efficiently index various 974 // internal maps, e.g. for memoization. Function cleanup passes like 975 // argument promotion create new functions. It is possible for a new 976 // function to be allocated at the address of a deleted function. We could 977 // index using names, but that's inefficient. Alternatively, we let the 978 // Advisor free the functions when it sees fit. 979 DeadF->getBasicBlockList().clear(); 980 M.getFunctionList().remove(DeadF); 981 982 ++NumDeleted; 983 } 984 985 if (!Changed) 986 return PreservedAnalyses::all(); 987 988 // Even if we change the IR, we update the core CGSCC data structures and so 989 // can preserve the proxy to the function analysis manager. 990 PreservedAnalyses PA; 991 PA.preserve<FunctionAnalysisManagerCGSCCProxy>(); 992 return PA; 993 } 994 995 ModuleInlinerWrapperPass::ModuleInlinerWrapperPass(InlineParams Params, 996 bool Debugging, 997 bool MandatoryFirst, 998 InliningAdvisorMode Mode, 999 unsigned MaxDevirtIterations) 1000 : Params(Params), Mode(Mode), MaxDevirtIterations(MaxDevirtIterations), 1001 PM(Debugging), MPM(Debugging) { 1002 // Run the inliner first. The theory is that we are walking bottom-up and so 1003 // the callees have already been fully optimized, and we want to inline them 1004 // into the callers so that our optimizations can reflect that. 1005 // For PreLinkThinLTO pass, we disable hot-caller heuristic for sample PGO 1006 // because it makes profile annotation in the backend inaccurate. 1007 if (MandatoryFirst) 1008 PM.addPass(InlinerPass(/*OnlyMandatory*/ true)); 1009 PM.addPass(InlinerPass()); 1010 } 1011 1012 PreservedAnalyses ModuleInlinerWrapperPass::run(Module &M, 1013 ModuleAnalysisManager &MAM) { 1014 auto &IAA = MAM.getResult<InlineAdvisorAnalysis>(M); 1015 if (!IAA.tryCreate(Params, Mode, CGSCCInlineReplayFile)) { 1016 M.getContext().emitError( 1017 "Could not setup Inlining Advisor for the requested " 1018 "mode and/or options"); 1019 return PreservedAnalyses::all(); 1020 } 1021 1022 // We wrap the CGSCC pipeline in a devirtualization repeater. This will try 1023 // to detect when we devirtualize indirect calls and iterate the SCC passes 1024 // in that case to try and catch knock-on inlining or function attrs 1025 // opportunities. Then we add it to the module pipeline by walking the SCCs 1026 // in postorder (or bottom-up). 1027 // If MaxDevirtIterations is 0, we just don't use the devirtualization 1028 // wrapper. 1029 if (MaxDevirtIterations == 0) 1030 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(std::move(PM))); 1031 else 1032 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor( 1033 createDevirtSCCRepeatedPass(std::move(PM), MaxDevirtIterations))); 1034 auto Ret = MPM.run(M, MAM); 1035 1036 IAA.clear(); 1037 return Ret; 1038 } 1039