1 //===- Inliner.cpp - Code common to all inliners --------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the mechanics required to implement inlining without 11 // missing any calls and updating the call graph. The decisions of which calls 12 // are profitable to inline are implemented elsewhere. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #include "llvm/Transforms/IPO/Inliner.h" 17 #include "llvm/ADT/DenseMap.h" 18 #include "llvm/ADT/None.h" 19 #include "llvm/ADT/Optional.h" 20 #include "llvm/ADT/STLExtras.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/AliasAnalysis.h" 27 #include "llvm/Analysis/AssumptionCache.h" 28 #include "llvm/Analysis/BasicAliasAnalysis.h" 29 #include "llvm/Analysis/BlockFrequencyInfo.h" 30 #include "llvm/Analysis/CGSCCPassManager.h" 31 #include "llvm/Analysis/CallGraph.h" 32 #include "llvm/Analysis/InlineCost.h" 33 #include "llvm/Analysis/LazyCallGraph.h" 34 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 35 #include "llvm/Analysis/ProfileSummaryInfo.h" 36 #include "llvm/Analysis/TargetLibraryInfo.h" 37 #include "llvm/Analysis/TargetTransformInfo.h" 38 #include "llvm/Transforms/Utils/Local.h" 39 #include "llvm/IR/Attributes.h" 40 #include "llvm/IR/BasicBlock.h" 41 #include "llvm/IR/CallSite.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/Cloning.h" 62 #include "llvm/Transforms/Utils/ImportedFunctionsInliningStatistics.h" 63 #include "llvm/Transforms/Utils/ModuleUtils.h" 64 #include <algorithm> 65 #include <cassert> 66 #include <functional> 67 #include <tuple> 68 #include <utility> 69 #include <vector> 70 71 using namespace llvm; 72 73 #define DEBUG_TYPE "inline" 74 75 STATISTIC(NumInlined, "Number of functions inlined"); 76 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined"); 77 STATISTIC(NumDeleted, "Number of functions deleted because all callers found"); 78 STATISTIC(NumMergedAllocas, "Number of allocas merged together"); 79 80 // This weirdly named statistic tracks the number of times that, when attempting 81 // to inline a function A into B, we analyze the callers of B in order to see 82 // if those would be more profitable and blocked inline steps. 83 STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed"); 84 85 /// Flag to disable manual alloca merging. 86 /// 87 /// Merging of allocas was originally done as a stack-size saving technique 88 /// prior to LLVM's code generator having support for stack coloring based on 89 /// lifetime markers. It is now in the process of being removed. To experiment 90 /// with disabling it and relying fully on lifetime marker based stack 91 /// coloring, you can pass this flag to LLVM. 92 static cl::opt<bool> 93 DisableInlinedAllocaMerging("disable-inlined-alloca-merging", 94 cl::init(false), cl::Hidden); 95 96 namespace { 97 98 enum class InlinerFunctionImportStatsOpts { 99 No = 0, 100 Basic = 1, 101 Verbose = 2, 102 }; 103 104 } // end anonymous namespace 105 106 static cl::opt<InlinerFunctionImportStatsOpts> InlinerFunctionImportStats( 107 "inliner-function-import-stats", 108 cl::init(InlinerFunctionImportStatsOpts::No), 109 cl::values(clEnumValN(InlinerFunctionImportStatsOpts::Basic, "basic", 110 "basic statistics"), 111 clEnumValN(InlinerFunctionImportStatsOpts::Verbose, "verbose", 112 "printing of statistics for each inlined function")), 113 cl::Hidden, cl::desc("Enable inliner stats for imported functions")); 114 115 LegacyInlinerBase::LegacyInlinerBase(char &ID) : CallGraphSCCPass(ID) {} 116 117 LegacyInlinerBase::LegacyInlinerBase(char &ID, bool InsertLifetime) 118 : CallGraphSCCPass(ID), InsertLifetime(InsertLifetime) {} 119 120 /// For this class, we declare that we require and preserve the call graph. 121 /// If the derived class implements this method, it should 122 /// always explicitly call the implementation here. 123 void LegacyInlinerBase::getAnalysisUsage(AnalysisUsage &AU) const { 124 AU.addRequired<AssumptionCacheTracker>(); 125 AU.addRequired<ProfileSummaryInfoWrapperPass>(); 126 AU.addRequired<TargetLibraryInfoWrapperPass>(); 127 getAAResultsAnalysisUsage(AU); 128 CallGraphSCCPass::getAnalysisUsage(AU); 129 } 130 131 using InlinedArrayAllocasTy = DenseMap<ArrayType *, std::vector<AllocaInst *>>; 132 133 /// Look at all of the allocas that we inlined through this call site. If we 134 /// have already inlined other allocas through other calls into this function, 135 /// then we know that they have disjoint lifetimes and that we can merge them. 136 /// 137 /// There are many heuristics possible for merging these allocas, and the 138 /// different options have different tradeoffs. One thing that we *really* 139 /// don't want to hurt is SRoA: once inlining happens, often allocas are no 140 /// longer address taken and so they can be promoted. 141 /// 142 /// Our "solution" for that is to only merge allocas whose outermost type is an 143 /// array type. These are usually not promoted because someone is using a 144 /// variable index into them. These are also often the most important ones to 145 /// merge. 146 /// 147 /// A better solution would be to have real memory lifetime markers in the IR 148 /// and not have the inliner do any merging of allocas at all. This would 149 /// allow the backend to do proper stack slot coloring of all allocas that 150 /// *actually make it to the backend*, which is really what we want. 151 /// 152 /// Because we don't have this information, we do this simple and useful hack. 153 static void mergeInlinedArrayAllocas( 154 Function *Caller, InlineFunctionInfo &IFI, 155 InlinedArrayAllocasTy &InlinedArrayAllocas, int InlineHistory) { 156 SmallPtrSet<AllocaInst *, 16> UsedAllocas; 157 158 // When processing our SCC, check to see if CS was inlined from some other 159 // call site. For example, if we're processing "A" in this code: 160 // A() { B() } 161 // B() { x = alloca ... C() } 162 // C() { y = alloca ... } 163 // Assume that C was not inlined into B initially, and so we're processing A 164 // and decide to inline B into A. Doing this makes an alloca available for 165 // reuse and makes a callsite (C) available for inlining. When we process 166 // the C call site we don't want to do any alloca merging between X and Y 167 // because their scopes are not disjoint. We could make this smarter by 168 // keeping track of the inline history for each alloca in the 169 // InlinedArrayAllocas but this isn't likely to be a significant win. 170 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC. 171 return; 172 173 // Loop over all the allocas we have so far and see if they can be merged with 174 // a previously inlined alloca. If not, remember that we had it. 175 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size(); AllocaNo != e; 176 ++AllocaNo) { 177 AllocaInst *AI = IFI.StaticAllocas[AllocaNo]; 178 179 // Don't bother trying to merge array allocations (they will usually be 180 // canonicalized to be an allocation *of* an array), or allocations whose 181 // type is not itself an array (because we're afraid of pessimizing SRoA). 182 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType()); 183 if (!ATy || AI->isArrayAllocation()) 184 continue; 185 186 // Get the list of all available allocas for this array type. 187 std::vector<AllocaInst *> &AllocasForType = InlinedArrayAllocas[ATy]; 188 189 // Loop over the allocas in AllocasForType to see if we can reuse one. Note 190 // that we have to be careful not to reuse the same "available" alloca for 191 // multiple different allocas that we just inlined, we use the 'UsedAllocas' 192 // set to keep track of which "available" allocas are being used by this 193 // function. Also, AllocasForType can be empty of course! 194 bool MergedAwayAlloca = false; 195 for (AllocaInst *AvailableAlloca : AllocasForType) { 196 unsigned Align1 = AI->getAlignment(), 197 Align2 = AvailableAlloca->getAlignment(); 198 199 // The available alloca has to be in the right function, not in some other 200 // function in this SCC. 201 if (AvailableAlloca->getParent() != AI->getParent()) 202 continue; 203 204 // If the inlined function already uses this alloca then we can't reuse 205 // it. 206 if (!UsedAllocas.insert(AvailableAlloca).second) 207 continue; 208 209 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare 210 // success! 211 LLVM_DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI 212 << "\n\t\tINTO: " << *AvailableAlloca << '\n'); 213 214 // Move affected dbg.declare calls immediately after the new alloca to 215 // avoid the situation when a dbg.declare precedes its alloca. 216 if (auto *L = LocalAsMetadata::getIfExists(AI)) 217 if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L)) 218 for (User *U : MDV->users()) 219 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U)) 220 DDI->moveBefore(AvailableAlloca->getNextNode()); 221 222 AI->replaceAllUsesWith(AvailableAlloca); 223 224 if (Align1 != Align2) { 225 if (!Align1 || !Align2) { 226 const DataLayout &DL = Caller->getParent()->getDataLayout(); 227 unsigned TypeAlign = DL.getABITypeAlignment(AI->getAllocatedType()); 228 229 Align1 = Align1 ? Align1 : TypeAlign; 230 Align2 = Align2 ? Align2 : TypeAlign; 231 } 232 233 if (Align1 > Align2) 234 AvailableAlloca->setAlignment(AI->getAlignment()); 235 } 236 237 AI->eraseFromParent(); 238 MergedAwayAlloca = true; 239 ++NumMergedAllocas; 240 IFI.StaticAllocas[AllocaNo] = nullptr; 241 break; 242 } 243 244 // If we already nuked the alloca, we're done with it. 245 if (MergedAwayAlloca) 246 continue; 247 248 // If we were unable to merge away the alloca either because there are no 249 // allocas of the right type available or because we reused them all 250 // already, remember that this alloca came from an inlined function and mark 251 // it used so we don't reuse it for other allocas from this inline 252 // operation. 253 AllocasForType.push_back(AI); 254 UsedAllocas.insert(AI); 255 } 256 } 257 258 /// If it is possible to inline the specified call site, 259 /// do so and update the CallGraph for this operation. 260 /// 261 /// This function also does some basic book-keeping to update the IR. The 262 /// InlinedArrayAllocas map keeps track of any allocas that are already 263 /// available from other functions inlined into the caller. If we are able to 264 /// inline this call site we attempt to reuse already available allocas or add 265 /// any new allocas to the set if not possible. 266 static bool InlineCallIfPossible( 267 CallSite CS, InlineFunctionInfo &IFI, 268 InlinedArrayAllocasTy &InlinedArrayAllocas, int InlineHistory, 269 bool InsertLifetime, function_ref<AAResults &(Function &)> &AARGetter, 270 ImportedFunctionsInliningStatistics &ImportedFunctionsStats) { 271 Function *Callee = CS.getCalledFunction(); 272 Function *Caller = CS.getCaller(); 273 274 AAResults &AAR = AARGetter(*Callee); 275 276 // Try to inline the function. Get the list of static allocas that were 277 // inlined. 278 if (!InlineFunction(CS, IFI, &AAR, InsertLifetime)) 279 return false; 280 281 if (InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No) 282 ImportedFunctionsStats.recordInline(*Caller, *Callee); 283 284 AttributeFuncs::mergeAttributesForInlining(*Caller, *Callee); 285 286 if (!DisableInlinedAllocaMerging) 287 mergeInlinedArrayAllocas(Caller, IFI, InlinedArrayAllocas, InlineHistory); 288 289 return true; 290 } 291 292 /// Return true if inlining of CS can block the caller from being 293 /// inlined which is proved to be more beneficial. \p IC is the 294 /// estimated inline cost associated with callsite \p CS. 295 /// \p TotalSecondaryCost will be set to the estimated cost of inlining the 296 /// caller if \p CS is suppressed for inlining. 297 static bool 298 shouldBeDeferred(Function *Caller, CallSite CS, InlineCost IC, 299 int &TotalSecondaryCost, 300 function_ref<InlineCost(CallSite CS)> GetInlineCost) { 301 // For now we only handle local or inline functions. 302 if (!Caller->hasLocalLinkage() && !Caller->hasLinkOnceODRLinkage()) 303 return false; 304 // Try to detect the case where the current inlining candidate caller (call 305 // it B) is a static or linkonce-ODR function and is an inlining candidate 306 // elsewhere, and the current candidate callee (call it C) is large enough 307 // that inlining it into B would make B too big to inline later. In these 308 // circumstances it may be best not to inline C into B, but to inline B into 309 // its callers. 310 // 311 // This only applies to static and linkonce-ODR functions because those are 312 // expected to be available for inlining in the translation units where they 313 // are used. Thus we will always have the opportunity to make local inlining 314 // decisions. Importantly the linkonce-ODR linkage covers inline functions 315 // and templates in C++. 316 // 317 // FIXME: All of this logic should be sunk into getInlineCost. It relies on 318 // the internal implementation of the inline cost metrics rather than 319 // treating them as truly abstract units etc. 320 TotalSecondaryCost = 0; 321 // The candidate cost to be imposed upon the current function. 322 int CandidateCost = IC.getCost() - 1; 323 // This bool tracks what happens if we do NOT inline C into B. 324 bool callerWillBeRemoved = Caller->hasLocalLinkage(); 325 // This bool tracks what happens if we DO inline C into B. 326 bool inliningPreventsSomeOuterInline = false; 327 for (User *U : Caller->users()) { 328 CallSite CS2(U); 329 330 // If this isn't a call to Caller (it could be some other sort 331 // of reference) skip it. Such references will prevent the caller 332 // from being removed. 333 if (!CS2 || CS2.getCalledFunction() != Caller) { 334 callerWillBeRemoved = false; 335 continue; 336 } 337 338 InlineCost IC2 = GetInlineCost(CS2); 339 ++NumCallerCallersAnalyzed; 340 if (!IC2) { 341 callerWillBeRemoved = false; 342 continue; 343 } 344 if (IC2.isAlways()) 345 continue; 346 347 // See if inlining of the original callsite would erase the cost delta of 348 // this callsite. We subtract off the penalty for the call instruction, 349 // which we would be deleting. 350 if (IC2.getCostDelta() <= CandidateCost) { 351 inliningPreventsSomeOuterInline = true; 352 TotalSecondaryCost += IC2.getCost(); 353 } 354 } 355 // If all outer calls to Caller would get inlined, the cost for the last 356 // one is set very low by getInlineCost, in anticipation that Caller will 357 // be removed entirely. We did not account for this above unless there 358 // is only one caller of Caller. 359 if (callerWillBeRemoved && !Caller->hasOneUse()) 360 TotalSecondaryCost -= InlineConstants::LastCallToStaticBonus; 361 362 if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost()) 363 return true; 364 365 return false; 366 } 367 368 /// Return the cost only if the inliner should attempt to inline at the given 369 /// CallSite. If we return the cost, we will emit an optimisation remark later 370 /// using that cost, so we won't do so from this function. 371 static Optional<InlineCost> 372 shouldInline(CallSite CS, function_ref<InlineCost(CallSite CS)> GetInlineCost, 373 OptimizationRemarkEmitter &ORE) { 374 using namespace ore; 375 376 InlineCost IC = GetInlineCost(CS); 377 Instruction *Call = CS.getInstruction(); 378 Function *Callee = CS.getCalledFunction(); 379 Function *Caller = CS.getCaller(); 380 381 if (IC.isAlways()) { 382 LLVM_DEBUG(dbgs() << " Inlining: cost=always" 383 << ", Call: " << *CS.getInstruction() << "\n"); 384 return IC; 385 } 386 387 if (IC.isNever()) { 388 LLVM_DEBUG(dbgs() << " NOT Inlining: cost=never" 389 << ", Call: " << *CS.getInstruction() << "\n"); 390 ORE.emit([&]() { 391 return OptimizationRemarkMissed(DEBUG_TYPE, "NeverInline", Call) 392 << NV("Callee", Callee) << " not inlined into " 393 << NV("Caller", Caller) 394 << " because it should never be inlined (cost=never)"; 395 }); 396 return None; 397 } 398 399 if (!IC) { 400 LLVM_DEBUG(dbgs() << " NOT Inlining: cost=" << IC.getCost() 401 << ", thres=" << IC.getThreshold() 402 << ", Call: " << *CS.getInstruction() << "\n"); 403 ORE.emit([&]() { 404 return OptimizationRemarkMissed(DEBUG_TYPE, "TooCostly", Call) 405 << NV("Callee", Callee) << " not inlined into " 406 << NV("Caller", Caller) << " because too costly to inline (cost=" 407 << NV("Cost", IC.getCost()) 408 << ", threshold=" << NV("Threshold", IC.getThreshold()) << ")"; 409 }); 410 return None; 411 } 412 413 int TotalSecondaryCost = 0; 414 if (shouldBeDeferred(Caller, CS, IC, TotalSecondaryCost, GetInlineCost)) { 415 LLVM_DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() 416 << " Cost = " << IC.getCost() 417 << ", outer Cost = " << TotalSecondaryCost << '\n'); 418 ORE.emit([&]() { 419 return OptimizationRemarkMissed(DEBUG_TYPE, "IncreaseCostInOtherContexts", 420 Call) 421 << "Not inlining. Cost of inlining " << NV("Callee", Callee) 422 << " increases the cost of inlining " << NV("Caller", Caller) 423 << " in other contexts"; 424 }); 425 426 // IC does not bool() to false, so get an InlineCost that will. 427 // This will not be inspected to make an error message. 428 return None; 429 } 430 431 LLVM_DEBUG(dbgs() << " Inlining: cost=" << IC.getCost() 432 << ", thres=" << IC.getThreshold() 433 << ", Call: " << *CS.getInstruction() << '\n'); 434 return IC; 435 } 436 437 /// Return true if the specified inline history ID 438 /// indicates an inline history that includes the specified function. 439 static bool InlineHistoryIncludes( 440 Function *F, int InlineHistoryID, 441 const SmallVectorImpl<std::pair<Function *, int>> &InlineHistory) { 442 while (InlineHistoryID != -1) { 443 assert(unsigned(InlineHistoryID) < InlineHistory.size() && 444 "Invalid inline history ID"); 445 if (InlineHistory[InlineHistoryID].first == F) 446 return true; 447 InlineHistoryID = InlineHistory[InlineHistoryID].second; 448 } 449 return false; 450 } 451 452 bool LegacyInlinerBase::doInitialization(CallGraph &CG) { 453 if (InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No) 454 ImportedFunctionsStats.setModuleInfo(CG.getModule()); 455 return false; // No changes to CallGraph. 456 } 457 458 bool LegacyInlinerBase::runOnSCC(CallGraphSCC &SCC) { 459 if (skipSCC(SCC)) 460 return false; 461 return inlineCalls(SCC); 462 } 463 464 static bool 465 inlineCallsImpl(CallGraphSCC &SCC, CallGraph &CG, 466 std::function<AssumptionCache &(Function &)> GetAssumptionCache, 467 ProfileSummaryInfo *PSI, TargetLibraryInfo &TLI, 468 bool InsertLifetime, 469 function_ref<InlineCost(CallSite CS)> GetInlineCost, 470 function_ref<AAResults &(Function &)> AARGetter, 471 ImportedFunctionsInliningStatistics &ImportedFunctionsStats) { 472 SmallPtrSet<Function *, 8> SCCFunctions; 473 LLVM_DEBUG(dbgs() << "Inliner visiting SCC:"); 474 for (CallGraphNode *Node : SCC) { 475 Function *F = Node->getFunction(); 476 if (F) 477 SCCFunctions.insert(F); 478 LLVM_DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE")); 479 } 480 481 // Scan through and identify all call sites ahead of time so that we only 482 // inline call sites in the original functions, not call sites that result 483 // from inlining other functions. 484 SmallVector<std::pair<CallSite, int>, 16> CallSites; 485 486 // When inlining a callee produces new call sites, we want to keep track of 487 // the fact that they were inlined from the callee. This allows us to avoid 488 // infinite inlining in some obscure cases. To represent this, we use an 489 // index into the InlineHistory vector. 490 SmallVector<std::pair<Function *, int>, 8> InlineHistory; 491 492 for (CallGraphNode *Node : SCC) { 493 Function *F = Node->getFunction(); 494 if (!F || F->isDeclaration()) 495 continue; 496 497 OptimizationRemarkEmitter ORE(F); 498 for (BasicBlock &BB : *F) 499 for (Instruction &I : BB) { 500 CallSite CS(cast<Value>(&I)); 501 // If this isn't a call, or it is a call to an intrinsic, it can 502 // never be inlined. 503 if (!CS || isa<IntrinsicInst>(I)) 504 continue; 505 506 // If this is a direct call to an external function, we can never inline 507 // it. If it is an indirect call, inlining may resolve it to be a 508 // direct call, so we keep it. 509 if (Function *Callee = CS.getCalledFunction()) 510 if (Callee->isDeclaration()) { 511 using namespace ore; 512 513 ORE.emit([&]() { 514 return OptimizationRemarkMissed(DEBUG_TYPE, "NoDefinition", &I) 515 << NV("Callee", Callee) << " will not be inlined into " 516 << NV("Caller", CS.getCaller()) 517 << " because its definition is unavailable" 518 << setIsVerbose(); 519 }); 520 continue; 521 } 522 523 CallSites.push_back(std::make_pair(CS, -1)); 524 } 525 } 526 527 LLVM_DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n"); 528 529 // If there are no calls in this function, exit early. 530 if (CallSites.empty()) 531 return false; 532 533 // Now that we have all of the call sites, move the ones to functions in the 534 // current SCC to the end of the list. 535 unsigned FirstCallInSCC = CallSites.size(); 536 for (unsigned i = 0; i < FirstCallInSCC; ++i) 537 if (Function *F = CallSites[i].first.getCalledFunction()) 538 if (SCCFunctions.count(F)) 539 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]); 540 541 InlinedArrayAllocasTy InlinedArrayAllocas; 542 InlineFunctionInfo InlineInfo(&CG, &GetAssumptionCache, PSI); 543 544 // Now that we have all of the call sites, loop over them and inline them if 545 // it looks profitable to do so. 546 bool Changed = false; 547 bool LocalChange; 548 do { 549 LocalChange = false; 550 // Iterate over the outer loop because inlining functions can cause indirect 551 // calls to become direct calls. 552 // CallSites may be modified inside so ranged for loop can not be used. 553 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) { 554 CallSite CS = CallSites[CSi].first; 555 556 Function *Caller = CS.getCaller(); 557 Function *Callee = CS.getCalledFunction(); 558 559 // We can only inline direct calls to non-declarations. 560 if (!Callee || Callee->isDeclaration()) 561 continue; 562 563 Instruction *Instr = CS.getInstruction(); 564 565 bool IsTriviallyDead = isInstructionTriviallyDead(Instr, &TLI); 566 567 int InlineHistoryID; 568 if (!IsTriviallyDead) { 569 // If this call site was obtained by inlining another function, verify 570 // that the include path for the function did not include the callee 571 // itself. If so, we'd be recursively inlining the same function, 572 // which would provide the same callsites, which would cause us to 573 // infinitely inline. 574 InlineHistoryID = CallSites[CSi].second; 575 if (InlineHistoryID != -1 && 576 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory)) 577 continue; 578 } 579 580 // FIXME for new PM: because of the old PM we currently generate ORE and 581 // in turn BFI on demand. With the new PM, the ORE dependency should 582 // just become a regular analysis dependency. 583 OptimizationRemarkEmitter ORE(Caller); 584 585 Optional<InlineCost> OIC = shouldInline(CS, GetInlineCost, ORE); 586 // If the policy determines that we should inline this function, 587 // delete the call instead. 588 if (!OIC) 589 continue; 590 591 // If this call site is dead and it is to a readonly function, we should 592 // just delete the call instead of trying to inline it, regardless of 593 // size. This happens because IPSCCP propagates the result out of the 594 // call and then we're left with the dead call. 595 if (IsTriviallyDead) { 596 LLVM_DEBUG(dbgs() << " -> Deleting dead call: " << *Instr << "\n"); 597 // Update the call graph by deleting the edge from Callee to Caller. 598 CG[Caller]->removeCallEdgeFor(CS); 599 Instr->eraseFromParent(); 600 ++NumCallsDeleted; 601 } else { 602 // Get DebugLoc to report. CS will be invalid after Inliner. 603 DebugLoc DLoc = CS->getDebugLoc(); 604 BasicBlock *Block = CS.getParent(); 605 606 // Attempt to inline the function. 607 using namespace ore; 608 609 if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas, 610 InlineHistoryID, InsertLifetime, AARGetter, 611 ImportedFunctionsStats)) { 612 ORE.emit([&]() { 613 return OptimizationRemarkMissed(DEBUG_TYPE, "NotInlined", DLoc, 614 Block) 615 << NV("Callee", Callee) << " will not be inlined into " 616 << NV("Caller", Caller); 617 }); 618 continue; 619 } 620 ++NumInlined; 621 622 ORE.emit([&]() { 623 bool AlwaysInline = OIC->isAlways(); 624 StringRef RemarkName = AlwaysInline ? "AlwaysInline" : "Inlined"; 625 OptimizationRemark R(DEBUG_TYPE, RemarkName, DLoc, Block); 626 R << NV("Callee", Callee) << " inlined into "; 627 R << NV("Caller", Caller); 628 if (AlwaysInline) 629 R << " with cost=always"; 630 else { 631 R << " with cost=" << NV("Cost", OIC->getCost()); 632 R << " (threshold=" << NV("Threshold", OIC->getThreshold()); 633 R << ")"; 634 } 635 return R; 636 }); 637 638 // If inlining this function gave us any new call sites, throw them 639 // onto our worklist to process. They are useful inline candidates. 640 if (!InlineInfo.InlinedCalls.empty()) { 641 // Create a new inline history entry for this, so that we remember 642 // that these new callsites came about due to inlining Callee. 643 int NewHistoryID = InlineHistory.size(); 644 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID)); 645 646 for (Value *Ptr : InlineInfo.InlinedCalls) 647 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID)); 648 } 649 } 650 651 // If we inlined or deleted the last possible call site to the function, 652 // delete the function body now. 653 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() && 654 // TODO: Can remove if in SCC now. 655 !SCCFunctions.count(Callee) && 656 // The function may be apparently dead, but if there are indirect 657 // callgraph references to the node, we cannot delete it yet, this 658 // could invalidate the CGSCC iterator. 659 CG[Callee]->getNumReferences() == 0) { 660 LLVM_DEBUG(dbgs() << " -> Deleting dead function: " 661 << Callee->getName() << "\n"); 662 CallGraphNode *CalleeNode = CG[Callee]; 663 664 // Remove any call graph edges from the callee to its callees. 665 CalleeNode->removeAllCalledFunctions(); 666 667 // Removing the node for callee from the call graph and delete it. 668 delete CG.removeFunctionFromModule(CalleeNode); 669 ++NumDeleted; 670 } 671 672 // Remove this call site from the list. If possible, use 673 // swap/pop_back for efficiency, but do not use it if doing so would 674 // move a call site to a function in this SCC before the 675 // 'FirstCallInSCC' barrier. 676 if (SCC.isSingular()) { 677 CallSites[CSi] = CallSites.back(); 678 CallSites.pop_back(); 679 } else { 680 CallSites.erase(CallSites.begin() + CSi); 681 } 682 --CSi; 683 684 Changed = true; 685 LocalChange = true; 686 } 687 } while (LocalChange); 688 689 return Changed; 690 } 691 692 bool LegacyInlinerBase::inlineCalls(CallGraphSCC &SCC) { 693 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph(); 694 ACT = &getAnalysis<AssumptionCacheTracker>(); 695 PSI = getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); 696 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(); 697 auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & { 698 return ACT->getAssumptionCache(F); 699 }; 700 return inlineCallsImpl(SCC, CG, GetAssumptionCache, PSI, TLI, InsertLifetime, 701 [this](CallSite CS) { return getInlineCost(CS); }, 702 LegacyAARGetter(*this), ImportedFunctionsStats); 703 } 704 705 /// Remove now-dead linkonce functions at the end of 706 /// processing to avoid breaking the SCC traversal. 707 bool LegacyInlinerBase::doFinalization(CallGraph &CG) { 708 if (InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No) 709 ImportedFunctionsStats.dump(InlinerFunctionImportStats == 710 InlinerFunctionImportStatsOpts::Verbose); 711 return removeDeadFunctions(CG); 712 } 713 714 /// Remove dead functions that are not included in DNR (Do Not Remove) list. 715 bool LegacyInlinerBase::removeDeadFunctions(CallGraph &CG, 716 bool AlwaysInlineOnly) { 717 SmallVector<CallGraphNode *, 16> FunctionsToRemove; 718 SmallVector<Function *, 16> DeadFunctionsInComdats; 719 720 auto RemoveCGN = [&](CallGraphNode *CGN) { 721 // Remove any call graph edges from the function to its callees. 722 CGN->removeAllCalledFunctions(); 723 724 // Remove any edges from the external node to the function's call graph 725 // node. These edges might have been made irrelegant due to 726 // optimization of the program. 727 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN); 728 729 // Removing the node for callee from the call graph and delete it. 730 FunctionsToRemove.push_back(CGN); 731 }; 732 733 // Scan for all of the functions, looking for ones that should now be removed 734 // from the program. Insert the dead ones in the FunctionsToRemove set. 735 for (const auto &I : CG) { 736 CallGraphNode *CGN = I.second.get(); 737 Function *F = CGN->getFunction(); 738 if (!F || F->isDeclaration()) 739 continue; 740 741 // Handle the case when this function is called and we only want to care 742 // about always-inline functions. This is a bit of a hack to share code 743 // between here and the InlineAlways pass. 744 if (AlwaysInlineOnly && !F->hasFnAttribute(Attribute::AlwaysInline)) 745 continue; 746 747 // If the only remaining users of the function are dead constants, remove 748 // them. 749 F->removeDeadConstantUsers(); 750 751 if (!F->isDefTriviallyDead()) 752 continue; 753 754 // It is unsafe to drop a function with discardable linkage from a COMDAT 755 // without also dropping the other members of the COMDAT. 756 // The inliner doesn't visit non-function entities which are in COMDAT 757 // groups so it is unsafe to do so *unless* the linkage is local. 758 if (!F->hasLocalLinkage()) { 759 if (F->hasComdat()) { 760 DeadFunctionsInComdats.push_back(F); 761 continue; 762 } 763 } 764 765 RemoveCGN(CGN); 766 } 767 if (!DeadFunctionsInComdats.empty()) { 768 // Filter out the functions whose comdats remain alive. 769 filterDeadComdatFunctions(CG.getModule(), DeadFunctionsInComdats); 770 // Remove the rest. 771 for (Function *F : DeadFunctionsInComdats) 772 RemoveCGN(CG[F]); 773 } 774 775 if (FunctionsToRemove.empty()) 776 return false; 777 778 // Now that we know which functions to delete, do so. We didn't want to do 779 // this inline, because that would invalidate our CallGraph::iterator 780 // objects. :( 781 // 782 // Note that it doesn't matter that we are iterating over a non-stable order 783 // here to do this, it doesn't matter which order the functions are deleted 784 // in. 785 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end()); 786 FunctionsToRemove.erase( 787 std::unique(FunctionsToRemove.begin(), FunctionsToRemove.end()), 788 FunctionsToRemove.end()); 789 for (CallGraphNode *CGN : FunctionsToRemove) { 790 delete CG.removeFunctionFromModule(CGN); 791 ++NumDeleted; 792 } 793 return true; 794 } 795 796 InlinerPass::~InlinerPass() { 797 if (ImportedFunctionsStats) { 798 assert(InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No); 799 ImportedFunctionsStats->dump(InlinerFunctionImportStats == 800 InlinerFunctionImportStatsOpts::Verbose); 801 } 802 } 803 804 PreservedAnalyses InlinerPass::run(LazyCallGraph::SCC &InitialC, 805 CGSCCAnalysisManager &AM, LazyCallGraph &CG, 806 CGSCCUpdateResult &UR) { 807 const ModuleAnalysisManager &MAM = 808 AM.getResult<ModuleAnalysisManagerCGSCCProxy>(InitialC, CG).getManager(); 809 bool Changed = false; 810 811 assert(InitialC.size() > 0 && "Cannot handle an empty SCC!"); 812 Module &M = *InitialC.begin()->getFunction().getParent(); 813 ProfileSummaryInfo *PSI = MAM.getCachedResult<ProfileSummaryAnalysis>(M); 814 815 if (!ImportedFunctionsStats && 816 InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No) { 817 ImportedFunctionsStats = 818 llvm::make_unique<ImportedFunctionsInliningStatistics>(); 819 ImportedFunctionsStats->setModuleInfo(M); 820 } 821 822 // We use a single common worklist for calls across the entire SCC. We 823 // process these in-order and append new calls introduced during inlining to 824 // the end. 825 // 826 // Note that this particular order of processing is actually critical to 827 // avoid very bad behaviors. Consider *highly connected* call graphs where 828 // each function contains a small amonut of code and a couple of calls to 829 // other functions. Because the LLVM inliner is fundamentally a bottom-up 830 // inliner, it can handle gracefully the fact that these all appear to be 831 // reasonable inlining candidates as it will flatten things until they become 832 // too big to inline, and then move on and flatten another batch. 833 // 834 // However, when processing call edges *within* an SCC we cannot rely on this 835 // bottom-up behavior. As a consequence, with heavily connected *SCCs* of 836 // functions we can end up incrementally inlining N calls into each of 837 // N functions because each incremental inlining decision looks good and we 838 // don't have a topological ordering to prevent explosions. 839 // 840 // To compensate for this, we don't process transitive edges made immediate 841 // by inlining until we've done one pass of inlining across the entire SCC. 842 // Large, highly connected SCCs still lead to some amount of code bloat in 843 // this model, but it is uniformly spread across all the functions in the SCC 844 // and eventually they all become too large to inline, rather than 845 // incrementally maknig a single function grow in a super linear fashion. 846 SmallVector<std::pair<CallSite, int>, 16> Calls; 847 848 FunctionAnalysisManager &FAM = 849 AM.getResult<FunctionAnalysisManagerCGSCCProxy>(InitialC, CG) 850 .getManager(); 851 852 // Populate the initial list of calls in this SCC. 853 for (auto &N : InitialC) { 854 auto &ORE = 855 FAM.getResult<OptimizationRemarkEmitterAnalysis>(N.getFunction()); 856 // We want to generally process call sites top-down in order for 857 // simplifications stemming from replacing the call with the returned value 858 // after inlining to be visible to subsequent inlining decisions. 859 // FIXME: Using instructions sequence is a really bad way to do this. 860 // Instead we should do an actual RPO walk of the function body. 861 for (Instruction &I : instructions(N.getFunction())) 862 if (auto CS = CallSite(&I)) 863 if (Function *Callee = CS.getCalledFunction()) { 864 if (!Callee->isDeclaration()) 865 Calls.push_back({CS, -1}); 866 else if (!isa<IntrinsicInst>(I)) { 867 using namespace ore; 868 ORE.emit([&]() { 869 return OptimizationRemarkMissed(DEBUG_TYPE, "NoDefinition", &I) 870 << NV("Callee", Callee) << " will not be inlined into " 871 << NV("Caller", CS.getCaller()) 872 << " because its definition is unavailable" 873 << setIsVerbose(); 874 }); 875 } 876 } 877 } 878 if (Calls.empty()) 879 return PreservedAnalyses::all(); 880 881 // Capture updatable variables for the current SCC and RefSCC. 882 auto *C = &InitialC; 883 auto *RC = &C->getOuterRefSCC(); 884 885 // When inlining a callee produces new call sites, we want to keep track of 886 // the fact that they were inlined from the callee. This allows us to avoid 887 // infinite inlining in some obscure cases. To represent this, we use an 888 // index into the InlineHistory vector. 889 SmallVector<std::pair<Function *, int>, 16> InlineHistory; 890 891 // Track a set vector of inlined callees so that we can augment the caller 892 // with all of their edges in the call graph before pruning out the ones that 893 // got simplified away. 894 SmallSetVector<Function *, 4> InlinedCallees; 895 896 // Track the dead functions to delete once finished with inlining calls. We 897 // defer deleting these to make it easier to handle the call graph updates. 898 SmallVector<Function *, 4> DeadFunctions; 899 900 // Loop forward over all of the calls. Note that we cannot cache the size as 901 // inlining can introduce new calls that need to be processed. 902 for (int i = 0; i < (int)Calls.size(); ++i) { 903 // We expect the calls to typically be batched with sequences of calls that 904 // have the same caller, so we first set up some shared infrastructure for 905 // this caller. We also do any pruning we can at this layer on the caller 906 // alone. 907 Function &F = *Calls[i].first.getCaller(); 908 LazyCallGraph::Node &N = *CG.lookup(F); 909 if (CG.lookupSCC(N) != C) 910 continue; 911 if (F.hasFnAttribute(Attribute::OptimizeNone)) 912 continue; 913 914 LLVM_DEBUG(dbgs() << "Inlining calls in: " << F.getName() << "\n"); 915 916 // Get a FunctionAnalysisManager via a proxy for this particular node. We 917 // do this each time we visit a node as the SCC may have changed and as 918 // we're going to mutate this particular function we want to make sure the 919 // proxy is in place to forward any invalidation events. We can use the 920 // manager we get here for looking up results for functions other than this 921 // node however because those functions aren't going to be mutated by this 922 // pass. 923 FunctionAnalysisManager &FAM = 924 AM.getResult<FunctionAnalysisManagerCGSCCProxy>(*C, CG) 925 .getManager(); 926 927 // Get the remarks emission analysis for the caller. 928 auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(F); 929 930 std::function<AssumptionCache &(Function &)> GetAssumptionCache = 931 [&](Function &F) -> AssumptionCache & { 932 return FAM.getResult<AssumptionAnalysis>(F); 933 }; 934 auto GetBFI = [&](Function &F) -> BlockFrequencyInfo & { 935 return FAM.getResult<BlockFrequencyAnalysis>(F); 936 }; 937 938 auto GetInlineCost = [&](CallSite CS) { 939 Function &Callee = *CS.getCalledFunction(); 940 auto &CalleeTTI = FAM.getResult<TargetIRAnalysis>(Callee); 941 return getInlineCost(CS, Params, CalleeTTI, GetAssumptionCache, {GetBFI}, 942 PSI, &ORE); 943 }; 944 945 // Now process as many calls as we have within this caller in the sequnece. 946 // We bail out as soon as the caller has to change so we can update the 947 // call graph and prepare the context of that new caller. 948 bool DidInline = false; 949 for (; i < (int)Calls.size() && Calls[i].first.getCaller() == &F; ++i) { 950 int InlineHistoryID; 951 CallSite CS; 952 std::tie(CS, InlineHistoryID) = Calls[i]; 953 Function &Callee = *CS.getCalledFunction(); 954 955 if (InlineHistoryID != -1 && 956 InlineHistoryIncludes(&Callee, InlineHistoryID, InlineHistory)) 957 continue; 958 959 // Check if this inlining may repeat breaking an SCC apart that has 960 // already been split once before. In that case, inlining here may 961 // trigger infinite inlining, much like is prevented within the inliner 962 // itself by the InlineHistory above, but spread across CGSCC iterations 963 // and thus hidden from the full inline history. 964 if (CG.lookupSCC(*CG.lookup(Callee)) == C && 965 UR.InlinedInternalEdges.count({&N, C})) { 966 LLVM_DEBUG(dbgs() << "Skipping inlining internal SCC edge from a node " 967 "previously split out of this SCC by inlining: " 968 << F.getName() << " -> " << Callee.getName() << "\n"); 969 continue; 970 } 971 972 Optional<InlineCost> OIC = shouldInline(CS, GetInlineCost, ORE); 973 // Check whether we want to inline this callsite. 974 if (!OIC) 975 continue; 976 977 // Setup the data structure used to plumb customization into the 978 // `InlineFunction` routine. 979 InlineFunctionInfo IFI( 980 /*cg=*/nullptr, &GetAssumptionCache, PSI, 981 &FAM.getResult<BlockFrequencyAnalysis>(*(CS.getCaller())), 982 &FAM.getResult<BlockFrequencyAnalysis>(Callee)); 983 984 // Get DebugLoc to report. CS will be invalid after Inliner. 985 DebugLoc DLoc = CS->getDebugLoc(); 986 BasicBlock *Block = CS.getParent(); 987 988 using namespace ore; 989 990 if (!InlineFunction(CS, IFI)) { 991 ORE.emit([&]() { 992 return OptimizationRemarkMissed(DEBUG_TYPE, "NotInlined", DLoc, Block) 993 << NV("Callee", &Callee) << " will not be inlined into " 994 << NV("Caller", &F); 995 }); 996 continue; 997 } 998 DidInline = true; 999 InlinedCallees.insert(&Callee); 1000 1001 ORE.emit([&]() { 1002 bool AlwaysInline = OIC->isAlways(); 1003 StringRef RemarkName = AlwaysInline ? "AlwaysInline" : "Inlined"; 1004 OptimizationRemark R(DEBUG_TYPE, RemarkName, DLoc, Block); 1005 R << NV("Callee", &Callee) << " inlined into "; 1006 R << NV("Caller", &F); 1007 if (AlwaysInline) 1008 R << " with cost=always"; 1009 else { 1010 R << " with cost=" << NV("Cost", OIC->getCost()); 1011 R << " (threshold=" << NV("Threshold", OIC->getThreshold()); 1012 R << ")"; 1013 } 1014 return R; 1015 }); 1016 1017 // Add any new callsites to defined functions to the worklist. 1018 if (!IFI.InlinedCallSites.empty()) { 1019 int NewHistoryID = InlineHistory.size(); 1020 InlineHistory.push_back({&Callee, InlineHistoryID}); 1021 for (CallSite &CS : reverse(IFI.InlinedCallSites)) 1022 if (Function *NewCallee = CS.getCalledFunction()) 1023 if (!NewCallee->isDeclaration()) 1024 Calls.push_back({CS, NewHistoryID}); 1025 } 1026 1027 if (InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No) 1028 ImportedFunctionsStats->recordInline(F, Callee); 1029 1030 // Merge the attributes based on the inlining. 1031 AttributeFuncs::mergeAttributesForInlining(F, Callee); 1032 1033 // For local functions, check whether this makes the callee trivially 1034 // dead. In that case, we can drop the body of the function eagerly 1035 // which may reduce the number of callers of other functions to one, 1036 // changing inline cost thresholds. 1037 if (Callee.hasLocalLinkage()) { 1038 // To check this we also need to nuke any dead constant uses (perhaps 1039 // made dead by this operation on other functions). 1040 Callee.removeDeadConstantUsers(); 1041 if (Callee.use_empty() && !CG.isLibFunction(Callee)) { 1042 Calls.erase( 1043 std::remove_if(Calls.begin() + i + 1, Calls.end(), 1044 [&Callee](const std::pair<CallSite, int> &Call) { 1045 return Call.first.getCaller() == &Callee; 1046 }), 1047 Calls.end()); 1048 // Clear the body and queue the function itself for deletion when we 1049 // finish inlining and call graph updates. 1050 // Note that after this point, it is an error to do anything other 1051 // than use the callee's address or delete it. 1052 Callee.dropAllReferences(); 1053 assert(find(DeadFunctions, &Callee) == DeadFunctions.end() && 1054 "Cannot put cause a function to become dead twice!"); 1055 DeadFunctions.push_back(&Callee); 1056 } 1057 } 1058 } 1059 1060 // Back the call index up by one to put us in a good position to go around 1061 // the outer loop. 1062 --i; 1063 1064 if (!DidInline) 1065 continue; 1066 Changed = true; 1067 1068 // Add all the inlined callees' edges as ref edges to the caller. These are 1069 // by definition trivial edges as we always have *some* transitive ref edge 1070 // chain. While in some cases these edges are direct calls inside the 1071 // callee, they have to be modeled in the inliner as reference edges as 1072 // there may be a reference edge anywhere along the chain from the current 1073 // caller to the callee that causes the whole thing to appear like 1074 // a (transitive) reference edge that will require promotion to a call edge 1075 // below. 1076 for (Function *InlinedCallee : InlinedCallees) { 1077 LazyCallGraph::Node &CalleeN = *CG.lookup(*InlinedCallee); 1078 for (LazyCallGraph::Edge &E : *CalleeN) 1079 RC->insertTrivialRefEdge(N, E.getNode()); 1080 } 1081 1082 // At this point, since we have made changes we have at least removed 1083 // a call instruction. However, in the process we do some incremental 1084 // simplification of the surrounding code. This simplification can 1085 // essentially do all of the same things as a function pass and we can 1086 // re-use the exact same logic for updating the call graph to reflect the 1087 // change. 1088 LazyCallGraph::SCC *OldC = C; 1089 C = &updateCGAndAnalysisManagerForFunctionPass(CG, *C, N, AM, UR); 1090 LLVM_DEBUG(dbgs() << "Updated inlining SCC: " << *C << "\n"); 1091 RC = &C->getOuterRefSCC(); 1092 1093 // If this causes an SCC to split apart into multiple smaller SCCs, there 1094 // is a subtle risk we need to prepare for. Other transformations may 1095 // expose an "infinite inlining" opportunity later, and because of the SCC 1096 // mutation, we will revisit this function and potentially re-inline. If we 1097 // do, and that re-inlining also has the potentially to mutate the SCC 1098 // structure, the infinite inlining problem can manifest through infinite 1099 // SCC splits and merges. To avoid this, we capture the originating caller 1100 // node and the SCC containing the call edge. This is a slight over 1101 // approximation of the possible inlining decisions that must be avoided, 1102 // but is relatively efficient to store. 1103 // FIXME: This seems like a very heavyweight way of retaining the inline 1104 // history, we should look for a more efficient way of tracking it. 1105 if (C != OldC && llvm::any_of(InlinedCallees, [&](Function *Callee) { 1106 return CG.lookupSCC(*CG.lookup(*Callee)) == OldC; 1107 })) { 1108 LLVM_DEBUG(dbgs() << "Inlined an internal call edge and split an SCC, " 1109 "retaining this to avoid infinite inlining.\n"); 1110 UR.InlinedInternalEdges.insert({&N, OldC}); 1111 } 1112 InlinedCallees.clear(); 1113 } 1114 1115 // Now that we've finished inlining all of the calls across this SCC, delete 1116 // all of the trivially dead functions, updating the call graph and the CGSCC 1117 // pass manager in the process. 1118 // 1119 // Note that this walks a pointer set which has non-deterministic order but 1120 // that is OK as all we do is delete things and add pointers to unordered 1121 // sets. 1122 for (Function *DeadF : DeadFunctions) { 1123 // Get the necessary information out of the call graph and nuke the 1124 // function there. Also, cclear out any cached analyses. 1125 auto &DeadC = *CG.lookupSCC(*CG.lookup(*DeadF)); 1126 FunctionAnalysisManager &FAM = 1127 AM.getResult<FunctionAnalysisManagerCGSCCProxy>(DeadC, CG) 1128 .getManager(); 1129 FAM.clear(*DeadF, DeadF->getName()); 1130 AM.clear(DeadC, DeadC.getName()); 1131 auto &DeadRC = DeadC.getOuterRefSCC(); 1132 CG.removeDeadFunction(*DeadF); 1133 1134 // Mark the relevant parts of the call graph as invalid so we don't visit 1135 // them. 1136 UR.InvalidatedSCCs.insert(&DeadC); 1137 UR.InvalidatedRefSCCs.insert(&DeadRC); 1138 1139 // And delete the actual function from the module. 1140 M.getFunctionList().erase(DeadF); 1141 } 1142 1143 if (!Changed) 1144 return PreservedAnalyses::all(); 1145 1146 // Even if we change the IR, we update the core CGSCC data structures and so 1147 // can preserve the proxy to the function analysis manager. 1148 PreservedAnalyses PA; 1149 PA.preserve<FunctionAnalysisManagerCGSCCProxy>(); 1150 return PA; 1151 } 1152