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/InlinerPass.h" 17 #include "llvm/ADT/SmallPtrSet.h" 18 #include "llvm/ADT/Statistic.h" 19 #include "llvm/Analysis/AliasAnalysis.h" 20 #include "llvm/Analysis/AssumptionCache.h" 21 #include "llvm/Analysis/CallGraph.h" 22 #include "llvm/Analysis/InlineCost.h" 23 #include "llvm/Analysis/TargetLibraryInfo.h" 24 #include "llvm/IR/CallSite.h" 25 #include "llvm/IR/DataLayout.h" 26 #include "llvm/IR/DiagnosticInfo.h" 27 #include "llvm/IR/Instructions.h" 28 #include "llvm/IR/IntrinsicInst.h" 29 #include "llvm/IR/Module.h" 30 #include "llvm/Support/CommandLine.h" 31 #include "llvm/Support/Debug.h" 32 #include "llvm/Support/raw_ostream.h" 33 #include "llvm/Transforms/Utils/Cloning.h" 34 #include "llvm/Transforms/Utils/Local.h" 35 using namespace llvm; 36 37 #define DEBUG_TYPE "inline" 38 39 STATISTIC(NumInlined, "Number of functions inlined"); 40 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined"); 41 STATISTIC(NumDeleted, "Number of functions deleted because all callers found"); 42 STATISTIC(NumMergedAllocas, "Number of allocas merged together"); 43 44 // This weirdly named statistic tracks the number of times that, when attempting 45 // to inline a function A into B, we analyze the callers of B in order to see 46 // if those would be more profitable and blocked inline steps. 47 STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed"); 48 49 static cl::opt<int> 50 InlineLimit("inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore, 51 cl::desc("Control the amount of inlining to perform (default = 225)")); 52 53 static cl::opt<int> 54 HintThreshold("inlinehint-threshold", cl::Hidden, cl::init(325), 55 cl::desc("Threshold for inlining functions with inline hint")); 56 57 // We instroduce this threshold to help performance of instrumentation based 58 // PGO before we actually hook up inliner with analysis passes such as BPI and 59 // BFI. 60 static cl::opt<int> 61 ColdThreshold("inlinecold-threshold", cl::Hidden, cl::init(225), 62 cl::desc("Threshold for inlining functions with cold attribute")); 63 64 // Threshold to use when optsize is specified (and there is no -inline-limit). 65 const int OptSizeThreshold = 75; 66 67 Inliner::Inliner(char &ID) 68 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit), InsertLifetime(true) {} 69 70 Inliner::Inliner(char &ID, int Threshold, bool InsertLifetime) 71 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit.getNumOccurrences() > 0 ? 72 InlineLimit : Threshold), 73 InsertLifetime(InsertLifetime) {} 74 75 /// For this class, we declare that we require and preserve the call graph. 76 /// If the derived class implements this method, it should 77 /// always explicitly call the implementation here. 78 void Inliner::getAnalysisUsage(AnalysisUsage &AU) const { 79 AU.addRequired<AliasAnalysis>(); 80 AU.addRequired<AssumptionCacheTracker>(); 81 CallGraphSCCPass::getAnalysisUsage(AU); 82 } 83 84 85 typedef DenseMap<ArrayType*, std::vector<AllocaInst*> > 86 InlinedArrayAllocasTy; 87 88 /// \brief If the inlined function had a higher stack protection level than the 89 /// calling function, then bump up the caller's stack protection level. 90 static void AdjustCallerSSPLevel(Function *Caller, Function *Callee) { 91 // If upgrading the SSP attribute, clear out the old SSP Attributes first. 92 // Having multiple SSP attributes doesn't actually hurt, but it adds useless 93 // clutter to the IR. 94 AttrBuilder B; 95 B.addAttribute(Attribute::StackProtect) 96 .addAttribute(Attribute::StackProtectStrong); 97 AttributeSet OldSSPAttr = AttributeSet::get(Caller->getContext(), 98 AttributeSet::FunctionIndex, 99 B); 100 101 if (Callee->hasFnAttribute(Attribute::StackProtectReq)) { 102 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr); 103 Caller->addFnAttr(Attribute::StackProtectReq); 104 } else if (Callee->hasFnAttribute(Attribute::StackProtectStrong) && 105 !Caller->hasFnAttribute(Attribute::StackProtectReq)) { 106 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr); 107 Caller->addFnAttr(Attribute::StackProtectStrong); 108 } else if (Callee->hasFnAttribute(Attribute::StackProtect) && 109 !Caller->hasFnAttribute(Attribute::StackProtectReq) && 110 !Caller->hasFnAttribute(Attribute::StackProtectStrong)) 111 Caller->addFnAttr(Attribute::StackProtect); 112 } 113 114 /// If it is possible to inline the specified call site, 115 /// do so and update the CallGraph for this operation. 116 /// 117 /// This function also does some basic book-keeping to update the IR. The 118 /// InlinedArrayAllocas map keeps track of any allocas that are already 119 /// available from other functions inlined into the caller. If we are able to 120 /// inline this call site we attempt to reuse already available allocas or add 121 /// any new allocas to the set if not possible. 122 static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI, 123 InlinedArrayAllocasTy &InlinedArrayAllocas, 124 int InlineHistory, bool InsertLifetime) { 125 Function *Callee = CS.getCalledFunction(); 126 Function *Caller = CS.getCaller(); 127 128 // Try to inline the function. Get the list of static allocas that were 129 // inlined. 130 if (!InlineFunction(CS, IFI, InsertLifetime)) 131 return false; 132 133 AdjustCallerSSPLevel(Caller, Callee); 134 135 // Look at all of the allocas that we inlined through this call site. If we 136 // have already inlined other allocas through other calls into this function, 137 // then we know that they have disjoint lifetimes and that we can merge them. 138 // 139 // There are many heuristics possible for merging these allocas, and the 140 // different options have different tradeoffs. One thing that we *really* 141 // don't want to hurt is SRoA: once inlining happens, often allocas are no 142 // longer address taken and so they can be promoted. 143 // 144 // Our "solution" for that is to only merge allocas whose outermost type is an 145 // array type. These are usually not promoted because someone is using a 146 // variable index into them. These are also often the most important ones to 147 // merge. 148 // 149 // A better solution would be to have real memory lifetime markers in the IR 150 // and not have the inliner do any merging of allocas at all. This would 151 // allow the backend to do proper stack slot coloring of all allocas that 152 // *actually make it to the backend*, which is really what we want. 153 // 154 // Because we don't have this information, we do this simple and useful hack. 155 // 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 true; 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(); 176 AllocaNo != e; ++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 (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) { 196 AllocaInst *AvailableAlloca = AllocasForType[i]; 197 198 unsigned Align1 = AI->getAlignment(), 199 Align2 = AvailableAlloca->getAlignment(); 200 201 // The available alloca has to be in the right function, not in some other 202 // function in this SCC. 203 if (AvailableAlloca->getParent() != AI->getParent()) 204 continue; 205 206 // If the inlined function already uses this alloca then we can't reuse 207 // it. 208 if (!UsedAllocas.insert(AvailableAlloca).second) 209 continue; 210 211 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare 212 // success! 213 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: " 214 << *AvailableAlloca << '\n'); 215 216 AI->replaceAllUsesWith(AvailableAlloca); 217 218 if (Align1 != Align2) { 219 if (!Align1 || !Align2) { 220 const DataLayout &DL = Caller->getParent()->getDataLayout(); 221 unsigned TypeAlign = DL.getABITypeAlignment(AI->getAllocatedType()); 222 223 Align1 = Align1 ? Align1 : TypeAlign; 224 Align2 = Align2 ? Align2 : TypeAlign; 225 } 226 227 if (Align1 > Align2) 228 AvailableAlloca->setAlignment(AI->getAlignment()); 229 } 230 231 AI->eraseFromParent(); 232 MergedAwayAlloca = true; 233 ++NumMergedAllocas; 234 IFI.StaticAllocas[AllocaNo] = nullptr; 235 break; 236 } 237 238 // If we already nuked the alloca, we're done with it. 239 if (MergedAwayAlloca) 240 continue; 241 242 // If we were unable to merge away the alloca either because there are no 243 // allocas of the right type available or because we reused them all 244 // already, remember that this alloca came from an inlined function and mark 245 // it used so we don't reuse it for other allocas from this inline 246 // operation. 247 AllocasForType.push_back(AI); 248 UsedAllocas.insert(AI); 249 } 250 251 return true; 252 } 253 254 unsigned Inliner::getInlineThreshold(CallSite CS) const { 255 int thres = InlineThreshold; // -inline-threshold or else selected by 256 // overall opt level 257 258 // If -inline-threshold is not given, listen to the optsize attribute when it 259 // would decrease the threshold. 260 Function *Caller = CS.getCaller(); 261 bool OptSize = Caller && !Caller->isDeclaration() && 262 Caller->hasFnAttribute(Attribute::OptimizeForSize); 263 if (!(InlineLimit.getNumOccurrences() > 0) && OptSize && 264 OptSizeThreshold < thres) 265 thres = OptSizeThreshold; 266 267 // Listen to the inlinehint attribute when it would increase the threshold 268 // and the caller does not need to minimize its size. 269 Function *Callee = CS.getCalledFunction(); 270 bool InlineHint = Callee && !Callee->isDeclaration() && 271 Callee->hasFnAttribute(Attribute::InlineHint); 272 if (InlineHint && HintThreshold > thres && 273 !Caller->hasFnAttribute(Attribute::MinSize)) 274 thres = HintThreshold; 275 276 // Listen to the cold attribute when it would decrease the threshold. 277 bool ColdCallee = Callee && !Callee->isDeclaration() && 278 Callee->hasFnAttribute(Attribute::Cold); 279 // Command line argument for InlineLimit will override the default 280 // ColdThreshold. If we have -inline-threshold but no -inlinecold-threshold, 281 // do not use the default cold threshold even if it is smaller. 282 if ((InlineLimit.getNumOccurrences() == 0 || 283 ColdThreshold.getNumOccurrences() > 0) && ColdCallee && 284 ColdThreshold < thres) 285 thres = ColdThreshold; 286 287 return thres; 288 } 289 290 static void emitAnalysis(CallSite CS, const Twine &Msg) { 291 Function *Caller = CS.getCaller(); 292 LLVMContext &Ctx = Caller->getContext(); 293 DebugLoc DLoc = CS.getInstruction()->getDebugLoc(); 294 emitOptimizationRemarkAnalysis(Ctx, DEBUG_TYPE, *Caller, DLoc, Msg); 295 } 296 297 /// Return true if the inliner should attempt to inline at the given CallSite. 298 bool Inliner::shouldInline(CallSite CS) { 299 InlineCost IC = getInlineCost(CS); 300 301 if (IC.isAlways()) { 302 DEBUG(dbgs() << " Inlining: cost=always" 303 << ", Call: " << *CS.getInstruction() << "\n"); 304 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName()) + 305 " should always be inlined (cost=always)"); 306 return true; 307 } 308 309 if (IC.isNever()) { 310 DEBUG(dbgs() << " NOT Inlining: cost=never" 311 << ", Call: " << *CS.getInstruction() << "\n"); 312 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() + 313 " should never be inlined (cost=never)")); 314 return false; 315 } 316 317 Function *Caller = CS.getCaller(); 318 if (!IC) { 319 DEBUG(dbgs() << " NOT Inlining: cost=" << IC.getCost() 320 << ", thres=" << (IC.getCostDelta() + IC.getCost()) 321 << ", Call: " << *CS.getInstruction() << "\n"); 322 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() + 323 " too costly to inline (cost=") + 324 Twine(IC.getCost()) + ", threshold=" + 325 Twine(IC.getCostDelta() + IC.getCost()) + ")"); 326 return false; 327 } 328 329 // Try to detect the case where the current inlining candidate caller (call 330 // it B) is a static or linkonce-ODR function and is an inlining candidate 331 // elsewhere, and the current candidate callee (call it C) is large enough 332 // that inlining it into B would make B too big to inline later. In these 333 // circumstances it may be best not to inline C into B, but to inline B into 334 // its callers. 335 // 336 // This only applies to static and linkonce-ODR functions because those are 337 // expected to be available for inlining in the translation units where they 338 // are used. Thus we will always have the opportunity to make local inlining 339 // decisions. Importantly the linkonce-ODR linkage covers inline functions 340 // and templates in C++. 341 // 342 // FIXME: All of this logic should be sunk into getInlineCost. It relies on 343 // the internal implementation of the inline cost metrics rather than 344 // treating them as truly abstract units etc. 345 if (Caller->hasLocalLinkage() || Caller->hasLinkOnceODRLinkage()) { 346 int TotalSecondaryCost = 0; 347 // The candidate cost to be imposed upon the current function. 348 int CandidateCost = IC.getCost() - (InlineConstants::CallPenalty + 1); 349 // This bool tracks what happens if we do NOT inline C into B. 350 bool callerWillBeRemoved = Caller->hasLocalLinkage(); 351 // This bool tracks what happens if we DO inline C into B. 352 bool inliningPreventsSomeOuterInline = false; 353 for (User *U : Caller->users()) { 354 CallSite CS2(U); 355 356 // If this isn't a call to Caller (it could be some other sort 357 // of reference) skip it. Such references will prevent the caller 358 // from being removed. 359 if (!CS2 || CS2.getCalledFunction() != Caller) { 360 callerWillBeRemoved = false; 361 continue; 362 } 363 364 InlineCost IC2 = getInlineCost(CS2); 365 ++NumCallerCallersAnalyzed; 366 if (!IC2) { 367 callerWillBeRemoved = false; 368 continue; 369 } 370 if (IC2.isAlways()) 371 continue; 372 373 // See if inlining or original callsite would erase the cost delta of 374 // this callsite. We subtract off the penalty for the call instruction, 375 // which we would be deleting. 376 if (IC2.getCostDelta() <= CandidateCost) { 377 inliningPreventsSomeOuterInline = true; 378 TotalSecondaryCost += IC2.getCost(); 379 } 380 } 381 // If all outer calls to Caller would get inlined, the cost for the last 382 // one is set very low by getInlineCost, in anticipation that Caller will 383 // be removed entirely. We did not account for this above unless there 384 // is only one caller of Caller. 385 if (callerWillBeRemoved && !Caller->use_empty()) 386 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus; 387 388 if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost()) { 389 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() << 390 " Cost = " << IC.getCost() << 391 ", outer Cost = " << TotalSecondaryCost << '\n'); 392 emitAnalysis( 393 CS, Twine("Not inlining. Cost of inlining " + 394 CS.getCalledFunction()->getName() + 395 " increases the cost of inlining " + 396 CS.getCaller()->getName() + " in other contexts")); 397 return false; 398 } 399 } 400 401 DEBUG(dbgs() << " Inlining: cost=" << IC.getCost() 402 << ", thres=" << (IC.getCostDelta() + IC.getCost()) 403 << ", Call: " << *CS.getInstruction() << '\n'); 404 emitAnalysis( 405 CS, CS.getCalledFunction()->getName() + Twine(" can be inlined into ") + 406 CS.getCaller()->getName() + " with cost=" + Twine(IC.getCost()) + 407 " (threshold=" + Twine(IC.getCostDelta() + IC.getCost()) + ")"); 408 return true; 409 } 410 411 /// Return true if the specified inline history ID 412 /// indicates an inline history that includes the specified function. 413 static bool InlineHistoryIncludes(Function *F, int InlineHistoryID, 414 const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) { 415 while (InlineHistoryID != -1) { 416 assert(unsigned(InlineHistoryID) < InlineHistory.size() && 417 "Invalid inline history ID"); 418 if (InlineHistory[InlineHistoryID].first == F) 419 return true; 420 InlineHistoryID = InlineHistory[InlineHistoryID].second; 421 } 422 return false; 423 } 424 425 bool Inliner::runOnSCC(CallGraphSCC &SCC) { 426 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph(); 427 AssumptionCacheTracker *ACT = &getAnalysis<AssumptionCacheTracker>(); 428 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); 429 const TargetLibraryInfo *TLI = TLIP ? &TLIP->getTLI() : nullptr; 430 AliasAnalysis *AA = &getAnalysis<AliasAnalysis>(); 431 432 SmallPtrSet<Function*, 8> SCCFunctions; 433 DEBUG(dbgs() << "Inliner visiting SCC:"); 434 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) { 435 Function *F = (*I)->getFunction(); 436 if (F) SCCFunctions.insert(F); 437 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE")); 438 } 439 440 // Scan through and identify all call sites ahead of time so that we only 441 // inline call sites in the original functions, not call sites that result 442 // from inlining other functions. 443 SmallVector<std::pair<CallSite, int>, 16> CallSites; 444 445 // When inlining a callee produces new call sites, we want to keep track of 446 // the fact that they were inlined from the callee. This allows us to avoid 447 // infinite inlining in some obscure cases. To represent this, we use an 448 // index into the InlineHistory vector. 449 SmallVector<std::pair<Function*, int>, 8> InlineHistory; 450 451 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) { 452 Function *F = (*I)->getFunction(); 453 if (!F) continue; 454 455 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 456 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 457 CallSite CS(cast<Value>(I)); 458 // If this isn't a call, or it is a call to an intrinsic, it can 459 // never be inlined. 460 if (!CS || isa<IntrinsicInst>(I)) 461 continue; 462 463 // If this is a direct call to an external function, we can never inline 464 // it. If it is an indirect call, inlining may resolve it to be a 465 // direct call, so we keep it. 466 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration()) 467 continue; 468 469 CallSites.push_back(std::make_pair(CS, -1)); 470 } 471 } 472 473 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n"); 474 475 // If there are no calls in this function, exit early. 476 if (CallSites.empty()) 477 return false; 478 479 // Now that we have all of the call sites, move the ones to functions in the 480 // current SCC to the end of the list. 481 unsigned FirstCallInSCC = CallSites.size(); 482 for (unsigned i = 0; i < FirstCallInSCC; ++i) 483 if (Function *F = CallSites[i].first.getCalledFunction()) 484 if (SCCFunctions.count(F)) 485 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]); 486 487 488 InlinedArrayAllocasTy InlinedArrayAllocas; 489 InlineFunctionInfo InlineInfo(&CG, AA, ACT); 490 491 // Now that we have all of the call sites, loop over them and inline them if 492 // it looks profitable to do so. 493 bool Changed = false; 494 bool LocalChange; 495 do { 496 LocalChange = false; 497 // Iterate over the outer loop because inlining functions can cause indirect 498 // calls to become direct calls. 499 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) { 500 CallSite CS = CallSites[CSi].first; 501 502 Function *Caller = CS.getCaller(); 503 Function *Callee = CS.getCalledFunction(); 504 505 // If this call site is dead and it is to a readonly function, we should 506 // just delete the call instead of trying to inline it, regardless of 507 // size. This happens because IPSCCP propagates the result out of the 508 // call and then we're left with the dead call. 509 if (isInstructionTriviallyDead(CS.getInstruction(), TLI)) { 510 DEBUG(dbgs() << " -> Deleting dead call: " 511 << *CS.getInstruction() << "\n"); 512 // Update the call graph by deleting the edge from Callee to Caller. 513 CG[Caller]->removeCallEdgeFor(CS); 514 CS.getInstruction()->eraseFromParent(); 515 ++NumCallsDeleted; 516 } else { 517 // We can only inline direct calls to non-declarations. 518 if (!Callee || Callee->isDeclaration()) continue; 519 520 // If this call site was obtained by inlining another function, verify 521 // that the include path for the function did not include the callee 522 // itself. If so, we'd be recursively inlining the same function, 523 // which would provide the same callsites, which would cause us to 524 // infinitely inline. 525 int InlineHistoryID = CallSites[CSi].second; 526 if (InlineHistoryID != -1 && 527 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory)) 528 continue; 529 530 LLVMContext &CallerCtx = Caller->getContext(); 531 532 // Get DebugLoc to report. CS will be invalid after Inliner. 533 DebugLoc DLoc = CS.getInstruction()->getDebugLoc(); 534 535 // If the policy determines that we should inline this function, 536 // try to do so. 537 if (!shouldInline(CS)) { 538 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc, 539 Twine(Callee->getName() + 540 " will not be inlined into " + 541 Caller->getName())); 542 continue; 543 } 544 545 // Attempt to inline the function. 546 if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas, 547 InlineHistoryID, InsertLifetime)) { 548 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc, 549 Twine(Callee->getName() + 550 " will not be inlined into " + 551 Caller->getName())); 552 continue; 553 } 554 ++NumInlined; 555 556 // Report the inline decision. 557 emitOptimizationRemark( 558 CallerCtx, DEBUG_TYPE, *Caller, DLoc, 559 Twine(Callee->getName() + " inlined into " + Caller->getName())); 560 561 // If inlining this function gave us any new call sites, throw them 562 // onto our worklist to process. They are useful inline candidates. 563 if (!InlineInfo.InlinedCalls.empty()) { 564 // Create a new inline history entry for this, so that we remember 565 // that these new callsites came about due to inlining Callee. 566 int NewHistoryID = InlineHistory.size(); 567 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID)); 568 569 for (unsigned i = 0, e = InlineInfo.InlinedCalls.size(); 570 i != e; ++i) { 571 Value *Ptr = InlineInfo.InlinedCalls[i]; 572 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID)); 573 } 574 } 575 } 576 577 // If we inlined or deleted the last possible call site to the function, 578 // delete the function body now. 579 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() && 580 // TODO: Can remove if in SCC now. 581 !SCCFunctions.count(Callee) && 582 583 // The function may be apparently dead, but if there are indirect 584 // callgraph references to the node, we cannot delete it yet, this 585 // could invalidate the CGSCC iterator. 586 CG[Callee]->getNumReferences() == 0) { 587 DEBUG(dbgs() << " -> Deleting dead function: " 588 << Callee->getName() << "\n"); 589 CallGraphNode *CalleeNode = CG[Callee]; 590 591 // Remove any call graph edges from the callee to its callees. 592 CalleeNode->removeAllCalledFunctions(); 593 594 // Removing the node for callee from the call graph and delete it. 595 delete CG.removeFunctionFromModule(CalleeNode); 596 ++NumDeleted; 597 } 598 599 // Remove this call site from the list. If possible, use 600 // swap/pop_back for efficiency, but do not use it if doing so would 601 // move a call site to a function in this SCC before the 602 // 'FirstCallInSCC' barrier. 603 if (SCC.isSingular()) { 604 CallSites[CSi] = CallSites.back(); 605 CallSites.pop_back(); 606 } else { 607 CallSites.erase(CallSites.begin()+CSi); 608 } 609 --CSi; 610 611 Changed = true; 612 LocalChange = true; 613 } 614 } while (LocalChange); 615 616 return Changed; 617 } 618 619 /// Remove now-dead linkonce functions at the end of 620 /// processing to avoid breaking the SCC traversal. 621 bool Inliner::doFinalization(CallGraph &CG) { 622 return removeDeadFunctions(CG); 623 } 624 625 /// Remove dead functions that are not included in DNR (Do Not Remove) list. 626 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) { 627 SmallVector<CallGraphNode*, 16> FunctionsToRemove; 628 629 // Scan for all of the functions, looking for ones that should now be removed 630 // from the program. Insert the dead ones in the FunctionsToRemove set. 631 for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) { 632 CallGraphNode *CGN = I->second; 633 Function *F = CGN->getFunction(); 634 if (!F || F->isDeclaration()) 635 continue; 636 637 // Handle the case when this function is called and we only want to care 638 // about always-inline functions. This is a bit of a hack to share code 639 // between here and the InlineAlways pass. 640 if (AlwaysInlineOnly && !F->hasFnAttribute(Attribute::AlwaysInline)) 641 continue; 642 643 // If the only remaining users of the function are dead constants, remove 644 // them. 645 F->removeDeadConstantUsers(); 646 647 if (!F->isDefTriviallyDead()) 648 continue; 649 650 // It is unsafe to drop a function with discardable linkage from a COMDAT 651 // without also dropping the other members of the COMDAT. 652 // The inliner doesn't visit non-function entities which are in COMDAT 653 // groups so it is unsafe to do so *unless* the linkage is local. 654 if (!F->hasLocalLinkage() && F->hasComdat()) 655 continue; 656 657 // Remove any call graph edges from the function to its callees. 658 CGN->removeAllCalledFunctions(); 659 660 // Remove any edges from the external node to the function's call graph 661 // node. These edges might have been made irrelegant due to 662 // optimization of the program. 663 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN); 664 665 // Removing the node for callee from the call graph and delete it. 666 FunctionsToRemove.push_back(CGN); 667 } 668 if (FunctionsToRemove.empty()) 669 return false; 670 671 // Now that we know which functions to delete, do so. We didn't want to do 672 // this inline, because that would invalidate our CallGraph::iterator 673 // objects. :( 674 // 675 // Note that it doesn't matter that we are iterating over a non-stable order 676 // here to do this, it doesn't matter which order the functions are deleted 677 // in. 678 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end()); 679 FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(), 680 FunctionsToRemove.end()), 681 FunctionsToRemove.end()); 682 for (SmallVectorImpl<CallGraphNode *>::iterator I = FunctionsToRemove.begin(), 683 E = FunctionsToRemove.end(); 684 I != E; ++I) { 685 delete CG.removeFunctionFromModule(*I); 686 ++NumDeleted; 687 } 688 return true; 689 } 690