1 //===- InlineFunction.cpp - Code to perform function inlining -------------===// 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 inlining of a function into a call site, resolving 11 // parameters and the return value as appropriate. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/Transforms/Utils/Cloning.h" 16 #include "llvm/ADT/SetVector.h" 17 #include "llvm/ADT/SmallPtrSet.h" 18 #include "llvm/ADT/SmallSet.h" 19 #include "llvm/ADT/SmallVector.h" 20 #include "llvm/ADT/StringExtras.h" 21 #include "llvm/Analysis/AliasAnalysis.h" 22 #include "llvm/Analysis/AssumptionCache.h" 23 #include "llvm/Analysis/BlockFrequencyInfo.h" 24 #include "llvm/Analysis/CallGraph.h" 25 #include "llvm/Analysis/CaptureTracking.h" 26 #include "llvm/Analysis/EHPersonalities.h" 27 #include "llvm/Analysis/InstructionSimplify.h" 28 #include "llvm/Analysis/ProfileSummaryInfo.h" 29 #include "llvm/Analysis/ValueTracking.h" 30 #include "llvm/IR/Attributes.h" 31 #include "llvm/IR/CallSite.h" 32 #include "llvm/IR/CFG.h" 33 #include "llvm/IR/Constants.h" 34 #include "llvm/IR/DataLayout.h" 35 #include "llvm/IR/DebugInfo.h" 36 #include "llvm/IR/DerivedTypes.h" 37 #include "llvm/IR/DIBuilder.h" 38 #include "llvm/IR/Dominators.h" 39 #include "llvm/IR/IRBuilder.h" 40 #include "llvm/IR/Instructions.h" 41 #include "llvm/IR/IntrinsicInst.h" 42 #include "llvm/IR/Intrinsics.h" 43 #include "llvm/IR/MDBuilder.h" 44 #include "llvm/IR/Module.h" 45 #include "llvm/Support/CommandLine.h" 46 #include "llvm/Transforms/Utils/Local.h" 47 #include <algorithm> 48 49 using namespace llvm; 50 51 static cl::opt<bool> 52 EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(true), 53 cl::Hidden, 54 cl::desc("Convert noalias attributes to metadata during inlining.")); 55 56 static cl::opt<bool> 57 PreserveAlignmentAssumptions("preserve-alignment-assumptions-during-inlining", 58 cl::init(true), cl::Hidden, 59 cl::desc("Convert align attributes to assumptions during inlining.")); 60 61 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI, 62 AAResults *CalleeAAR, bool InsertLifetime) { 63 return InlineFunction(CallSite(CI), IFI, CalleeAAR, InsertLifetime); 64 } 65 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI, 66 AAResults *CalleeAAR, bool InsertLifetime) { 67 return InlineFunction(CallSite(II), IFI, CalleeAAR, InsertLifetime); 68 } 69 70 namespace { 71 /// A class for recording information about inlining a landing pad. 72 class LandingPadInliningInfo { 73 BasicBlock *OuterResumeDest; ///< Destination of the invoke's unwind. 74 BasicBlock *InnerResumeDest; ///< Destination for the callee's resume. 75 LandingPadInst *CallerLPad; ///< LandingPadInst associated with the invoke. 76 PHINode *InnerEHValuesPHI; ///< PHI for EH values from landingpad insts. 77 SmallVector<Value*, 8> UnwindDestPHIValues; 78 79 public: 80 LandingPadInliningInfo(InvokeInst *II) 81 : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(nullptr), 82 CallerLPad(nullptr), InnerEHValuesPHI(nullptr) { 83 // If there are PHI nodes in the unwind destination block, we need to keep 84 // track of which values came into them from the invoke before removing 85 // the edge from this block. 86 llvm::BasicBlock *InvokeBB = II->getParent(); 87 BasicBlock::iterator I = OuterResumeDest->begin(); 88 for (; isa<PHINode>(I); ++I) { 89 // Save the value to use for this edge. 90 PHINode *PHI = cast<PHINode>(I); 91 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB)); 92 } 93 94 CallerLPad = cast<LandingPadInst>(I); 95 } 96 97 /// The outer unwind destination is the target of 98 /// unwind edges introduced for calls within the inlined function. 99 BasicBlock *getOuterResumeDest() const { 100 return OuterResumeDest; 101 } 102 103 BasicBlock *getInnerResumeDest(); 104 105 LandingPadInst *getLandingPadInst() const { return CallerLPad; } 106 107 /// Forward the 'resume' instruction to the caller's landing pad block. 108 /// When the landing pad block has only one predecessor, this is 109 /// a simple branch. When there is more than one predecessor, we need to 110 /// split the landing pad block after the landingpad instruction and jump 111 /// to there. 112 void forwardResume(ResumeInst *RI, 113 SmallPtrSetImpl<LandingPadInst*> &InlinedLPads); 114 115 /// Add incoming-PHI values to the unwind destination block for the given 116 /// basic block, using the values for the original invoke's source block. 117 void addIncomingPHIValuesFor(BasicBlock *BB) const { 118 addIncomingPHIValuesForInto(BB, OuterResumeDest); 119 } 120 121 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const { 122 BasicBlock::iterator I = dest->begin(); 123 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 124 PHINode *phi = cast<PHINode>(I); 125 phi->addIncoming(UnwindDestPHIValues[i], src); 126 } 127 } 128 }; 129 } // anonymous namespace 130 131 /// Get or create a target for the branch from ResumeInsts. 132 BasicBlock *LandingPadInliningInfo::getInnerResumeDest() { 133 if (InnerResumeDest) return InnerResumeDest; 134 135 // Split the landing pad. 136 BasicBlock::iterator SplitPoint = ++CallerLPad->getIterator(); 137 InnerResumeDest = 138 OuterResumeDest->splitBasicBlock(SplitPoint, 139 OuterResumeDest->getName() + ".body"); 140 141 // The number of incoming edges we expect to the inner landing pad. 142 const unsigned PHICapacity = 2; 143 144 // Create corresponding new PHIs for all the PHIs in the outer landing pad. 145 Instruction *InsertPoint = &InnerResumeDest->front(); 146 BasicBlock::iterator I = OuterResumeDest->begin(); 147 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 148 PHINode *OuterPHI = cast<PHINode>(I); 149 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity, 150 OuterPHI->getName() + ".lpad-body", 151 InsertPoint); 152 OuterPHI->replaceAllUsesWith(InnerPHI); 153 InnerPHI->addIncoming(OuterPHI, OuterResumeDest); 154 } 155 156 // Create a PHI for the exception values. 157 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity, 158 "eh.lpad-body", InsertPoint); 159 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI); 160 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest); 161 162 // All done. 163 return InnerResumeDest; 164 } 165 166 /// Forward the 'resume' instruction to the caller's landing pad block. 167 /// When the landing pad block has only one predecessor, this is a simple 168 /// branch. When there is more than one predecessor, we need to split the 169 /// landing pad block after the landingpad instruction and jump to there. 170 void LandingPadInliningInfo::forwardResume( 171 ResumeInst *RI, SmallPtrSetImpl<LandingPadInst *> &InlinedLPads) { 172 BasicBlock *Dest = getInnerResumeDest(); 173 BasicBlock *Src = RI->getParent(); 174 175 BranchInst::Create(Dest, Src); 176 177 // Update the PHIs in the destination. They were inserted in an order which 178 // makes this work. 179 addIncomingPHIValuesForInto(Src, Dest); 180 181 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src); 182 RI->eraseFromParent(); 183 } 184 185 /// Helper for getUnwindDestToken/getUnwindDestTokenHelper. 186 static Value *getParentPad(Value *EHPad) { 187 if (auto *FPI = dyn_cast<FuncletPadInst>(EHPad)) 188 return FPI->getParentPad(); 189 return cast<CatchSwitchInst>(EHPad)->getParentPad(); 190 } 191 192 typedef DenseMap<Instruction *, Value *> UnwindDestMemoTy; 193 194 /// Helper for getUnwindDestToken that does the descendant-ward part of 195 /// the search. 196 static Value *getUnwindDestTokenHelper(Instruction *EHPad, 197 UnwindDestMemoTy &MemoMap) { 198 SmallVector<Instruction *, 8> Worklist(1, EHPad); 199 200 while (!Worklist.empty()) { 201 Instruction *CurrentPad = Worklist.pop_back_val(); 202 // We only put pads on the worklist that aren't in the MemoMap. When 203 // we find an unwind dest for a pad we may update its ancestors, but 204 // the queue only ever contains uncles/great-uncles/etc. of CurrentPad, 205 // so they should never get updated while queued on the worklist. 206 assert(!MemoMap.count(CurrentPad)); 207 Value *UnwindDestToken = nullptr; 208 if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(CurrentPad)) { 209 if (CatchSwitch->hasUnwindDest()) { 210 UnwindDestToken = CatchSwitch->getUnwindDest()->getFirstNonPHI(); 211 } else { 212 // Catchswitch doesn't have a 'nounwind' variant, and one might be 213 // annotated as "unwinds to caller" when really it's nounwind (see 214 // e.g. SimplifyCFGOpt::SimplifyUnreachable), so we can't infer the 215 // parent's unwind dest from this. We can check its catchpads' 216 // descendants, since they might include a cleanuppad with an 217 // "unwinds to caller" cleanupret, which can be trusted. 218 for (auto HI = CatchSwitch->handler_begin(), 219 HE = CatchSwitch->handler_end(); 220 HI != HE && !UnwindDestToken; ++HI) { 221 BasicBlock *HandlerBlock = *HI; 222 auto *CatchPad = cast<CatchPadInst>(HandlerBlock->getFirstNonPHI()); 223 for (User *Child : CatchPad->users()) { 224 // Intentionally ignore invokes here -- since the catchswitch is 225 // marked "unwind to caller", it would be a verifier error if it 226 // contained an invoke which unwinds out of it, so any invoke we'd 227 // encounter must unwind to some child of the catch. 228 if (!isa<CleanupPadInst>(Child) && !isa<CatchSwitchInst>(Child)) 229 continue; 230 231 Instruction *ChildPad = cast<Instruction>(Child); 232 auto Memo = MemoMap.find(ChildPad); 233 if (Memo == MemoMap.end()) { 234 // Haven't figured out this child pad yet; queue it. 235 Worklist.push_back(ChildPad); 236 continue; 237 } 238 // We've already checked this child, but might have found that 239 // it offers no proof either way. 240 Value *ChildUnwindDestToken = Memo->second; 241 if (!ChildUnwindDestToken) 242 continue; 243 // We already know the child's unwind dest, which can either 244 // be ConstantTokenNone to indicate unwind to caller, or can 245 // be another child of the catchpad. Only the former indicates 246 // the unwind dest of the catchswitch. 247 if (isa<ConstantTokenNone>(ChildUnwindDestToken)) { 248 UnwindDestToken = ChildUnwindDestToken; 249 break; 250 } 251 assert(getParentPad(ChildUnwindDestToken) == CatchPad); 252 } 253 } 254 } 255 } else { 256 auto *CleanupPad = cast<CleanupPadInst>(CurrentPad); 257 for (User *U : CleanupPad->users()) { 258 if (auto *CleanupRet = dyn_cast<CleanupReturnInst>(U)) { 259 if (BasicBlock *RetUnwindDest = CleanupRet->getUnwindDest()) 260 UnwindDestToken = RetUnwindDest->getFirstNonPHI(); 261 else 262 UnwindDestToken = ConstantTokenNone::get(CleanupPad->getContext()); 263 break; 264 } 265 Value *ChildUnwindDestToken; 266 if (auto *Invoke = dyn_cast<InvokeInst>(U)) { 267 ChildUnwindDestToken = Invoke->getUnwindDest()->getFirstNonPHI(); 268 } else if (isa<CleanupPadInst>(U) || isa<CatchSwitchInst>(U)) { 269 Instruction *ChildPad = cast<Instruction>(U); 270 auto Memo = MemoMap.find(ChildPad); 271 if (Memo == MemoMap.end()) { 272 // Haven't resolved this child yet; queue it and keep searching. 273 Worklist.push_back(ChildPad); 274 continue; 275 } 276 // We've checked this child, but still need to ignore it if it 277 // had no proof either way. 278 ChildUnwindDestToken = Memo->second; 279 if (!ChildUnwindDestToken) 280 continue; 281 } else { 282 // Not a relevant user of the cleanuppad 283 continue; 284 } 285 // In a well-formed program, the child/invoke must either unwind to 286 // an(other) child of the cleanup, or exit the cleanup. In the 287 // first case, continue searching. 288 if (isa<Instruction>(ChildUnwindDestToken) && 289 getParentPad(ChildUnwindDestToken) == CleanupPad) 290 continue; 291 UnwindDestToken = ChildUnwindDestToken; 292 break; 293 } 294 } 295 // If we haven't found an unwind dest for CurrentPad, we may have queued its 296 // children, so move on to the next in the worklist. 297 if (!UnwindDestToken) 298 continue; 299 300 // Now we know that CurrentPad unwinds to UnwindDestToken. It also exits 301 // any ancestors of CurrentPad up to but not including UnwindDestToken's 302 // parent pad. Record this in the memo map, and check to see if the 303 // original EHPad being queried is one of the ones exited. 304 Value *UnwindParent; 305 if (auto *UnwindPad = dyn_cast<Instruction>(UnwindDestToken)) 306 UnwindParent = getParentPad(UnwindPad); 307 else 308 UnwindParent = nullptr; 309 bool ExitedOriginalPad = false; 310 for (Instruction *ExitedPad = CurrentPad; 311 ExitedPad && ExitedPad != UnwindParent; 312 ExitedPad = dyn_cast<Instruction>(getParentPad(ExitedPad))) { 313 // Skip over catchpads since they just follow their catchswitches. 314 if (isa<CatchPadInst>(ExitedPad)) 315 continue; 316 MemoMap[ExitedPad] = UnwindDestToken; 317 ExitedOriginalPad |= (ExitedPad == EHPad); 318 } 319 320 if (ExitedOriginalPad) 321 return UnwindDestToken; 322 323 // Continue the search. 324 } 325 326 // No definitive information is contained within this funclet. 327 return nullptr; 328 } 329 330 /// Given an EH pad, find where it unwinds. If it unwinds to an EH pad, 331 /// return that pad instruction. If it unwinds to caller, return 332 /// ConstantTokenNone. If it does not have a definitive unwind destination, 333 /// return nullptr. 334 /// 335 /// This routine gets invoked for calls in funclets in inlinees when inlining 336 /// an invoke. Since many funclets don't have calls inside them, it's queried 337 /// on-demand rather than building a map of pads to unwind dests up front. 338 /// Determining a funclet's unwind dest may require recursively searching its 339 /// descendants, and also ancestors and cousins if the descendants don't provide 340 /// an answer. Since most funclets will have their unwind dest immediately 341 /// available as the unwind dest of a catchswitch or cleanupret, this routine 342 /// searches top-down from the given pad and then up. To avoid worst-case 343 /// quadratic run-time given that approach, it uses a memo map to avoid 344 /// re-processing funclet trees. The callers that rewrite the IR as they go 345 /// take advantage of this, for correctness, by checking/forcing rewritten 346 /// pads' entries to match the original callee view. 347 static Value *getUnwindDestToken(Instruction *EHPad, 348 UnwindDestMemoTy &MemoMap) { 349 // Catchpads unwind to the same place as their catchswitch; 350 // redirct any queries on catchpads so the code below can 351 // deal with just catchswitches and cleanuppads. 352 if (auto *CPI = dyn_cast<CatchPadInst>(EHPad)) 353 EHPad = CPI->getCatchSwitch(); 354 355 // Check if we've already determined the unwind dest for this pad. 356 auto Memo = MemoMap.find(EHPad); 357 if (Memo != MemoMap.end()) 358 return Memo->second; 359 360 // Search EHPad and, if necessary, its descendants. 361 Value *UnwindDestToken = getUnwindDestTokenHelper(EHPad, MemoMap); 362 assert((UnwindDestToken == nullptr) != (MemoMap.count(EHPad) != 0)); 363 if (UnwindDestToken) 364 return UnwindDestToken; 365 366 // No information is available for this EHPad from itself or any of its 367 // descendants. An unwind all the way out to a pad in the caller would 368 // need also to agree with the unwind dest of the parent funclet, so 369 // search up the chain to try to find a funclet with information. Put 370 // null entries in the memo map to avoid re-processing as we go up. 371 MemoMap[EHPad] = nullptr; 372 #ifndef NDEBUG 373 SmallPtrSet<Instruction *, 4> TempMemos; 374 TempMemos.insert(EHPad); 375 #endif 376 Instruction *LastUselessPad = EHPad; 377 Value *AncestorToken; 378 for (AncestorToken = getParentPad(EHPad); 379 auto *AncestorPad = dyn_cast<Instruction>(AncestorToken); 380 AncestorToken = getParentPad(AncestorToken)) { 381 // Skip over catchpads since they just follow their catchswitches. 382 if (isa<CatchPadInst>(AncestorPad)) 383 continue; 384 // If the MemoMap had an entry mapping AncestorPad to nullptr, since we 385 // haven't yet called getUnwindDestTokenHelper for AncestorPad in this 386 // call to getUnwindDestToken, that would mean that AncestorPad had no 387 // information in itself, its descendants, or its ancestors. If that 388 // were the case, then we should also have recorded the lack of information 389 // for the descendant that we're coming from. So assert that we don't 390 // find a null entry in the MemoMap for AncestorPad. 391 assert(!MemoMap.count(AncestorPad) || MemoMap[AncestorPad]); 392 auto AncestorMemo = MemoMap.find(AncestorPad); 393 if (AncestorMemo == MemoMap.end()) { 394 UnwindDestToken = getUnwindDestTokenHelper(AncestorPad, MemoMap); 395 } else { 396 UnwindDestToken = AncestorMemo->second; 397 } 398 if (UnwindDestToken) 399 break; 400 LastUselessPad = AncestorPad; 401 MemoMap[LastUselessPad] = nullptr; 402 #ifndef NDEBUG 403 TempMemos.insert(LastUselessPad); 404 #endif 405 } 406 407 // We know that getUnwindDestTokenHelper was called on LastUselessPad and 408 // returned nullptr (and likewise for EHPad and any of its ancestors up to 409 // LastUselessPad), so LastUselessPad has no information from below. Since 410 // getUnwindDestTokenHelper must investigate all downward paths through 411 // no-information nodes to prove that a node has no information like this, 412 // and since any time it finds information it records it in the MemoMap for 413 // not just the immediately-containing funclet but also any ancestors also 414 // exited, it must be the case that, walking downward from LastUselessPad, 415 // visiting just those nodes which have not been mapped to an unwind dest 416 // by getUnwindDestTokenHelper (the nullptr TempMemos notwithstanding, since 417 // they are just used to keep getUnwindDestTokenHelper from repeating work), 418 // any node visited must have been exhaustively searched with no information 419 // for it found. 420 SmallVector<Instruction *, 8> Worklist(1, LastUselessPad); 421 while (!Worklist.empty()) { 422 Instruction *UselessPad = Worklist.pop_back_val(); 423 auto Memo = MemoMap.find(UselessPad); 424 if (Memo != MemoMap.end() && Memo->second) { 425 // Here the name 'UselessPad' is a bit of a misnomer, because we've found 426 // that it is a funclet that does have information about unwinding to 427 // a particular destination; its parent was a useless pad. 428 // Since its parent has no information, the unwind edge must not escape 429 // the parent, and must target a sibling of this pad. This local unwind 430 // gives us no information about EHPad. Leave it and the subtree rooted 431 // at it alone. 432 assert(getParentPad(Memo->second) == getParentPad(UselessPad)); 433 continue; 434 } 435 // We know we don't have information for UselesPad. If it has an entry in 436 // the MemoMap (mapping it to nullptr), it must be one of the TempMemos 437 // added on this invocation of getUnwindDestToken; if a previous invocation 438 // recorded nullptr, it would have had to prove that the ancestors of 439 // UselessPad, which include LastUselessPad, had no information, and that 440 // in turn would have required proving that the descendants of 441 // LastUselesPad, which include EHPad, have no information about 442 // LastUselessPad, which would imply that EHPad was mapped to nullptr in 443 // the MemoMap on that invocation, which isn't the case if we got here. 444 assert(!MemoMap.count(UselessPad) || TempMemos.count(UselessPad)); 445 // Assert as we enumerate users that 'UselessPad' doesn't have any unwind 446 // information that we'd be contradicting by making a map entry for it 447 // (which is something that getUnwindDestTokenHelper must have proved for 448 // us to get here). Just assert on is direct users here; the checks in 449 // this downward walk at its descendants will verify that they don't have 450 // any unwind edges that exit 'UselessPad' either (i.e. they either have no 451 // unwind edges or unwind to a sibling). 452 MemoMap[UselessPad] = UnwindDestToken; 453 if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(UselessPad)) { 454 assert(CatchSwitch->getUnwindDest() == nullptr && "Expected useless pad"); 455 for (BasicBlock *HandlerBlock : CatchSwitch->handlers()) { 456 auto *CatchPad = HandlerBlock->getFirstNonPHI(); 457 for (User *U : CatchPad->users()) { 458 assert( 459 (!isa<InvokeInst>(U) || 460 (getParentPad( 461 cast<InvokeInst>(U)->getUnwindDest()->getFirstNonPHI()) == 462 CatchPad)) && 463 "Expected useless pad"); 464 if (isa<CatchSwitchInst>(U) || isa<CleanupPadInst>(U)) 465 Worklist.push_back(cast<Instruction>(U)); 466 } 467 } 468 } else { 469 assert(isa<CleanupPadInst>(UselessPad)); 470 for (User *U : UselessPad->users()) { 471 assert(!isa<CleanupReturnInst>(U) && "Expected useless pad"); 472 assert((!isa<InvokeInst>(U) || 473 (getParentPad( 474 cast<InvokeInst>(U)->getUnwindDest()->getFirstNonPHI()) == 475 UselessPad)) && 476 "Expected useless pad"); 477 if (isa<CatchSwitchInst>(U) || isa<CleanupPadInst>(U)) 478 Worklist.push_back(cast<Instruction>(U)); 479 } 480 } 481 } 482 483 return UnwindDestToken; 484 } 485 486 /// When we inline a basic block into an invoke, 487 /// we have to turn all of the calls that can throw into invokes. 488 /// This function analyze BB to see if there are any calls, and if so, 489 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI 490 /// nodes in that block with the values specified in InvokeDestPHIValues. 491 static BasicBlock *HandleCallsInBlockInlinedThroughInvoke( 492 BasicBlock *BB, BasicBlock *UnwindEdge, 493 UnwindDestMemoTy *FuncletUnwindMap = nullptr) { 494 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 495 Instruction *I = &*BBI++; 496 497 // We only need to check for function calls: inlined invoke 498 // instructions require no special handling. 499 CallInst *CI = dyn_cast<CallInst>(I); 500 501 if (!CI || CI->doesNotThrow() || isa<InlineAsm>(CI->getCalledValue())) 502 continue; 503 504 // We do not need to (and in fact, cannot) convert possibly throwing calls 505 // to @llvm.experimental_deoptimize (resp. @llvm.experimental.guard) into 506 // invokes. The caller's "segment" of the deoptimization continuation 507 // attached to the newly inlined @llvm.experimental_deoptimize 508 // (resp. @llvm.experimental.guard) call should contain the exception 509 // handling logic, if any. 510 if (auto *F = CI->getCalledFunction()) 511 if (F->getIntrinsicID() == Intrinsic::experimental_deoptimize || 512 F->getIntrinsicID() == Intrinsic::experimental_guard) 513 continue; 514 515 if (auto FuncletBundle = CI->getOperandBundle(LLVMContext::OB_funclet)) { 516 // This call is nested inside a funclet. If that funclet has an unwind 517 // destination within the inlinee, then unwinding out of this call would 518 // be UB. Rewriting this call to an invoke which targets the inlined 519 // invoke's unwind dest would give the call's parent funclet multiple 520 // unwind destinations, which is something that subsequent EH table 521 // generation can't handle and that the veirifer rejects. So when we 522 // see such a call, leave it as a call. 523 auto *FuncletPad = cast<Instruction>(FuncletBundle->Inputs[0]); 524 Value *UnwindDestToken = 525 getUnwindDestToken(FuncletPad, *FuncletUnwindMap); 526 if (UnwindDestToken && !isa<ConstantTokenNone>(UnwindDestToken)) 527 continue; 528 #ifndef NDEBUG 529 Instruction *MemoKey; 530 if (auto *CatchPad = dyn_cast<CatchPadInst>(FuncletPad)) 531 MemoKey = CatchPad->getCatchSwitch(); 532 else 533 MemoKey = FuncletPad; 534 assert(FuncletUnwindMap->count(MemoKey) && 535 (*FuncletUnwindMap)[MemoKey] == UnwindDestToken && 536 "must get memoized to avoid confusing later searches"); 537 #endif // NDEBUG 538 } 539 540 changeToInvokeAndSplitBasicBlock(CI, UnwindEdge); 541 return BB; 542 } 543 return nullptr; 544 } 545 546 /// If we inlined an invoke site, we need to convert calls 547 /// in the body of the inlined function into invokes. 548 /// 549 /// II is the invoke instruction being inlined. FirstNewBlock is the first 550 /// block of the inlined code (the last block is the end of the function), 551 /// and InlineCodeInfo is information about the code that got inlined. 552 static void HandleInlinedLandingPad(InvokeInst *II, BasicBlock *FirstNewBlock, 553 ClonedCodeInfo &InlinedCodeInfo) { 554 BasicBlock *InvokeDest = II->getUnwindDest(); 555 556 Function *Caller = FirstNewBlock->getParent(); 557 558 // The inlined code is currently at the end of the function, scan from the 559 // start of the inlined code to its end, checking for stuff we need to 560 // rewrite. 561 LandingPadInliningInfo Invoke(II); 562 563 // Get all of the inlined landing pad instructions. 564 SmallPtrSet<LandingPadInst*, 16> InlinedLPads; 565 for (Function::iterator I = FirstNewBlock->getIterator(), E = Caller->end(); 566 I != E; ++I) 567 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) 568 InlinedLPads.insert(II->getLandingPadInst()); 569 570 // Append the clauses from the outer landing pad instruction into the inlined 571 // landing pad instructions. 572 LandingPadInst *OuterLPad = Invoke.getLandingPadInst(); 573 for (LandingPadInst *InlinedLPad : InlinedLPads) { 574 unsigned OuterNum = OuterLPad->getNumClauses(); 575 InlinedLPad->reserveClauses(OuterNum); 576 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx) 577 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx)); 578 if (OuterLPad->isCleanup()) 579 InlinedLPad->setCleanup(true); 580 } 581 582 for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end(); 583 BB != E; ++BB) { 584 if (InlinedCodeInfo.ContainsCalls) 585 if (BasicBlock *NewBB = HandleCallsInBlockInlinedThroughInvoke( 586 &*BB, Invoke.getOuterResumeDest())) 587 // Update any PHI nodes in the exceptional block to indicate that there 588 // is now a new entry in them. 589 Invoke.addIncomingPHIValuesFor(NewBB); 590 591 // Forward any resumes that are remaining here. 592 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) 593 Invoke.forwardResume(RI, InlinedLPads); 594 } 595 596 // Now that everything is happy, we have one final detail. The PHI nodes in 597 // the exception destination block still have entries due to the original 598 // invoke instruction. Eliminate these entries (which might even delete the 599 // PHI node) now. 600 InvokeDest->removePredecessor(II->getParent()); 601 } 602 603 /// If we inlined an invoke site, we need to convert calls 604 /// in the body of the inlined function into invokes. 605 /// 606 /// II is the invoke instruction being inlined. FirstNewBlock is the first 607 /// block of the inlined code (the last block is the end of the function), 608 /// and InlineCodeInfo is information about the code that got inlined. 609 static void HandleInlinedEHPad(InvokeInst *II, BasicBlock *FirstNewBlock, 610 ClonedCodeInfo &InlinedCodeInfo) { 611 BasicBlock *UnwindDest = II->getUnwindDest(); 612 Function *Caller = FirstNewBlock->getParent(); 613 614 assert(UnwindDest->getFirstNonPHI()->isEHPad() && "unexpected BasicBlock!"); 615 616 // If there are PHI nodes in the unwind destination block, we need to keep 617 // track of which values came into them from the invoke before removing the 618 // edge from this block. 619 SmallVector<Value *, 8> UnwindDestPHIValues; 620 llvm::BasicBlock *InvokeBB = II->getParent(); 621 for (Instruction &I : *UnwindDest) { 622 // Save the value to use for this edge. 623 PHINode *PHI = dyn_cast<PHINode>(&I); 624 if (!PHI) 625 break; 626 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB)); 627 } 628 629 // Add incoming-PHI values to the unwind destination block for the given basic 630 // block, using the values for the original invoke's source block. 631 auto UpdatePHINodes = [&](BasicBlock *Src) { 632 BasicBlock::iterator I = UnwindDest->begin(); 633 for (Value *V : UnwindDestPHIValues) { 634 PHINode *PHI = cast<PHINode>(I); 635 PHI->addIncoming(V, Src); 636 ++I; 637 } 638 }; 639 640 // This connects all the instructions which 'unwind to caller' to the invoke 641 // destination. 642 UnwindDestMemoTy FuncletUnwindMap; 643 for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end(); 644 BB != E; ++BB) { 645 if (auto *CRI = dyn_cast<CleanupReturnInst>(BB->getTerminator())) { 646 if (CRI->unwindsToCaller()) { 647 auto *CleanupPad = CRI->getCleanupPad(); 648 CleanupReturnInst::Create(CleanupPad, UnwindDest, CRI); 649 CRI->eraseFromParent(); 650 UpdatePHINodes(&*BB); 651 // Finding a cleanupret with an unwind destination would confuse 652 // subsequent calls to getUnwindDestToken, so map the cleanuppad 653 // to short-circuit any such calls and recognize this as an "unwind 654 // to caller" cleanup. 655 assert(!FuncletUnwindMap.count(CleanupPad) || 656 isa<ConstantTokenNone>(FuncletUnwindMap[CleanupPad])); 657 FuncletUnwindMap[CleanupPad] = 658 ConstantTokenNone::get(Caller->getContext()); 659 } 660 } 661 662 Instruction *I = BB->getFirstNonPHI(); 663 if (!I->isEHPad()) 664 continue; 665 666 Instruction *Replacement = nullptr; 667 if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I)) { 668 if (CatchSwitch->unwindsToCaller()) { 669 Value *UnwindDestToken; 670 if (auto *ParentPad = 671 dyn_cast<Instruction>(CatchSwitch->getParentPad())) { 672 // This catchswitch is nested inside another funclet. If that 673 // funclet has an unwind destination within the inlinee, then 674 // unwinding out of this catchswitch would be UB. Rewriting this 675 // catchswitch to unwind to the inlined invoke's unwind dest would 676 // give the parent funclet multiple unwind destinations, which is 677 // something that subsequent EH table generation can't handle and 678 // that the veirifer rejects. So when we see such a call, leave it 679 // as "unwind to caller". 680 UnwindDestToken = getUnwindDestToken(ParentPad, FuncletUnwindMap); 681 if (UnwindDestToken && !isa<ConstantTokenNone>(UnwindDestToken)) 682 continue; 683 } else { 684 // This catchswitch has no parent to inherit constraints from, and 685 // none of its descendants can have an unwind edge that exits it and 686 // targets another funclet in the inlinee. It may or may not have a 687 // descendant that definitively has an unwind to caller. In either 688 // case, we'll have to assume that any unwinds out of it may need to 689 // be routed to the caller, so treat it as though it has a definitive 690 // unwind to caller. 691 UnwindDestToken = ConstantTokenNone::get(Caller->getContext()); 692 } 693 auto *NewCatchSwitch = CatchSwitchInst::Create( 694 CatchSwitch->getParentPad(), UnwindDest, 695 CatchSwitch->getNumHandlers(), CatchSwitch->getName(), 696 CatchSwitch); 697 for (BasicBlock *PadBB : CatchSwitch->handlers()) 698 NewCatchSwitch->addHandler(PadBB); 699 // Propagate info for the old catchswitch over to the new one in 700 // the unwind map. This also serves to short-circuit any subsequent 701 // checks for the unwind dest of this catchswitch, which would get 702 // confused if they found the outer handler in the callee. 703 FuncletUnwindMap[NewCatchSwitch] = UnwindDestToken; 704 Replacement = NewCatchSwitch; 705 } 706 } else if (!isa<FuncletPadInst>(I)) { 707 llvm_unreachable("unexpected EHPad!"); 708 } 709 710 if (Replacement) { 711 Replacement->takeName(I); 712 I->replaceAllUsesWith(Replacement); 713 I->eraseFromParent(); 714 UpdatePHINodes(&*BB); 715 } 716 } 717 718 if (InlinedCodeInfo.ContainsCalls) 719 for (Function::iterator BB = FirstNewBlock->getIterator(), 720 E = Caller->end(); 721 BB != E; ++BB) 722 if (BasicBlock *NewBB = HandleCallsInBlockInlinedThroughInvoke( 723 &*BB, UnwindDest, &FuncletUnwindMap)) 724 // Update any PHI nodes in the exceptional block to indicate that there 725 // is now a new entry in them. 726 UpdatePHINodes(NewBB); 727 728 // Now that everything is happy, we have one final detail. The PHI nodes in 729 // the exception destination block still have entries due to the original 730 // invoke instruction. Eliminate these entries (which might even delete the 731 // PHI node) now. 732 UnwindDest->removePredecessor(InvokeBB); 733 } 734 735 /// When inlining a call site that has !llvm.mem.parallel_loop_access metadata, 736 /// that metadata should be propagated to all memory-accessing cloned 737 /// instructions. 738 static void PropagateParallelLoopAccessMetadata(CallSite CS, 739 ValueToValueMapTy &VMap) { 740 MDNode *M = 741 CS.getInstruction()->getMetadata(LLVMContext::MD_mem_parallel_loop_access); 742 if (!M) 743 return; 744 745 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end(); 746 VMI != VMIE; ++VMI) { 747 if (!VMI->second) 748 continue; 749 750 Instruction *NI = dyn_cast<Instruction>(VMI->second); 751 if (!NI) 752 continue; 753 754 if (MDNode *PM = NI->getMetadata(LLVMContext::MD_mem_parallel_loop_access)) { 755 M = MDNode::concatenate(PM, M); 756 NI->setMetadata(LLVMContext::MD_mem_parallel_loop_access, M); 757 } else if (NI->mayReadOrWriteMemory()) { 758 NI->setMetadata(LLVMContext::MD_mem_parallel_loop_access, M); 759 } 760 } 761 } 762 763 /// When inlining a function that contains noalias scope metadata, 764 /// this metadata needs to be cloned so that the inlined blocks 765 /// have different "unique scopes" at every call site. Were this not done, then 766 /// aliasing scopes from a function inlined into a caller multiple times could 767 /// not be differentiated (and this would lead to miscompiles because the 768 /// non-aliasing property communicated by the metadata could have 769 /// call-site-specific control dependencies). 770 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) { 771 const Function *CalledFunc = CS.getCalledFunction(); 772 SetVector<const MDNode *> MD; 773 774 // Note: We could only clone the metadata if it is already used in the 775 // caller. I'm omitting that check here because it might confuse 776 // inter-procedural alias analysis passes. We can revisit this if it becomes 777 // an efficiency or overhead problem. 778 779 for (const BasicBlock &I : *CalledFunc) 780 for (const Instruction &J : I) { 781 if (const MDNode *M = J.getMetadata(LLVMContext::MD_alias_scope)) 782 MD.insert(M); 783 if (const MDNode *M = J.getMetadata(LLVMContext::MD_noalias)) 784 MD.insert(M); 785 } 786 787 if (MD.empty()) 788 return; 789 790 // Walk the existing metadata, adding the complete (perhaps cyclic) chain to 791 // the set. 792 SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end()); 793 while (!Queue.empty()) { 794 const MDNode *M = cast<MDNode>(Queue.pop_back_val()); 795 for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i) 796 if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i))) 797 if (MD.insert(M1)) 798 Queue.push_back(M1); 799 } 800 801 // Now we have a complete set of all metadata in the chains used to specify 802 // the noalias scopes and the lists of those scopes. 803 SmallVector<TempMDTuple, 16> DummyNodes; 804 DenseMap<const MDNode *, TrackingMDNodeRef> MDMap; 805 for (const MDNode *I : MD) { 806 DummyNodes.push_back(MDTuple::getTemporary(CalledFunc->getContext(), None)); 807 MDMap[I].reset(DummyNodes.back().get()); 808 } 809 810 // Create new metadata nodes to replace the dummy nodes, replacing old 811 // metadata references with either a dummy node or an already-created new 812 // node. 813 for (const MDNode *I : MD) { 814 SmallVector<Metadata *, 4> NewOps; 815 for (unsigned i = 0, ie = I->getNumOperands(); i != ie; ++i) { 816 const Metadata *V = I->getOperand(i); 817 if (const MDNode *M = dyn_cast<MDNode>(V)) 818 NewOps.push_back(MDMap[M]); 819 else 820 NewOps.push_back(const_cast<Metadata *>(V)); 821 } 822 823 MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps); 824 MDTuple *TempM = cast<MDTuple>(MDMap[I]); 825 assert(TempM->isTemporary() && "Expected temporary node"); 826 827 TempM->replaceAllUsesWith(NewM); 828 } 829 830 // Now replace the metadata in the new inlined instructions with the 831 // repacements from the map. 832 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end(); 833 VMI != VMIE; ++VMI) { 834 if (!VMI->second) 835 continue; 836 837 Instruction *NI = dyn_cast<Instruction>(VMI->second); 838 if (!NI) 839 continue; 840 841 if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) { 842 MDNode *NewMD = MDMap[M]; 843 // If the call site also had alias scope metadata (a list of scopes to 844 // which instructions inside it might belong), propagate those scopes to 845 // the inlined instructions. 846 if (MDNode *CSM = 847 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope)) 848 NewMD = MDNode::concatenate(NewMD, CSM); 849 NI->setMetadata(LLVMContext::MD_alias_scope, NewMD); 850 } else if (NI->mayReadOrWriteMemory()) { 851 if (MDNode *M = 852 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope)) 853 NI->setMetadata(LLVMContext::MD_alias_scope, M); 854 } 855 856 if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) { 857 MDNode *NewMD = MDMap[M]; 858 // If the call site also had noalias metadata (a list of scopes with 859 // which instructions inside it don't alias), propagate those scopes to 860 // the inlined instructions. 861 if (MDNode *CSM = 862 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias)) 863 NewMD = MDNode::concatenate(NewMD, CSM); 864 NI->setMetadata(LLVMContext::MD_noalias, NewMD); 865 } else if (NI->mayReadOrWriteMemory()) { 866 if (MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_noalias)) 867 NI->setMetadata(LLVMContext::MD_noalias, M); 868 } 869 } 870 } 871 872 /// If the inlined function has noalias arguments, 873 /// then add new alias scopes for each noalias argument, tag the mapped noalias 874 /// parameters with noalias metadata specifying the new scope, and tag all 875 /// non-derived loads, stores and memory intrinsics with the new alias scopes. 876 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap, 877 const DataLayout &DL, AAResults *CalleeAAR) { 878 if (!EnableNoAliasConversion) 879 return; 880 881 const Function *CalledFunc = CS.getCalledFunction(); 882 SmallVector<const Argument *, 4> NoAliasArgs; 883 884 for (const Argument &Arg : CalledFunc->args()) 885 if (Arg.hasNoAliasAttr() && !Arg.use_empty()) 886 NoAliasArgs.push_back(&Arg); 887 888 if (NoAliasArgs.empty()) 889 return; 890 891 // To do a good job, if a noalias variable is captured, we need to know if 892 // the capture point dominates the particular use we're considering. 893 DominatorTree DT; 894 DT.recalculate(const_cast<Function&>(*CalledFunc)); 895 896 // noalias indicates that pointer values based on the argument do not alias 897 // pointer values which are not based on it. So we add a new "scope" for each 898 // noalias function argument. Accesses using pointers based on that argument 899 // become part of that alias scope, accesses using pointers not based on that 900 // argument are tagged as noalias with that scope. 901 902 DenseMap<const Argument *, MDNode *> NewScopes; 903 MDBuilder MDB(CalledFunc->getContext()); 904 905 // Create a new scope domain for this function. 906 MDNode *NewDomain = 907 MDB.createAnonymousAliasScopeDomain(CalledFunc->getName()); 908 for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) { 909 const Argument *A = NoAliasArgs[i]; 910 911 std::string Name = CalledFunc->getName(); 912 if (A->hasName()) { 913 Name += ": %"; 914 Name += A->getName(); 915 } else { 916 Name += ": argument "; 917 Name += utostr(i); 918 } 919 920 // Note: We always create a new anonymous root here. This is true regardless 921 // of the linkage of the callee because the aliasing "scope" is not just a 922 // property of the callee, but also all control dependencies in the caller. 923 MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name); 924 NewScopes.insert(std::make_pair(A, NewScope)); 925 } 926 927 // Iterate over all new instructions in the map; for all memory-access 928 // instructions, add the alias scope metadata. 929 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end(); 930 VMI != VMIE; ++VMI) { 931 if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) { 932 if (!VMI->second) 933 continue; 934 935 Instruction *NI = dyn_cast<Instruction>(VMI->second); 936 if (!NI) 937 continue; 938 939 bool IsArgMemOnlyCall = false, IsFuncCall = false; 940 SmallVector<const Value *, 2> PtrArgs; 941 942 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) 943 PtrArgs.push_back(LI->getPointerOperand()); 944 else if (const StoreInst *SI = dyn_cast<StoreInst>(I)) 945 PtrArgs.push_back(SI->getPointerOperand()); 946 else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I)) 947 PtrArgs.push_back(VAAI->getPointerOperand()); 948 else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I)) 949 PtrArgs.push_back(CXI->getPointerOperand()); 950 else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) 951 PtrArgs.push_back(RMWI->getPointerOperand()); 952 else if (ImmutableCallSite ICS = ImmutableCallSite(I)) { 953 // If we know that the call does not access memory, then we'll still 954 // know that about the inlined clone of this call site, and we don't 955 // need to add metadata. 956 if (ICS.doesNotAccessMemory()) 957 continue; 958 959 IsFuncCall = true; 960 if (CalleeAAR) { 961 FunctionModRefBehavior MRB = CalleeAAR->getModRefBehavior(ICS); 962 if (MRB == FMRB_OnlyAccessesArgumentPointees || 963 MRB == FMRB_OnlyReadsArgumentPointees) 964 IsArgMemOnlyCall = true; 965 } 966 967 for (Value *Arg : ICS.args()) { 968 // We need to check the underlying objects of all arguments, not just 969 // the pointer arguments, because we might be passing pointers as 970 // integers, etc. 971 // However, if we know that the call only accesses pointer arguments, 972 // then we only need to check the pointer arguments. 973 if (IsArgMemOnlyCall && !Arg->getType()->isPointerTy()) 974 continue; 975 976 PtrArgs.push_back(Arg); 977 } 978 } 979 980 // If we found no pointers, then this instruction is not suitable for 981 // pairing with an instruction to receive aliasing metadata. 982 // However, if this is a call, this we might just alias with none of the 983 // noalias arguments. 984 if (PtrArgs.empty() && !IsFuncCall) 985 continue; 986 987 // It is possible that there is only one underlying object, but you 988 // need to go through several PHIs to see it, and thus could be 989 // repeated in the Objects list. 990 SmallPtrSet<const Value *, 4> ObjSet; 991 SmallVector<Metadata *, 4> Scopes, NoAliases; 992 993 SmallSetVector<const Argument *, 4> NAPtrArgs; 994 for (const Value *V : PtrArgs) { 995 SmallVector<Value *, 4> Objects; 996 GetUnderlyingObjects(const_cast<Value*>(V), 997 Objects, DL, /* LI = */ nullptr); 998 999 for (Value *O : Objects) 1000 ObjSet.insert(O); 1001 } 1002 1003 // Figure out if we're derived from anything that is not a noalias 1004 // argument. 1005 bool CanDeriveViaCapture = false, UsesAliasingPtr = false; 1006 for (const Value *V : ObjSet) { 1007 // Is this value a constant that cannot be derived from any pointer 1008 // value (we need to exclude constant expressions, for example, that 1009 // are formed from arithmetic on global symbols). 1010 bool IsNonPtrConst = isa<ConstantInt>(V) || isa<ConstantFP>(V) || 1011 isa<ConstantPointerNull>(V) || 1012 isa<ConstantDataVector>(V) || isa<UndefValue>(V); 1013 if (IsNonPtrConst) 1014 continue; 1015 1016 // If this is anything other than a noalias argument, then we cannot 1017 // completely describe the aliasing properties using alias.scope 1018 // metadata (and, thus, won't add any). 1019 if (const Argument *A = dyn_cast<Argument>(V)) { 1020 if (!A->hasNoAliasAttr()) 1021 UsesAliasingPtr = true; 1022 } else { 1023 UsesAliasingPtr = true; 1024 } 1025 1026 // If this is not some identified function-local object (which cannot 1027 // directly alias a noalias argument), or some other argument (which, 1028 // by definition, also cannot alias a noalias argument), then we could 1029 // alias a noalias argument that has been captured). 1030 if (!isa<Argument>(V) && 1031 !isIdentifiedFunctionLocal(const_cast<Value*>(V))) 1032 CanDeriveViaCapture = true; 1033 } 1034 1035 // A function call can always get captured noalias pointers (via other 1036 // parameters, globals, etc.). 1037 if (IsFuncCall && !IsArgMemOnlyCall) 1038 CanDeriveViaCapture = true; 1039 1040 // First, we want to figure out all of the sets with which we definitely 1041 // don't alias. Iterate over all noalias set, and add those for which: 1042 // 1. The noalias argument is not in the set of objects from which we 1043 // definitely derive. 1044 // 2. The noalias argument has not yet been captured. 1045 // An arbitrary function that might load pointers could see captured 1046 // noalias arguments via other noalias arguments or globals, and so we 1047 // must always check for prior capture. 1048 for (const Argument *A : NoAliasArgs) { 1049 if (!ObjSet.count(A) && (!CanDeriveViaCapture || 1050 // It might be tempting to skip the 1051 // PointerMayBeCapturedBefore check if 1052 // A->hasNoCaptureAttr() is true, but this is 1053 // incorrect because nocapture only guarantees 1054 // that no copies outlive the function, not 1055 // that the value cannot be locally captured. 1056 !PointerMayBeCapturedBefore(A, 1057 /* ReturnCaptures */ false, 1058 /* StoreCaptures */ false, I, &DT))) 1059 NoAliases.push_back(NewScopes[A]); 1060 } 1061 1062 if (!NoAliases.empty()) 1063 NI->setMetadata(LLVMContext::MD_noalias, 1064 MDNode::concatenate( 1065 NI->getMetadata(LLVMContext::MD_noalias), 1066 MDNode::get(CalledFunc->getContext(), NoAliases))); 1067 1068 // Next, we want to figure out all of the sets to which we might belong. 1069 // We might belong to a set if the noalias argument is in the set of 1070 // underlying objects. If there is some non-noalias argument in our list 1071 // of underlying objects, then we cannot add a scope because the fact 1072 // that some access does not alias with any set of our noalias arguments 1073 // cannot itself guarantee that it does not alias with this access 1074 // (because there is some pointer of unknown origin involved and the 1075 // other access might also depend on this pointer). We also cannot add 1076 // scopes to arbitrary functions unless we know they don't access any 1077 // non-parameter pointer-values. 1078 bool CanAddScopes = !UsesAliasingPtr; 1079 if (CanAddScopes && IsFuncCall) 1080 CanAddScopes = IsArgMemOnlyCall; 1081 1082 if (CanAddScopes) 1083 for (const Argument *A : NoAliasArgs) { 1084 if (ObjSet.count(A)) 1085 Scopes.push_back(NewScopes[A]); 1086 } 1087 1088 if (!Scopes.empty()) 1089 NI->setMetadata( 1090 LLVMContext::MD_alias_scope, 1091 MDNode::concatenate(NI->getMetadata(LLVMContext::MD_alias_scope), 1092 MDNode::get(CalledFunc->getContext(), Scopes))); 1093 } 1094 } 1095 } 1096 1097 /// If the inlined function has non-byval align arguments, then 1098 /// add @llvm.assume-based alignment assumptions to preserve this information. 1099 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) { 1100 if (!PreserveAlignmentAssumptions || !IFI.GetAssumptionCache) 1101 return; 1102 1103 AssumptionCache *AC = &(*IFI.GetAssumptionCache)(*CS.getCaller()); 1104 auto &DL = CS.getCaller()->getParent()->getDataLayout(); 1105 1106 // To avoid inserting redundant assumptions, we should check for assumptions 1107 // already in the caller. To do this, we might need a DT of the caller. 1108 DominatorTree DT; 1109 bool DTCalculated = false; 1110 1111 Function *CalledFunc = CS.getCalledFunction(); 1112 for (Argument &Arg : CalledFunc->args()) { 1113 unsigned Align = Arg.getType()->isPointerTy() ? Arg.getParamAlignment() : 0; 1114 if (Align && !Arg.hasByValOrInAllocaAttr() && !Arg.hasNUses(0)) { 1115 if (!DTCalculated) { 1116 DT.recalculate(*CS.getCaller()); 1117 DTCalculated = true; 1118 } 1119 1120 // If we can already prove the asserted alignment in the context of the 1121 // caller, then don't bother inserting the assumption. 1122 Value *ArgVal = CS.getArgument(Arg.getArgNo()); 1123 if (getKnownAlignment(ArgVal, DL, CS.getInstruction(), AC, &DT) >= Align) 1124 continue; 1125 1126 CallInst *NewAsmp = IRBuilder<>(CS.getInstruction()) 1127 .CreateAlignmentAssumption(DL, ArgVal, Align); 1128 AC->registerAssumption(NewAsmp); 1129 } 1130 } 1131 } 1132 1133 /// Once we have cloned code over from a callee into the caller, 1134 /// update the specified callgraph to reflect the changes we made. 1135 /// Note that it's possible that not all code was copied over, so only 1136 /// some edges of the callgraph may remain. 1137 static void UpdateCallGraphAfterInlining(CallSite CS, 1138 Function::iterator FirstNewBlock, 1139 ValueToValueMapTy &VMap, 1140 InlineFunctionInfo &IFI) { 1141 CallGraph &CG = *IFI.CG; 1142 const Function *Caller = CS.getCaller(); 1143 const Function *Callee = CS.getCalledFunction(); 1144 CallGraphNode *CalleeNode = CG[Callee]; 1145 CallGraphNode *CallerNode = CG[Caller]; 1146 1147 // Since we inlined some uninlined call sites in the callee into the caller, 1148 // add edges from the caller to all of the callees of the callee. 1149 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 1150 1151 // Consider the case where CalleeNode == CallerNode. 1152 CallGraphNode::CalledFunctionsVector CallCache; 1153 if (CalleeNode == CallerNode) { 1154 CallCache.assign(I, E); 1155 I = CallCache.begin(); 1156 E = CallCache.end(); 1157 } 1158 1159 for (; I != E; ++I) { 1160 const Value *OrigCall = I->first; 1161 1162 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); 1163 // Only copy the edge if the call was inlined! 1164 if (VMI == VMap.end() || VMI->second == nullptr) 1165 continue; 1166 1167 // If the call was inlined, but then constant folded, there is no edge to 1168 // add. Check for this case. 1169 Instruction *NewCall = dyn_cast<Instruction>(VMI->second); 1170 if (!NewCall) 1171 continue; 1172 1173 // We do not treat intrinsic calls like real function calls because we 1174 // expect them to become inline code; do not add an edge for an intrinsic. 1175 CallSite CS = CallSite(NewCall); 1176 if (CS && CS.getCalledFunction() && CS.getCalledFunction()->isIntrinsic()) 1177 continue; 1178 1179 // Remember that this call site got inlined for the client of 1180 // InlineFunction. 1181 IFI.InlinedCalls.push_back(NewCall); 1182 1183 // It's possible that inlining the callsite will cause it to go from an 1184 // indirect to a direct call by resolving a function pointer. If this 1185 // happens, set the callee of the new call site to a more precise 1186 // destination. This can also happen if the call graph node of the caller 1187 // was just unnecessarily imprecise. 1188 if (!I->second->getFunction()) 1189 if (Function *F = CallSite(NewCall).getCalledFunction()) { 1190 // Indirect call site resolved to direct call. 1191 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); 1192 1193 continue; 1194 } 1195 1196 CallerNode->addCalledFunction(CallSite(NewCall), I->second); 1197 } 1198 1199 // Update the call graph by deleting the edge from Callee to Caller. We must 1200 // do this after the loop above in case Caller and Callee are the same. 1201 CallerNode->removeCallEdgeFor(CS); 1202 } 1203 1204 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M, 1205 BasicBlock *InsertBlock, 1206 InlineFunctionInfo &IFI) { 1207 Type *AggTy = cast<PointerType>(Src->getType())->getElementType(); 1208 IRBuilder<> Builder(InsertBlock, InsertBlock->begin()); 1209 1210 Value *Size = Builder.getInt64(M->getDataLayout().getTypeStoreSize(AggTy)); 1211 1212 // Always generate a memcpy of alignment 1 here because we don't know 1213 // the alignment of the src pointer. Other optimizations can infer 1214 // better alignment. 1215 Builder.CreateMemCpy(Dst, Src, Size, /*Align=*/1); 1216 } 1217 1218 /// When inlining a call site that has a byval argument, 1219 /// we have to make the implicit memcpy explicit by adding it. 1220 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, 1221 const Function *CalledFunc, 1222 InlineFunctionInfo &IFI, 1223 unsigned ByValAlignment) { 1224 PointerType *ArgTy = cast<PointerType>(Arg->getType()); 1225 Type *AggTy = ArgTy->getElementType(); 1226 1227 Function *Caller = TheCall->getFunction(); 1228 1229 // If the called function is readonly, then it could not mutate the caller's 1230 // copy of the byval'd memory. In this case, it is safe to elide the copy and 1231 // temporary. 1232 if (CalledFunc->onlyReadsMemory()) { 1233 // If the byval argument has a specified alignment that is greater than the 1234 // passed in pointer, then we either have to round up the input pointer or 1235 // give up on this transformation. 1236 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. 1237 return Arg; 1238 1239 AssumptionCache *AC = 1240 IFI.GetAssumptionCache ? &(*IFI.GetAssumptionCache)(*Caller) : nullptr; 1241 const DataLayout &DL = Caller->getParent()->getDataLayout(); 1242 1243 // If the pointer is already known to be sufficiently aligned, or if we can 1244 // round it up to a larger alignment, then we don't need a temporary. 1245 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, DL, TheCall, AC) >= 1246 ByValAlignment) 1247 return Arg; 1248 1249 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad 1250 // for code quality, but rarely happens and is required for correctness. 1251 } 1252 1253 // Create the alloca. If we have DataLayout, use nice alignment. 1254 unsigned Align = 1255 Caller->getParent()->getDataLayout().getPrefTypeAlignment(AggTy); 1256 1257 // If the byval had an alignment specified, we *must* use at least that 1258 // alignment, as it is required by the byval argument (and uses of the 1259 // pointer inside the callee). 1260 Align = std::max(Align, ByValAlignment); 1261 1262 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(), 1263 &*Caller->begin()->begin()); 1264 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca)); 1265 1266 // Uses of the argument in the function should use our new alloca 1267 // instead. 1268 return NewAlloca; 1269 } 1270 1271 // Check whether this Value is used by a lifetime intrinsic. 1272 static bool isUsedByLifetimeMarker(Value *V) { 1273 for (User *U : V->users()) { 1274 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { 1275 switch (II->getIntrinsicID()) { 1276 default: break; 1277 case Intrinsic::lifetime_start: 1278 case Intrinsic::lifetime_end: 1279 return true; 1280 } 1281 } 1282 } 1283 return false; 1284 } 1285 1286 // Check whether the given alloca already has 1287 // lifetime.start or lifetime.end intrinsics. 1288 static bool hasLifetimeMarkers(AllocaInst *AI) { 1289 Type *Ty = AI->getType(); 1290 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(), 1291 Ty->getPointerAddressSpace()); 1292 if (Ty == Int8PtrTy) 1293 return isUsedByLifetimeMarker(AI); 1294 1295 // Do a scan to find all the casts to i8*. 1296 for (User *U : AI->users()) { 1297 if (U->getType() != Int8PtrTy) continue; 1298 if (U->stripPointerCasts() != AI) continue; 1299 if (isUsedByLifetimeMarker(U)) 1300 return true; 1301 } 1302 return false; 1303 } 1304 1305 /// Rebuild the entire inlined-at chain for this instruction so that the top of 1306 /// the chain now is inlined-at the new call site. 1307 static DebugLoc 1308 updateInlinedAtInfo(const DebugLoc &DL, DILocation *InlinedAtNode, 1309 LLVMContext &Ctx, 1310 DenseMap<const DILocation *, DILocation *> &IANodes) { 1311 SmallVector<DILocation *, 3> InlinedAtLocations; 1312 DILocation *Last = InlinedAtNode; 1313 DILocation *CurInlinedAt = DL; 1314 1315 // Gather all the inlined-at nodes 1316 while (DILocation *IA = CurInlinedAt->getInlinedAt()) { 1317 // Skip any we've already built nodes for 1318 if (DILocation *Found = IANodes[IA]) { 1319 Last = Found; 1320 break; 1321 } 1322 1323 InlinedAtLocations.push_back(IA); 1324 CurInlinedAt = IA; 1325 } 1326 1327 // Starting from the top, rebuild the nodes to point to the new inlined-at 1328 // location (then rebuilding the rest of the chain behind it) and update the 1329 // map of already-constructed inlined-at nodes. 1330 for (const DILocation *MD : reverse(InlinedAtLocations)) { 1331 Last = IANodes[MD] = DILocation::getDistinct( 1332 Ctx, MD->getLine(), MD->getColumn(), MD->getScope(), Last); 1333 } 1334 1335 // And finally create the normal location for this instruction, referring to 1336 // the new inlined-at chain. 1337 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(), Last); 1338 } 1339 1340 /// Return the result of AI->isStaticAlloca() if AI were moved to the entry 1341 /// block. Allocas used in inalloca calls and allocas of dynamic array size 1342 /// cannot be static. 1343 static bool allocaWouldBeStaticInEntry(const AllocaInst *AI ) { 1344 return isa<Constant>(AI->getArraySize()) && !AI->isUsedWithInAlloca(); 1345 } 1346 1347 /// Update inlined instructions' line numbers to 1348 /// to encode location where these instructions are inlined. 1349 static void fixupLineNumbers(Function *Fn, Function::iterator FI, 1350 Instruction *TheCall, bool CalleeHasDebugInfo) { 1351 const DebugLoc &TheCallDL = TheCall->getDebugLoc(); 1352 if (!TheCallDL) 1353 return; 1354 1355 auto &Ctx = Fn->getContext(); 1356 DILocation *InlinedAtNode = TheCallDL; 1357 1358 // Create a unique call site, not to be confused with any other call from the 1359 // same location. 1360 InlinedAtNode = DILocation::getDistinct( 1361 Ctx, InlinedAtNode->getLine(), InlinedAtNode->getColumn(), 1362 InlinedAtNode->getScope(), InlinedAtNode->getInlinedAt()); 1363 1364 // Cache the inlined-at nodes as they're built so they are reused, without 1365 // this every instruction's inlined-at chain would become distinct from each 1366 // other. 1367 DenseMap<const DILocation *, DILocation *> IANodes; 1368 1369 for (; FI != Fn->end(); ++FI) { 1370 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end(); 1371 BI != BE; ++BI) { 1372 if (DebugLoc DL = BI->getDebugLoc()) { 1373 BI->setDebugLoc( 1374 updateInlinedAtInfo(DL, InlinedAtNode, BI->getContext(), IANodes)); 1375 continue; 1376 } 1377 1378 if (CalleeHasDebugInfo) 1379 continue; 1380 1381 // If the inlined instruction has no line number, make it look as if it 1382 // originates from the call location. This is important for 1383 // ((__always_inline__, __nodebug__)) functions which must use caller 1384 // location for all instructions in their function body. 1385 1386 // Don't update static allocas, as they may get moved later. 1387 if (auto *AI = dyn_cast<AllocaInst>(BI)) 1388 if (allocaWouldBeStaticInEntry(AI)) 1389 continue; 1390 1391 BI->setDebugLoc(TheCallDL); 1392 } 1393 } 1394 } 1395 /// Update the block frequencies of the caller after a callee has been inlined. 1396 /// 1397 /// Each block cloned into the caller has its block frequency scaled by the 1398 /// ratio of CallSiteFreq/CalleeEntryFreq. This ensures that the cloned copy of 1399 /// callee's entry block gets the same frequency as the callsite block and the 1400 /// relative frequencies of all cloned blocks remain the same after cloning. 1401 static void updateCallerBFI(BasicBlock *CallSiteBlock, 1402 const ValueToValueMapTy &VMap, 1403 BlockFrequencyInfo *CallerBFI, 1404 BlockFrequencyInfo *CalleeBFI, 1405 const BasicBlock &CalleeEntryBlock) { 1406 SmallPtrSet<BasicBlock *, 16> ClonedBBs; 1407 for (auto const &Entry : VMap) { 1408 if (!isa<BasicBlock>(Entry.first) || !Entry.second) 1409 continue; 1410 auto *OrigBB = cast<BasicBlock>(Entry.first); 1411 auto *ClonedBB = cast<BasicBlock>(Entry.second); 1412 uint64_t Freq = CalleeBFI->getBlockFreq(OrigBB).getFrequency(); 1413 if (!ClonedBBs.insert(ClonedBB).second) { 1414 // Multiple blocks in the callee might get mapped to one cloned block in 1415 // the caller since we prune the callee as we clone it. When that happens, 1416 // we want to use the maximum among the original blocks' frequencies. 1417 uint64_t NewFreq = CallerBFI->getBlockFreq(ClonedBB).getFrequency(); 1418 if (NewFreq > Freq) 1419 Freq = NewFreq; 1420 } 1421 CallerBFI->setBlockFreq(ClonedBB, Freq); 1422 } 1423 BasicBlock *EntryClone = cast<BasicBlock>(VMap.lookup(&CalleeEntryBlock)); 1424 CallerBFI->setBlockFreqAndScale( 1425 EntryClone, CallerBFI->getBlockFreq(CallSiteBlock).getFrequency(), 1426 ClonedBBs); 1427 } 1428 1429 /// Update the branch metadata for cloned call instructions. 1430 static void updateCallProfile(Function *Callee, const ValueToValueMapTy &VMap, 1431 const Optional<uint64_t> &CalleeEntryCount, 1432 const Instruction *TheCall) { 1433 if (!CalleeEntryCount.hasValue() || CalleeEntryCount.getValue() < 1) 1434 return; 1435 Optional<uint64_t> CallSiteCount = 1436 ProfileSummaryInfo::getProfileCount(TheCall, nullptr); 1437 uint64_t CallCount = 1438 std::min(CallSiteCount.hasValue() ? CallSiteCount.getValue() : 0, 1439 CalleeEntryCount.getValue()); 1440 1441 for (auto const &Entry : VMap) 1442 if (isa<CallInst>(Entry.first)) 1443 if (auto *CI = dyn_cast_or_null<CallInst>(Entry.second)) 1444 CI->updateProfWeight(CallCount, CalleeEntryCount.getValue()); 1445 for (BasicBlock &BB : *Callee) 1446 // No need to update the callsite if it is pruned during inlining. 1447 if (VMap.count(&BB)) 1448 for (Instruction &I : BB) 1449 if (CallInst *CI = dyn_cast<CallInst>(&I)) 1450 CI->updateProfWeight(CalleeEntryCount.getValue() - CallCount, 1451 CalleeEntryCount.getValue()); 1452 } 1453 1454 /// Update the entry count of callee after inlining. 1455 /// 1456 /// The callsite's block count is subtracted from the callee's function entry 1457 /// count. 1458 static void updateCalleeCount(BlockFrequencyInfo *CallerBFI, BasicBlock *CallBB, 1459 Instruction *CallInst, Function *Callee) { 1460 // If the callee has a original count of N, and the estimated count of 1461 // callsite is M, the new callee count is set to N - M. M is estimated from 1462 // the caller's entry count, its entry block frequency and the block frequency 1463 // of the callsite. 1464 Optional<uint64_t> CalleeCount = Callee->getEntryCount(); 1465 if (!CalleeCount.hasValue()) 1466 return; 1467 Optional<uint64_t> CallCount = 1468 ProfileSummaryInfo::getProfileCount(CallInst, CallerBFI); 1469 if (!CallCount.hasValue()) 1470 return; 1471 // Since CallSiteCount is an estimate, it could exceed the original callee 1472 // count and has to be set to 0. 1473 if (CallCount.getValue() > CalleeCount.getValue()) 1474 Callee->setEntryCount(0); 1475 else 1476 Callee->setEntryCount(CalleeCount.getValue() - CallCount.getValue()); 1477 } 1478 1479 /// This function inlines the called function into the basic block of the 1480 /// caller. This returns false if it is not possible to inline this call. 1481 /// The program is still in a well defined state if this occurs though. 1482 /// 1483 /// Note that this only does one level of inlining. For example, if the 1484 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 1485 /// exists in the instruction stream. Similarly this will inline a recursive 1486 /// function by one level. 1487 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI, 1488 AAResults *CalleeAAR, bool InsertLifetime) { 1489 Instruction *TheCall = CS.getInstruction(); 1490 assert(TheCall->getParent() && TheCall->getFunction() 1491 && "Instruction not in function!"); 1492 1493 // If IFI has any state in it, zap it before we fill it in. 1494 IFI.reset(); 1495 1496 Function *CalledFunc = CS.getCalledFunction(); 1497 if (!CalledFunc || // Can't inline external function or indirect 1498 CalledFunc->isDeclaration() || // call, or call to a vararg function! 1499 CalledFunc->getFunctionType()->isVarArg()) return false; 1500 1501 // The inliner does not know how to inline through calls with operand bundles 1502 // in general ... 1503 if (CS.hasOperandBundles()) { 1504 for (int i = 0, e = CS.getNumOperandBundles(); i != e; ++i) { 1505 uint32_t Tag = CS.getOperandBundleAt(i).getTagID(); 1506 // ... but it knows how to inline through "deopt" operand bundles ... 1507 if (Tag == LLVMContext::OB_deopt) 1508 continue; 1509 // ... and "funclet" operand bundles. 1510 if (Tag == LLVMContext::OB_funclet) 1511 continue; 1512 1513 return false; 1514 } 1515 } 1516 1517 // If the call to the callee cannot throw, set the 'nounwind' flag on any 1518 // calls that we inline. 1519 bool MarkNoUnwind = CS.doesNotThrow(); 1520 1521 BasicBlock *OrigBB = TheCall->getParent(); 1522 Function *Caller = OrigBB->getParent(); 1523 1524 // GC poses two hazards to inlining, which only occur when the callee has GC: 1525 // 1. If the caller has no GC, then the callee's GC must be propagated to the 1526 // caller. 1527 // 2. If the caller has a differing GC, it is invalid to inline. 1528 if (CalledFunc->hasGC()) { 1529 if (!Caller->hasGC()) 1530 Caller->setGC(CalledFunc->getGC()); 1531 else if (CalledFunc->getGC() != Caller->getGC()) 1532 return false; 1533 } 1534 1535 // Get the personality function from the callee if it contains a landing pad. 1536 Constant *CalledPersonality = 1537 CalledFunc->hasPersonalityFn() 1538 ? CalledFunc->getPersonalityFn()->stripPointerCasts() 1539 : nullptr; 1540 1541 // Find the personality function used by the landing pads of the caller. If it 1542 // exists, then check to see that it matches the personality function used in 1543 // the callee. 1544 Constant *CallerPersonality = 1545 Caller->hasPersonalityFn() 1546 ? Caller->getPersonalityFn()->stripPointerCasts() 1547 : nullptr; 1548 if (CalledPersonality) { 1549 if (!CallerPersonality) 1550 Caller->setPersonalityFn(CalledPersonality); 1551 // If the personality functions match, then we can perform the 1552 // inlining. Otherwise, we can't inline. 1553 // TODO: This isn't 100% true. Some personality functions are proper 1554 // supersets of others and can be used in place of the other. 1555 else if (CalledPersonality != CallerPersonality) 1556 return false; 1557 } 1558 1559 // We need to figure out which funclet the callsite was in so that we may 1560 // properly nest the callee. 1561 Instruction *CallSiteEHPad = nullptr; 1562 if (CallerPersonality) { 1563 EHPersonality Personality = classifyEHPersonality(CallerPersonality); 1564 if (isFuncletEHPersonality(Personality)) { 1565 Optional<OperandBundleUse> ParentFunclet = 1566 CS.getOperandBundle(LLVMContext::OB_funclet); 1567 if (ParentFunclet) 1568 CallSiteEHPad = cast<FuncletPadInst>(ParentFunclet->Inputs.front()); 1569 1570 // OK, the inlining site is legal. What about the target function? 1571 1572 if (CallSiteEHPad) { 1573 if (Personality == EHPersonality::MSVC_CXX) { 1574 // The MSVC personality cannot tolerate catches getting inlined into 1575 // cleanup funclets. 1576 if (isa<CleanupPadInst>(CallSiteEHPad)) { 1577 // Ok, the call site is within a cleanuppad. Let's check the callee 1578 // for catchpads. 1579 for (const BasicBlock &CalledBB : *CalledFunc) { 1580 if (isa<CatchSwitchInst>(CalledBB.getFirstNonPHI())) 1581 return false; 1582 } 1583 } 1584 } else if (isAsynchronousEHPersonality(Personality)) { 1585 // SEH is even less tolerant, there may not be any sort of exceptional 1586 // funclet in the callee. 1587 for (const BasicBlock &CalledBB : *CalledFunc) { 1588 if (CalledBB.isEHPad()) 1589 return false; 1590 } 1591 } 1592 } 1593 } 1594 } 1595 1596 // Determine if we are dealing with a call in an EHPad which does not unwind 1597 // to caller. 1598 bool EHPadForCallUnwindsLocally = false; 1599 if (CallSiteEHPad && CS.isCall()) { 1600 UnwindDestMemoTy FuncletUnwindMap; 1601 Value *CallSiteUnwindDestToken = 1602 getUnwindDestToken(CallSiteEHPad, FuncletUnwindMap); 1603 1604 EHPadForCallUnwindsLocally = 1605 CallSiteUnwindDestToken && 1606 !isa<ConstantTokenNone>(CallSiteUnwindDestToken); 1607 } 1608 1609 // Get an iterator to the last basic block in the function, which will have 1610 // the new function inlined after it. 1611 Function::iterator LastBlock = --Caller->end(); 1612 1613 // Make sure to capture all of the return instructions from the cloned 1614 // function. 1615 SmallVector<ReturnInst*, 8> Returns; 1616 ClonedCodeInfo InlinedFunctionInfo; 1617 Function::iterator FirstNewBlock; 1618 1619 { // Scope to destroy VMap after cloning. 1620 ValueToValueMapTy VMap; 1621 // Keep a list of pair (dst, src) to emit byval initializations. 1622 SmallVector<std::pair<Value*, Value*>, 4> ByValInit; 1623 1624 auto &DL = Caller->getParent()->getDataLayout(); 1625 1626 assert(CalledFunc->arg_size() == CS.arg_size() && 1627 "No varargs calls can be inlined!"); 1628 1629 // Calculate the vector of arguments to pass into the function cloner, which 1630 // matches up the formal to the actual argument values. 1631 CallSite::arg_iterator AI = CS.arg_begin(); 1632 unsigned ArgNo = 0; 1633 for (Function::arg_iterator I = CalledFunc->arg_begin(), 1634 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 1635 Value *ActualArg = *AI; 1636 1637 // When byval arguments actually inlined, we need to make the copy implied 1638 // by them explicit. However, we don't do this if the callee is readonly 1639 // or readnone, because the copy would be unneeded: the callee doesn't 1640 // modify the struct. 1641 if (CS.isByValArgument(ArgNo)) { 1642 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, 1643 CalledFunc->getParamAlignment(ArgNo+1)); 1644 if (ActualArg != *AI) 1645 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI)); 1646 } 1647 1648 VMap[&*I] = ActualArg; 1649 } 1650 1651 // Add alignment assumptions if necessary. We do this before the inlined 1652 // instructions are actually cloned into the caller so that we can easily 1653 // check what will be known at the start of the inlined code. 1654 AddAlignmentAssumptions(CS, IFI); 1655 1656 // We want the inliner to prune the code as it copies. We would LOVE to 1657 // have no dead or constant instructions leftover after inlining occurs 1658 // (which can happen, e.g., because an argument was constant), but we'll be 1659 // happy with whatever the cloner can do. 1660 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 1661 /*ModuleLevelChanges=*/false, Returns, ".i", 1662 &InlinedFunctionInfo, TheCall); 1663 // Remember the first block that is newly cloned over. 1664 FirstNewBlock = LastBlock; ++FirstNewBlock; 1665 1666 if (IFI.CallerBFI != nullptr && IFI.CalleeBFI != nullptr) 1667 // Update the BFI of blocks cloned into the caller. 1668 updateCallerBFI(OrigBB, VMap, IFI.CallerBFI, IFI.CalleeBFI, 1669 CalledFunc->front()); 1670 1671 updateCallProfile(CalledFunc, VMap, CalledFunc->getEntryCount(), TheCall); 1672 // Update the profile count of callee. 1673 updateCalleeCount(IFI.CallerBFI, OrigBB, TheCall, CalledFunc); 1674 1675 // Inject byval arguments initialization. 1676 for (std::pair<Value*, Value*> &Init : ByValInit) 1677 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(), 1678 &*FirstNewBlock, IFI); 1679 1680 Optional<OperandBundleUse> ParentDeopt = 1681 CS.getOperandBundle(LLVMContext::OB_deopt); 1682 if (ParentDeopt) { 1683 SmallVector<OperandBundleDef, 2> OpDefs; 1684 1685 for (auto &VH : InlinedFunctionInfo.OperandBundleCallSites) { 1686 Instruction *I = dyn_cast_or_null<Instruction>(VH); 1687 if (!I) continue; // instruction was DCE'd or RAUW'ed to undef 1688 1689 OpDefs.clear(); 1690 1691 CallSite ICS(I); 1692 OpDefs.reserve(ICS.getNumOperandBundles()); 1693 1694 for (unsigned i = 0, e = ICS.getNumOperandBundles(); i < e; ++i) { 1695 auto ChildOB = ICS.getOperandBundleAt(i); 1696 if (ChildOB.getTagID() != LLVMContext::OB_deopt) { 1697 // If the inlined call has other operand bundles, let them be 1698 OpDefs.emplace_back(ChildOB); 1699 continue; 1700 } 1701 1702 // It may be useful to separate this logic (of handling operand 1703 // bundles) out to a separate "policy" component if this gets crowded. 1704 // Prepend the parent's deoptimization continuation to the newly 1705 // inlined call's deoptimization continuation. 1706 std::vector<Value *> MergedDeoptArgs; 1707 MergedDeoptArgs.reserve(ParentDeopt->Inputs.size() + 1708 ChildOB.Inputs.size()); 1709 1710 MergedDeoptArgs.insert(MergedDeoptArgs.end(), 1711 ParentDeopt->Inputs.begin(), 1712 ParentDeopt->Inputs.end()); 1713 MergedDeoptArgs.insert(MergedDeoptArgs.end(), ChildOB.Inputs.begin(), 1714 ChildOB.Inputs.end()); 1715 1716 OpDefs.emplace_back("deopt", std::move(MergedDeoptArgs)); 1717 } 1718 1719 Instruction *NewI = nullptr; 1720 if (isa<CallInst>(I)) 1721 NewI = CallInst::Create(cast<CallInst>(I), OpDefs, I); 1722 else 1723 NewI = InvokeInst::Create(cast<InvokeInst>(I), OpDefs, I); 1724 1725 // Note: the RAUW does the appropriate fixup in VMap, so we need to do 1726 // this even if the call returns void. 1727 I->replaceAllUsesWith(NewI); 1728 1729 VH = nullptr; 1730 I->eraseFromParent(); 1731 } 1732 } 1733 1734 // Update the callgraph if requested. 1735 if (IFI.CG) 1736 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); 1737 1738 // For 'nodebug' functions, the associated DISubprogram is always null. 1739 // Conservatively avoid propagating the callsite debug location to 1740 // instructions inlined from a function whose DISubprogram is not null. 1741 fixupLineNumbers(Caller, FirstNewBlock, TheCall, 1742 CalledFunc->getSubprogram() != nullptr); 1743 1744 // Clone existing noalias metadata if necessary. 1745 CloneAliasScopeMetadata(CS, VMap); 1746 1747 // Add noalias metadata if necessary. 1748 AddAliasScopeMetadata(CS, VMap, DL, CalleeAAR); 1749 1750 // Propagate llvm.mem.parallel_loop_access if necessary. 1751 PropagateParallelLoopAccessMetadata(CS, VMap); 1752 1753 // Register any cloned assumptions. 1754 if (IFI.GetAssumptionCache) 1755 for (BasicBlock &NewBlock : 1756 make_range(FirstNewBlock->getIterator(), Caller->end())) 1757 for (Instruction &I : NewBlock) { 1758 if (auto *II = dyn_cast<IntrinsicInst>(&I)) 1759 if (II->getIntrinsicID() == Intrinsic::assume) 1760 (*IFI.GetAssumptionCache)(*Caller).registerAssumption(II); 1761 } 1762 } 1763 1764 // If there are any alloca instructions in the block that used to be the entry 1765 // block for the callee, move them to the entry block of the caller. First 1766 // calculate which instruction they should be inserted before. We insert the 1767 // instructions at the end of the current alloca list. 1768 { 1769 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 1770 for (BasicBlock::iterator I = FirstNewBlock->begin(), 1771 E = FirstNewBlock->end(); I != E; ) { 1772 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 1773 if (!AI) continue; 1774 1775 // If the alloca is now dead, remove it. This often occurs due to code 1776 // specialization. 1777 if (AI->use_empty()) { 1778 AI->eraseFromParent(); 1779 continue; 1780 } 1781 1782 if (!allocaWouldBeStaticInEntry(AI)) 1783 continue; 1784 1785 // Keep track of the static allocas that we inline into the caller. 1786 IFI.StaticAllocas.push_back(AI); 1787 1788 // Scan for the block of allocas that we can move over, and move them 1789 // all at once. 1790 while (isa<AllocaInst>(I) && 1791 allocaWouldBeStaticInEntry(cast<AllocaInst>(I))) { 1792 IFI.StaticAllocas.push_back(cast<AllocaInst>(I)); 1793 ++I; 1794 } 1795 1796 // Transfer all of the allocas over in a block. Using splice means 1797 // that the instructions aren't removed from the symbol table, then 1798 // reinserted. 1799 Caller->getEntryBlock().getInstList().splice( 1800 InsertPoint, FirstNewBlock->getInstList(), AI->getIterator(), I); 1801 } 1802 // Move any dbg.declares describing the allocas into the entry basic block. 1803 DIBuilder DIB(*Caller->getParent()); 1804 for (auto &AI : IFI.StaticAllocas) 1805 replaceDbgDeclareForAlloca(AI, AI, DIB, /*Deref=*/false); 1806 } 1807 1808 bool InlinedMustTailCalls = false, InlinedDeoptimizeCalls = false; 1809 if (InlinedFunctionInfo.ContainsCalls) { 1810 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None; 1811 if (CallInst *CI = dyn_cast<CallInst>(TheCall)) 1812 CallSiteTailKind = CI->getTailCallKind(); 1813 1814 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; 1815 ++BB) { 1816 for (Instruction &I : *BB) { 1817 CallInst *CI = dyn_cast<CallInst>(&I); 1818 if (!CI) 1819 continue; 1820 1821 if (Function *F = CI->getCalledFunction()) 1822 InlinedDeoptimizeCalls |= 1823 F->getIntrinsicID() == Intrinsic::experimental_deoptimize; 1824 1825 // We need to reduce the strength of any inlined tail calls. For 1826 // musttail, we have to avoid introducing potential unbounded stack 1827 // growth. For example, if functions 'f' and 'g' are mutually recursive 1828 // with musttail, we can inline 'g' into 'f' so long as we preserve 1829 // musttail on the cloned call to 'f'. If either the inlined call site 1830 // or the cloned call site is *not* musttail, the program already has 1831 // one frame of stack growth, so it's safe to remove musttail. Here is 1832 // a table of example transformations: 1833 // 1834 // f -> musttail g -> musttail f ==> f -> musttail f 1835 // f -> musttail g -> tail f ==> f -> tail f 1836 // f -> g -> musttail f ==> f -> f 1837 // f -> g -> tail f ==> f -> f 1838 CallInst::TailCallKind ChildTCK = CI->getTailCallKind(); 1839 ChildTCK = std::min(CallSiteTailKind, ChildTCK); 1840 CI->setTailCallKind(ChildTCK); 1841 InlinedMustTailCalls |= CI->isMustTailCall(); 1842 1843 // Calls inlined through a 'nounwind' call site should be marked 1844 // 'nounwind'. 1845 if (MarkNoUnwind) 1846 CI->setDoesNotThrow(); 1847 } 1848 } 1849 } 1850 1851 // Leave lifetime markers for the static alloca's, scoping them to the 1852 // function we just inlined. 1853 if (InsertLifetime && !IFI.StaticAllocas.empty()) { 1854 IRBuilder<> builder(&FirstNewBlock->front()); 1855 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) { 1856 AllocaInst *AI = IFI.StaticAllocas[ai]; 1857 // Don't mark swifterror allocas. They can't have bitcast uses. 1858 if (AI->isSwiftError()) 1859 continue; 1860 1861 // If the alloca is already scoped to something smaller than the whole 1862 // function then there's no need to add redundant, less accurate markers. 1863 if (hasLifetimeMarkers(AI)) 1864 continue; 1865 1866 // Try to determine the size of the allocation. 1867 ConstantInt *AllocaSize = nullptr; 1868 if (ConstantInt *AIArraySize = 1869 dyn_cast<ConstantInt>(AI->getArraySize())) { 1870 auto &DL = Caller->getParent()->getDataLayout(); 1871 Type *AllocaType = AI->getAllocatedType(); 1872 uint64_t AllocaTypeSize = DL.getTypeAllocSize(AllocaType); 1873 uint64_t AllocaArraySize = AIArraySize->getLimitedValue(); 1874 1875 // Don't add markers for zero-sized allocas. 1876 if (AllocaArraySize == 0) 1877 continue; 1878 1879 // Check that array size doesn't saturate uint64_t and doesn't 1880 // overflow when it's multiplied by type size. 1881 if (AllocaArraySize != ~0ULL && 1882 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) { 1883 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()), 1884 AllocaArraySize * AllocaTypeSize); 1885 } 1886 } 1887 1888 builder.CreateLifetimeStart(AI, AllocaSize); 1889 for (ReturnInst *RI : Returns) { 1890 // Don't insert llvm.lifetime.end calls between a musttail or deoptimize 1891 // call and a return. The return kills all local allocas. 1892 if (InlinedMustTailCalls && 1893 RI->getParent()->getTerminatingMustTailCall()) 1894 continue; 1895 if (InlinedDeoptimizeCalls && 1896 RI->getParent()->getTerminatingDeoptimizeCall()) 1897 continue; 1898 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize); 1899 } 1900 } 1901 } 1902 1903 // If the inlined code contained dynamic alloca instructions, wrap the inlined 1904 // code with llvm.stacksave/llvm.stackrestore intrinsics. 1905 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 1906 Module *M = Caller->getParent(); 1907 // Get the two intrinsics we care about. 1908 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 1909 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 1910 1911 // Insert the llvm.stacksave. 1912 CallInst *SavedPtr = IRBuilder<>(&*FirstNewBlock, FirstNewBlock->begin()) 1913 .CreateCall(StackSave, {}, "savedstack"); 1914 1915 // Insert a call to llvm.stackrestore before any return instructions in the 1916 // inlined function. 1917 for (ReturnInst *RI : Returns) { 1918 // Don't insert llvm.stackrestore calls between a musttail or deoptimize 1919 // call and a return. The return will restore the stack pointer. 1920 if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall()) 1921 continue; 1922 if (InlinedDeoptimizeCalls && RI->getParent()->getTerminatingDeoptimizeCall()) 1923 continue; 1924 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr); 1925 } 1926 } 1927 1928 // If we are inlining for an invoke instruction, we must make sure to rewrite 1929 // any call instructions into invoke instructions. This is sensitive to which 1930 // funclet pads were top-level in the inlinee, so must be done before 1931 // rewriting the "parent pad" links. 1932 if (auto *II = dyn_cast<InvokeInst>(TheCall)) { 1933 BasicBlock *UnwindDest = II->getUnwindDest(); 1934 Instruction *FirstNonPHI = UnwindDest->getFirstNonPHI(); 1935 if (isa<LandingPadInst>(FirstNonPHI)) { 1936 HandleInlinedLandingPad(II, &*FirstNewBlock, InlinedFunctionInfo); 1937 } else { 1938 HandleInlinedEHPad(II, &*FirstNewBlock, InlinedFunctionInfo); 1939 } 1940 } 1941 1942 // Update the lexical scopes of the new funclets and callsites. 1943 // Anything that had 'none' as its parent is now nested inside the callsite's 1944 // EHPad. 1945 1946 if (CallSiteEHPad) { 1947 for (Function::iterator BB = FirstNewBlock->getIterator(), 1948 E = Caller->end(); 1949 BB != E; ++BB) { 1950 // Add bundle operands to any top-level call sites. 1951 SmallVector<OperandBundleDef, 1> OpBundles; 1952 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E;) { 1953 Instruction *I = &*BBI++; 1954 CallSite CS(I); 1955 if (!CS) 1956 continue; 1957 1958 // Skip call sites which are nounwind intrinsics. 1959 auto *CalledFn = 1960 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts()); 1961 if (CalledFn && CalledFn->isIntrinsic() && CS.doesNotThrow()) 1962 continue; 1963 1964 // Skip call sites which already have a "funclet" bundle. 1965 if (CS.getOperandBundle(LLVMContext::OB_funclet)) 1966 continue; 1967 1968 CS.getOperandBundlesAsDefs(OpBundles); 1969 OpBundles.emplace_back("funclet", CallSiteEHPad); 1970 1971 Instruction *NewInst; 1972 if (CS.isCall()) 1973 NewInst = CallInst::Create(cast<CallInst>(I), OpBundles, I); 1974 else 1975 NewInst = InvokeInst::Create(cast<InvokeInst>(I), OpBundles, I); 1976 NewInst->takeName(I); 1977 I->replaceAllUsesWith(NewInst); 1978 I->eraseFromParent(); 1979 1980 OpBundles.clear(); 1981 } 1982 1983 // It is problematic if the inlinee has a cleanupret which unwinds to 1984 // caller and we inline it into a call site which doesn't unwind but into 1985 // an EH pad that does. Such an edge must be dynamically unreachable. 1986 // As such, we replace the cleanupret with unreachable. 1987 if (auto *CleanupRet = dyn_cast<CleanupReturnInst>(BB->getTerminator())) 1988 if (CleanupRet->unwindsToCaller() && EHPadForCallUnwindsLocally) 1989 changeToUnreachable(CleanupRet, /*UseLLVMTrap=*/false); 1990 1991 Instruction *I = BB->getFirstNonPHI(); 1992 if (!I->isEHPad()) 1993 continue; 1994 1995 if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I)) { 1996 if (isa<ConstantTokenNone>(CatchSwitch->getParentPad())) 1997 CatchSwitch->setParentPad(CallSiteEHPad); 1998 } else { 1999 auto *FPI = cast<FuncletPadInst>(I); 2000 if (isa<ConstantTokenNone>(FPI->getParentPad())) 2001 FPI->setParentPad(CallSiteEHPad); 2002 } 2003 } 2004 } 2005 2006 if (InlinedDeoptimizeCalls) { 2007 // We need to at least remove the deoptimizing returns from the Return set, 2008 // so that the control flow from those returns does not get merged into the 2009 // caller (but terminate it instead). If the caller's return type does not 2010 // match the callee's return type, we also need to change the return type of 2011 // the intrinsic. 2012 if (Caller->getReturnType() == TheCall->getType()) { 2013 auto NewEnd = remove_if(Returns, [](ReturnInst *RI) { 2014 return RI->getParent()->getTerminatingDeoptimizeCall() != nullptr; 2015 }); 2016 Returns.erase(NewEnd, Returns.end()); 2017 } else { 2018 SmallVector<ReturnInst *, 8> NormalReturns; 2019 Function *NewDeoptIntrinsic = Intrinsic::getDeclaration( 2020 Caller->getParent(), Intrinsic::experimental_deoptimize, 2021 {Caller->getReturnType()}); 2022 2023 for (ReturnInst *RI : Returns) { 2024 CallInst *DeoptCall = RI->getParent()->getTerminatingDeoptimizeCall(); 2025 if (!DeoptCall) { 2026 NormalReturns.push_back(RI); 2027 continue; 2028 } 2029 2030 // The calling convention on the deoptimize call itself may be bogus, 2031 // since the code we're inlining may have undefined behavior (and may 2032 // never actually execute at runtime); but all 2033 // @llvm.experimental.deoptimize declarations have to have the same 2034 // calling convention in a well-formed module. 2035 auto CallingConv = DeoptCall->getCalledFunction()->getCallingConv(); 2036 NewDeoptIntrinsic->setCallingConv(CallingConv); 2037 auto *CurBB = RI->getParent(); 2038 RI->eraseFromParent(); 2039 2040 SmallVector<Value *, 4> CallArgs(DeoptCall->arg_begin(), 2041 DeoptCall->arg_end()); 2042 2043 SmallVector<OperandBundleDef, 1> OpBundles; 2044 DeoptCall->getOperandBundlesAsDefs(OpBundles); 2045 DeoptCall->eraseFromParent(); 2046 assert(!OpBundles.empty() && 2047 "Expected at least the deopt operand bundle"); 2048 2049 IRBuilder<> Builder(CurBB); 2050 CallInst *NewDeoptCall = 2051 Builder.CreateCall(NewDeoptIntrinsic, CallArgs, OpBundles); 2052 NewDeoptCall->setCallingConv(CallingConv); 2053 if (NewDeoptCall->getType()->isVoidTy()) 2054 Builder.CreateRetVoid(); 2055 else 2056 Builder.CreateRet(NewDeoptCall); 2057 } 2058 2059 // Leave behind the normal returns so we can merge control flow. 2060 std::swap(Returns, NormalReturns); 2061 } 2062 } 2063 2064 // Handle any inlined musttail call sites. In order for a new call site to be 2065 // musttail, the source of the clone and the inlined call site must have been 2066 // musttail. Therefore it's safe to return without merging control into the 2067 // phi below. 2068 if (InlinedMustTailCalls) { 2069 // Check if we need to bitcast the result of any musttail calls. 2070 Type *NewRetTy = Caller->getReturnType(); 2071 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy; 2072 2073 // Handle the returns preceded by musttail calls separately. 2074 SmallVector<ReturnInst *, 8> NormalReturns; 2075 for (ReturnInst *RI : Returns) { 2076 CallInst *ReturnedMustTail = 2077 RI->getParent()->getTerminatingMustTailCall(); 2078 if (!ReturnedMustTail) { 2079 NormalReturns.push_back(RI); 2080 continue; 2081 } 2082 if (!NeedBitCast) 2083 continue; 2084 2085 // Delete the old return and any preceding bitcast. 2086 BasicBlock *CurBB = RI->getParent(); 2087 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue()); 2088 RI->eraseFromParent(); 2089 if (OldCast) 2090 OldCast->eraseFromParent(); 2091 2092 // Insert a new bitcast and return with the right type. 2093 IRBuilder<> Builder(CurBB); 2094 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy)); 2095 } 2096 2097 // Leave behind the normal returns so we can merge control flow. 2098 std::swap(Returns, NormalReturns); 2099 } 2100 2101 // Now that all of the transforms on the inlined code have taken place but 2102 // before we splice the inlined code into the CFG and lose track of which 2103 // blocks were actually inlined, collect the call sites. We only do this if 2104 // call graph updates weren't requested, as those provide value handle based 2105 // tracking of inlined call sites instead. 2106 if (InlinedFunctionInfo.ContainsCalls && !IFI.CG) { 2107 // Otherwise just collect the raw call sites that were inlined. 2108 for (BasicBlock &NewBB : 2109 make_range(FirstNewBlock->getIterator(), Caller->end())) 2110 for (Instruction &I : NewBB) 2111 if (auto CS = CallSite(&I)) 2112 IFI.InlinedCallSites.push_back(CS); 2113 } 2114 2115 // If we cloned in _exactly one_ basic block, and if that block ends in a 2116 // return instruction, we splice the body of the inlined callee directly into 2117 // the calling basic block. 2118 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 2119 // Move all of the instructions right before the call. 2120 OrigBB->getInstList().splice(TheCall->getIterator(), 2121 FirstNewBlock->getInstList(), 2122 FirstNewBlock->begin(), FirstNewBlock->end()); 2123 // Remove the cloned basic block. 2124 Caller->getBasicBlockList().pop_back(); 2125 2126 // If the call site was an invoke instruction, add a branch to the normal 2127 // destination. 2128 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 2129 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 2130 NewBr->setDebugLoc(Returns[0]->getDebugLoc()); 2131 } 2132 2133 // If the return instruction returned a value, replace uses of the call with 2134 // uses of the returned value. 2135 if (!TheCall->use_empty()) { 2136 ReturnInst *R = Returns[0]; 2137 if (TheCall == R->getReturnValue()) 2138 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 2139 else 2140 TheCall->replaceAllUsesWith(R->getReturnValue()); 2141 } 2142 // Since we are now done with the Call/Invoke, we can delete it. 2143 TheCall->eraseFromParent(); 2144 2145 // Since we are now done with the return instruction, delete it also. 2146 Returns[0]->eraseFromParent(); 2147 2148 // We are now done with the inlining. 2149 return true; 2150 } 2151 2152 // Otherwise, we have the normal case, of more than one block to inline or 2153 // multiple return sites. 2154 2155 // We want to clone the entire callee function into the hole between the 2156 // "starter" and "ender" blocks. How we accomplish this depends on whether 2157 // this is an invoke instruction or a call instruction. 2158 BasicBlock *AfterCallBB; 2159 BranchInst *CreatedBranchToNormalDest = nullptr; 2160 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 2161 2162 // Add an unconditional branch to make this look like the CallInst case... 2163 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall); 2164 2165 // Split the basic block. This guarantees that no PHI nodes will have to be 2166 // updated due to new incoming edges, and make the invoke case more 2167 // symmetric to the call case. 2168 AfterCallBB = 2169 OrigBB->splitBasicBlock(CreatedBranchToNormalDest->getIterator(), 2170 CalledFunc->getName() + ".exit"); 2171 2172 } else { // It's a call 2173 // If this is a call instruction, we need to split the basic block that 2174 // the call lives in. 2175 // 2176 AfterCallBB = OrigBB->splitBasicBlock(TheCall->getIterator(), 2177 CalledFunc->getName() + ".exit"); 2178 } 2179 2180 if (IFI.CallerBFI) { 2181 // Copy original BB's block frequency to AfterCallBB 2182 IFI.CallerBFI->setBlockFreq( 2183 AfterCallBB, IFI.CallerBFI->getBlockFreq(OrigBB).getFrequency()); 2184 } 2185 2186 // Change the branch that used to go to AfterCallBB to branch to the first 2187 // basic block of the inlined function. 2188 // 2189 TerminatorInst *Br = OrigBB->getTerminator(); 2190 assert(Br && Br->getOpcode() == Instruction::Br && 2191 "splitBasicBlock broken!"); 2192 Br->setOperand(0, &*FirstNewBlock); 2193 2194 // Now that the function is correct, make it a little bit nicer. In 2195 // particular, move the basic blocks inserted from the end of the function 2196 // into the space made by splitting the source basic block. 2197 Caller->getBasicBlockList().splice(AfterCallBB->getIterator(), 2198 Caller->getBasicBlockList(), FirstNewBlock, 2199 Caller->end()); 2200 2201 // Handle all of the return instructions that we just cloned in, and eliminate 2202 // any users of the original call/invoke instruction. 2203 Type *RTy = CalledFunc->getReturnType(); 2204 2205 PHINode *PHI = nullptr; 2206 if (Returns.size() > 1) { 2207 // The PHI node should go at the front of the new basic block to merge all 2208 // possible incoming values. 2209 if (!TheCall->use_empty()) { 2210 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(), 2211 &AfterCallBB->front()); 2212 // Anything that used the result of the function call should now use the 2213 // PHI node as their operand. 2214 TheCall->replaceAllUsesWith(PHI); 2215 } 2216 2217 // Loop over all of the return instructions adding entries to the PHI node 2218 // as appropriate. 2219 if (PHI) { 2220 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 2221 ReturnInst *RI = Returns[i]; 2222 assert(RI->getReturnValue()->getType() == PHI->getType() && 2223 "Ret value not consistent in function!"); 2224 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 2225 } 2226 } 2227 2228 // Add a branch to the merge points and remove return instructions. 2229 DebugLoc Loc; 2230 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 2231 ReturnInst *RI = Returns[i]; 2232 BranchInst* BI = BranchInst::Create(AfterCallBB, RI); 2233 Loc = RI->getDebugLoc(); 2234 BI->setDebugLoc(Loc); 2235 RI->eraseFromParent(); 2236 } 2237 // We need to set the debug location to *somewhere* inside the 2238 // inlined function. The line number may be nonsensical, but the 2239 // instruction will at least be associated with the right 2240 // function. 2241 if (CreatedBranchToNormalDest) 2242 CreatedBranchToNormalDest->setDebugLoc(Loc); 2243 } else if (!Returns.empty()) { 2244 // Otherwise, if there is exactly one return value, just replace anything 2245 // using the return value of the call with the computed value. 2246 if (!TheCall->use_empty()) { 2247 if (TheCall == Returns[0]->getReturnValue()) 2248 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 2249 else 2250 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 2251 } 2252 2253 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 2254 BasicBlock *ReturnBB = Returns[0]->getParent(); 2255 ReturnBB->replaceAllUsesWith(AfterCallBB); 2256 2257 // Splice the code from the return block into the block that it will return 2258 // to, which contains the code that was after the call. 2259 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 2260 ReturnBB->getInstList()); 2261 2262 if (CreatedBranchToNormalDest) 2263 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc()); 2264 2265 // Delete the return instruction now and empty ReturnBB now. 2266 Returns[0]->eraseFromParent(); 2267 ReturnBB->eraseFromParent(); 2268 } else if (!TheCall->use_empty()) { 2269 // No returns, but something is using the return value of the call. Just 2270 // nuke the result. 2271 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 2272 } 2273 2274 // Since we are now done with the Call/Invoke, we can delete it. 2275 TheCall->eraseFromParent(); 2276 2277 // If we inlined any musttail calls and the original return is now 2278 // unreachable, delete it. It can only contain a bitcast and ret. 2279 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB)) 2280 AfterCallBB->eraseFromParent(); 2281 2282 // We should always be able to fold the entry block of the function into the 2283 // single predecessor of the block... 2284 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 2285 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 2286 2287 // Splice the code entry block into calling block, right before the 2288 // unconditional branch. 2289 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 2290 OrigBB->getInstList().splice(Br->getIterator(), CalleeEntry->getInstList()); 2291 2292 // Remove the unconditional branch. 2293 OrigBB->getInstList().erase(Br); 2294 2295 // Now we can remove the CalleeEntry block, which is now empty. 2296 Caller->getBasicBlockList().erase(CalleeEntry); 2297 2298 // If we inserted a phi node, check to see if it has a single value (e.g. all 2299 // the entries are the same or undef). If so, remove the PHI so it doesn't 2300 // block other optimizations. 2301 if (PHI) { 2302 AssumptionCache *AC = 2303 IFI.GetAssumptionCache ? &(*IFI.GetAssumptionCache)(*Caller) : nullptr; 2304 auto &DL = Caller->getParent()->getDataLayout(); 2305 if (Value *V = SimplifyInstruction(PHI, DL, nullptr, nullptr, AC)) { 2306 PHI->replaceAllUsesWith(V); 2307 PHI->eraseFromParent(); 2308 } 2309 } 2310 2311 return true; 2312 } 2313