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/ADT/DenseMap.h" 16 #include "llvm/ADT/None.h" 17 #include "llvm/ADT/Optional.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/SetVector.h" 20 #include "llvm/ADT/SmallPtrSet.h" 21 #include "llvm/ADT/SmallVector.h" 22 #include "llvm/ADT/StringExtras.h" 23 #include "llvm/ADT/iterator_range.h" 24 #include "llvm/Analysis/AliasAnalysis.h" 25 #include "llvm/Analysis/AssumptionCache.h" 26 #include "llvm/Analysis/BlockFrequencyInfo.h" 27 #include "llvm/Analysis/CallGraph.h" 28 #include "llvm/Analysis/CaptureTracking.h" 29 #include "llvm/Analysis/EHPersonalities.h" 30 #include "llvm/Analysis/InstructionSimplify.h" 31 #include "llvm/Analysis/ProfileSummaryInfo.h" 32 #include "llvm/Transforms/Utils/Local.h" 33 #include "llvm/Analysis/ValueTracking.h" 34 #include "llvm/IR/Argument.h" 35 #include "llvm/IR/BasicBlock.h" 36 #include "llvm/IR/CFG.h" 37 #include "llvm/IR/CallSite.h" 38 #include "llvm/IR/Constant.h" 39 #include "llvm/IR/Constants.h" 40 #include "llvm/IR/DIBuilder.h" 41 #include "llvm/IR/DataLayout.h" 42 #include "llvm/IR/DebugInfoMetadata.h" 43 #include "llvm/IR/DebugLoc.h" 44 #include "llvm/IR/DerivedTypes.h" 45 #include "llvm/IR/Dominators.h" 46 #include "llvm/IR/Function.h" 47 #include "llvm/IR/IRBuilder.h" 48 #include "llvm/IR/InstrTypes.h" 49 #include "llvm/IR/Instruction.h" 50 #include "llvm/IR/Instructions.h" 51 #include "llvm/IR/IntrinsicInst.h" 52 #include "llvm/IR/Intrinsics.h" 53 #include "llvm/IR/LLVMContext.h" 54 #include "llvm/IR/MDBuilder.h" 55 #include "llvm/IR/Metadata.h" 56 #include "llvm/IR/Module.h" 57 #include "llvm/IR/Type.h" 58 #include "llvm/IR/User.h" 59 #include "llvm/IR/Value.h" 60 #include "llvm/Support/Casting.h" 61 #include "llvm/Support/CommandLine.h" 62 #include "llvm/Support/ErrorHandling.h" 63 #include "llvm/Transforms/Utils/Cloning.h" 64 #include "llvm/Transforms/Utils/ValueMapper.h" 65 #include <algorithm> 66 #include <cassert> 67 #include <cstdint> 68 #include <iterator> 69 #include <limits> 70 #include <string> 71 #include <utility> 72 #include <vector> 73 74 using namespace llvm; 75 using ProfileCount = Function::ProfileCount; 76 77 static cl::opt<bool> 78 EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(true), 79 cl::Hidden, 80 cl::desc("Convert noalias attributes to metadata during inlining.")); 81 82 static cl::opt<bool> 83 PreserveAlignmentAssumptions("preserve-alignment-assumptions-during-inlining", 84 cl::init(true), cl::Hidden, 85 cl::desc("Convert align attributes to assumptions during inlining.")); 86 87 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI, 88 AAResults *CalleeAAR, bool InsertLifetime) { 89 return InlineFunction(CallSite(CI), IFI, CalleeAAR, InsertLifetime); 90 } 91 92 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI, 93 AAResults *CalleeAAR, bool InsertLifetime) { 94 return InlineFunction(CallSite(II), IFI, CalleeAAR, InsertLifetime); 95 } 96 97 namespace { 98 99 /// A class for recording information about inlining a landing pad. 100 class LandingPadInliningInfo { 101 /// Destination of the invoke's unwind. 102 BasicBlock *OuterResumeDest; 103 104 /// Destination for the callee's resume. 105 BasicBlock *InnerResumeDest = nullptr; 106 107 /// LandingPadInst associated with the invoke. 108 LandingPadInst *CallerLPad = nullptr; 109 110 /// PHI for EH values from landingpad insts. 111 PHINode *InnerEHValuesPHI = nullptr; 112 113 SmallVector<Value*, 8> UnwindDestPHIValues; 114 115 public: 116 LandingPadInliningInfo(InvokeInst *II) 117 : OuterResumeDest(II->getUnwindDest()) { 118 // If there are PHI nodes in the unwind destination block, we need to keep 119 // track of which values came into them from the invoke before removing 120 // the edge from this block. 121 BasicBlock *InvokeBB = II->getParent(); 122 BasicBlock::iterator I = OuterResumeDest->begin(); 123 for (; isa<PHINode>(I); ++I) { 124 // Save the value to use for this edge. 125 PHINode *PHI = cast<PHINode>(I); 126 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB)); 127 } 128 129 CallerLPad = cast<LandingPadInst>(I); 130 } 131 132 /// The outer unwind destination is the target of 133 /// unwind edges introduced for calls within the inlined function. 134 BasicBlock *getOuterResumeDest() const { 135 return OuterResumeDest; 136 } 137 138 BasicBlock *getInnerResumeDest(); 139 140 LandingPadInst *getLandingPadInst() const { return CallerLPad; } 141 142 /// Forward the 'resume' instruction to the caller's landing pad block. 143 /// When the landing pad block has only one predecessor, this is 144 /// a simple branch. When there is more than one predecessor, we need to 145 /// split the landing pad block after the landingpad instruction and jump 146 /// to there. 147 void forwardResume(ResumeInst *RI, 148 SmallPtrSetImpl<LandingPadInst*> &InlinedLPads); 149 150 /// Add incoming-PHI values to the unwind destination block for the given 151 /// basic block, using the values for the original invoke's source block. 152 void addIncomingPHIValuesFor(BasicBlock *BB) const { 153 addIncomingPHIValuesForInto(BB, OuterResumeDest); 154 } 155 156 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const { 157 BasicBlock::iterator I = dest->begin(); 158 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 159 PHINode *phi = cast<PHINode>(I); 160 phi->addIncoming(UnwindDestPHIValues[i], src); 161 } 162 } 163 }; 164 165 } // end anonymous namespace 166 167 /// Get or create a target for the branch from ResumeInsts. 168 BasicBlock *LandingPadInliningInfo::getInnerResumeDest() { 169 if (InnerResumeDest) return InnerResumeDest; 170 171 // Split the landing pad. 172 BasicBlock::iterator SplitPoint = ++CallerLPad->getIterator(); 173 InnerResumeDest = 174 OuterResumeDest->splitBasicBlock(SplitPoint, 175 OuterResumeDest->getName() + ".body"); 176 177 // The number of incoming edges we expect to the inner landing pad. 178 const unsigned PHICapacity = 2; 179 180 // Create corresponding new PHIs for all the PHIs in the outer landing pad. 181 Instruction *InsertPoint = &InnerResumeDest->front(); 182 BasicBlock::iterator I = OuterResumeDest->begin(); 183 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 184 PHINode *OuterPHI = cast<PHINode>(I); 185 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity, 186 OuterPHI->getName() + ".lpad-body", 187 InsertPoint); 188 OuterPHI->replaceAllUsesWith(InnerPHI); 189 InnerPHI->addIncoming(OuterPHI, OuterResumeDest); 190 } 191 192 // Create a PHI for the exception values. 193 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity, 194 "eh.lpad-body", InsertPoint); 195 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI); 196 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest); 197 198 // All done. 199 return InnerResumeDest; 200 } 201 202 /// Forward the 'resume' instruction to the caller's landing pad block. 203 /// When the landing pad block has only one predecessor, this is a simple 204 /// branch. When there is more than one predecessor, we need to split the 205 /// landing pad block after the landingpad instruction and jump to there. 206 void LandingPadInliningInfo::forwardResume( 207 ResumeInst *RI, SmallPtrSetImpl<LandingPadInst *> &InlinedLPads) { 208 BasicBlock *Dest = getInnerResumeDest(); 209 BasicBlock *Src = RI->getParent(); 210 211 BranchInst::Create(Dest, Src); 212 213 // Update the PHIs in the destination. They were inserted in an order which 214 // makes this work. 215 addIncomingPHIValuesForInto(Src, Dest); 216 217 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src); 218 RI->eraseFromParent(); 219 } 220 221 /// Helper for getUnwindDestToken/getUnwindDestTokenHelper. 222 static Value *getParentPad(Value *EHPad) { 223 if (auto *FPI = dyn_cast<FuncletPadInst>(EHPad)) 224 return FPI->getParentPad(); 225 return cast<CatchSwitchInst>(EHPad)->getParentPad(); 226 } 227 228 using UnwindDestMemoTy = DenseMap<Instruction *, Value *>; 229 230 /// Helper for getUnwindDestToken that does the descendant-ward part of 231 /// the search. 232 static Value *getUnwindDestTokenHelper(Instruction *EHPad, 233 UnwindDestMemoTy &MemoMap) { 234 SmallVector<Instruction *, 8> Worklist(1, EHPad); 235 236 while (!Worklist.empty()) { 237 Instruction *CurrentPad = Worklist.pop_back_val(); 238 // We only put pads on the worklist that aren't in the MemoMap. When 239 // we find an unwind dest for a pad we may update its ancestors, but 240 // the queue only ever contains uncles/great-uncles/etc. of CurrentPad, 241 // so they should never get updated while queued on the worklist. 242 assert(!MemoMap.count(CurrentPad)); 243 Value *UnwindDestToken = nullptr; 244 if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(CurrentPad)) { 245 if (CatchSwitch->hasUnwindDest()) { 246 UnwindDestToken = CatchSwitch->getUnwindDest()->getFirstNonPHI(); 247 } else { 248 // Catchswitch doesn't have a 'nounwind' variant, and one might be 249 // annotated as "unwinds to caller" when really it's nounwind (see 250 // e.g. SimplifyCFGOpt::SimplifyUnreachable), so we can't infer the 251 // parent's unwind dest from this. We can check its catchpads' 252 // descendants, since they might include a cleanuppad with an 253 // "unwinds to caller" cleanupret, which can be trusted. 254 for (auto HI = CatchSwitch->handler_begin(), 255 HE = CatchSwitch->handler_end(); 256 HI != HE && !UnwindDestToken; ++HI) { 257 BasicBlock *HandlerBlock = *HI; 258 auto *CatchPad = cast<CatchPadInst>(HandlerBlock->getFirstNonPHI()); 259 for (User *Child : CatchPad->users()) { 260 // Intentionally ignore invokes here -- since the catchswitch is 261 // marked "unwind to caller", it would be a verifier error if it 262 // contained an invoke which unwinds out of it, so any invoke we'd 263 // encounter must unwind to some child of the catch. 264 if (!isa<CleanupPadInst>(Child) && !isa<CatchSwitchInst>(Child)) 265 continue; 266 267 Instruction *ChildPad = cast<Instruction>(Child); 268 auto Memo = MemoMap.find(ChildPad); 269 if (Memo == MemoMap.end()) { 270 // Haven't figured out this child pad yet; queue it. 271 Worklist.push_back(ChildPad); 272 continue; 273 } 274 // We've already checked this child, but might have found that 275 // it offers no proof either way. 276 Value *ChildUnwindDestToken = Memo->second; 277 if (!ChildUnwindDestToken) 278 continue; 279 // We already know the child's unwind dest, which can either 280 // be ConstantTokenNone to indicate unwind to caller, or can 281 // be another child of the catchpad. Only the former indicates 282 // the unwind dest of the catchswitch. 283 if (isa<ConstantTokenNone>(ChildUnwindDestToken)) { 284 UnwindDestToken = ChildUnwindDestToken; 285 break; 286 } 287 assert(getParentPad(ChildUnwindDestToken) == CatchPad); 288 } 289 } 290 } 291 } else { 292 auto *CleanupPad = cast<CleanupPadInst>(CurrentPad); 293 for (User *U : CleanupPad->users()) { 294 if (auto *CleanupRet = dyn_cast<CleanupReturnInst>(U)) { 295 if (BasicBlock *RetUnwindDest = CleanupRet->getUnwindDest()) 296 UnwindDestToken = RetUnwindDest->getFirstNonPHI(); 297 else 298 UnwindDestToken = ConstantTokenNone::get(CleanupPad->getContext()); 299 break; 300 } 301 Value *ChildUnwindDestToken; 302 if (auto *Invoke = dyn_cast<InvokeInst>(U)) { 303 ChildUnwindDestToken = Invoke->getUnwindDest()->getFirstNonPHI(); 304 } else if (isa<CleanupPadInst>(U) || isa<CatchSwitchInst>(U)) { 305 Instruction *ChildPad = cast<Instruction>(U); 306 auto Memo = MemoMap.find(ChildPad); 307 if (Memo == MemoMap.end()) { 308 // Haven't resolved this child yet; queue it and keep searching. 309 Worklist.push_back(ChildPad); 310 continue; 311 } 312 // We've checked this child, but still need to ignore it if it 313 // had no proof either way. 314 ChildUnwindDestToken = Memo->second; 315 if (!ChildUnwindDestToken) 316 continue; 317 } else { 318 // Not a relevant user of the cleanuppad 319 continue; 320 } 321 // In a well-formed program, the child/invoke must either unwind to 322 // an(other) child of the cleanup, or exit the cleanup. In the 323 // first case, continue searching. 324 if (isa<Instruction>(ChildUnwindDestToken) && 325 getParentPad(ChildUnwindDestToken) == CleanupPad) 326 continue; 327 UnwindDestToken = ChildUnwindDestToken; 328 break; 329 } 330 } 331 // If we haven't found an unwind dest for CurrentPad, we may have queued its 332 // children, so move on to the next in the worklist. 333 if (!UnwindDestToken) 334 continue; 335 336 // Now we know that CurrentPad unwinds to UnwindDestToken. It also exits 337 // any ancestors of CurrentPad up to but not including UnwindDestToken's 338 // parent pad. Record this in the memo map, and check to see if the 339 // original EHPad being queried is one of the ones exited. 340 Value *UnwindParent; 341 if (auto *UnwindPad = dyn_cast<Instruction>(UnwindDestToken)) 342 UnwindParent = getParentPad(UnwindPad); 343 else 344 UnwindParent = nullptr; 345 bool ExitedOriginalPad = false; 346 for (Instruction *ExitedPad = CurrentPad; 347 ExitedPad && ExitedPad != UnwindParent; 348 ExitedPad = dyn_cast<Instruction>(getParentPad(ExitedPad))) { 349 // Skip over catchpads since they just follow their catchswitches. 350 if (isa<CatchPadInst>(ExitedPad)) 351 continue; 352 MemoMap[ExitedPad] = UnwindDestToken; 353 ExitedOriginalPad |= (ExitedPad == EHPad); 354 } 355 356 if (ExitedOriginalPad) 357 return UnwindDestToken; 358 359 // Continue the search. 360 } 361 362 // No definitive information is contained within this funclet. 363 return nullptr; 364 } 365 366 /// Given an EH pad, find where it unwinds. If it unwinds to an EH pad, 367 /// return that pad instruction. If it unwinds to caller, return 368 /// ConstantTokenNone. If it does not have a definitive unwind destination, 369 /// return nullptr. 370 /// 371 /// This routine gets invoked for calls in funclets in inlinees when inlining 372 /// an invoke. Since many funclets don't have calls inside them, it's queried 373 /// on-demand rather than building a map of pads to unwind dests up front. 374 /// Determining a funclet's unwind dest may require recursively searching its 375 /// descendants, and also ancestors and cousins if the descendants don't provide 376 /// an answer. Since most funclets will have their unwind dest immediately 377 /// available as the unwind dest of a catchswitch or cleanupret, this routine 378 /// searches top-down from the given pad and then up. To avoid worst-case 379 /// quadratic run-time given that approach, it uses a memo map to avoid 380 /// re-processing funclet trees. The callers that rewrite the IR as they go 381 /// take advantage of this, for correctness, by checking/forcing rewritten 382 /// pads' entries to match the original callee view. 383 static Value *getUnwindDestToken(Instruction *EHPad, 384 UnwindDestMemoTy &MemoMap) { 385 // Catchpads unwind to the same place as their catchswitch; 386 // redirct any queries on catchpads so the code below can 387 // deal with just catchswitches and cleanuppads. 388 if (auto *CPI = dyn_cast<CatchPadInst>(EHPad)) 389 EHPad = CPI->getCatchSwitch(); 390 391 // Check if we've already determined the unwind dest for this pad. 392 auto Memo = MemoMap.find(EHPad); 393 if (Memo != MemoMap.end()) 394 return Memo->second; 395 396 // Search EHPad and, if necessary, its descendants. 397 Value *UnwindDestToken = getUnwindDestTokenHelper(EHPad, MemoMap); 398 assert((UnwindDestToken == nullptr) != (MemoMap.count(EHPad) != 0)); 399 if (UnwindDestToken) 400 return UnwindDestToken; 401 402 // No information is available for this EHPad from itself or any of its 403 // descendants. An unwind all the way out to a pad in the caller would 404 // need also to agree with the unwind dest of the parent funclet, so 405 // search up the chain to try to find a funclet with information. Put 406 // null entries in the memo map to avoid re-processing as we go up. 407 MemoMap[EHPad] = nullptr; 408 #ifndef NDEBUG 409 SmallPtrSet<Instruction *, 4> TempMemos; 410 TempMemos.insert(EHPad); 411 #endif 412 Instruction *LastUselessPad = EHPad; 413 Value *AncestorToken; 414 for (AncestorToken = getParentPad(EHPad); 415 auto *AncestorPad = dyn_cast<Instruction>(AncestorToken); 416 AncestorToken = getParentPad(AncestorToken)) { 417 // Skip over catchpads since they just follow their catchswitches. 418 if (isa<CatchPadInst>(AncestorPad)) 419 continue; 420 // If the MemoMap had an entry mapping AncestorPad to nullptr, since we 421 // haven't yet called getUnwindDestTokenHelper for AncestorPad in this 422 // call to getUnwindDestToken, that would mean that AncestorPad had no 423 // information in itself, its descendants, or its ancestors. If that 424 // were the case, then we should also have recorded the lack of information 425 // for the descendant that we're coming from. So assert that we don't 426 // find a null entry in the MemoMap for AncestorPad. 427 assert(!MemoMap.count(AncestorPad) || MemoMap[AncestorPad]); 428 auto AncestorMemo = MemoMap.find(AncestorPad); 429 if (AncestorMemo == MemoMap.end()) { 430 UnwindDestToken = getUnwindDestTokenHelper(AncestorPad, MemoMap); 431 } else { 432 UnwindDestToken = AncestorMemo->second; 433 } 434 if (UnwindDestToken) 435 break; 436 LastUselessPad = AncestorPad; 437 MemoMap[LastUselessPad] = nullptr; 438 #ifndef NDEBUG 439 TempMemos.insert(LastUselessPad); 440 #endif 441 } 442 443 // We know that getUnwindDestTokenHelper was called on LastUselessPad and 444 // returned nullptr (and likewise for EHPad and any of its ancestors up to 445 // LastUselessPad), so LastUselessPad has no information from below. Since 446 // getUnwindDestTokenHelper must investigate all downward paths through 447 // no-information nodes to prove that a node has no information like this, 448 // and since any time it finds information it records it in the MemoMap for 449 // not just the immediately-containing funclet but also any ancestors also 450 // exited, it must be the case that, walking downward from LastUselessPad, 451 // visiting just those nodes which have not been mapped to an unwind dest 452 // by getUnwindDestTokenHelper (the nullptr TempMemos notwithstanding, since 453 // they are just used to keep getUnwindDestTokenHelper from repeating work), 454 // any node visited must have been exhaustively searched with no information 455 // for it found. 456 SmallVector<Instruction *, 8> Worklist(1, LastUselessPad); 457 while (!Worklist.empty()) { 458 Instruction *UselessPad = Worklist.pop_back_val(); 459 auto Memo = MemoMap.find(UselessPad); 460 if (Memo != MemoMap.end() && Memo->second) { 461 // Here the name 'UselessPad' is a bit of a misnomer, because we've found 462 // that it is a funclet that does have information about unwinding to 463 // a particular destination; its parent was a useless pad. 464 // Since its parent has no information, the unwind edge must not escape 465 // the parent, and must target a sibling of this pad. This local unwind 466 // gives us no information about EHPad. Leave it and the subtree rooted 467 // at it alone. 468 assert(getParentPad(Memo->second) == getParentPad(UselessPad)); 469 continue; 470 } 471 // We know we don't have information for UselesPad. If it has an entry in 472 // the MemoMap (mapping it to nullptr), it must be one of the TempMemos 473 // added on this invocation of getUnwindDestToken; if a previous invocation 474 // recorded nullptr, it would have had to prove that the ancestors of 475 // UselessPad, which include LastUselessPad, had no information, and that 476 // in turn would have required proving that the descendants of 477 // LastUselesPad, which include EHPad, have no information about 478 // LastUselessPad, which would imply that EHPad was mapped to nullptr in 479 // the MemoMap on that invocation, which isn't the case if we got here. 480 assert(!MemoMap.count(UselessPad) || TempMemos.count(UselessPad)); 481 // Assert as we enumerate users that 'UselessPad' doesn't have any unwind 482 // information that we'd be contradicting by making a map entry for it 483 // (which is something that getUnwindDestTokenHelper must have proved for 484 // us to get here). Just assert on is direct users here; the checks in 485 // this downward walk at its descendants will verify that they don't have 486 // any unwind edges that exit 'UselessPad' either (i.e. they either have no 487 // unwind edges or unwind to a sibling). 488 MemoMap[UselessPad] = UnwindDestToken; 489 if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(UselessPad)) { 490 assert(CatchSwitch->getUnwindDest() == nullptr && "Expected useless pad"); 491 for (BasicBlock *HandlerBlock : CatchSwitch->handlers()) { 492 auto *CatchPad = HandlerBlock->getFirstNonPHI(); 493 for (User *U : CatchPad->users()) { 494 assert( 495 (!isa<InvokeInst>(U) || 496 (getParentPad( 497 cast<InvokeInst>(U)->getUnwindDest()->getFirstNonPHI()) == 498 CatchPad)) && 499 "Expected useless pad"); 500 if (isa<CatchSwitchInst>(U) || isa<CleanupPadInst>(U)) 501 Worklist.push_back(cast<Instruction>(U)); 502 } 503 } 504 } else { 505 assert(isa<CleanupPadInst>(UselessPad)); 506 for (User *U : UselessPad->users()) { 507 assert(!isa<CleanupReturnInst>(U) && "Expected useless pad"); 508 assert((!isa<InvokeInst>(U) || 509 (getParentPad( 510 cast<InvokeInst>(U)->getUnwindDest()->getFirstNonPHI()) == 511 UselessPad)) && 512 "Expected useless pad"); 513 if (isa<CatchSwitchInst>(U) || isa<CleanupPadInst>(U)) 514 Worklist.push_back(cast<Instruction>(U)); 515 } 516 } 517 } 518 519 return UnwindDestToken; 520 } 521 522 /// When we inline a basic block into an invoke, 523 /// we have to turn all of the calls that can throw into invokes. 524 /// This function analyze BB to see if there are any calls, and if so, 525 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI 526 /// nodes in that block with the values specified in InvokeDestPHIValues. 527 static BasicBlock *HandleCallsInBlockInlinedThroughInvoke( 528 BasicBlock *BB, BasicBlock *UnwindEdge, 529 UnwindDestMemoTy *FuncletUnwindMap = nullptr) { 530 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 531 Instruction *I = &*BBI++; 532 533 // We only need to check for function calls: inlined invoke 534 // instructions require no special handling. 535 CallInst *CI = dyn_cast<CallInst>(I); 536 537 if (!CI || CI->doesNotThrow() || isa<InlineAsm>(CI->getCalledValue())) 538 continue; 539 540 // We do not need to (and in fact, cannot) convert possibly throwing calls 541 // to @llvm.experimental_deoptimize (resp. @llvm.experimental.guard) into 542 // invokes. The caller's "segment" of the deoptimization continuation 543 // attached to the newly inlined @llvm.experimental_deoptimize 544 // (resp. @llvm.experimental.guard) call should contain the exception 545 // handling logic, if any. 546 if (auto *F = CI->getCalledFunction()) 547 if (F->getIntrinsicID() == Intrinsic::experimental_deoptimize || 548 F->getIntrinsicID() == Intrinsic::experimental_guard) 549 continue; 550 551 if (auto FuncletBundle = CI->getOperandBundle(LLVMContext::OB_funclet)) { 552 // This call is nested inside a funclet. If that funclet has an unwind 553 // destination within the inlinee, then unwinding out of this call would 554 // be UB. Rewriting this call to an invoke which targets the inlined 555 // invoke's unwind dest would give the call's parent funclet multiple 556 // unwind destinations, which is something that subsequent EH table 557 // generation can't handle and that the veirifer rejects. So when we 558 // see such a call, leave it as a call. 559 auto *FuncletPad = cast<Instruction>(FuncletBundle->Inputs[0]); 560 Value *UnwindDestToken = 561 getUnwindDestToken(FuncletPad, *FuncletUnwindMap); 562 if (UnwindDestToken && !isa<ConstantTokenNone>(UnwindDestToken)) 563 continue; 564 #ifndef NDEBUG 565 Instruction *MemoKey; 566 if (auto *CatchPad = dyn_cast<CatchPadInst>(FuncletPad)) 567 MemoKey = CatchPad->getCatchSwitch(); 568 else 569 MemoKey = FuncletPad; 570 assert(FuncletUnwindMap->count(MemoKey) && 571 (*FuncletUnwindMap)[MemoKey] == UnwindDestToken && 572 "must get memoized to avoid confusing later searches"); 573 #endif // NDEBUG 574 } 575 576 changeToInvokeAndSplitBasicBlock(CI, UnwindEdge); 577 return BB; 578 } 579 return nullptr; 580 } 581 582 /// If we inlined an invoke site, we need to convert calls 583 /// in the body of the inlined function into invokes. 584 /// 585 /// II is the invoke instruction being inlined. FirstNewBlock is the first 586 /// block of the inlined code (the last block is the end of the function), 587 /// and InlineCodeInfo is information about the code that got inlined. 588 static void HandleInlinedLandingPad(InvokeInst *II, BasicBlock *FirstNewBlock, 589 ClonedCodeInfo &InlinedCodeInfo) { 590 BasicBlock *InvokeDest = II->getUnwindDest(); 591 592 Function *Caller = FirstNewBlock->getParent(); 593 594 // The inlined code is currently at the end of the function, scan from the 595 // start of the inlined code to its end, checking for stuff we need to 596 // rewrite. 597 LandingPadInliningInfo Invoke(II); 598 599 // Get all of the inlined landing pad instructions. 600 SmallPtrSet<LandingPadInst*, 16> InlinedLPads; 601 for (Function::iterator I = FirstNewBlock->getIterator(), E = Caller->end(); 602 I != E; ++I) 603 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) 604 InlinedLPads.insert(II->getLandingPadInst()); 605 606 // Append the clauses from the outer landing pad instruction into the inlined 607 // landing pad instructions. 608 LandingPadInst *OuterLPad = Invoke.getLandingPadInst(); 609 for (LandingPadInst *InlinedLPad : InlinedLPads) { 610 unsigned OuterNum = OuterLPad->getNumClauses(); 611 InlinedLPad->reserveClauses(OuterNum); 612 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx) 613 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx)); 614 if (OuterLPad->isCleanup()) 615 InlinedLPad->setCleanup(true); 616 } 617 618 for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end(); 619 BB != E; ++BB) { 620 if (InlinedCodeInfo.ContainsCalls) 621 if (BasicBlock *NewBB = HandleCallsInBlockInlinedThroughInvoke( 622 &*BB, Invoke.getOuterResumeDest())) 623 // Update any PHI nodes in the exceptional block to indicate that there 624 // is now a new entry in them. 625 Invoke.addIncomingPHIValuesFor(NewBB); 626 627 // Forward any resumes that are remaining here. 628 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) 629 Invoke.forwardResume(RI, InlinedLPads); 630 } 631 632 // Now that everything is happy, we have one final detail. The PHI nodes in 633 // the exception destination block still have entries due to the original 634 // invoke instruction. Eliminate these entries (which might even delete the 635 // PHI node) now. 636 InvokeDest->removePredecessor(II->getParent()); 637 } 638 639 /// If we inlined an invoke site, we need to convert calls 640 /// in the body of the inlined function into invokes. 641 /// 642 /// II is the invoke instruction being inlined. FirstNewBlock is the first 643 /// block of the inlined code (the last block is the end of the function), 644 /// and InlineCodeInfo is information about the code that got inlined. 645 static void HandleInlinedEHPad(InvokeInst *II, BasicBlock *FirstNewBlock, 646 ClonedCodeInfo &InlinedCodeInfo) { 647 BasicBlock *UnwindDest = II->getUnwindDest(); 648 Function *Caller = FirstNewBlock->getParent(); 649 650 assert(UnwindDest->getFirstNonPHI()->isEHPad() && "unexpected BasicBlock!"); 651 652 // If there are PHI nodes in the unwind destination block, we need to keep 653 // track of which values came into them from the invoke before removing the 654 // edge from this block. 655 SmallVector<Value *, 8> UnwindDestPHIValues; 656 BasicBlock *InvokeBB = II->getParent(); 657 for (Instruction &I : *UnwindDest) { 658 // Save the value to use for this edge. 659 PHINode *PHI = dyn_cast<PHINode>(&I); 660 if (!PHI) 661 break; 662 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB)); 663 } 664 665 // Add incoming-PHI values to the unwind destination block for the given basic 666 // block, using the values for the original invoke's source block. 667 auto UpdatePHINodes = [&](BasicBlock *Src) { 668 BasicBlock::iterator I = UnwindDest->begin(); 669 for (Value *V : UnwindDestPHIValues) { 670 PHINode *PHI = cast<PHINode>(I); 671 PHI->addIncoming(V, Src); 672 ++I; 673 } 674 }; 675 676 // This connects all the instructions which 'unwind to caller' to the invoke 677 // destination. 678 UnwindDestMemoTy FuncletUnwindMap; 679 for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end(); 680 BB != E; ++BB) { 681 if (auto *CRI = dyn_cast<CleanupReturnInst>(BB->getTerminator())) { 682 if (CRI->unwindsToCaller()) { 683 auto *CleanupPad = CRI->getCleanupPad(); 684 CleanupReturnInst::Create(CleanupPad, UnwindDest, CRI); 685 CRI->eraseFromParent(); 686 UpdatePHINodes(&*BB); 687 // Finding a cleanupret with an unwind destination would confuse 688 // subsequent calls to getUnwindDestToken, so map the cleanuppad 689 // to short-circuit any such calls and recognize this as an "unwind 690 // to caller" cleanup. 691 assert(!FuncletUnwindMap.count(CleanupPad) || 692 isa<ConstantTokenNone>(FuncletUnwindMap[CleanupPad])); 693 FuncletUnwindMap[CleanupPad] = 694 ConstantTokenNone::get(Caller->getContext()); 695 } 696 } 697 698 Instruction *I = BB->getFirstNonPHI(); 699 if (!I->isEHPad()) 700 continue; 701 702 Instruction *Replacement = nullptr; 703 if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I)) { 704 if (CatchSwitch->unwindsToCaller()) { 705 Value *UnwindDestToken; 706 if (auto *ParentPad = 707 dyn_cast<Instruction>(CatchSwitch->getParentPad())) { 708 // This catchswitch is nested inside another funclet. If that 709 // funclet has an unwind destination within the inlinee, then 710 // unwinding out of this catchswitch would be UB. Rewriting this 711 // catchswitch to unwind to the inlined invoke's unwind dest would 712 // give the parent funclet multiple unwind destinations, which is 713 // something that subsequent EH table generation can't handle and 714 // that the veirifer rejects. So when we see such a call, leave it 715 // as "unwind to caller". 716 UnwindDestToken = getUnwindDestToken(ParentPad, FuncletUnwindMap); 717 if (UnwindDestToken && !isa<ConstantTokenNone>(UnwindDestToken)) 718 continue; 719 } else { 720 // This catchswitch has no parent to inherit constraints from, and 721 // none of its descendants can have an unwind edge that exits it and 722 // targets another funclet in the inlinee. It may or may not have a 723 // descendant that definitively has an unwind to caller. In either 724 // case, we'll have to assume that any unwinds out of it may need to 725 // be routed to the caller, so treat it as though it has a definitive 726 // unwind to caller. 727 UnwindDestToken = ConstantTokenNone::get(Caller->getContext()); 728 } 729 auto *NewCatchSwitch = CatchSwitchInst::Create( 730 CatchSwitch->getParentPad(), UnwindDest, 731 CatchSwitch->getNumHandlers(), CatchSwitch->getName(), 732 CatchSwitch); 733 for (BasicBlock *PadBB : CatchSwitch->handlers()) 734 NewCatchSwitch->addHandler(PadBB); 735 // Propagate info for the old catchswitch over to the new one in 736 // the unwind map. This also serves to short-circuit any subsequent 737 // checks for the unwind dest of this catchswitch, which would get 738 // confused if they found the outer handler in the callee. 739 FuncletUnwindMap[NewCatchSwitch] = UnwindDestToken; 740 Replacement = NewCatchSwitch; 741 } 742 } else if (!isa<FuncletPadInst>(I)) { 743 llvm_unreachable("unexpected EHPad!"); 744 } 745 746 if (Replacement) { 747 Replacement->takeName(I); 748 I->replaceAllUsesWith(Replacement); 749 I->eraseFromParent(); 750 UpdatePHINodes(&*BB); 751 } 752 } 753 754 if (InlinedCodeInfo.ContainsCalls) 755 for (Function::iterator BB = FirstNewBlock->getIterator(), 756 E = Caller->end(); 757 BB != E; ++BB) 758 if (BasicBlock *NewBB = HandleCallsInBlockInlinedThroughInvoke( 759 &*BB, UnwindDest, &FuncletUnwindMap)) 760 // Update any PHI nodes in the exceptional block to indicate that there 761 // is now a new entry in them. 762 UpdatePHINodes(NewBB); 763 764 // Now that everything is happy, we have one final detail. The PHI nodes in 765 // the exception destination block still have entries due to the original 766 // invoke instruction. Eliminate these entries (which might even delete the 767 // PHI node) now. 768 UnwindDest->removePredecessor(InvokeBB); 769 } 770 771 /// When inlining a call site that has !llvm.mem.parallel_loop_access metadata, 772 /// that metadata should be propagated to all memory-accessing cloned 773 /// instructions. 774 static void PropagateParallelLoopAccessMetadata(CallSite CS, 775 ValueToValueMapTy &VMap) { 776 MDNode *M = 777 CS.getInstruction()->getMetadata(LLVMContext::MD_mem_parallel_loop_access); 778 if (!M) 779 return; 780 781 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end(); 782 VMI != VMIE; ++VMI) { 783 if (!VMI->second) 784 continue; 785 786 Instruction *NI = dyn_cast<Instruction>(VMI->second); 787 if (!NI) 788 continue; 789 790 if (MDNode *PM = NI->getMetadata(LLVMContext::MD_mem_parallel_loop_access)) { 791 M = MDNode::concatenate(PM, M); 792 NI->setMetadata(LLVMContext::MD_mem_parallel_loop_access, M); 793 } else if (NI->mayReadOrWriteMemory()) { 794 NI->setMetadata(LLVMContext::MD_mem_parallel_loop_access, M); 795 } 796 } 797 } 798 799 /// When inlining a function that contains noalias scope metadata, 800 /// this metadata needs to be cloned so that the inlined blocks 801 /// have different "unique scopes" at every call site. Were this not done, then 802 /// aliasing scopes from a function inlined into a caller multiple times could 803 /// not be differentiated (and this would lead to miscompiles because the 804 /// non-aliasing property communicated by the metadata could have 805 /// call-site-specific control dependencies). 806 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) { 807 const Function *CalledFunc = CS.getCalledFunction(); 808 SetVector<const MDNode *> MD; 809 810 // Note: We could only clone the metadata if it is already used in the 811 // caller. I'm omitting that check here because it might confuse 812 // inter-procedural alias analysis passes. We can revisit this if it becomes 813 // an efficiency or overhead problem. 814 815 for (const BasicBlock &I : *CalledFunc) 816 for (const Instruction &J : I) { 817 if (const MDNode *M = J.getMetadata(LLVMContext::MD_alias_scope)) 818 MD.insert(M); 819 if (const MDNode *M = J.getMetadata(LLVMContext::MD_noalias)) 820 MD.insert(M); 821 } 822 823 if (MD.empty()) 824 return; 825 826 // Walk the existing metadata, adding the complete (perhaps cyclic) chain to 827 // the set. 828 SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end()); 829 while (!Queue.empty()) { 830 const MDNode *M = cast<MDNode>(Queue.pop_back_val()); 831 for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i) 832 if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i))) 833 if (MD.insert(M1)) 834 Queue.push_back(M1); 835 } 836 837 // Now we have a complete set of all metadata in the chains used to specify 838 // the noalias scopes and the lists of those scopes. 839 SmallVector<TempMDTuple, 16> DummyNodes; 840 DenseMap<const MDNode *, TrackingMDNodeRef> MDMap; 841 for (const MDNode *I : MD) { 842 DummyNodes.push_back(MDTuple::getTemporary(CalledFunc->getContext(), None)); 843 MDMap[I].reset(DummyNodes.back().get()); 844 } 845 846 // Create new metadata nodes to replace the dummy nodes, replacing old 847 // metadata references with either a dummy node or an already-created new 848 // node. 849 for (const MDNode *I : MD) { 850 SmallVector<Metadata *, 4> NewOps; 851 for (unsigned i = 0, ie = I->getNumOperands(); i != ie; ++i) { 852 const Metadata *V = I->getOperand(i); 853 if (const MDNode *M = dyn_cast<MDNode>(V)) 854 NewOps.push_back(MDMap[M]); 855 else 856 NewOps.push_back(const_cast<Metadata *>(V)); 857 } 858 859 MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps); 860 MDTuple *TempM = cast<MDTuple>(MDMap[I]); 861 assert(TempM->isTemporary() && "Expected temporary node"); 862 863 TempM->replaceAllUsesWith(NewM); 864 } 865 866 // Now replace the metadata in the new inlined instructions with the 867 // repacements from the map. 868 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end(); 869 VMI != VMIE; ++VMI) { 870 if (!VMI->second) 871 continue; 872 873 Instruction *NI = dyn_cast<Instruction>(VMI->second); 874 if (!NI) 875 continue; 876 877 if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) { 878 MDNode *NewMD = MDMap[M]; 879 // If the call site also had alias scope metadata (a list of scopes to 880 // which instructions inside it might belong), propagate those scopes to 881 // the inlined instructions. 882 if (MDNode *CSM = 883 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope)) 884 NewMD = MDNode::concatenate(NewMD, CSM); 885 NI->setMetadata(LLVMContext::MD_alias_scope, NewMD); 886 } else if (NI->mayReadOrWriteMemory()) { 887 if (MDNode *M = 888 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope)) 889 NI->setMetadata(LLVMContext::MD_alias_scope, M); 890 } 891 892 if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) { 893 MDNode *NewMD = MDMap[M]; 894 // If the call site also had noalias metadata (a list of scopes with 895 // which instructions inside it don't alias), propagate those scopes to 896 // the inlined instructions. 897 if (MDNode *CSM = 898 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias)) 899 NewMD = MDNode::concatenate(NewMD, CSM); 900 NI->setMetadata(LLVMContext::MD_noalias, NewMD); 901 } else if (NI->mayReadOrWriteMemory()) { 902 if (MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_noalias)) 903 NI->setMetadata(LLVMContext::MD_noalias, M); 904 } 905 } 906 } 907 908 /// If the inlined function has noalias arguments, 909 /// then add new alias scopes for each noalias argument, tag the mapped noalias 910 /// parameters with noalias metadata specifying the new scope, and tag all 911 /// non-derived loads, stores and memory intrinsics with the new alias scopes. 912 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap, 913 const DataLayout &DL, AAResults *CalleeAAR) { 914 if (!EnableNoAliasConversion) 915 return; 916 917 const Function *CalledFunc = CS.getCalledFunction(); 918 SmallVector<const Argument *, 4> NoAliasArgs; 919 920 for (const Argument &Arg : CalledFunc->args()) 921 if (Arg.hasNoAliasAttr() && !Arg.use_empty()) 922 NoAliasArgs.push_back(&Arg); 923 924 if (NoAliasArgs.empty()) 925 return; 926 927 // To do a good job, if a noalias variable is captured, we need to know if 928 // the capture point dominates the particular use we're considering. 929 DominatorTree DT; 930 DT.recalculate(const_cast<Function&>(*CalledFunc)); 931 932 // noalias indicates that pointer values based on the argument do not alias 933 // pointer values which are not based on it. So we add a new "scope" for each 934 // noalias function argument. Accesses using pointers based on that argument 935 // become part of that alias scope, accesses using pointers not based on that 936 // argument are tagged as noalias with that scope. 937 938 DenseMap<const Argument *, MDNode *> NewScopes; 939 MDBuilder MDB(CalledFunc->getContext()); 940 941 // Create a new scope domain for this function. 942 MDNode *NewDomain = 943 MDB.createAnonymousAliasScopeDomain(CalledFunc->getName()); 944 for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) { 945 const Argument *A = NoAliasArgs[i]; 946 947 std::string Name = CalledFunc->getName(); 948 if (A->hasName()) { 949 Name += ": %"; 950 Name += A->getName(); 951 } else { 952 Name += ": argument "; 953 Name += utostr(i); 954 } 955 956 // Note: We always create a new anonymous root here. This is true regardless 957 // of the linkage of the callee because the aliasing "scope" is not just a 958 // property of the callee, but also all control dependencies in the caller. 959 MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name); 960 NewScopes.insert(std::make_pair(A, NewScope)); 961 } 962 963 // Iterate over all new instructions in the map; for all memory-access 964 // instructions, add the alias scope metadata. 965 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end(); 966 VMI != VMIE; ++VMI) { 967 if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) { 968 if (!VMI->second) 969 continue; 970 971 Instruction *NI = dyn_cast<Instruction>(VMI->second); 972 if (!NI) 973 continue; 974 975 bool IsArgMemOnlyCall = false, IsFuncCall = false; 976 SmallVector<const Value *, 2> PtrArgs; 977 978 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) 979 PtrArgs.push_back(LI->getPointerOperand()); 980 else if (const StoreInst *SI = dyn_cast<StoreInst>(I)) 981 PtrArgs.push_back(SI->getPointerOperand()); 982 else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I)) 983 PtrArgs.push_back(VAAI->getPointerOperand()); 984 else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I)) 985 PtrArgs.push_back(CXI->getPointerOperand()); 986 else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) 987 PtrArgs.push_back(RMWI->getPointerOperand()); 988 else if (ImmutableCallSite ICS = ImmutableCallSite(I)) { 989 // If we know that the call does not access memory, then we'll still 990 // know that about the inlined clone of this call site, and we don't 991 // need to add metadata. 992 if (ICS.doesNotAccessMemory()) 993 continue; 994 995 IsFuncCall = true; 996 if (CalleeAAR) { 997 FunctionModRefBehavior MRB = CalleeAAR->getModRefBehavior(ICS); 998 if (MRB == FMRB_OnlyAccessesArgumentPointees || 999 MRB == FMRB_OnlyReadsArgumentPointees) 1000 IsArgMemOnlyCall = true; 1001 } 1002 1003 for (Value *Arg : ICS.args()) { 1004 // We need to check the underlying objects of all arguments, not just 1005 // the pointer arguments, because we might be passing pointers as 1006 // integers, etc. 1007 // However, if we know that the call only accesses pointer arguments, 1008 // then we only need to check the pointer arguments. 1009 if (IsArgMemOnlyCall && !Arg->getType()->isPointerTy()) 1010 continue; 1011 1012 PtrArgs.push_back(Arg); 1013 } 1014 } 1015 1016 // If we found no pointers, then this instruction is not suitable for 1017 // pairing with an instruction to receive aliasing metadata. 1018 // However, if this is a call, this we might just alias with none of the 1019 // noalias arguments. 1020 if (PtrArgs.empty() && !IsFuncCall) 1021 continue; 1022 1023 // It is possible that there is only one underlying object, but you 1024 // need to go through several PHIs to see it, and thus could be 1025 // repeated in the Objects list. 1026 SmallPtrSet<const Value *, 4> ObjSet; 1027 SmallVector<Metadata *, 4> Scopes, NoAliases; 1028 1029 SmallSetVector<const Argument *, 4> NAPtrArgs; 1030 for (const Value *V : PtrArgs) { 1031 SmallVector<Value *, 4> Objects; 1032 GetUnderlyingObjects(const_cast<Value*>(V), 1033 Objects, DL, /* LI = */ nullptr); 1034 1035 for (Value *O : Objects) 1036 ObjSet.insert(O); 1037 } 1038 1039 // Figure out if we're derived from anything that is not a noalias 1040 // argument. 1041 bool CanDeriveViaCapture = false, UsesAliasingPtr = false; 1042 for (const Value *V : ObjSet) { 1043 // Is this value a constant that cannot be derived from any pointer 1044 // value (we need to exclude constant expressions, for example, that 1045 // are formed from arithmetic on global symbols). 1046 bool IsNonPtrConst = isa<ConstantInt>(V) || isa<ConstantFP>(V) || 1047 isa<ConstantPointerNull>(V) || 1048 isa<ConstantDataVector>(V) || isa<UndefValue>(V); 1049 if (IsNonPtrConst) 1050 continue; 1051 1052 // If this is anything other than a noalias argument, then we cannot 1053 // completely describe the aliasing properties using alias.scope 1054 // metadata (and, thus, won't add any). 1055 if (const Argument *A = dyn_cast<Argument>(V)) { 1056 if (!A->hasNoAliasAttr()) 1057 UsesAliasingPtr = true; 1058 } else { 1059 UsesAliasingPtr = true; 1060 } 1061 1062 // If this is not some identified function-local object (which cannot 1063 // directly alias a noalias argument), or some other argument (which, 1064 // by definition, also cannot alias a noalias argument), then we could 1065 // alias a noalias argument that has been captured). 1066 if (!isa<Argument>(V) && 1067 !isIdentifiedFunctionLocal(const_cast<Value*>(V))) 1068 CanDeriveViaCapture = true; 1069 } 1070 1071 // A function call can always get captured noalias pointers (via other 1072 // parameters, globals, etc.). 1073 if (IsFuncCall && !IsArgMemOnlyCall) 1074 CanDeriveViaCapture = true; 1075 1076 // First, we want to figure out all of the sets with which we definitely 1077 // don't alias. Iterate over all noalias set, and add those for which: 1078 // 1. The noalias argument is not in the set of objects from which we 1079 // definitely derive. 1080 // 2. The noalias argument has not yet been captured. 1081 // An arbitrary function that might load pointers could see captured 1082 // noalias arguments via other noalias arguments or globals, and so we 1083 // must always check for prior capture. 1084 for (const Argument *A : NoAliasArgs) { 1085 if (!ObjSet.count(A) && (!CanDeriveViaCapture || 1086 // It might be tempting to skip the 1087 // PointerMayBeCapturedBefore check if 1088 // A->hasNoCaptureAttr() is true, but this is 1089 // incorrect because nocapture only guarantees 1090 // that no copies outlive the function, not 1091 // that the value cannot be locally captured. 1092 !PointerMayBeCapturedBefore(A, 1093 /* ReturnCaptures */ false, 1094 /* StoreCaptures */ false, I, &DT))) 1095 NoAliases.push_back(NewScopes[A]); 1096 } 1097 1098 if (!NoAliases.empty()) 1099 NI->setMetadata(LLVMContext::MD_noalias, 1100 MDNode::concatenate( 1101 NI->getMetadata(LLVMContext::MD_noalias), 1102 MDNode::get(CalledFunc->getContext(), NoAliases))); 1103 1104 // Next, we want to figure out all of the sets to which we might belong. 1105 // We might belong to a set if the noalias argument is in the set of 1106 // underlying objects. If there is some non-noalias argument in our list 1107 // of underlying objects, then we cannot add a scope because the fact 1108 // that some access does not alias with any set of our noalias arguments 1109 // cannot itself guarantee that it does not alias with this access 1110 // (because there is some pointer of unknown origin involved and the 1111 // other access might also depend on this pointer). We also cannot add 1112 // scopes to arbitrary functions unless we know they don't access any 1113 // non-parameter pointer-values. 1114 bool CanAddScopes = !UsesAliasingPtr; 1115 if (CanAddScopes && IsFuncCall) 1116 CanAddScopes = IsArgMemOnlyCall; 1117 1118 if (CanAddScopes) 1119 for (const Argument *A : NoAliasArgs) { 1120 if (ObjSet.count(A)) 1121 Scopes.push_back(NewScopes[A]); 1122 } 1123 1124 if (!Scopes.empty()) 1125 NI->setMetadata( 1126 LLVMContext::MD_alias_scope, 1127 MDNode::concatenate(NI->getMetadata(LLVMContext::MD_alias_scope), 1128 MDNode::get(CalledFunc->getContext(), Scopes))); 1129 } 1130 } 1131 } 1132 1133 /// If the inlined function has non-byval align arguments, then 1134 /// add @llvm.assume-based alignment assumptions to preserve this information. 1135 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) { 1136 if (!PreserveAlignmentAssumptions || !IFI.GetAssumptionCache) 1137 return; 1138 1139 AssumptionCache *AC = &(*IFI.GetAssumptionCache)(*CS.getCaller()); 1140 auto &DL = CS.getCaller()->getParent()->getDataLayout(); 1141 1142 // To avoid inserting redundant assumptions, we should check for assumptions 1143 // already in the caller. To do this, we might need a DT of the caller. 1144 DominatorTree DT; 1145 bool DTCalculated = false; 1146 1147 Function *CalledFunc = CS.getCalledFunction(); 1148 for (Argument &Arg : CalledFunc->args()) { 1149 unsigned Align = Arg.getType()->isPointerTy() ? Arg.getParamAlignment() : 0; 1150 if (Align && !Arg.hasByValOrInAllocaAttr() && !Arg.hasNUses(0)) { 1151 if (!DTCalculated) { 1152 DT.recalculate(*CS.getCaller()); 1153 DTCalculated = true; 1154 } 1155 1156 // If we can already prove the asserted alignment in the context of the 1157 // caller, then don't bother inserting the assumption. 1158 Value *ArgVal = CS.getArgument(Arg.getArgNo()); 1159 if (getKnownAlignment(ArgVal, DL, CS.getInstruction(), AC, &DT) >= Align) 1160 continue; 1161 1162 CallInst *NewAsmp = IRBuilder<>(CS.getInstruction()) 1163 .CreateAlignmentAssumption(DL, ArgVal, Align); 1164 AC->registerAssumption(NewAsmp); 1165 } 1166 } 1167 } 1168 1169 /// Once we have cloned code over from a callee into the caller, 1170 /// update the specified callgraph to reflect the changes we made. 1171 /// Note that it's possible that not all code was copied over, so only 1172 /// some edges of the callgraph may remain. 1173 static void UpdateCallGraphAfterInlining(CallSite CS, 1174 Function::iterator FirstNewBlock, 1175 ValueToValueMapTy &VMap, 1176 InlineFunctionInfo &IFI) { 1177 CallGraph &CG = *IFI.CG; 1178 const Function *Caller = CS.getCaller(); 1179 const Function *Callee = CS.getCalledFunction(); 1180 CallGraphNode *CalleeNode = CG[Callee]; 1181 CallGraphNode *CallerNode = CG[Caller]; 1182 1183 // Since we inlined some uninlined call sites in the callee into the caller, 1184 // add edges from the caller to all of the callees of the callee. 1185 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 1186 1187 // Consider the case where CalleeNode == CallerNode. 1188 CallGraphNode::CalledFunctionsVector CallCache; 1189 if (CalleeNode == CallerNode) { 1190 CallCache.assign(I, E); 1191 I = CallCache.begin(); 1192 E = CallCache.end(); 1193 } 1194 1195 for (; I != E; ++I) { 1196 const Value *OrigCall = I->first; 1197 1198 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); 1199 // Only copy the edge if the call was inlined! 1200 if (VMI == VMap.end() || VMI->second == nullptr) 1201 continue; 1202 1203 // If the call was inlined, but then constant folded, there is no edge to 1204 // add. Check for this case. 1205 Instruction *NewCall = dyn_cast<Instruction>(VMI->second); 1206 if (!NewCall) 1207 continue; 1208 1209 // We do not treat intrinsic calls like real function calls because we 1210 // expect them to become inline code; do not add an edge for an intrinsic. 1211 CallSite CS = CallSite(NewCall); 1212 if (CS && CS.getCalledFunction() && CS.getCalledFunction()->isIntrinsic()) 1213 continue; 1214 1215 // Remember that this call site got inlined for the client of 1216 // InlineFunction. 1217 IFI.InlinedCalls.push_back(NewCall); 1218 1219 // It's possible that inlining the callsite will cause it to go from an 1220 // indirect to a direct call by resolving a function pointer. If this 1221 // happens, set the callee of the new call site to a more precise 1222 // destination. This can also happen if the call graph node of the caller 1223 // was just unnecessarily imprecise. 1224 if (!I->second->getFunction()) 1225 if (Function *F = CallSite(NewCall).getCalledFunction()) { 1226 // Indirect call site resolved to direct call. 1227 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); 1228 1229 continue; 1230 } 1231 1232 CallerNode->addCalledFunction(CallSite(NewCall), I->second); 1233 } 1234 1235 // Update the call graph by deleting the edge from Callee to Caller. We must 1236 // do this after the loop above in case Caller and Callee are the same. 1237 CallerNode->removeCallEdgeFor(CS); 1238 } 1239 1240 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M, 1241 BasicBlock *InsertBlock, 1242 InlineFunctionInfo &IFI) { 1243 Type *AggTy = cast<PointerType>(Src->getType())->getElementType(); 1244 IRBuilder<> Builder(InsertBlock, InsertBlock->begin()); 1245 1246 Value *Size = Builder.getInt64(M->getDataLayout().getTypeStoreSize(AggTy)); 1247 1248 // Always generate a memcpy of alignment 1 here because we don't know 1249 // the alignment of the src pointer. Other optimizations can infer 1250 // better alignment. 1251 Builder.CreateMemCpy(Dst, /*DstAlign*/1, Src, /*SrcAlign*/1, Size); 1252 } 1253 1254 /// When inlining a call site that has a byval argument, 1255 /// we have to make the implicit memcpy explicit by adding it. 1256 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, 1257 const Function *CalledFunc, 1258 InlineFunctionInfo &IFI, 1259 unsigned ByValAlignment) { 1260 PointerType *ArgTy = cast<PointerType>(Arg->getType()); 1261 Type *AggTy = ArgTy->getElementType(); 1262 1263 Function *Caller = TheCall->getFunction(); 1264 const DataLayout &DL = Caller->getParent()->getDataLayout(); 1265 1266 // If the called function is readonly, then it could not mutate the caller's 1267 // copy of the byval'd memory. In this case, it is safe to elide the copy and 1268 // temporary. 1269 if (CalledFunc->onlyReadsMemory()) { 1270 // If the byval argument has a specified alignment that is greater than the 1271 // passed in pointer, then we either have to round up the input pointer or 1272 // give up on this transformation. 1273 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. 1274 return Arg; 1275 1276 AssumptionCache *AC = 1277 IFI.GetAssumptionCache ? &(*IFI.GetAssumptionCache)(*Caller) : nullptr; 1278 1279 // If the pointer is already known to be sufficiently aligned, or if we can 1280 // round it up to a larger alignment, then we don't need a temporary. 1281 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, DL, TheCall, AC) >= 1282 ByValAlignment) 1283 return Arg; 1284 1285 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad 1286 // for code quality, but rarely happens and is required for correctness. 1287 } 1288 1289 // Create the alloca. If we have DataLayout, use nice alignment. 1290 unsigned Align = DL.getPrefTypeAlignment(AggTy); 1291 1292 // If the byval had an alignment specified, we *must* use at least that 1293 // alignment, as it is required by the byval argument (and uses of the 1294 // pointer inside the callee). 1295 Align = std::max(Align, ByValAlignment); 1296 1297 Value *NewAlloca = new AllocaInst(AggTy, DL.getAllocaAddrSpace(), 1298 nullptr, Align, Arg->getName(), 1299 &*Caller->begin()->begin()); 1300 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca)); 1301 1302 // Uses of the argument in the function should use our new alloca 1303 // instead. 1304 return NewAlloca; 1305 } 1306 1307 // Check whether this Value is used by a lifetime intrinsic. 1308 static bool isUsedByLifetimeMarker(Value *V) { 1309 for (User *U : V->users()) { 1310 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { 1311 switch (II->getIntrinsicID()) { 1312 default: break; 1313 case Intrinsic::lifetime_start: 1314 case Intrinsic::lifetime_end: 1315 return true; 1316 } 1317 } 1318 } 1319 return false; 1320 } 1321 1322 // Check whether the given alloca already has 1323 // lifetime.start or lifetime.end intrinsics. 1324 static bool hasLifetimeMarkers(AllocaInst *AI) { 1325 Type *Ty = AI->getType(); 1326 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(), 1327 Ty->getPointerAddressSpace()); 1328 if (Ty == Int8PtrTy) 1329 return isUsedByLifetimeMarker(AI); 1330 1331 // Do a scan to find all the casts to i8*. 1332 for (User *U : AI->users()) { 1333 if (U->getType() != Int8PtrTy) continue; 1334 if (U->stripPointerCasts() != AI) continue; 1335 if (isUsedByLifetimeMarker(U)) 1336 return true; 1337 } 1338 return false; 1339 } 1340 1341 /// Return the result of AI->isStaticAlloca() if AI were moved to the entry 1342 /// block. Allocas used in inalloca calls and allocas of dynamic array size 1343 /// cannot be static. 1344 static bool allocaWouldBeStaticInEntry(const AllocaInst *AI ) { 1345 return isa<Constant>(AI->getArraySize()) && !AI->isUsedWithInAlloca(); 1346 } 1347 1348 /// Update inlined instructions' line numbers to 1349 /// to encode location where these instructions are inlined. 1350 static void fixupLineNumbers(Function *Fn, Function::iterator FI, 1351 Instruction *TheCall, bool CalleeHasDebugInfo) { 1352 const DebugLoc &TheCallDL = TheCall->getDebugLoc(); 1353 if (!TheCallDL) 1354 return; 1355 1356 auto &Ctx = Fn->getContext(); 1357 DILocation *InlinedAtNode = TheCallDL; 1358 1359 // Create a unique call site, not to be confused with any other call from the 1360 // same location. 1361 InlinedAtNode = DILocation::getDistinct( 1362 Ctx, InlinedAtNode->getLine(), InlinedAtNode->getColumn(), 1363 InlinedAtNode->getScope(), InlinedAtNode->getInlinedAt()); 1364 1365 // Cache the inlined-at nodes as they're built so they are reused, without 1366 // this every instruction's inlined-at chain would become distinct from each 1367 // other. 1368 DenseMap<const MDNode *, MDNode *> IANodes; 1369 1370 for (; FI != Fn->end(); ++FI) { 1371 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end(); 1372 BI != BE; ++BI) { 1373 if (DebugLoc DL = BI->getDebugLoc()) { 1374 auto IA = DebugLoc::appendInlinedAt(DL, InlinedAtNode, BI->getContext(), 1375 IANodes); 1376 auto IDL = DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(), IA); 1377 BI->setDebugLoc(IDL); 1378 continue; 1379 } 1380 1381 if (CalleeHasDebugInfo) 1382 continue; 1383 1384 // If the inlined instruction has no line number, make it look as if it 1385 // originates from the call location. This is important for 1386 // ((__always_inline__, __nodebug__)) functions which must use caller 1387 // location for all instructions in their function body. 1388 1389 // Don't update static allocas, as they may get moved later. 1390 if (auto *AI = dyn_cast<AllocaInst>(BI)) 1391 if (allocaWouldBeStaticInEntry(AI)) 1392 continue; 1393 1394 BI->setDebugLoc(TheCallDL); 1395 } 1396 } 1397 } 1398 1399 /// Update the block frequencies of the caller after a callee has been inlined. 1400 /// 1401 /// Each block cloned into the caller has its block frequency scaled by the 1402 /// ratio of CallSiteFreq/CalleeEntryFreq. This ensures that the cloned copy of 1403 /// callee's entry block gets the same frequency as the callsite block and the 1404 /// relative frequencies of all cloned blocks remain the same after cloning. 1405 static void updateCallerBFI(BasicBlock *CallSiteBlock, 1406 const ValueToValueMapTy &VMap, 1407 BlockFrequencyInfo *CallerBFI, 1408 BlockFrequencyInfo *CalleeBFI, 1409 const BasicBlock &CalleeEntryBlock) { 1410 SmallPtrSet<BasicBlock *, 16> ClonedBBs; 1411 for (auto const &Entry : VMap) { 1412 if (!isa<BasicBlock>(Entry.first) || !Entry.second) 1413 continue; 1414 auto *OrigBB = cast<BasicBlock>(Entry.first); 1415 auto *ClonedBB = cast<BasicBlock>(Entry.second); 1416 uint64_t Freq = CalleeBFI->getBlockFreq(OrigBB).getFrequency(); 1417 if (!ClonedBBs.insert(ClonedBB).second) { 1418 // Multiple blocks in the callee might get mapped to one cloned block in 1419 // the caller since we prune the callee as we clone it. When that happens, 1420 // we want to use the maximum among the original blocks' frequencies. 1421 uint64_t NewFreq = CallerBFI->getBlockFreq(ClonedBB).getFrequency(); 1422 if (NewFreq > Freq) 1423 Freq = NewFreq; 1424 } 1425 CallerBFI->setBlockFreq(ClonedBB, Freq); 1426 } 1427 BasicBlock *EntryClone = cast<BasicBlock>(VMap.lookup(&CalleeEntryBlock)); 1428 CallerBFI->setBlockFreqAndScale( 1429 EntryClone, CallerBFI->getBlockFreq(CallSiteBlock).getFrequency(), 1430 ClonedBBs); 1431 } 1432 1433 /// Update the branch metadata for cloned call instructions. 1434 static void updateCallProfile(Function *Callee, const ValueToValueMapTy &VMap, 1435 const ProfileCount &CalleeEntryCount, 1436 const Instruction *TheCall, 1437 ProfileSummaryInfo *PSI, 1438 BlockFrequencyInfo *CallerBFI) { 1439 if (!CalleeEntryCount.hasValue() || CalleeEntryCount.isSynthetic() || 1440 CalleeEntryCount.getCount() < 1) 1441 return; 1442 auto CallSiteCount = PSI ? PSI->getProfileCount(TheCall, CallerBFI) : None; 1443 uint64_t CallCount = 1444 std::min(CallSiteCount.hasValue() ? CallSiteCount.getValue() : 0, 1445 CalleeEntryCount.getCount()); 1446 1447 for (auto const &Entry : VMap) 1448 if (isa<CallInst>(Entry.first)) 1449 if (auto *CI = dyn_cast_or_null<CallInst>(Entry.second)) 1450 CI->updateProfWeight(CallCount, CalleeEntryCount.getCount()); 1451 for (BasicBlock &BB : *Callee) 1452 // No need to update the callsite if it is pruned during inlining. 1453 if (VMap.count(&BB)) 1454 for (Instruction &I : BB) 1455 if (CallInst *CI = dyn_cast<CallInst>(&I)) 1456 CI->updateProfWeight(CalleeEntryCount.getCount() - CallCount, 1457 CalleeEntryCount.getCount()); 1458 } 1459 1460 /// Update the entry count of callee after inlining. 1461 /// 1462 /// The callsite's block count is subtracted from the callee's function entry 1463 /// count. 1464 static void updateCalleeCount(BlockFrequencyInfo *CallerBFI, BasicBlock *CallBB, 1465 Instruction *CallInst, Function *Callee, 1466 ProfileSummaryInfo *PSI) { 1467 // If the callee has a original count of N, and the estimated count of 1468 // callsite is M, the new callee count is set to N - M. M is estimated from 1469 // the caller's entry count, its entry block frequency and the block frequency 1470 // of the callsite. 1471 auto CalleeCount = Callee->getEntryCount(); 1472 if (!CalleeCount.hasValue() || !PSI) 1473 return; 1474 auto CallCount = PSI->getProfileCount(CallInst, CallerBFI); 1475 if (!CallCount.hasValue()) 1476 return; 1477 // Since CallSiteCount is an estimate, it could exceed the original callee 1478 // count and has to be set to 0. 1479 if (CallCount.getValue() > CalleeCount.getCount()) 1480 CalleeCount.setCount(0); 1481 else 1482 CalleeCount.setCount(CalleeCount.getCount() - CallCount.getValue()); 1483 Callee->setEntryCount(CalleeCount); 1484 } 1485 1486 /// This function inlines the called function into the basic block of the 1487 /// caller. This returns false if it is not possible to inline this call. 1488 /// The program is still in a well defined state if this occurs though. 1489 /// 1490 /// Note that this only does one level of inlining. For example, if the 1491 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 1492 /// exists in the instruction stream. Similarly this will inline a recursive 1493 /// function by one level. 1494 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI, 1495 AAResults *CalleeAAR, bool InsertLifetime, 1496 Function *ForwardVarArgsTo) { 1497 Instruction *TheCall = CS.getInstruction(); 1498 assert(TheCall->getParent() && TheCall->getFunction() 1499 && "Instruction not in function!"); 1500 1501 // If IFI has any state in it, zap it before we fill it in. 1502 IFI.reset(); 1503 1504 Function *CalledFunc = CS.getCalledFunction(); 1505 if (!CalledFunc || // Can't inline external function or indirect 1506 CalledFunc->isDeclaration()) // call! 1507 return false; 1508 1509 // The inliner does not know how to inline through calls with operand bundles 1510 // in general ... 1511 if (CS.hasOperandBundles()) { 1512 for (int i = 0, e = CS.getNumOperandBundles(); i != e; ++i) { 1513 uint32_t Tag = CS.getOperandBundleAt(i).getTagID(); 1514 // ... but it knows how to inline through "deopt" operand bundles ... 1515 if (Tag == LLVMContext::OB_deopt) 1516 continue; 1517 // ... and "funclet" operand bundles. 1518 if (Tag == LLVMContext::OB_funclet) 1519 continue; 1520 1521 return false; 1522 } 1523 } 1524 1525 // If the call to the callee cannot throw, set the 'nounwind' flag on any 1526 // calls that we inline. 1527 bool MarkNoUnwind = CS.doesNotThrow(); 1528 1529 BasicBlock *OrigBB = TheCall->getParent(); 1530 Function *Caller = OrigBB->getParent(); 1531 1532 // GC poses two hazards to inlining, which only occur when the callee has GC: 1533 // 1. If the caller has no GC, then the callee's GC must be propagated to the 1534 // caller. 1535 // 2. If the caller has a differing GC, it is invalid to inline. 1536 if (CalledFunc->hasGC()) { 1537 if (!Caller->hasGC()) 1538 Caller->setGC(CalledFunc->getGC()); 1539 else if (CalledFunc->getGC() != Caller->getGC()) 1540 return false; 1541 } 1542 1543 // Get the personality function from the callee if it contains a landing pad. 1544 Constant *CalledPersonality = 1545 CalledFunc->hasPersonalityFn() 1546 ? CalledFunc->getPersonalityFn()->stripPointerCasts() 1547 : nullptr; 1548 1549 // Find the personality function used by the landing pads of the caller. If it 1550 // exists, then check to see that it matches the personality function used in 1551 // the callee. 1552 Constant *CallerPersonality = 1553 Caller->hasPersonalityFn() 1554 ? Caller->getPersonalityFn()->stripPointerCasts() 1555 : nullptr; 1556 if (CalledPersonality) { 1557 if (!CallerPersonality) 1558 Caller->setPersonalityFn(CalledPersonality); 1559 // If the personality functions match, then we can perform the 1560 // inlining. Otherwise, we can't inline. 1561 // TODO: This isn't 100% true. Some personality functions are proper 1562 // supersets of others and can be used in place of the other. 1563 else if (CalledPersonality != CallerPersonality) 1564 return false; 1565 } 1566 1567 // We need to figure out which funclet the callsite was in so that we may 1568 // properly nest the callee. 1569 Instruction *CallSiteEHPad = nullptr; 1570 if (CallerPersonality) { 1571 EHPersonality Personality = classifyEHPersonality(CallerPersonality); 1572 if (isScopedEHPersonality(Personality)) { 1573 Optional<OperandBundleUse> ParentFunclet = 1574 CS.getOperandBundle(LLVMContext::OB_funclet); 1575 if (ParentFunclet) 1576 CallSiteEHPad = cast<FuncletPadInst>(ParentFunclet->Inputs.front()); 1577 1578 // OK, the inlining site is legal. What about the target function? 1579 1580 if (CallSiteEHPad) { 1581 if (Personality == EHPersonality::MSVC_CXX) { 1582 // The MSVC personality cannot tolerate catches getting inlined into 1583 // cleanup funclets. 1584 if (isa<CleanupPadInst>(CallSiteEHPad)) { 1585 // Ok, the call site is within a cleanuppad. Let's check the callee 1586 // for catchpads. 1587 for (const BasicBlock &CalledBB : *CalledFunc) { 1588 if (isa<CatchSwitchInst>(CalledBB.getFirstNonPHI())) 1589 return false; 1590 } 1591 } 1592 } else if (isAsynchronousEHPersonality(Personality)) { 1593 // SEH is even less tolerant, there may not be any sort of exceptional 1594 // funclet in the callee. 1595 for (const BasicBlock &CalledBB : *CalledFunc) { 1596 if (CalledBB.isEHPad()) 1597 return false; 1598 } 1599 } 1600 } 1601 } 1602 } 1603 1604 // Determine if we are dealing with a call in an EHPad which does not unwind 1605 // to caller. 1606 bool EHPadForCallUnwindsLocally = false; 1607 if (CallSiteEHPad && CS.isCall()) { 1608 UnwindDestMemoTy FuncletUnwindMap; 1609 Value *CallSiteUnwindDestToken = 1610 getUnwindDestToken(CallSiteEHPad, FuncletUnwindMap); 1611 1612 EHPadForCallUnwindsLocally = 1613 CallSiteUnwindDestToken && 1614 !isa<ConstantTokenNone>(CallSiteUnwindDestToken); 1615 } 1616 1617 // Get an iterator to the last basic block in the function, which will have 1618 // the new function inlined after it. 1619 Function::iterator LastBlock = --Caller->end(); 1620 1621 // Make sure to capture all of the return instructions from the cloned 1622 // function. 1623 SmallVector<ReturnInst*, 8> Returns; 1624 ClonedCodeInfo InlinedFunctionInfo; 1625 Function::iterator FirstNewBlock; 1626 1627 { // Scope to destroy VMap after cloning. 1628 ValueToValueMapTy VMap; 1629 // Keep a list of pair (dst, src) to emit byval initializations. 1630 SmallVector<std::pair<Value*, Value*>, 4> ByValInit; 1631 1632 auto &DL = Caller->getParent()->getDataLayout(); 1633 1634 // Calculate the vector of arguments to pass into the function cloner, which 1635 // matches up the formal to the actual argument values. 1636 CallSite::arg_iterator AI = CS.arg_begin(); 1637 unsigned ArgNo = 0; 1638 for (Function::arg_iterator I = CalledFunc->arg_begin(), 1639 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 1640 Value *ActualArg = *AI; 1641 1642 // When byval arguments actually inlined, we need to make the copy implied 1643 // by them explicit. However, we don't do this if the callee is readonly 1644 // or readnone, because the copy would be unneeded: the callee doesn't 1645 // modify the struct. 1646 if (CS.isByValArgument(ArgNo)) { 1647 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, 1648 CalledFunc->getParamAlignment(ArgNo)); 1649 if (ActualArg != *AI) 1650 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI)); 1651 } 1652 1653 VMap[&*I] = ActualArg; 1654 } 1655 1656 // Add alignment assumptions if necessary. We do this before the inlined 1657 // instructions are actually cloned into the caller so that we can easily 1658 // check what will be known at the start of the inlined code. 1659 AddAlignmentAssumptions(CS, IFI); 1660 1661 // We want the inliner to prune the code as it copies. We would LOVE to 1662 // have no dead or constant instructions leftover after inlining occurs 1663 // (which can happen, e.g., because an argument was constant), but we'll be 1664 // happy with whatever the cloner can do. 1665 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 1666 /*ModuleLevelChanges=*/false, Returns, ".i", 1667 &InlinedFunctionInfo, TheCall); 1668 // Remember the first block that is newly cloned over. 1669 FirstNewBlock = LastBlock; ++FirstNewBlock; 1670 1671 if (IFI.CallerBFI != nullptr && IFI.CalleeBFI != nullptr) 1672 // Update the BFI of blocks cloned into the caller. 1673 updateCallerBFI(OrigBB, VMap, IFI.CallerBFI, IFI.CalleeBFI, 1674 CalledFunc->front()); 1675 1676 updateCallProfile(CalledFunc, VMap, CalledFunc->getEntryCount(), TheCall, 1677 IFI.PSI, IFI.CallerBFI); 1678 // Update the profile count of callee. 1679 updateCalleeCount(IFI.CallerBFI, OrigBB, TheCall, CalledFunc, IFI.PSI); 1680 1681 // Inject byval arguments initialization. 1682 for (std::pair<Value*, Value*> &Init : ByValInit) 1683 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(), 1684 &*FirstNewBlock, IFI); 1685 1686 Optional<OperandBundleUse> ParentDeopt = 1687 CS.getOperandBundle(LLVMContext::OB_deopt); 1688 if (ParentDeopt) { 1689 SmallVector<OperandBundleDef, 2> OpDefs; 1690 1691 for (auto &VH : InlinedFunctionInfo.OperandBundleCallSites) { 1692 Instruction *I = dyn_cast_or_null<Instruction>(VH); 1693 if (!I) continue; // instruction was DCE'd or RAUW'ed to undef 1694 1695 OpDefs.clear(); 1696 1697 CallSite ICS(I); 1698 OpDefs.reserve(ICS.getNumOperandBundles()); 1699 1700 for (unsigned i = 0, e = ICS.getNumOperandBundles(); i < e; ++i) { 1701 auto ChildOB = ICS.getOperandBundleAt(i); 1702 if (ChildOB.getTagID() != LLVMContext::OB_deopt) { 1703 // If the inlined call has other operand bundles, let them be 1704 OpDefs.emplace_back(ChildOB); 1705 continue; 1706 } 1707 1708 // It may be useful to separate this logic (of handling operand 1709 // bundles) out to a separate "policy" component if this gets crowded. 1710 // Prepend the parent's deoptimization continuation to the newly 1711 // inlined call's deoptimization continuation. 1712 std::vector<Value *> MergedDeoptArgs; 1713 MergedDeoptArgs.reserve(ParentDeopt->Inputs.size() + 1714 ChildOB.Inputs.size()); 1715 1716 MergedDeoptArgs.insert(MergedDeoptArgs.end(), 1717 ParentDeopt->Inputs.begin(), 1718 ParentDeopt->Inputs.end()); 1719 MergedDeoptArgs.insert(MergedDeoptArgs.end(), ChildOB.Inputs.begin(), 1720 ChildOB.Inputs.end()); 1721 1722 OpDefs.emplace_back("deopt", std::move(MergedDeoptArgs)); 1723 } 1724 1725 Instruction *NewI = nullptr; 1726 if (isa<CallInst>(I)) 1727 NewI = CallInst::Create(cast<CallInst>(I), OpDefs, I); 1728 else 1729 NewI = InvokeInst::Create(cast<InvokeInst>(I), OpDefs, I); 1730 1731 // Note: the RAUW does the appropriate fixup in VMap, so we need to do 1732 // this even if the call returns void. 1733 I->replaceAllUsesWith(NewI); 1734 1735 VH = nullptr; 1736 I->eraseFromParent(); 1737 } 1738 } 1739 1740 // Update the callgraph if requested. 1741 if (IFI.CG) 1742 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); 1743 1744 // For 'nodebug' functions, the associated DISubprogram is always null. 1745 // Conservatively avoid propagating the callsite debug location to 1746 // instructions inlined from a function whose DISubprogram is not null. 1747 fixupLineNumbers(Caller, FirstNewBlock, TheCall, 1748 CalledFunc->getSubprogram() != nullptr); 1749 1750 // Clone existing noalias metadata if necessary. 1751 CloneAliasScopeMetadata(CS, VMap); 1752 1753 // Add noalias metadata if necessary. 1754 AddAliasScopeMetadata(CS, VMap, DL, CalleeAAR); 1755 1756 // Propagate llvm.mem.parallel_loop_access if necessary. 1757 PropagateParallelLoopAccessMetadata(CS, VMap); 1758 1759 // Register any cloned assumptions. 1760 if (IFI.GetAssumptionCache) 1761 for (BasicBlock &NewBlock : 1762 make_range(FirstNewBlock->getIterator(), Caller->end())) 1763 for (Instruction &I : NewBlock) { 1764 if (auto *II = dyn_cast<IntrinsicInst>(&I)) 1765 if (II->getIntrinsicID() == Intrinsic::assume) 1766 (*IFI.GetAssumptionCache)(*Caller).registerAssumption(II); 1767 } 1768 } 1769 1770 // If there are any alloca instructions in the block that used to be the entry 1771 // block for the callee, move them to the entry block of the caller. First 1772 // calculate which instruction they should be inserted before. We insert the 1773 // instructions at the end of the current alloca list. 1774 { 1775 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 1776 for (BasicBlock::iterator I = FirstNewBlock->begin(), 1777 E = FirstNewBlock->end(); I != E; ) { 1778 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 1779 if (!AI) continue; 1780 1781 // If the alloca is now dead, remove it. This often occurs due to code 1782 // specialization. 1783 if (AI->use_empty()) { 1784 AI->eraseFromParent(); 1785 continue; 1786 } 1787 1788 if (!allocaWouldBeStaticInEntry(AI)) 1789 continue; 1790 1791 // Keep track of the static allocas that we inline into the caller. 1792 IFI.StaticAllocas.push_back(AI); 1793 1794 // Scan for the block of allocas that we can move over, and move them 1795 // all at once. 1796 while (isa<AllocaInst>(I) && 1797 allocaWouldBeStaticInEntry(cast<AllocaInst>(I))) { 1798 IFI.StaticAllocas.push_back(cast<AllocaInst>(I)); 1799 ++I; 1800 } 1801 1802 // Transfer all of the allocas over in a block. Using splice means 1803 // that the instructions aren't removed from the symbol table, then 1804 // reinserted. 1805 Caller->getEntryBlock().getInstList().splice( 1806 InsertPoint, FirstNewBlock->getInstList(), AI->getIterator(), I); 1807 } 1808 // Move any dbg.declares describing the allocas into the entry basic block. 1809 DIBuilder DIB(*Caller->getParent()); 1810 for (auto &AI : IFI.StaticAllocas) 1811 replaceDbgDeclareForAlloca(AI, AI, DIB, DIExpression::NoDeref, 0, 1812 DIExpression::NoDeref); 1813 } 1814 1815 SmallVector<Value*,4> VarArgsToForward; 1816 SmallVector<AttributeSet, 4> VarArgsAttrs; 1817 for (unsigned i = CalledFunc->getFunctionType()->getNumParams(); 1818 i < CS.getNumArgOperands(); i++) { 1819 VarArgsToForward.push_back(CS.getArgOperand(i)); 1820 VarArgsAttrs.push_back(CS.getAttributes().getParamAttributes(i)); 1821 } 1822 1823 bool InlinedMustTailCalls = false, InlinedDeoptimizeCalls = false; 1824 if (InlinedFunctionInfo.ContainsCalls) { 1825 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None; 1826 if (CallInst *CI = dyn_cast<CallInst>(TheCall)) 1827 CallSiteTailKind = CI->getTailCallKind(); 1828 1829 // For inlining purposes, the "notail" marker is the same as no marker. 1830 if (CallSiteTailKind == CallInst::TCK_NoTail) 1831 CallSiteTailKind = CallInst::TCK_None; 1832 1833 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; 1834 ++BB) { 1835 for (auto II = BB->begin(); II != BB->end();) { 1836 Instruction &I = *II++; 1837 CallInst *CI = dyn_cast<CallInst>(&I); 1838 if (!CI) 1839 continue; 1840 1841 // Forward varargs from inlined call site to calls to the 1842 // ForwardVarArgsTo function, if requested, and to musttail calls. 1843 if (!VarArgsToForward.empty() && 1844 ((ForwardVarArgsTo && 1845 CI->getCalledFunction() == ForwardVarArgsTo) || 1846 CI->isMustTailCall())) { 1847 // Collect attributes for non-vararg parameters. 1848 AttributeList Attrs = CI->getAttributes(); 1849 SmallVector<AttributeSet, 8> ArgAttrs; 1850 if (!Attrs.isEmpty() || !VarArgsAttrs.empty()) { 1851 for (unsigned ArgNo = 0; 1852 ArgNo < CI->getFunctionType()->getNumParams(); ++ArgNo) 1853 ArgAttrs.push_back(Attrs.getParamAttributes(ArgNo)); 1854 } 1855 1856 // Add VarArg attributes. 1857 ArgAttrs.append(VarArgsAttrs.begin(), VarArgsAttrs.end()); 1858 Attrs = AttributeList::get(CI->getContext(), Attrs.getFnAttributes(), 1859 Attrs.getRetAttributes(), ArgAttrs); 1860 // Add VarArgs to existing parameters. 1861 SmallVector<Value *, 6> Params(CI->arg_operands()); 1862 Params.append(VarArgsToForward.begin(), VarArgsToForward.end()); 1863 CallInst *NewCI = 1864 CallInst::Create(CI->getCalledFunction() ? CI->getCalledFunction() 1865 : CI->getCalledValue(), 1866 Params, "", CI); 1867 NewCI->setDebugLoc(CI->getDebugLoc()); 1868 NewCI->setAttributes(Attrs); 1869 NewCI->setCallingConv(CI->getCallingConv()); 1870 CI->replaceAllUsesWith(NewCI); 1871 CI->eraseFromParent(); 1872 CI = NewCI; 1873 } 1874 1875 if (Function *F = CI->getCalledFunction()) 1876 InlinedDeoptimizeCalls |= 1877 F->getIntrinsicID() == Intrinsic::experimental_deoptimize; 1878 1879 // We need to reduce the strength of any inlined tail calls. For 1880 // musttail, we have to avoid introducing potential unbounded stack 1881 // growth. For example, if functions 'f' and 'g' are mutually recursive 1882 // with musttail, we can inline 'g' into 'f' so long as we preserve 1883 // musttail on the cloned call to 'f'. If either the inlined call site 1884 // or the cloned call site is *not* musttail, the program already has 1885 // one frame of stack growth, so it's safe to remove musttail. Here is 1886 // a table of example transformations: 1887 // 1888 // f -> musttail g -> musttail f ==> f -> musttail f 1889 // f -> musttail g -> tail f ==> f -> tail f 1890 // f -> g -> musttail f ==> f -> f 1891 // f -> g -> tail f ==> f -> f 1892 // 1893 // Inlined notail calls should remain notail calls. 1894 CallInst::TailCallKind ChildTCK = CI->getTailCallKind(); 1895 if (ChildTCK != CallInst::TCK_NoTail) 1896 ChildTCK = std::min(CallSiteTailKind, ChildTCK); 1897 CI->setTailCallKind(ChildTCK); 1898 InlinedMustTailCalls |= CI->isMustTailCall(); 1899 1900 // Calls inlined through a 'nounwind' call site should be marked 1901 // 'nounwind'. 1902 if (MarkNoUnwind) 1903 CI->setDoesNotThrow(); 1904 } 1905 } 1906 } 1907 1908 // Leave lifetime markers for the static alloca's, scoping them to the 1909 // function we just inlined. 1910 if (InsertLifetime && !IFI.StaticAllocas.empty()) { 1911 IRBuilder<> builder(&FirstNewBlock->front()); 1912 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) { 1913 AllocaInst *AI = IFI.StaticAllocas[ai]; 1914 // Don't mark swifterror allocas. They can't have bitcast uses. 1915 if (AI->isSwiftError()) 1916 continue; 1917 1918 // If the alloca is already scoped to something smaller than the whole 1919 // function then there's no need to add redundant, less accurate markers. 1920 if (hasLifetimeMarkers(AI)) 1921 continue; 1922 1923 // Try to determine the size of the allocation. 1924 ConstantInt *AllocaSize = nullptr; 1925 if (ConstantInt *AIArraySize = 1926 dyn_cast<ConstantInt>(AI->getArraySize())) { 1927 auto &DL = Caller->getParent()->getDataLayout(); 1928 Type *AllocaType = AI->getAllocatedType(); 1929 uint64_t AllocaTypeSize = DL.getTypeAllocSize(AllocaType); 1930 uint64_t AllocaArraySize = AIArraySize->getLimitedValue(); 1931 1932 // Don't add markers for zero-sized allocas. 1933 if (AllocaArraySize == 0) 1934 continue; 1935 1936 // Check that array size doesn't saturate uint64_t and doesn't 1937 // overflow when it's multiplied by type size. 1938 if (AllocaArraySize != std::numeric_limits<uint64_t>::max() && 1939 std::numeric_limits<uint64_t>::max() / AllocaArraySize >= 1940 AllocaTypeSize) { 1941 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()), 1942 AllocaArraySize * AllocaTypeSize); 1943 } 1944 } 1945 1946 builder.CreateLifetimeStart(AI, AllocaSize); 1947 for (ReturnInst *RI : Returns) { 1948 // Don't insert llvm.lifetime.end calls between a musttail or deoptimize 1949 // call and a return. The return kills all local allocas. 1950 if (InlinedMustTailCalls && 1951 RI->getParent()->getTerminatingMustTailCall()) 1952 continue; 1953 if (InlinedDeoptimizeCalls && 1954 RI->getParent()->getTerminatingDeoptimizeCall()) 1955 continue; 1956 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize); 1957 } 1958 } 1959 } 1960 1961 // If the inlined code contained dynamic alloca instructions, wrap the inlined 1962 // code with llvm.stacksave/llvm.stackrestore intrinsics. 1963 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 1964 Module *M = Caller->getParent(); 1965 // Get the two intrinsics we care about. 1966 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 1967 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 1968 1969 // Insert the llvm.stacksave. 1970 CallInst *SavedPtr = IRBuilder<>(&*FirstNewBlock, FirstNewBlock->begin()) 1971 .CreateCall(StackSave, {}, "savedstack"); 1972 1973 // Insert a call to llvm.stackrestore before any return instructions in the 1974 // inlined function. 1975 for (ReturnInst *RI : Returns) { 1976 // Don't insert llvm.stackrestore calls between a musttail or deoptimize 1977 // call and a return. The return will restore the stack pointer. 1978 if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall()) 1979 continue; 1980 if (InlinedDeoptimizeCalls && RI->getParent()->getTerminatingDeoptimizeCall()) 1981 continue; 1982 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr); 1983 } 1984 } 1985 1986 // If we are inlining for an invoke instruction, we must make sure to rewrite 1987 // any call instructions into invoke instructions. This is sensitive to which 1988 // funclet pads were top-level in the inlinee, so must be done before 1989 // rewriting the "parent pad" links. 1990 if (auto *II = dyn_cast<InvokeInst>(TheCall)) { 1991 BasicBlock *UnwindDest = II->getUnwindDest(); 1992 Instruction *FirstNonPHI = UnwindDest->getFirstNonPHI(); 1993 if (isa<LandingPadInst>(FirstNonPHI)) { 1994 HandleInlinedLandingPad(II, &*FirstNewBlock, InlinedFunctionInfo); 1995 } else { 1996 HandleInlinedEHPad(II, &*FirstNewBlock, InlinedFunctionInfo); 1997 } 1998 } 1999 2000 // Update the lexical scopes of the new funclets and callsites. 2001 // Anything that had 'none' as its parent is now nested inside the callsite's 2002 // EHPad. 2003 2004 if (CallSiteEHPad) { 2005 for (Function::iterator BB = FirstNewBlock->getIterator(), 2006 E = Caller->end(); 2007 BB != E; ++BB) { 2008 // Add bundle operands to any top-level call sites. 2009 SmallVector<OperandBundleDef, 1> OpBundles; 2010 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E;) { 2011 Instruction *I = &*BBI++; 2012 CallSite CS(I); 2013 if (!CS) 2014 continue; 2015 2016 // Skip call sites which are nounwind intrinsics. 2017 auto *CalledFn = 2018 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts()); 2019 if (CalledFn && CalledFn->isIntrinsic() && CS.doesNotThrow()) 2020 continue; 2021 2022 // Skip call sites which already have a "funclet" bundle. 2023 if (CS.getOperandBundle(LLVMContext::OB_funclet)) 2024 continue; 2025 2026 CS.getOperandBundlesAsDefs(OpBundles); 2027 OpBundles.emplace_back("funclet", CallSiteEHPad); 2028 2029 Instruction *NewInst; 2030 if (CS.isCall()) 2031 NewInst = CallInst::Create(cast<CallInst>(I), OpBundles, I); 2032 else 2033 NewInst = InvokeInst::Create(cast<InvokeInst>(I), OpBundles, I); 2034 NewInst->takeName(I); 2035 I->replaceAllUsesWith(NewInst); 2036 I->eraseFromParent(); 2037 2038 OpBundles.clear(); 2039 } 2040 2041 // It is problematic if the inlinee has a cleanupret which unwinds to 2042 // caller and we inline it into a call site which doesn't unwind but into 2043 // an EH pad that does. Such an edge must be dynamically unreachable. 2044 // As such, we replace the cleanupret with unreachable. 2045 if (auto *CleanupRet = dyn_cast<CleanupReturnInst>(BB->getTerminator())) 2046 if (CleanupRet->unwindsToCaller() && EHPadForCallUnwindsLocally) 2047 changeToUnreachable(CleanupRet, /*UseLLVMTrap=*/false); 2048 2049 Instruction *I = BB->getFirstNonPHI(); 2050 if (!I->isEHPad()) 2051 continue; 2052 2053 if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I)) { 2054 if (isa<ConstantTokenNone>(CatchSwitch->getParentPad())) 2055 CatchSwitch->setParentPad(CallSiteEHPad); 2056 } else { 2057 auto *FPI = cast<FuncletPadInst>(I); 2058 if (isa<ConstantTokenNone>(FPI->getParentPad())) 2059 FPI->setParentPad(CallSiteEHPad); 2060 } 2061 } 2062 } 2063 2064 if (InlinedDeoptimizeCalls) { 2065 // We need to at least remove the deoptimizing returns from the Return set, 2066 // so that the control flow from those returns does not get merged into the 2067 // caller (but terminate it instead). If the caller's return type does not 2068 // match the callee's return type, we also need to change the return type of 2069 // the intrinsic. 2070 if (Caller->getReturnType() == TheCall->getType()) { 2071 auto NewEnd = llvm::remove_if(Returns, [](ReturnInst *RI) { 2072 return RI->getParent()->getTerminatingDeoptimizeCall() != nullptr; 2073 }); 2074 Returns.erase(NewEnd, Returns.end()); 2075 } else { 2076 SmallVector<ReturnInst *, 8> NormalReturns; 2077 Function *NewDeoptIntrinsic = Intrinsic::getDeclaration( 2078 Caller->getParent(), Intrinsic::experimental_deoptimize, 2079 {Caller->getReturnType()}); 2080 2081 for (ReturnInst *RI : Returns) { 2082 CallInst *DeoptCall = RI->getParent()->getTerminatingDeoptimizeCall(); 2083 if (!DeoptCall) { 2084 NormalReturns.push_back(RI); 2085 continue; 2086 } 2087 2088 // The calling convention on the deoptimize call itself may be bogus, 2089 // since the code we're inlining may have undefined behavior (and may 2090 // never actually execute at runtime); but all 2091 // @llvm.experimental.deoptimize declarations have to have the same 2092 // calling convention in a well-formed module. 2093 auto CallingConv = DeoptCall->getCalledFunction()->getCallingConv(); 2094 NewDeoptIntrinsic->setCallingConv(CallingConv); 2095 auto *CurBB = RI->getParent(); 2096 RI->eraseFromParent(); 2097 2098 SmallVector<Value *, 4> CallArgs(DeoptCall->arg_begin(), 2099 DeoptCall->arg_end()); 2100 2101 SmallVector<OperandBundleDef, 1> OpBundles; 2102 DeoptCall->getOperandBundlesAsDefs(OpBundles); 2103 DeoptCall->eraseFromParent(); 2104 assert(!OpBundles.empty() && 2105 "Expected at least the deopt operand bundle"); 2106 2107 IRBuilder<> Builder(CurBB); 2108 CallInst *NewDeoptCall = 2109 Builder.CreateCall(NewDeoptIntrinsic, CallArgs, OpBundles); 2110 NewDeoptCall->setCallingConv(CallingConv); 2111 if (NewDeoptCall->getType()->isVoidTy()) 2112 Builder.CreateRetVoid(); 2113 else 2114 Builder.CreateRet(NewDeoptCall); 2115 } 2116 2117 // Leave behind the normal returns so we can merge control flow. 2118 std::swap(Returns, NormalReturns); 2119 } 2120 } 2121 2122 // Handle any inlined musttail call sites. In order for a new call site to be 2123 // musttail, the source of the clone and the inlined call site must have been 2124 // musttail. Therefore it's safe to return without merging control into the 2125 // phi below. 2126 if (InlinedMustTailCalls) { 2127 // Check if we need to bitcast the result of any musttail calls. 2128 Type *NewRetTy = Caller->getReturnType(); 2129 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy; 2130 2131 // Handle the returns preceded by musttail calls separately. 2132 SmallVector<ReturnInst *, 8> NormalReturns; 2133 for (ReturnInst *RI : Returns) { 2134 CallInst *ReturnedMustTail = 2135 RI->getParent()->getTerminatingMustTailCall(); 2136 if (!ReturnedMustTail) { 2137 NormalReturns.push_back(RI); 2138 continue; 2139 } 2140 if (!NeedBitCast) 2141 continue; 2142 2143 // Delete the old return and any preceding bitcast. 2144 BasicBlock *CurBB = RI->getParent(); 2145 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue()); 2146 RI->eraseFromParent(); 2147 if (OldCast) 2148 OldCast->eraseFromParent(); 2149 2150 // Insert a new bitcast and return with the right type. 2151 IRBuilder<> Builder(CurBB); 2152 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy)); 2153 } 2154 2155 // Leave behind the normal returns so we can merge control flow. 2156 std::swap(Returns, NormalReturns); 2157 } 2158 2159 // Now that all of the transforms on the inlined code have taken place but 2160 // before we splice the inlined code into the CFG and lose track of which 2161 // blocks were actually inlined, collect the call sites. We only do this if 2162 // call graph updates weren't requested, as those provide value handle based 2163 // tracking of inlined call sites instead. 2164 if (InlinedFunctionInfo.ContainsCalls && !IFI.CG) { 2165 // Otherwise just collect the raw call sites that were inlined. 2166 for (BasicBlock &NewBB : 2167 make_range(FirstNewBlock->getIterator(), Caller->end())) 2168 for (Instruction &I : NewBB) 2169 if (auto CS = CallSite(&I)) 2170 IFI.InlinedCallSites.push_back(CS); 2171 } 2172 2173 // If we cloned in _exactly one_ basic block, and if that block ends in a 2174 // return instruction, we splice the body of the inlined callee directly into 2175 // the calling basic block. 2176 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 2177 // Move all of the instructions right before the call. 2178 OrigBB->getInstList().splice(TheCall->getIterator(), 2179 FirstNewBlock->getInstList(), 2180 FirstNewBlock->begin(), FirstNewBlock->end()); 2181 // Remove the cloned basic block. 2182 Caller->getBasicBlockList().pop_back(); 2183 2184 // If the call site was an invoke instruction, add a branch to the normal 2185 // destination. 2186 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 2187 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 2188 NewBr->setDebugLoc(Returns[0]->getDebugLoc()); 2189 } 2190 2191 // If the return instruction returned a value, replace uses of the call with 2192 // uses of the returned value. 2193 if (!TheCall->use_empty()) { 2194 ReturnInst *R = Returns[0]; 2195 if (TheCall == R->getReturnValue()) 2196 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 2197 else 2198 TheCall->replaceAllUsesWith(R->getReturnValue()); 2199 } 2200 // Since we are now done with the Call/Invoke, we can delete it. 2201 TheCall->eraseFromParent(); 2202 2203 // Since we are now done with the return instruction, delete it also. 2204 Returns[0]->eraseFromParent(); 2205 2206 // We are now done with the inlining. 2207 return true; 2208 } 2209 2210 // Otherwise, we have the normal case, of more than one block to inline or 2211 // multiple return sites. 2212 2213 // We want to clone the entire callee function into the hole between the 2214 // "starter" and "ender" blocks. How we accomplish this depends on whether 2215 // this is an invoke instruction or a call instruction. 2216 BasicBlock *AfterCallBB; 2217 BranchInst *CreatedBranchToNormalDest = nullptr; 2218 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 2219 2220 // Add an unconditional branch to make this look like the CallInst case... 2221 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall); 2222 2223 // Split the basic block. This guarantees that no PHI nodes will have to be 2224 // updated due to new incoming edges, and make the invoke case more 2225 // symmetric to the call case. 2226 AfterCallBB = 2227 OrigBB->splitBasicBlock(CreatedBranchToNormalDest->getIterator(), 2228 CalledFunc->getName() + ".exit"); 2229 2230 } else { // It's a call 2231 // If this is a call instruction, we need to split the basic block that 2232 // the call lives in. 2233 // 2234 AfterCallBB = OrigBB->splitBasicBlock(TheCall->getIterator(), 2235 CalledFunc->getName() + ".exit"); 2236 } 2237 2238 if (IFI.CallerBFI) { 2239 // Copy original BB's block frequency to AfterCallBB 2240 IFI.CallerBFI->setBlockFreq( 2241 AfterCallBB, IFI.CallerBFI->getBlockFreq(OrigBB).getFrequency()); 2242 } 2243 2244 // Change the branch that used to go to AfterCallBB to branch to the first 2245 // basic block of the inlined function. 2246 // 2247 TerminatorInst *Br = OrigBB->getTerminator(); 2248 assert(Br && Br->getOpcode() == Instruction::Br && 2249 "splitBasicBlock broken!"); 2250 Br->setOperand(0, &*FirstNewBlock); 2251 2252 // Now that the function is correct, make it a little bit nicer. In 2253 // particular, move the basic blocks inserted from the end of the function 2254 // into the space made by splitting the source basic block. 2255 Caller->getBasicBlockList().splice(AfterCallBB->getIterator(), 2256 Caller->getBasicBlockList(), FirstNewBlock, 2257 Caller->end()); 2258 2259 // Handle all of the return instructions that we just cloned in, and eliminate 2260 // any users of the original call/invoke instruction. 2261 Type *RTy = CalledFunc->getReturnType(); 2262 2263 PHINode *PHI = nullptr; 2264 if (Returns.size() > 1) { 2265 // The PHI node should go at the front of the new basic block to merge all 2266 // possible incoming values. 2267 if (!TheCall->use_empty()) { 2268 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(), 2269 &AfterCallBB->front()); 2270 // Anything that used the result of the function call should now use the 2271 // PHI node as their operand. 2272 TheCall->replaceAllUsesWith(PHI); 2273 } 2274 2275 // Loop over all of the return instructions adding entries to the PHI node 2276 // as appropriate. 2277 if (PHI) { 2278 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 2279 ReturnInst *RI = Returns[i]; 2280 assert(RI->getReturnValue()->getType() == PHI->getType() && 2281 "Ret value not consistent in function!"); 2282 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 2283 } 2284 } 2285 2286 // Add a branch to the merge points and remove return instructions. 2287 DebugLoc Loc; 2288 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 2289 ReturnInst *RI = Returns[i]; 2290 BranchInst* BI = BranchInst::Create(AfterCallBB, RI); 2291 Loc = RI->getDebugLoc(); 2292 BI->setDebugLoc(Loc); 2293 RI->eraseFromParent(); 2294 } 2295 // We need to set the debug location to *somewhere* inside the 2296 // inlined function. The line number may be nonsensical, but the 2297 // instruction will at least be associated with the right 2298 // function. 2299 if (CreatedBranchToNormalDest) 2300 CreatedBranchToNormalDest->setDebugLoc(Loc); 2301 } else if (!Returns.empty()) { 2302 // Otherwise, if there is exactly one return value, just replace anything 2303 // using the return value of the call with the computed value. 2304 if (!TheCall->use_empty()) { 2305 if (TheCall == Returns[0]->getReturnValue()) 2306 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 2307 else 2308 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 2309 } 2310 2311 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 2312 BasicBlock *ReturnBB = Returns[0]->getParent(); 2313 ReturnBB->replaceAllUsesWith(AfterCallBB); 2314 2315 // Splice the code from the return block into the block that it will return 2316 // to, which contains the code that was after the call. 2317 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 2318 ReturnBB->getInstList()); 2319 2320 if (CreatedBranchToNormalDest) 2321 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc()); 2322 2323 // Delete the return instruction now and empty ReturnBB now. 2324 Returns[0]->eraseFromParent(); 2325 ReturnBB->eraseFromParent(); 2326 } else if (!TheCall->use_empty()) { 2327 // No returns, but something is using the return value of the call. Just 2328 // nuke the result. 2329 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 2330 } 2331 2332 // Since we are now done with the Call/Invoke, we can delete it. 2333 TheCall->eraseFromParent(); 2334 2335 // If we inlined any musttail calls and the original return is now 2336 // unreachable, delete it. It can only contain a bitcast and ret. 2337 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB)) 2338 AfterCallBB->eraseFromParent(); 2339 2340 // We should always be able to fold the entry block of the function into the 2341 // single predecessor of the block... 2342 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 2343 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 2344 2345 // Splice the code entry block into calling block, right before the 2346 // unconditional branch. 2347 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 2348 OrigBB->getInstList().splice(Br->getIterator(), CalleeEntry->getInstList()); 2349 2350 // Remove the unconditional branch. 2351 OrigBB->getInstList().erase(Br); 2352 2353 // Now we can remove the CalleeEntry block, which is now empty. 2354 Caller->getBasicBlockList().erase(CalleeEntry); 2355 2356 // If we inserted a phi node, check to see if it has a single value (e.g. all 2357 // the entries are the same or undef). If so, remove the PHI so it doesn't 2358 // block other optimizations. 2359 if (PHI) { 2360 AssumptionCache *AC = 2361 IFI.GetAssumptionCache ? &(*IFI.GetAssumptionCache)(*Caller) : nullptr; 2362 auto &DL = Caller->getParent()->getDataLayout(); 2363 if (Value *V = SimplifyInstruction(PHI, {DL, nullptr, nullptr, AC})) { 2364 PHI->replaceAllUsesWith(V); 2365 PHI->eraseFromParent(); 2366 } 2367 } 2368 2369 return true; 2370 } 2371