InlineFunction.cpp revision dce4a407a24b04eebc6a376f8e62b41aaa7b071f
1//===- InlineFunction.cpp - Code to perform function inlining -------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements inlining of a function into a call site, resolving 11// parameters and the return value as appropriate. 12// 13//===----------------------------------------------------------------------===// 14 15#include "llvm/Transforms/Utils/Cloning.h" 16#include "llvm/ADT/SmallVector.h" 17#include "llvm/ADT/StringExtras.h" 18#include "llvm/Analysis/CallGraph.h" 19#include "llvm/Analysis/InstructionSimplify.h" 20#include "llvm/IR/Attributes.h" 21#include "llvm/IR/CallSite.h" 22#include "llvm/IR/CFG.h" 23#include "llvm/IR/Constants.h" 24#include "llvm/IR/DataLayout.h" 25#include "llvm/IR/DebugInfo.h" 26#include "llvm/IR/DerivedTypes.h" 27#include "llvm/IR/IRBuilder.h" 28#include "llvm/IR/Instructions.h" 29#include "llvm/IR/IntrinsicInst.h" 30#include "llvm/IR/Intrinsics.h" 31#include "llvm/IR/Module.h" 32#include "llvm/Transforms/Utils/Local.h" 33using namespace llvm; 34 35bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI, 36 bool InsertLifetime) { 37 return InlineFunction(CallSite(CI), IFI, InsertLifetime); 38} 39bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI, 40 bool InsertLifetime) { 41 return InlineFunction(CallSite(II), IFI, InsertLifetime); 42} 43 44namespace { 45 /// A class for recording information about inlining through an invoke. 46 class InvokeInliningInfo { 47 BasicBlock *OuterResumeDest; ///< Destination of the invoke's unwind. 48 BasicBlock *InnerResumeDest; ///< Destination for the callee's resume. 49 LandingPadInst *CallerLPad; ///< LandingPadInst associated with the invoke. 50 PHINode *InnerEHValuesPHI; ///< PHI for EH values from landingpad insts. 51 SmallVector<Value*, 8> UnwindDestPHIValues; 52 53 public: 54 InvokeInliningInfo(InvokeInst *II) 55 : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(nullptr), 56 CallerLPad(nullptr), InnerEHValuesPHI(nullptr) { 57 // If there are PHI nodes in the unwind destination block, we need to keep 58 // track of which values came into them from the invoke before removing 59 // the edge from this block. 60 llvm::BasicBlock *InvokeBB = II->getParent(); 61 BasicBlock::iterator I = OuterResumeDest->begin(); 62 for (; isa<PHINode>(I); ++I) { 63 // Save the value to use for this edge. 64 PHINode *PHI = cast<PHINode>(I); 65 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB)); 66 } 67 68 CallerLPad = cast<LandingPadInst>(I); 69 } 70 71 /// getOuterResumeDest - The outer unwind destination is the target of 72 /// unwind edges introduced for calls within the inlined function. 73 BasicBlock *getOuterResumeDest() const { 74 return OuterResumeDest; 75 } 76 77 BasicBlock *getInnerResumeDest(); 78 79 LandingPadInst *getLandingPadInst() const { return CallerLPad; } 80 81 /// forwardResume - Forward the 'resume' instruction to the caller's landing 82 /// pad block. When the landing pad block has only one predecessor, this is 83 /// a simple branch. When there is more than one predecessor, we need to 84 /// split the landing pad block after the landingpad instruction and jump 85 /// to there. 86 void forwardResume(ResumeInst *RI, 87 SmallPtrSet<LandingPadInst*, 16> &InlinedLPads); 88 89 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind 90 /// destination block for the given basic block, using the values for the 91 /// original invoke's source block. 92 void addIncomingPHIValuesFor(BasicBlock *BB) const { 93 addIncomingPHIValuesForInto(BB, OuterResumeDest); 94 } 95 96 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const { 97 BasicBlock::iterator I = dest->begin(); 98 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 99 PHINode *phi = cast<PHINode>(I); 100 phi->addIncoming(UnwindDestPHIValues[i], src); 101 } 102 } 103 }; 104} 105 106/// getInnerResumeDest - Get or create a target for the branch from ResumeInsts. 107BasicBlock *InvokeInliningInfo::getInnerResumeDest() { 108 if (InnerResumeDest) return InnerResumeDest; 109 110 // Split the landing pad. 111 BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint; 112 InnerResumeDest = 113 OuterResumeDest->splitBasicBlock(SplitPoint, 114 OuterResumeDest->getName() + ".body"); 115 116 // The number of incoming edges we expect to the inner landing pad. 117 const unsigned PHICapacity = 2; 118 119 // Create corresponding new PHIs for all the PHIs in the outer landing pad. 120 BasicBlock::iterator InsertPoint = InnerResumeDest->begin(); 121 BasicBlock::iterator I = OuterResumeDest->begin(); 122 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 123 PHINode *OuterPHI = cast<PHINode>(I); 124 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity, 125 OuterPHI->getName() + ".lpad-body", 126 InsertPoint); 127 OuterPHI->replaceAllUsesWith(InnerPHI); 128 InnerPHI->addIncoming(OuterPHI, OuterResumeDest); 129 } 130 131 // Create a PHI for the exception values. 132 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity, 133 "eh.lpad-body", InsertPoint); 134 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI); 135 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest); 136 137 // All done. 138 return InnerResumeDest; 139} 140 141/// forwardResume - Forward the 'resume' instruction to the caller's landing pad 142/// block. When the landing pad block has only one predecessor, this is a simple 143/// branch. When there is more than one predecessor, we need to split the 144/// landing pad block after the landingpad instruction and jump to there. 145void InvokeInliningInfo::forwardResume(ResumeInst *RI, 146 SmallPtrSet<LandingPadInst*, 16> &InlinedLPads) { 147 BasicBlock *Dest = getInnerResumeDest(); 148 BasicBlock *Src = RI->getParent(); 149 150 BranchInst::Create(Dest, Src); 151 152 // Update the PHIs in the destination. They were inserted in an order which 153 // makes this work. 154 addIncomingPHIValuesForInto(Src, Dest); 155 156 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src); 157 RI->eraseFromParent(); 158} 159 160/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into 161/// an invoke, we have to turn all of the calls that can throw into 162/// invokes. This function analyze BB to see if there are any calls, and if so, 163/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI 164/// nodes in that block with the values specified in InvokeDestPHIValues. 165static void HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, 166 InvokeInliningInfo &Invoke) { 167 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 168 Instruction *I = BBI++; 169 170 // We only need to check for function calls: inlined invoke 171 // instructions require no special handling. 172 CallInst *CI = dyn_cast<CallInst>(I); 173 174 // If this call cannot unwind, don't convert it to an invoke. 175 // Inline asm calls cannot throw. 176 if (!CI || CI->doesNotThrow() || isa<InlineAsm>(CI->getCalledValue())) 177 continue; 178 179 // Convert this function call into an invoke instruction. First, split the 180 // basic block. 181 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 182 183 // Delete the unconditional branch inserted by splitBasicBlock 184 BB->getInstList().pop_back(); 185 186 // Create the new invoke instruction. 187 ImmutableCallSite CS(CI); 188 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end()); 189 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split, 190 Invoke.getOuterResumeDest(), 191 InvokeArgs, CI->getName(), BB); 192 II->setCallingConv(CI->getCallingConv()); 193 II->setAttributes(CI->getAttributes()); 194 195 // Make sure that anything using the call now uses the invoke! This also 196 // updates the CallGraph if present, because it uses a WeakVH. 197 CI->replaceAllUsesWith(II); 198 199 // Delete the original call 200 Split->getInstList().pop_front(); 201 202 // Update any PHI nodes in the exceptional block to indicate that there is 203 // now a new entry in them. 204 Invoke.addIncomingPHIValuesFor(BB); 205 return; 206 } 207} 208 209/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 210/// in the body of the inlined function into invokes. 211/// 212/// II is the invoke instruction being inlined. FirstNewBlock is the first 213/// block of the inlined code (the last block is the end of the function), 214/// and InlineCodeInfo is information about the code that got inlined. 215static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 216 ClonedCodeInfo &InlinedCodeInfo) { 217 BasicBlock *InvokeDest = II->getUnwindDest(); 218 219 Function *Caller = FirstNewBlock->getParent(); 220 221 // The inlined code is currently at the end of the function, scan from the 222 // start of the inlined code to its end, checking for stuff we need to 223 // rewrite. 224 InvokeInliningInfo Invoke(II); 225 226 // Get all of the inlined landing pad instructions. 227 SmallPtrSet<LandingPadInst*, 16> InlinedLPads; 228 for (Function::iterator I = FirstNewBlock, E = Caller->end(); I != E; ++I) 229 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) 230 InlinedLPads.insert(II->getLandingPadInst()); 231 232 // Append the clauses from the outer landing pad instruction into the inlined 233 // landing pad instructions. 234 LandingPadInst *OuterLPad = Invoke.getLandingPadInst(); 235 for (SmallPtrSet<LandingPadInst*, 16>::iterator I = InlinedLPads.begin(), 236 E = InlinedLPads.end(); I != E; ++I) { 237 LandingPadInst *InlinedLPad = *I; 238 unsigned OuterNum = OuterLPad->getNumClauses(); 239 InlinedLPad->reserveClauses(OuterNum); 240 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx) 241 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx)); 242 if (OuterLPad->isCleanup()) 243 InlinedLPad->setCleanup(true); 244 } 245 246 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){ 247 if (InlinedCodeInfo.ContainsCalls) 248 HandleCallsInBlockInlinedThroughInvoke(BB, Invoke); 249 250 // Forward any resumes that are remaining here. 251 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) 252 Invoke.forwardResume(RI, InlinedLPads); 253 } 254 255 // Now that everything is happy, we have one final detail. The PHI nodes in 256 // the exception destination block still have entries due to the original 257 // invoke instruction. Eliminate these entries (which might even delete the 258 // PHI node) now. 259 InvokeDest->removePredecessor(II->getParent()); 260} 261 262/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 263/// into the caller, update the specified callgraph to reflect the changes we 264/// made. Note that it's possible that not all code was copied over, so only 265/// some edges of the callgraph may remain. 266static void UpdateCallGraphAfterInlining(CallSite CS, 267 Function::iterator FirstNewBlock, 268 ValueToValueMapTy &VMap, 269 InlineFunctionInfo &IFI) { 270 CallGraph &CG = *IFI.CG; 271 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 272 const Function *Callee = CS.getCalledFunction(); 273 CallGraphNode *CalleeNode = CG[Callee]; 274 CallGraphNode *CallerNode = CG[Caller]; 275 276 // Since we inlined some uninlined call sites in the callee into the caller, 277 // add edges from the caller to all of the callees of the callee. 278 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 279 280 // Consider the case where CalleeNode == CallerNode. 281 CallGraphNode::CalledFunctionsVector CallCache; 282 if (CalleeNode == CallerNode) { 283 CallCache.assign(I, E); 284 I = CallCache.begin(); 285 E = CallCache.end(); 286 } 287 288 for (; I != E; ++I) { 289 const Value *OrigCall = I->first; 290 291 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); 292 // Only copy the edge if the call was inlined! 293 if (VMI == VMap.end() || VMI->second == nullptr) 294 continue; 295 296 // If the call was inlined, but then constant folded, there is no edge to 297 // add. Check for this case. 298 Instruction *NewCall = dyn_cast<Instruction>(VMI->second); 299 if (!NewCall) continue; 300 301 // Remember that this call site got inlined for the client of 302 // InlineFunction. 303 IFI.InlinedCalls.push_back(NewCall); 304 305 // It's possible that inlining the callsite will cause it to go from an 306 // indirect to a direct call by resolving a function pointer. If this 307 // happens, set the callee of the new call site to a more precise 308 // destination. This can also happen if the call graph node of the caller 309 // was just unnecessarily imprecise. 310 if (!I->second->getFunction()) 311 if (Function *F = CallSite(NewCall).getCalledFunction()) { 312 // Indirect call site resolved to direct call. 313 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); 314 315 continue; 316 } 317 318 CallerNode->addCalledFunction(CallSite(NewCall), I->second); 319 } 320 321 // Update the call graph by deleting the edge from Callee to Caller. We must 322 // do this after the loop above in case Caller and Callee are the same. 323 CallerNode->removeCallEdgeFor(CS); 324} 325 326static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M, 327 BasicBlock *InsertBlock, 328 InlineFunctionInfo &IFI) { 329 LLVMContext &Context = Src->getContext(); 330 Type *VoidPtrTy = Type::getInt8PtrTy(Context); 331 Type *AggTy = cast<PointerType>(Src->getType())->getElementType(); 332 Type *Tys[3] = { VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context) }; 333 Function *MemCpyFn = Intrinsic::getDeclaration(M, Intrinsic::memcpy, Tys); 334 IRBuilder<> builder(InsertBlock->begin()); 335 Value *DstCast = builder.CreateBitCast(Dst, VoidPtrTy, "tmp"); 336 Value *SrcCast = builder.CreateBitCast(Src, VoidPtrTy, "tmp"); 337 338 Value *Size; 339 if (IFI.DL == nullptr) 340 Size = ConstantExpr::getSizeOf(AggTy); 341 else 342 Size = ConstantInt::get(Type::getInt64Ty(Context), 343 IFI.DL->getTypeStoreSize(AggTy)); 344 345 // Always generate a memcpy of alignment 1 here because we don't know 346 // the alignment of the src pointer. Other optimizations can infer 347 // better alignment. 348 Value *CallArgs[] = { 349 DstCast, SrcCast, Size, 350 ConstantInt::get(Type::getInt32Ty(Context), 1), 351 ConstantInt::getFalse(Context) // isVolatile 352 }; 353 builder.CreateCall(MemCpyFn, CallArgs); 354} 355 356/// HandleByValArgument - When inlining a call site that has a byval argument, 357/// we have to make the implicit memcpy explicit by adding it. 358static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, 359 const Function *CalledFunc, 360 InlineFunctionInfo &IFI, 361 unsigned ByValAlignment) { 362 PointerType *ArgTy = cast<PointerType>(Arg->getType()); 363 Type *AggTy = ArgTy->getElementType(); 364 365 // If the called function is readonly, then it could not mutate the caller's 366 // copy of the byval'd memory. In this case, it is safe to elide the copy and 367 // temporary. 368 if (CalledFunc->onlyReadsMemory()) { 369 // If the byval argument has a specified alignment that is greater than the 370 // passed in pointer, then we either have to round up the input pointer or 371 // give up on this transformation. 372 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. 373 return Arg; 374 375 // If the pointer is already known to be sufficiently aligned, or if we can 376 // round it up to a larger alignment, then we don't need a temporary. 377 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, 378 IFI.DL) >= ByValAlignment) 379 return Arg; 380 381 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad 382 // for code quality, but rarely happens and is required for correctness. 383 } 384 385 // Create the alloca. If we have DataLayout, use nice alignment. 386 unsigned Align = 1; 387 if (IFI.DL) 388 Align = IFI.DL->getPrefTypeAlignment(AggTy); 389 390 // If the byval had an alignment specified, we *must* use at least that 391 // alignment, as it is required by the byval argument (and uses of the 392 // pointer inside the callee). 393 Align = std::max(Align, ByValAlignment); 394 395 Function *Caller = TheCall->getParent()->getParent(); 396 397 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(), 398 &*Caller->begin()->begin()); 399 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca)); 400 401 // Uses of the argument in the function should use our new alloca 402 // instead. 403 return NewAlloca; 404} 405 406// isUsedByLifetimeMarker - Check whether this Value is used by a lifetime 407// intrinsic. 408static bool isUsedByLifetimeMarker(Value *V) { 409 for (User *U : V->users()) { 410 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { 411 switch (II->getIntrinsicID()) { 412 default: break; 413 case Intrinsic::lifetime_start: 414 case Intrinsic::lifetime_end: 415 return true; 416 } 417 } 418 } 419 return false; 420} 421 422// hasLifetimeMarkers - Check whether the given alloca already has 423// lifetime.start or lifetime.end intrinsics. 424static bool hasLifetimeMarkers(AllocaInst *AI) { 425 Type *Ty = AI->getType(); 426 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(), 427 Ty->getPointerAddressSpace()); 428 if (Ty == Int8PtrTy) 429 return isUsedByLifetimeMarker(AI); 430 431 // Do a scan to find all the casts to i8*. 432 for (User *U : AI->users()) { 433 if (U->getType() != Int8PtrTy) continue; 434 if (U->stripPointerCasts() != AI) continue; 435 if (isUsedByLifetimeMarker(U)) 436 return true; 437 } 438 return false; 439} 440 441/// updateInlinedAtInfo - Helper function used by fixupLineNumbers to 442/// recursively update InlinedAtEntry of a DebugLoc. 443static DebugLoc updateInlinedAtInfo(const DebugLoc &DL, 444 const DebugLoc &InlinedAtDL, 445 LLVMContext &Ctx) { 446 if (MDNode *IA = DL.getInlinedAt(Ctx)) { 447 DebugLoc NewInlinedAtDL 448 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx); 449 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), 450 NewInlinedAtDL.getAsMDNode(Ctx)); 451 } 452 453 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), 454 InlinedAtDL.getAsMDNode(Ctx)); 455} 456 457/// fixupLineNumbers - Update inlined instructions' line numbers to 458/// to encode location where these instructions are inlined. 459static void fixupLineNumbers(Function *Fn, Function::iterator FI, 460 Instruction *TheCall) { 461 DebugLoc TheCallDL = TheCall->getDebugLoc(); 462 if (TheCallDL.isUnknown()) 463 return; 464 465 for (; FI != Fn->end(); ++FI) { 466 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end(); 467 BI != BE; ++BI) { 468 DebugLoc DL = BI->getDebugLoc(); 469 if (!DL.isUnknown()) { 470 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext())); 471 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) { 472 LLVMContext &Ctx = BI->getContext(); 473 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx); 474 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(), 475 InlinedAt, Ctx)); 476 } 477 } 478 } 479 } 480} 481 482/// Returns a musttail call instruction if one immediately precedes the given 483/// return instruction with an optional bitcast instruction between them. 484static CallInst *getPrecedingMustTailCall(ReturnInst *RI) { 485 Instruction *Prev = RI->getPrevNode(); 486 if (!Prev) 487 return nullptr; 488 489 if (Value *RV = RI->getReturnValue()) { 490 if (RV != Prev) 491 return nullptr; 492 493 // Look through the optional bitcast. 494 if (auto *BI = dyn_cast<BitCastInst>(Prev)) { 495 RV = BI->getOperand(0); 496 Prev = BI->getPrevNode(); 497 if (!Prev || RV != Prev) 498 return nullptr; 499 } 500 } 501 502 if (auto *CI = dyn_cast<CallInst>(Prev)) { 503 if (CI->isMustTailCall()) 504 return CI; 505 } 506 return nullptr; 507} 508 509/// InlineFunction - This function inlines the called function into the basic 510/// block of the caller. This returns false if it is not possible to inline 511/// this call. The program is still in a well defined state if this occurs 512/// though. 513/// 514/// Note that this only does one level of inlining. For example, if the 515/// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 516/// exists in the instruction stream. Similarly this will inline a recursive 517/// function by one level. 518bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI, 519 bool InsertLifetime) { 520 Instruction *TheCall = CS.getInstruction(); 521 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 522 "Instruction not in function!"); 523 524 // If IFI has any state in it, zap it before we fill it in. 525 IFI.reset(); 526 527 const Function *CalledFunc = CS.getCalledFunction(); 528 if (!CalledFunc || // Can't inline external function or indirect 529 CalledFunc->isDeclaration() || // call, or call to a vararg function! 530 CalledFunc->getFunctionType()->isVarArg()) return false; 531 532 // If the call to the callee cannot throw, set the 'nounwind' flag on any 533 // calls that we inline. 534 bool MarkNoUnwind = CS.doesNotThrow(); 535 536 BasicBlock *OrigBB = TheCall->getParent(); 537 Function *Caller = OrigBB->getParent(); 538 539 // GC poses two hazards to inlining, which only occur when the callee has GC: 540 // 1. If the caller has no GC, then the callee's GC must be propagated to the 541 // caller. 542 // 2. If the caller has a differing GC, it is invalid to inline. 543 if (CalledFunc->hasGC()) { 544 if (!Caller->hasGC()) 545 Caller->setGC(CalledFunc->getGC()); 546 else if (CalledFunc->getGC() != Caller->getGC()) 547 return false; 548 } 549 550 // Get the personality function from the callee if it contains a landing pad. 551 Value *CalleePersonality = nullptr; 552 for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end(); 553 I != E; ++I) 554 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) { 555 const BasicBlock *BB = II->getUnwindDest(); 556 const LandingPadInst *LP = BB->getLandingPadInst(); 557 CalleePersonality = LP->getPersonalityFn(); 558 break; 559 } 560 561 // Find the personality function used by the landing pads of the caller. If it 562 // exists, then check to see that it matches the personality function used in 563 // the callee. 564 if (CalleePersonality) { 565 for (Function::const_iterator I = Caller->begin(), E = Caller->end(); 566 I != E; ++I) 567 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) { 568 const BasicBlock *BB = II->getUnwindDest(); 569 const LandingPadInst *LP = BB->getLandingPadInst(); 570 571 // If the personality functions match, then we can perform the 572 // inlining. Otherwise, we can't inline. 573 // TODO: This isn't 100% true. Some personality functions are proper 574 // supersets of others and can be used in place of the other. 575 if (LP->getPersonalityFn() != CalleePersonality) 576 return false; 577 578 break; 579 } 580 } 581 582 // Get an iterator to the last basic block in the function, which will have 583 // the new function inlined after it. 584 Function::iterator LastBlock = &Caller->back(); 585 586 // Make sure to capture all of the return instructions from the cloned 587 // function. 588 SmallVector<ReturnInst*, 8> Returns; 589 ClonedCodeInfo InlinedFunctionInfo; 590 Function::iterator FirstNewBlock; 591 592 { // Scope to destroy VMap after cloning. 593 ValueToValueMapTy VMap; 594 // Keep a list of pair (dst, src) to emit byval initializations. 595 SmallVector<std::pair<Value*, Value*>, 4> ByValInit; 596 597 assert(CalledFunc->arg_size() == CS.arg_size() && 598 "No varargs calls can be inlined!"); 599 600 // Calculate the vector of arguments to pass into the function cloner, which 601 // matches up the formal to the actual argument values. 602 CallSite::arg_iterator AI = CS.arg_begin(); 603 unsigned ArgNo = 0; 604 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 605 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 606 Value *ActualArg = *AI; 607 608 // When byval arguments actually inlined, we need to make the copy implied 609 // by them explicit. However, we don't do this if the callee is readonly 610 // or readnone, because the copy would be unneeded: the callee doesn't 611 // modify the struct. 612 if (CS.isByValArgument(ArgNo)) { 613 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, 614 CalledFunc->getParamAlignment(ArgNo+1)); 615 if (ActualArg != *AI) 616 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI)); 617 } 618 619 VMap[I] = ActualArg; 620 } 621 622 // We want the inliner to prune the code as it copies. We would LOVE to 623 // have no dead or constant instructions leftover after inlining occurs 624 // (which can happen, e.g., because an argument was constant), but we'll be 625 // happy with whatever the cloner can do. 626 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 627 /*ModuleLevelChanges=*/false, Returns, ".i", 628 &InlinedFunctionInfo, IFI.DL, TheCall); 629 630 // Remember the first block that is newly cloned over. 631 FirstNewBlock = LastBlock; ++FirstNewBlock; 632 633 // Inject byval arguments initialization. 634 for (std::pair<Value*, Value*> &Init : ByValInit) 635 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(), 636 FirstNewBlock, IFI); 637 638 // Update the callgraph if requested. 639 if (IFI.CG) 640 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); 641 642 // Update inlined instructions' line number information. 643 fixupLineNumbers(Caller, FirstNewBlock, TheCall); 644 } 645 646 // If there are any alloca instructions in the block that used to be the entry 647 // block for the callee, move them to the entry block of the caller. First 648 // calculate which instruction they should be inserted before. We insert the 649 // instructions at the end of the current alloca list. 650 { 651 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 652 for (BasicBlock::iterator I = FirstNewBlock->begin(), 653 E = FirstNewBlock->end(); I != E; ) { 654 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 655 if (!AI) continue; 656 657 // If the alloca is now dead, remove it. This often occurs due to code 658 // specialization. 659 if (AI->use_empty()) { 660 AI->eraseFromParent(); 661 continue; 662 } 663 664 if (!isa<Constant>(AI->getArraySize())) 665 continue; 666 667 // Keep track of the static allocas that we inline into the caller. 668 IFI.StaticAllocas.push_back(AI); 669 670 // Scan for the block of allocas that we can move over, and move them 671 // all at once. 672 while (isa<AllocaInst>(I) && 673 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 674 IFI.StaticAllocas.push_back(cast<AllocaInst>(I)); 675 ++I; 676 } 677 678 // Transfer all of the allocas over in a block. Using splice means 679 // that the instructions aren't removed from the symbol table, then 680 // reinserted. 681 Caller->getEntryBlock().getInstList().splice(InsertPoint, 682 FirstNewBlock->getInstList(), 683 AI, I); 684 } 685 } 686 687 bool InlinedMustTailCalls = false; 688 if (InlinedFunctionInfo.ContainsCalls) { 689 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None; 690 if (CallInst *CI = dyn_cast<CallInst>(TheCall)) 691 CallSiteTailKind = CI->getTailCallKind(); 692 693 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; 694 ++BB) { 695 for (Instruction &I : *BB) { 696 CallInst *CI = dyn_cast<CallInst>(&I); 697 if (!CI) 698 continue; 699 700 // We need to reduce the strength of any inlined tail calls. For 701 // musttail, we have to avoid introducing potential unbounded stack 702 // growth. For example, if functions 'f' and 'g' are mutually recursive 703 // with musttail, we can inline 'g' into 'f' so long as we preserve 704 // musttail on the cloned call to 'f'. If either the inlined call site 705 // or the cloned call site is *not* musttail, the program already has 706 // one frame of stack growth, so it's safe to remove musttail. Here is 707 // a table of example transformations: 708 // 709 // f -> musttail g -> musttail f ==> f -> musttail f 710 // f -> musttail g -> tail f ==> f -> tail f 711 // f -> g -> musttail f ==> f -> f 712 // f -> g -> tail f ==> f -> f 713 CallInst::TailCallKind ChildTCK = CI->getTailCallKind(); 714 ChildTCK = std::min(CallSiteTailKind, ChildTCK); 715 CI->setTailCallKind(ChildTCK); 716 InlinedMustTailCalls |= CI->isMustTailCall(); 717 718 // Calls inlined through a 'nounwind' call site should be marked 719 // 'nounwind'. 720 if (MarkNoUnwind) 721 CI->setDoesNotThrow(); 722 } 723 } 724 } 725 726 // Leave lifetime markers for the static alloca's, scoping them to the 727 // function we just inlined. 728 if (InsertLifetime && !IFI.StaticAllocas.empty()) { 729 IRBuilder<> builder(FirstNewBlock->begin()); 730 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) { 731 AllocaInst *AI = IFI.StaticAllocas[ai]; 732 733 // If the alloca is already scoped to something smaller than the whole 734 // function then there's no need to add redundant, less accurate markers. 735 if (hasLifetimeMarkers(AI)) 736 continue; 737 738 // Try to determine the size of the allocation. 739 ConstantInt *AllocaSize = nullptr; 740 if (ConstantInt *AIArraySize = 741 dyn_cast<ConstantInt>(AI->getArraySize())) { 742 if (IFI.DL) { 743 Type *AllocaType = AI->getAllocatedType(); 744 uint64_t AllocaTypeSize = IFI.DL->getTypeAllocSize(AllocaType); 745 uint64_t AllocaArraySize = AIArraySize->getLimitedValue(); 746 assert(AllocaArraySize > 0 && "array size of AllocaInst is zero"); 747 // Check that array size doesn't saturate uint64_t and doesn't 748 // overflow when it's multiplied by type size. 749 if (AllocaArraySize != ~0ULL && 750 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) { 751 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()), 752 AllocaArraySize * AllocaTypeSize); 753 } 754 } 755 } 756 757 builder.CreateLifetimeStart(AI, AllocaSize); 758 for (ReturnInst *RI : Returns) { 759 // Don't insert llvm.lifetime.end calls between a musttail call and a 760 // return. The return kills all local allocas. 761 if (InlinedMustTailCalls && getPrecedingMustTailCall(RI)) 762 continue; 763 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize); 764 } 765 } 766 } 767 768 // If the inlined code contained dynamic alloca instructions, wrap the inlined 769 // code with llvm.stacksave/llvm.stackrestore intrinsics. 770 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 771 Module *M = Caller->getParent(); 772 // Get the two intrinsics we care about. 773 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 774 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 775 776 // Insert the llvm.stacksave. 777 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin()) 778 .CreateCall(StackSave, "savedstack"); 779 780 // Insert a call to llvm.stackrestore before any return instructions in the 781 // inlined function. 782 for (ReturnInst *RI : Returns) { 783 // Don't insert llvm.stackrestore calls between a musttail call and a 784 // return. The return will restore the stack pointer. 785 if (InlinedMustTailCalls && getPrecedingMustTailCall(RI)) 786 continue; 787 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr); 788 } 789 } 790 791 // If we are inlining for an invoke instruction, we must make sure to rewrite 792 // any call instructions into invoke instructions. 793 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 794 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 795 796 // Handle any inlined musttail call sites. In order for a new call site to be 797 // musttail, the source of the clone and the inlined call site must have been 798 // musttail. Therefore it's safe to return without merging control into the 799 // phi below. 800 if (InlinedMustTailCalls) { 801 // Check if we need to bitcast the result of any musttail calls. 802 Type *NewRetTy = Caller->getReturnType(); 803 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy; 804 805 // Handle the returns preceded by musttail calls separately. 806 SmallVector<ReturnInst *, 8> NormalReturns; 807 for (ReturnInst *RI : Returns) { 808 CallInst *ReturnedMustTail = getPrecedingMustTailCall(RI); 809 if (!ReturnedMustTail) { 810 NormalReturns.push_back(RI); 811 continue; 812 } 813 if (!NeedBitCast) 814 continue; 815 816 // Delete the old return and any preceding bitcast. 817 BasicBlock *CurBB = RI->getParent(); 818 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue()); 819 RI->eraseFromParent(); 820 if (OldCast) 821 OldCast->eraseFromParent(); 822 823 // Insert a new bitcast and return with the right type. 824 IRBuilder<> Builder(CurBB); 825 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy)); 826 } 827 828 // Leave behind the normal returns so we can merge control flow. 829 std::swap(Returns, NormalReturns); 830 } 831 832 // If we cloned in _exactly one_ basic block, and if that block ends in a 833 // return instruction, we splice the body of the inlined callee directly into 834 // the calling basic block. 835 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 836 // Move all of the instructions right before the call. 837 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 838 FirstNewBlock->begin(), FirstNewBlock->end()); 839 // Remove the cloned basic block. 840 Caller->getBasicBlockList().pop_back(); 841 842 // If the call site was an invoke instruction, add a branch to the normal 843 // destination. 844 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 845 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 846 NewBr->setDebugLoc(Returns[0]->getDebugLoc()); 847 } 848 849 // If the return instruction returned a value, replace uses of the call with 850 // uses of the returned value. 851 if (!TheCall->use_empty()) { 852 ReturnInst *R = Returns[0]; 853 if (TheCall == R->getReturnValue()) 854 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 855 else 856 TheCall->replaceAllUsesWith(R->getReturnValue()); 857 } 858 // Since we are now done with the Call/Invoke, we can delete it. 859 TheCall->eraseFromParent(); 860 861 // Since we are now done with the return instruction, delete it also. 862 Returns[0]->eraseFromParent(); 863 864 // We are now done with the inlining. 865 return true; 866 } 867 868 // Otherwise, we have the normal case, of more than one block to inline or 869 // multiple return sites. 870 871 // We want to clone the entire callee function into the hole between the 872 // "starter" and "ender" blocks. How we accomplish this depends on whether 873 // this is an invoke instruction or a call instruction. 874 BasicBlock *AfterCallBB; 875 BranchInst *CreatedBranchToNormalDest = nullptr; 876 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 877 878 // Add an unconditional branch to make this look like the CallInst case... 879 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall); 880 881 // Split the basic block. This guarantees that no PHI nodes will have to be 882 // updated due to new incoming edges, and make the invoke case more 883 // symmetric to the call case. 884 AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest, 885 CalledFunc->getName()+".exit"); 886 887 } else { // It's a call 888 // If this is a call instruction, we need to split the basic block that 889 // the call lives in. 890 // 891 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 892 CalledFunc->getName()+".exit"); 893 } 894 895 // Change the branch that used to go to AfterCallBB to branch to the first 896 // basic block of the inlined function. 897 // 898 TerminatorInst *Br = OrigBB->getTerminator(); 899 assert(Br && Br->getOpcode() == Instruction::Br && 900 "splitBasicBlock broken!"); 901 Br->setOperand(0, FirstNewBlock); 902 903 904 // Now that the function is correct, make it a little bit nicer. In 905 // particular, move the basic blocks inserted from the end of the function 906 // into the space made by splitting the source basic block. 907 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 908 FirstNewBlock, Caller->end()); 909 910 // Handle all of the return instructions that we just cloned in, and eliminate 911 // any users of the original call/invoke instruction. 912 Type *RTy = CalledFunc->getReturnType(); 913 914 PHINode *PHI = nullptr; 915 if (Returns.size() > 1) { 916 // The PHI node should go at the front of the new basic block to merge all 917 // possible incoming values. 918 if (!TheCall->use_empty()) { 919 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(), 920 AfterCallBB->begin()); 921 // Anything that used the result of the function call should now use the 922 // PHI node as their operand. 923 TheCall->replaceAllUsesWith(PHI); 924 } 925 926 // Loop over all of the return instructions adding entries to the PHI node 927 // as appropriate. 928 if (PHI) { 929 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 930 ReturnInst *RI = Returns[i]; 931 assert(RI->getReturnValue()->getType() == PHI->getType() && 932 "Ret value not consistent in function!"); 933 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 934 } 935 } 936 937 938 // Add a branch to the merge points and remove return instructions. 939 DebugLoc Loc; 940 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 941 ReturnInst *RI = Returns[i]; 942 BranchInst* BI = BranchInst::Create(AfterCallBB, RI); 943 Loc = RI->getDebugLoc(); 944 BI->setDebugLoc(Loc); 945 RI->eraseFromParent(); 946 } 947 // We need to set the debug location to *somewhere* inside the 948 // inlined function. The line number may be nonsensical, but the 949 // instruction will at least be associated with the right 950 // function. 951 if (CreatedBranchToNormalDest) 952 CreatedBranchToNormalDest->setDebugLoc(Loc); 953 } else if (!Returns.empty()) { 954 // Otherwise, if there is exactly one return value, just replace anything 955 // using the return value of the call with the computed value. 956 if (!TheCall->use_empty()) { 957 if (TheCall == Returns[0]->getReturnValue()) 958 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 959 else 960 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 961 } 962 963 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 964 BasicBlock *ReturnBB = Returns[0]->getParent(); 965 ReturnBB->replaceAllUsesWith(AfterCallBB); 966 967 // Splice the code from the return block into the block that it will return 968 // to, which contains the code that was after the call. 969 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 970 ReturnBB->getInstList()); 971 972 if (CreatedBranchToNormalDest) 973 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc()); 974 975 // Delete the return instruction now and empty ReturnBB now. 976 Returns[0]->eraseFromParent(); 977 ReturnBB->eraseFromParent(); 978 } else if (!TheCall->use_empty()) { 979 // No returns, but something is using the return value of the call. Just 980 // nuke the result. 981 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 982 } 983 984 // Since we are now done with the Call/Invoke, we can delete it. 985 TheCall->eraseFromParent(); 986 987 // If we inlined any musttail calls and the original return is now 988 // unreachable, delete it. It can only contain a bitcast and ret. 989 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB)) 990 AfterCallBB->eraseFromParent(); 991 992 // We should always be able to fold the entry block of the function into the 993 // single predecessor of the block... 994 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 995 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 996 997 // Splice the code entry block into calling block, right before the 998 // unconditional branch. 999 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 1000 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 1001 1002 // Remove the unconditional branch. 1003 OrigBB->getInstList().erase(Br); 1004 1005 // Now we can remove the CalleeEntry block, which is now empty. 1006 Caller->getBasicBlockList().erase(CalleeEntry); 1007 1008 // If we inserted a phi node, check to see if it has a single value (e.g. all 1009 // the entries are the same or undef). If so, remove the PHI so it doesn't 1010 // block other optimizations. 1011 if (PHI) { 1012 if (Value *V = SimplifyInstruction(PHI, IFI.DL)) { 1013 PHI->replaceAllUsesWith(V); 1014 PHI->eraseFromParent(); 1015 } 1016 } 1017 1018 return true; 1019} 1020