InlineFunction.cpp revision cd81d94322a39503e4a3e87b6ee03d4fcb3465fb
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->setDebugLoc(CI->getDebugLoc()); 193 II->setCallingConv(CI->getCallingConv()); 194 II->setAttributes(CI->getAttributes()); 195 196 // Make sure that anything using the call now uses the invoke! This also 197 // updates the CallGraph if present, because it uses a WeakVH. 198 CI->replaceAllUsesWith(II); 199 200 // Delete the original call 201 Split->getInstList().pop_front(); 202 203 // Update any PHI nodes in the exceptional block to indicate that there is 204 // now a new entry in them. 205 Invoke.addIncomingPHIValuesFor(BB); 206 return; 207 } 208} 209 210/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 211/// in the body of the inlined function into invokes. 212/// 213/// II is the invoke instruction being inlined. FirstNewBlock is the first 214/// block of the inlined code (the last block is the end of the function), 215/// and InlineCodeInfo is information about the code that got inlined. 216static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 217 ClonedCodeInfo &InlinedCodeInfo) { 218 BasicBlock *InvokeDest = II->getUnwindDest(); 219 220 Function *Caller = FirstNewBlock->getParent(); 221 222 // The inlined code is currently at the end of the function, scan from the 223 // start of the inlined code to its end, checking for stuff we need to 224 // rewrite. 225 InvokeInliningInfo Invoke(II); 226 227 // Get all of the inlined landing pad instructions. 228 SmallPtrSet<LandingPadInst*, 16> InlinedLPads; 229 for (Function::iterator I = FirstNewBlock, E = Caller->end(); I != E; ++I) 230 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) 231 InlinedLPads.insert(II->getLandingPadInst()); 232 233 // Append the clauses from the outer landing pad instruction into the inlined 234 // landing pad instructions. 235 LandingPadInst *OuterLPad = Invoke.getLandingPadInst(); 236 for (SmallPtrSet<LandingPadInst*, 16>::iterator I = InlinedLPads.begin(), 237 E = InlinedLPads.end(); I != E; ++I) { 238 LandingPadInst *InlinedLPad = *I; 239 unsigned OuterNum = OuterLPad->getNumClauses(); 240 InlinedLPad->reserveClauses(OuterNum); 241 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx) 242 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx)); 243 if (OuterLPad->isCleanup()) 244 InlinedLPad->setCleanup(true); 245 } 246 247 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){ 248 if (InlinedCodeInfo.ContainsCalls) 249 HandleCallsInBlockInlinedThroughInvoke(BB, Invoke); 250 251 // Forward any resumes that are remaining here. 252 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) 253 Invoke.forwardResume(RI, InlinedLPads); 254 } 255 256 // Now that everything is happy, we have one final detail. The PHI nodes in 257 // the exception destination block still have entries due to the original 258 // invoke instruction. Eliminate these entries (which might even delete the 259 // PHI node) now. 260 InvokeDest->removePredecessor(II->getParent()); 261} 262 263/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 264/// into the caller, update the specified callgraph to reflect the changes we 265/// made. Note that it's possible that not all code was copied over, so only 266/// some edges of the callgraph may remain. 267static void UpdateCallGraphAfterInlining(CallSite CS, 268 Function::iterator FirstNewBlock, 269 ValueToValueMapTy &VMap, 270 InlineFunctionInfo &IFI) { 271 CallGraph &CG = *IFI.CG; 272 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 273 const Function *Callee = CS.getCalledFunction(); 274 CallGraphNode *CalleeNode = CG[Callee]; 275 CallGraphNode *CallerNode = CG[Caller]; 276 277 // Since we inlined some uninlined call sites in the callee into the caller, 278 // add edges from the caller to all of the callees of the callee. 279 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 280 281 // Consider the case where CalleeNode == CallerNode. 282 CallGraphNode::CalledFunctionsVector CallCache; 283 if (CalleeNode == CallerNode) { 284 CallCache.assign(I, E); 285 I = CallCache.begin(); 286 E = CallCache.end(); 287 } 288 289 for (; I != E; ++I) { 290 const Value *OrigCall = I->first; 291 292 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); 293 // Only copy the edge if the call was inlined! 294 if (VMI == VMap.end() || VMI->second == nullptr) 295 continue; 296 297 // If the call was inlined, but then constant folded, there is no edge to 298 // add. Check for this case. 299 Instruction *NewCall = dyn_cast<Instruction>(VMI->second); 300 if (!NewCall) continue; 301 302 // Remember that this call site got inlined for the client of 303 // InlineFunction. 304 IFI.InlinedCalls.push_back(NewCall); 305 306 // It's possible that inlining the callsite will cause it to go from an 307 // indirect to a direct call by resolving a function pointer. If this 308 // happens, set the callee of the new call site to a more precise 309 // destination. This can also happen if the call graph node of the caller 310 // was just unnecessarily imprecise. 311 if (!I->second->getFunction()) 312 if (Function *F = CallSite(NewCall).getCalledFunction()) { 313 // Indirect call site resolved to direct call. 314 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); 315 316 continue; 317 } 318 319 CallerNode->addCalledFunction(CallSite(NewCall), I->second); 320 } 321 322 // Update the call graph by deleting the edge from Callee to Caller. We must 323 // do this after the loop above in case Caller and Callee are the same. 324 CallerNode->removeCallEdgeFor(CS); 325} 326 327static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M, 328 BasicBlock *InsertBlock, 329 InlineFunctionInfo &IFI) { 330 LLVMContext &Context = Src->getContext(); 331 Type *VoidPtrTy = Type::getInt8PtrTy(Context); 332 Type *AggTy = cast<PointerType>(Src->getType())->getElementType(); 333 Type *Tys[3] = { VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context) }; 334 Function *MemCpyFn = Intrinsic::getDeclaration(M, Intrinsic::memcpy, Tys); 335 IRBuilder<> builder(InsertBlock->begin()); 336 Value *DstCast = builder.CreateBitCast(Dst, VoidPtrTy, "tmp"); 337 Value *SrcCast = builder.CreateBitCast(Src, VoidPtrTy, "tmp"); 338 339 Value *Size; 340 if (IFI.DL == nullptr) 341 Size = ConstantExpr::getSizeOf(AggTy); 342 else 343 Size = ConstantInt::get(Type::getInt64Ty(Context), 344 IFI.DL->getTypeStoreSize(AggTy)); 345 346 // Always generate a memcpy of alignment 1 here because we don't know 347 // the alignment of the src pointer. Other optimizations can infer 348 // better alignment. 349 Value *CallArgs[] = { 350 DstCast, SrcCast, Size, 351 ConstantInt::get(Type::getInt32Ty(Context), 1), 352 ConstantInt::getFalse(Context) // isVolatile 353 }; 354 builder.CreateCall(MemCpyFn, CallArgs); 355} 356 357/// HandleByValArgument - When inlining a call site that has a byval argument, 358/// we have to make the implicit memcpy explicit by adding it. 359static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, 360 const Function *CalledFunc, 361 InlineFunctionInfo &IFI, 362 unsigned ByValAlignment) { 363 PointerType *ArgTy = cast<PointerType>(Arg->getType()); 364 Type *AggTy = ArgTy->getElementType(); 365 366 // If the called function is readonly, then it could not mutate the caller's 367 // copy of the byval'd memory. In this case, it is safe to elide the copy and 368 // temporary. 369 if (CalledFunc->onlyReadsMemory()) { 370 // If the byval argument has a specified alignment that is greater than the 371 // passed in pointer, then we either have to round up the input pointer or 372 // give up on this transformation. 373 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. 374 return Arg; 375 376 // If the pointer is already known to be sufficiently aligned, or if we can 377 // round it up to a larger alignment, then we don't need a temporary. 378 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, 379 IFI.DL) >= ByValAlignment) 380 return Arg; 381 382 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad 383 // for code quality, but rarely happens and is required for correctness. 384 } 385 386 // Create the alloca. If we have DataLayout, use nice alignment. 387 unsigned Align = 1; 388 if (IFI.DL) 389 Align = IFI.DL->getPrefTypeAlignment(AggTy); 390 391 // If the byval had an alignment specified, we *must* use at least that 392 // alignment, as it is required by the byval argument (and uses of the 393 // pointer inside the callee). 394 Align = std::max(Align, ByValAlignment); 395 396 Function *Caller = TheCall->getParent()->getParent(); 397 398 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(), 399 &*Caller->begin()->begin()); 400 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca)); 401 402 // Uses of the argument in the function should use our new alloca 403 // instead. 404 return NewAlloca; 405} 406 407// isUsedByLifetimeMarker - Check whether this Value is used by a lifetime 408// intrinsic. 409static bool isUsedByLifetimeMarker(Value *V) { 410 for (User *U : V->users()) { 411 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { 412 switch (II->getIntrinsicID()) { 413 default: break; 414 case Intrinsic::lifetime_start: 415 case Intrinsic::lifetime_end: 416 return true; 417 } 418 } 419 } 420 return false; 421} 422 423// hasLifetimeMarkers - Check whether the given alloca already has 424// lifetime.start or lifetime.end intrinsics. 425static bool hasLifetimeMarkers(AllocaInst *AI) { 426 Type *Ty = AI->getType(); 427 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(), 428 Ty->getPointerAddressSpace()); 429 if (Ty == Int8PtrTy) 430 return isUsedByLifetimeMarker(AI); 431 432 // Do a scan to find all the casts to i8*. 433 for (User *U : AI->users()) { 434 if (U->getType() != Int8PtrTy) continue; 435 if (U->stripPointerCasts() != AI) continue; 436 if (isUsedByLifetimeMarker(U)) 437 return true; 438 } 439 return false; 440} 441 442/// updateInlinedAtInfo - Helper function used by fixupLineNumbers to 443/// recursively update InlinedAtEntry of a DebugLoc. 444static DebugLoc updateInlinedAtInfo(const DebugLoc &DL, 445 const DebugLoc &InlinedAtDL, 446 LLVMContext &Ctx) { 447 if (MDNode *IA = DL.getInlinedAt(Ctx)) { 448 DebugLoc NewInlinedAtDL 449 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx); 450 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), 451 NewInlinedAtDL.getAsMDNode(Ctx)); 452 } 453 454 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), 455 InlinedAtDL.getAsMDNode(Ctx)); 456} 457 458/// fixupLineNumbers - Update inlined instructions' line numbers to 459/// to encode location where these instructions are inlined. 460static void fixupLineNumbers(Function *Fn, Function::iterator FI, 461 Instruction *TheCall) { 462 DebugLoc TheCallDL = TheCall->getDebugLoc(); 463 if (TheCallDL.isUnknown()) 464 return; 465 466 for (; FI != Fn->end(); ++FI) { 467 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end(); 468 BI != BE; ++BI) { 469 DebugLoc DL = BI->getDebugLoc(); 470 if (DL.isUnknown()) { 471 // If the inlined instruction has no line number, make it look as if it 472 // originates from the call location. This is important for 473 // ((__always_inline__, __nodebug__)) functions which must use caller 474 // location for all instructions in their function body. 475 BI->setDebugLoc(TheCallDL); 476 } else { 477 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext())); 478 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) { 479 LLVMContext &Ctx = BI->getContext(); 480 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx); 481 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(), 482 InlinedAt, Ctx)); 483 } 484 } 485 } 486 } 487} 488 489/// Returns a musttail call instruction if one immediately precedes the given 490/// return instruction with an optional bitcast instruction between them. 491static CallInst *getPrecedingMustTailCall(ReturnInst *RI) { 492 Instruction *Prev = RI->getPrevNode(); 493 if (!Prev) 494 return nullptr; 495 496 if (Value *RV = RI->getReturnValue()) { 497 if (RV != Prev) 498 return nullptr; 499 500 // Look through the optional bitcast. 501 if (auto *BI = dyn_cast<BitCastInst>(Prev)) { 502 RV = BI->getOperand(0); 503 Prev = BI->getPrevNode(); 504 if (!Prev || RV != Prev) 505 return nullptr; 506 } 507 } 508 509 if (auto *CI = dyn_cast<CallInst>(Prev)) { 510 if (CI->isMustTailCall()) 511 return CI; 512 } 513 return nullptr; 514} 515 516/// InlineFunction - This function inlines the called function into the basic 517/// block of the caller. This returns false if it is not possible to inline 518/// this call. The program is still in a well defined state if this occurs 519/// though. 520/// 521/// Note that this only does one level of inlining. For example, if the 522/// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 523/// exists in the instruction stream. Similarly this will inline a recursive 524/// function by one level. 525bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI, 526 bool InsertLifetime) { 527 Instruction *TheCall = CS.getInstruction(); 528 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 529 "Instruction not in function!"); 530 531 // If IFI has any state in it, zap it before we fill it in. 532 IFI.reset(); 533 534 const Function *CalledFunc = CS.getCalledFunction(); 535 if (!CalledFunc || // Can't inline external function or indirect 536 CalledFunc->isDeclaration() || // call, or call to a vararg function! 537 CalledFunc->getFunctionType()->isVarArg()) return false; 538 539 // If the call to the callee cannot throw, set the 'nounwind' flag on any 540 // calls that we inline. 541 bool MarkNoUnwind = CS.doesNotThrow(); 542 543 BasicBlock *OrigBB = TheCall->getParent(); 544 Function *Caller = OrigBB->getParent(); 545 546 // GC poses two hazards to inlining, which only occur when the callee has GC: 547 // 1. If the caller has no GC, then the callee's GC must be propagated to the 548 // caller. 549 // 2. If the caller has a differing GC, it is invalid to inline. 550 if (CalledFunc->hasGC()) { 551 if (!Caller->hasGC()) 552 Caller->setGC(CalledFunc->getGC()); 553 else if (CalledFunc->getGC() != Caller->getGC()) 554 return false; 555 } 556 557 // Get the personality function from the callee if it contains a landing pad. 558 Value *CalleePersonality = nullptr; 559 for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end(); 560 I != E; ++I) 561 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) { 562 const BasicBlock *BB = II->getUnwindDest(); 563 const LandingPadInst *LP = BB->getLandingPadInst(); 564 CalleePersonality = LP->getPersonalityFn(); 565 break; 566 } 567 568 // Find the personality function used by the landing pads of the caller. If it 569 // exists, then check to see that it matches the personality function used in 570 // the callee. 571 if (CalleePersonality) { 572 for (Function::const_iterator I = Caller->begin(), E = Caller->end(); 573 I != E; ++I) 574 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) { 575 const BasicBlock *BB = II->getUnwindDest(); 576 const LandingPadInst *LP = BB->getLandingPadInst(); 577 578 // If the personality functions match, then we can perform the 579 // inlining. Otherwise, we can't inline. 580 // TODO: This isn't 100% true. Some personality functions are proper 581 // supersets of others and can be used in place of the other. 582 if (LP->getPersonalityFn() != CalleePersonality) 583 return false; 584 585 break; 586 } 587 } 588 589 // Get an iterator to the last basic block in the function, which will have 590 // the new function inlined after it. 591 Function::iterator LastBlock = &Caller->back(); 592 593 // Make sure to capture all of the return instructions from the cloned 594 // function. 595 SmallVector<ReturnInst*, 8> Returns; 596 ClonedCodeInfo InlinedFunctionInfo; 597 Function::iterator FirstNewBlock; 598 599 { // Scope to destroy VMap after cloning. 600 ValueToValueMapTy VMap; 601 // Keep a list of pair (dst, src) to emit byval initializations. 602 SmallVector<std::pair<Value*, Value*>, 4> ByValInit; 603 604 assert(CalledFunc->arg_size() == CS.arg_size() && 605 "No varargs calls can be inlined!"); 606 607 // Calculate the vector of arguments to pass into the function cloner, which 608 // matches up the formal to the actual argument values. 609 CallSite::arg_iterator AI = CS.arg_begin(); 610 unsigned ArgNo = 0; 611 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 612 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 613 Value *ActualArg = *AI; 614 615 // When byval arguments actually inlined, we need to make the copy implied 616 // by them explicit. However, we don't do this if the callee is readonly 617 // or readnone, because the copy would be unneeded: the callee doesn't 618 // modify the struct. 619 if (CS.isByValArgument(ArgNo)) { 620 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, 621 CalledFunc->getParamAlignment(ArgNo+1)); 622 if (ActualArg != *AI) 623 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI)); 624 } 625 626 VMap[I] = ActualArg; 627 } 628 629 // We want the inliner to prune the code as it copies. We would LOVE to 630 // have no dead or constant instructions leftover after inlining occurs 631 // (which can happen, e.g., because an argument was constant), but we'll be 632 // happy with whatever the cloner can do. 633 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 634 /*ModuleLevelChanges=*/false, Returns, ".i", 635 &InlinedFunctionInfo, IFI.DL, TheCall); 636 637 // Remember the first block that is newly cloned over. 638 FirstNewBlock = LastBlock; ++FirstNewBlock; 639 640 // Inject byval arguments initialization. 641 for (std::pair<Value*, Value*> &Init : ByValInit) 642 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(), 643 FirstNewBlock, IFI); 644 645 // Update the callgraph if requested. 646 if (IFI.CG) 647 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); 648 649 // Update inlined instructions' line number information. 650 fixupLineNumbers(Caller, FirstNewBlock, TheCall); 651 } 652 653 // If there are any alloca instructions in the block that used to be the entry 654 // block for the callee, move them to the entry block of the caller. First 655 // calculate which instruction they should be inserted before. We insert the 656 // instructions at the end of the current alloca list. 657 { 658 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 659 for (BasicBlock::iterator I = FirstNewBlock->begin(), 660 E = FirstNewBlock->end(); I != E; ) { 661 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 662 if (!AI) continue; 663 664 // If the alloca is now dead, remove it. This often occurs due to code 665 // specialization. 666 if (AI->use_empty()) { 667 AI->eraseFromParent(); 668 continue; 669 } 670 671 if (!isa<Constant>(AI->getArraySize())) 672 continue; 673 674 // Keep track of the static allocas that we inline into the caller. 675 IFI.StaticAllocas.push_back(AI); 676 677 // Scan for the block of allocas that we can move over, and move them 678 // all at once. 679 while (isa<AllocaInst>(I) && 680 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 681 IFI.StaticAllocas.push_back(cast<AllocaInst>(I)); 682 ++I; 683 } 684 685 // Transfer all of the allocas over in a block. Using splice means 686 // that the instructions aren't removed from the symbol table, then 687 // reinserted. 688 Caller->getEntryBlock().getInstList().splice(InsertPoint, 689 FirstNewBlock->getInstList(), 690 AI, I); 691 } 692 } 693 694 bool InlinedMustTailCalls = false; 695 if (InlinedFunctionInfo.ContainsCalls) { 696 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None; 697 if (CallInst *CI = dyn_cast<CallInst>(TheCall)) 698 CallSiteTailKind = CI->getTailCallKind(); 699 700 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; 701 ++BB) { 702 for (Instruction &I : *BB) { 703 CallInst *CI = dyn_cast<CallInst>(&I); 704 if (!CI) 705 continue; 706 707 // We need to reduce the strength of any inlined tail calls. For 708 // musttail, we have to avoid introducing potential unbounded stack 709 // growth. For example, if functions 'f' and 'g' are mutually recursive 710 // with musttail, we can inline 'g' into 'f' so long as we preserve 711 // musttail on the cloned call to 'f'. If either the inlined call site 712 // or the cloned call site is *not* musttail, the program already has 713 // one frame of stack growth, so it's safe to remove musttail. Here is 714 // a table of example transformations: 715 // 716 // f -> musttail g -> musttail f ==> f -> musttail f 717 // f -> musttail g -> tail f ==> f -> tail f 718 // f -> g -> musttail f ==> f -> f 719 // f -> g -> tail f ==> f -> f 720 CallInst::TailCallKind ChildTCK = CI->getTailCallKind(); 721 ChildTCK = std::min(CallSiteTailKind, ChildTCK); 722 CI->setTailCallKind(ChildTCK); 723 InlinedMustTailCalls |= CI->isMustTailCall(); 724 725 // Calls inlined through a 'nounwind' call site should be marked 726 // 'nounwind'. 727 if (MarkNoUnwind) 728 CI->setDoesNotThrow(); 729 } 730 } 731 } 732 733 // Leave lifetime markers for the static alloca's, scoping them to the 734 // function we just inlined. 735 if (InsertLifetime && !IFI.StaticAllocas.empty()) { 736 IRBuilder<> builder(FirstNewBlock->begin()); 737 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) { 738 AllocaInst *AI = IFI.StaticAllocas[ai]; 739 740 // If the alloca is already scoped to something smaller than the whole 741 // function then there's no need to add redundant, less accurate markers. 742 if (hasLifetimeMarkers(AI)) 743 continue; 744 745 // Try to determine the size of the allocation. 746 ConstantInt *AllocaSize = nullptr; 747 if (ConstantInt *AIArraySize = 748 dyn_cast<ConstantInt>(AI->getArraySize())) { 749 if (IFI.DL) { 750 Type *AllocaType = AI->getAllocatedType(); 751 uint64_t AllocaTypeSize = IFI.DL->getTypeAllocSize(AllocaType); 752 uint64_t AllocaArraySize = AIArraySize->getLimitedValue(); 753 assert(AllocaArraySize > 0 && "array size of AllocaInst is zero"); 754 // Check that array size doesn't saturate uint64_t and doesn't 755 // overflow when it's multiplied by type size. 756 if (AllocaArraySize != ~0ULL && 757 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) { 758 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()), 759 AllocaArraySize * AllocaTypeSize); 760 } 761 } 762 } 763 764 builder.CreateLifetimeStart(AI, AllocaSize); 765 for (ReturnInst *RI : Returns) { 766 // Don't insert llvm.lifetime.end calls between a musttail call and a 767 // return. The return kills all local allocas. 768 if (InlinedMustTailCalls && getPrecedingMustTailCall(RI)) 769 continue; 770 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize); 771 } 772 } 773 } 774 775 // If the inlined code contained dynamic alloca instructions, wrap the inlined 776 // code with llvm.stacksave/llvm.stackrestore intrinsics. 777 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 778 Module *M = Caller->getParent(); 779 // Get the two intrinsics we care about. 780 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 781 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 782 783 // Insert the llvm.stacksave. 784 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin()) 785 .CreateCall(StackSave, "savedstack"); 786 787 // Insert a call to llvm.stackrestore before any return instructions in the 788 // inlined function. 789 for (ReturnInst *RI : Returns) { 790 // Don't insert llvm.stackrestore calls between a musttail call and a 791 // return. The return will restore the stack pointer. 792 if (InlinedMustTailCalls && getPrecedingMustTailCall(RI)) 793 continue; 794 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr); 795 } 796 } 797 798 // If we are inlining for an invoke instruction, we must make sure to rewrite 799 // any call instructions into invoke instructions. 800 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 801 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 802 803 // Handle any inlined musttail call sites. In order for a new call site to be 804 // musttail, the source of the clone and the inlined call site must have been 805 // musttail. Therefore it's safe to return without merging control into the 806 // phi below. 807 if (InlinedMustTailCalls) { 808 // Check if we need to bitcast the result of any musttail calls. 809 Type *NewRetTy = Caller->getReturnType(); 810 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy; 811 812 // Handle the returns preceded by musttail calls separately. 813 SmallVector<ReturnInst *, 8> NormalReturns; 814 for (ReturnInst *RI : Returns) { 815 CallInst *ReturnedMustTail = getPrecedingMustTailCall(RI); 816 if (!ReturnedMustTail) { 817 NormalReturns.push_back(RI); 818 continue; 819 } 820 if (!NeedBitCast) 821 continue; 822 823 // Delete the old return and any preceding bitcast. 824 BasicBlock *CurBB = RI->getParent(); 825 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue()); 826 RI->eraseFromParent(); 827 if (OldCast) 828 OldCast->eraseFromParent(); 829 830 // Insert a new bitcast and return with the right type. 831 IRBuilder<> Builder(CurBB); 832 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy)); 833 } 834 835 // Leave behind the normal returns so we can merge control flow. 836 std::swap(Returns, NormalReturns); 837 } 838 839 // If we cloned in _exactly one_ basic block, and if that block ends in a 840 // return instruction, we splice the body of the inlined callee directly into 841 // the calling basic block. 842 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 843 // Move all of the instructions right before the call. 844 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 845 FirstNewBlock->begin(), FirstNewBlock->end()); 846 // Remove the cloned basic block. 847 Caller->getBasicBlockList().pop_back(); 848 849 // If the call site was an invoke instruction, add a branch to the normal 850 // destination. 851 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 852 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 853 NewBr->setDebugLoc(Returns[0]->getDebugLoc()); 854 } 855 856 // If the return instruction returned a value, replace uses of the call with 857 // uses of the returned value. 858 if (!TheCall->use_empty()) { 859 ReturnInst *R = Returns[0]; 860 if (TheCall == R->getReturnValue()) 861 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 862 else 863 TheCall->replaceAllUsesWith(R->getReturnValue()); 864 } 865 // Since we are now done with the Call/Invoke, we can delete it. 866 TheCall->eraseFromParent(); 867 868 // Since we are now done with the return instruction, delete it also. 869 Returns[0]->eraseFromParent(); 870 871 // We are now done with the inlining. 872 return true; 873 } 874 875 // Otherwise, we have the normal case, of more than one block to inline or 876 // multiple return sites. 877 878 // We want to clone the entire callee function into the hole between the 879 // "starter" and "ender" blocks. How we accomplish this depends on whether 880 // this is an invoke instruction or a call instruction. 881 BasicBlock *AfterCallBB; 882 BranchInst *CreatedBranchToNormalDest = nullptr; 883 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 884 885 // Add an unconditional branch to make this look like the CallInst case... 886 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall); 887 888 // Split the basic block. This guarantees that no PHI nodes will have to be 889 // updated due to new incoming edges, and make the invoke case more 890 // symmetric to the call case. 891 AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest, 892 CalledFunc->getName()+".exit"); 893 894 } else { // It's a call 895 // If this is a call instruction, we need to split the basic block that 896 // the call lives in. 897 // 898 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 899 CalledFunc->getName()+".exit"); 900 } 901 902 // Change the branch that used to go to AfterCallBB to branch to the first 903 // basic block of the inlined function. 904 // 905 TerminatorInst *Br = OrigBB->getTerminator(); 906 assert(Br && Br->getOpcode() == Instruction::Br && 907 "splitBasicBlock broken!"); 908 Br->setOperand(0, FirstNewBlock); 909 910 911 // Now that the function is correct, make it a little bit nicer. In 912 // particular, move the basic blocks inserted from the end of the function 913 // into the space made by splitting the source basic block. 914 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 915 FirstNewBlock, Caller->end()); 916 917 // Handle all of the return instructions that we just cloned in, and eliminate 918 // any users of the original call/invoke instruction. 919 Type *RTy = CalledFunc->getReturnType(); 920 921 PHINode *PHI = nullptr; 922 if (Returns.size() > 1) { 923 // The PHI node should go at the front of the new basic block to merge all 924 // possible incoming values. 925 if (!TheCall->use_empty()) { 926 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(), 927 AfterCallBB->begin()); 928 // Anything that used the result of the function call should now use the 929 // PHI node as their operand. 930 TheCall->replaceAllUsesWith(PHI); 931 } 932 933 // Loop over all of the return instructions adding entries to the PHI node 934 // as appropriate. 935 if (PHI) { 936 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 937 ReturnInst *RI = Returns[i]; 938 assert(RI->getReturnValue()->getType() == PHI->getType() && 939 "Ret value not consistent in function!"); 940 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 941 } 942 } 943 944 945 // Add a branch to the merge points and remove return instructions. 946 DebugLoc Loc; 947 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 948 ReturnInst *RI = Returns[i]; 949 BranchInst* BI = BranchInst::Create(AfterCallBB, RI); 950 Loc = RI->getDebugLoc(); 951 BI->setDebugLoc(Loc); 952 RI->eraseFromParent(); 953 } 954 // We need to set the debug location to *somewhere* inside the 955 // inlined function. The line number may be nonsensical, but the 956 // instruction will at least be associated with the right 957 // function. 958 if (CreatedBranchToNormalDest) 959 CreatedBranchToNormalDest->setDebugLoc(Loc); 960 } else if (!Returns.empty()) { 961 // Otherwise, if there is exactly one return value, just replace anything 962 // using the return value of the call with the computed value. 963 if (!TheCall->use_empty()) { 964 if (TheCall == Returns[0]->getReturnValue()) 965 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 966 else 967 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 968 } 969 970 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 971 BasicBlock *ReturnBB = Returns[0]->getParent(); 972 ReturnBB->replaceAllUsesWith(AfterCallBB); 973 974 // Splice the code from the return block into the block that it will return 975 // to, which contains the code that was after the call. 976 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 977 ReturnBB->getInstList()); 978 979 if (CreatedBranchToNormalDest) 980 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc()); 981 982 // Delete the return instruction now and empty ReturnBB now. 983 Returns[0]->eraseFromParent(); 984 ReturnBB->eraseFromParent(); 985 } else if (!TheCall->use_empty()) { 986 // No returns, but something is using the return value of the call. Just 987 // nuke the result. 988 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 989 } 990 991 // Since we are now done with the Call/Invoke, we can delete it. 992 TheCall->eraseFromParent(); 993 994 // If we inlined any musttail calls and the original return is now 995 // unreachable, delete it. It can only contain a bitcast and ret. 996 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB)) 997 AfterCallBB->eraseFromParent(); 998 999 // We should always be able to fold the entry block of the function into the 1000 // single predecessor of the block... 1001 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 1002 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 1003 1004 // Splice the code entry block into calling block, right before the 1005 // unconditional branch. 1006 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 1007 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 1008 1009 // Remove the unconditional branch. 1010 OrigBB->getInstList().erase(Br); 1011 1012 // Now we can remove the CalleeEntry block, which is now empty. 1013 Caller->getBasicBlockList().erase(CalleeEntry); 1014 1015 // If we inserted a phi node, check to see if it has a single value (e.g. all 1016 // the entries are the same or undef). If so, remove the PHI so it doesn't 1017 // block other optimizations. 1018 if (PHI) { 1019 if (Value *V = SimplifyInstruction(PHI, IFI.DL)) { 1020 PHI->replaceAllUsesWith(V); 1021 PHI->eraseFromParent(); 1022 } 1023 } 1024 1025 return true; 1026} 1027