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