InlineFunction.cpp revision 6e69442416c23c3c56a335120662557d6bed2fae
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/Constants.h" 17#include "llvm/DerivedTypes.h" 18#include "llvm/Module.h" 19#include "llvm/Instructions.h" 20#include "llvm/IntrinsicInst.h" 21#include "llvm/Intrinsics.h" 22#include "llvm/Attributes.h" 23#include "llvm/Analysis/CallGraph.h" 24#include "llvm/Analysis/DebugInfo.h" 25#include "llvm/Analysis/InstructionSimplify.h" 26#include "llvm/Target/TargetData.h" 27#include "llvm/Transforms/Utils/Local.h" 28#include "llvm/ADT/SmallVector.h" 29#include "llvm/ADT/StringExtras.h" 30#include "llvm/Support/CallSite.h" 31#include "llvm/Support/IRBuilder.h" 32using namespace llvm; 33 34bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI) { 35 return InlineFunction(CallSite(CI), IFI); 36} 37bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI) { 38 return InlineFunction(CallSite(II), IFI); 39} 40 41namespace { 42 /// A class for recording information about inlining through an invoke. 43 class InvokeInliningInfo { 44 BasicBlock *OuterUnwindDest; 45 EHSelectorInst *OuterSelector; 46 BasicBlock *InnerUnwindDest; 47 PHINode *InnerExceptionPHI; 48 PHINode *InnerSelectorPHI; 49 SmallVector<Value*, 8> UnwindDestPHIValues; 50 51 // FIXME: New EH - These will replace the analogous ones above. 52 BasicBlock *OuterResumeDest; //< Destination of the invoke's unwind. 53 BasicBlock *InnerResumeDest; //< Destination for the callee's resume. 54 LandingPadInst *CallerLPad; //< LandingPadInst associated with the invoke. 55 PHINode *InnerEHValuesPHI; //< PHI for EH values from landingpad insts. 56 57 public: 58 InvokeInliningInfo(InvokeInst *II) 59 : OuterUnwindDest(II->getUnwindDest()), OuterSelector(0), 60 InnerUnwindDest(0), InnerExceptionPHI(0), InnerSelectorPHI(0), 61 OuterResumeDest(II->getUnwindDest()), InnerResumeDest(0), 62 CallerLPad(0), InnerEHValuesPHI(0) { 63 // If there are PHI nodes in the unwind destination block, we need to keep 64 // track of which values came into them from the invoke before removing 65 // the edge from this block. 66 llvm::BasicBlock *InvokeBB = II->getParent(); 67 BasicBlock::iterator I = OuterUnwindDest->begin(); 68 for (; isa<PHINode>(I); ++I) { 69 // Save the value to use for this edge. 70 PHINode *PHI = cast<PHINode>(I); 71 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB)); 72 } 73 74 CallerLPad = cast<LandingPadInst>(I); 75 } 76 77 /// The outer unwind destination is the target of unwind edges 78 /// introduced for calls within the inlined function. 79 BasicBlock *getOuterUnwindDest() const { 80 return OuterUnwindDest; 81 } 82 83 BasicBlock *getInnerUnwindDest(); 84 85 LandingPadInst *getLandingPadInst() const { return CallerLPad; } 86 87 /// forwardResume - Forward the 'resume' instruction to the caller's landing 88 /// pad block. When the landing pad block has only one predecessor, this is 89 /// a simple branch. When there is more than one predecessor, we need to 90 /// split the landing pad block after the landingpad instruction and jump 91 /// to there. 92 void forwardResume(ResumeInst *RI); 93 94 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind 95 /// destination block for the given basic block, using the values for the 96 /// original invoke's source block. 97 void addIncomingPHIValuesFor(BasicBlock *BB) const { 98 addIncomingPHIValuesForInto(BB, OuterUnwindDest); 99 } 100 101 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const { 102 BasicBlock::iterator I = dest->begin(); 103 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 104 PHINode *phi = cast<PHINode>(I); 105 phi->addIncoming(UnwindDestPHIValues[i], src); 106 } 107 } 108 }; 109} 110 111/// getInnerUnwindDest - Get or create a target for the branch from ResumeInsts. 112BasicBlock *InvokeInliningInfo::getInnerUnwindDest() { 113 if (InnerResumeDest) return InnerResumeDest; 114 115 // Split the landing pad. 116 BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint; 117 InnerResumeDest = 118 OuterResumeDest->splitBasicBlock(SplitPoint, 119 OuterResumeDest->getName() + ".body"); 120 121 // The number of incoming edges we expect to the inner landing pad. 122 const unsigned PHICapacity = 2; 123 124 // Create corresponding new PHIs for all the PHIs in the outer landing pad. 125 BasicBlock::iterator InsertPoint = InnerResumeDest->begin(); 126 BasicBlock::iterator I = OuterResumeDest->begin(); 127 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 128 PHINode *OuterPHI = cast<PHINode>(I); 129 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity, 130 OuterPHI->getName() + ".lpad-body", 131 InsertPoint); 132 OuterPHI->replaceAllUsesWith(InnerPHI); 133 InnerPHI->addIncoming(OuterPHI, OuterResumeDest); 134 } 135 136 // Create a PHI for the exception values. 137 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity, 138 "eh.lpad-body", InsertPoint); 139 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI); 140 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest); 141 142 // All done. 143 return InnerResumeDest; 144} 145 146/// forwardResume - Forward the 'resume' instruction to the caller's landing pad 147/// block. When the landing pad block has only one predecessor, this is a simple 148/// branch. When there is more than one predecessor, we need to split the 149/// landing pad block after the landingpad instruction and jump to there. 150void InvokeInliningInfo::forwardResume(ResumeInst *RI) { 151 BasicBlock *Dest = getInnerUnwindDest(); 152 BasicBlock *Src = RI->getParent(); 153 154 BranchInst::Create(Dest, Src); 155 156 // Update the PHIs in the destination. They were inserted in an order which 157 // makes this work. 158 addIncomingPHIValuesForInto(Src, Dest); 159 160 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src); 161 RI->eraseFromParent(); 162} 163 164/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into 165/// an invoke, we have to turn all of the calls that can throw into 166/// invokes. This function analyze BB to see if there are any calls, and if so, 167/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI 168/// nodes in that block with the values specified in InvokeDestPHIValues. 169/// 170/// Returns true to indicate that the next block should be skipped. 171static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, 172 InvokeInliningInfo &Invoke) { 173 LandingPadInst *LPI = Invoke.getLandingPadInst(); 174 175 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 176 Instruction *I = BBI++; 177 178 if (LandingPadInst *L = dyn_cast<LandingPadInst>(I)) { 179 unsigned NumClauses = LPI->getNumClauses(); 180 L->reserveClauses(NumClauses); 181 for (unsigned i = 0; i != NumClauses; ++i) 182 L->addClause(LPI->getClause(i)); 183 } 184 185 // We only need to check for function calls: inlined invoke 186 // instructions require no special handling. 187 CallInst *CI = dyn_cast<CallInst>(I); 188 189 // If this call cannot unwind, don't convert it to an invoke. 190 if (!CI || CI->doesNotThrow()) 191 continue; 192 193 // Convert this function call into an invoke instruction. First, split the 194 // basic block. 195 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 196 197 // Delete the unconditional branch inserted by splitBasicBlock 198 BB->getInstList().pop_back(); 199 200 // Create the new invoke instruction. 201 ImmutableCallSite CS(CI); 202 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end()); 203 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split, 204 Invoke.getOuterUnwindDest(), 205 InvokeArgs, CI->getName(), BB); 206 II->setCallingConv(CI->getCallingConv()); 207 II->setAttributes(CI->getAttributes()); 208 209 // Make sure that anything using the call now uses the invoke! This also 210 // updates the CallGraph if present, because it uses a WeakVH. 211 CI->replaceAllUsesWith(II); 212 213 // Delete the original call 214 Split->getInstList().pop_front(); 215 216 // Update any PHI nodes in the exceptional block to indicate that there is 217 // now a new entry in them. 218 Invoke.addIncomingPHIValuesFor(BB); 219 return false; 220 } 221 222 return false; 223} 224 225/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 226/// in the body of the inlined function into invokes and turn unwind 227/// instructions into branches to the invoke unwind dest. 228/// 229/// II is the invoke instruction being inlined. FirstNewBlock is the first 230/// block of the inlined code (the last block is the end of the function), 231/// and InlineCodeInfo is information about the code that got inlined. 232static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 233 ClonedCodeInfo &InlinedCodeInfo) { 234 BasicBlock *InvokeDest = II->getUnwindDest(); 235 236 Function *Caller = FirstNewBlock->getParent(); 237 238 // The inlined code is currently at the end of the function, scan from the 239 // start of the inlined code to its end, checking for stuff we need to 240 // rewrite. If the code doesn't have calls or unwinds, we know there is 241 // nothing to rewrite. 242 if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) { 243 // Now that everything is happy, we have one final detail. The PHI nodes in 244 // the exception destination block still have entries due to the original 245 // invoke instruction. Eliminate these entries (which might even delete the 246 // PHI node) now. 247 InvokeDest->removePredecessor(II->getParent()); 248 return; 249 } 250 251 InvokeInliningInfo Invoke(II); 252 253 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){ 254 if (InlinedCodeInfo.ContainsCalls) 255 if (HandleCallsInBlockInlinedThroughInvoke(BB, Invoke)) { 256 // Honor a request to skip the next block. We don't need to 257 // consider UnwindInsts in this case either. 258 ++BB; 259 continue; 260 } 261 262 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 263 // An UnwindInst requires special handling when it gets inlined into an 264 // invoke site. Once this happens, we know that the unwind would cause 265 // a control transfer to the invoke exception destination, so we can 266 // transform it into a direct branch to the exception destination. 267 BranchInst::Create(InvokeDest, UI); 268 269 // Delete the unwind instruction! 270 UI->eraseFromParent(); 271 272 // Update any PHI nodes in the exceptional block to indicate that 273 // there is now a new entry in them. 274 Invoke.addIncomingPHIValuesFor(BB); 275 } 276 277 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) 278 Invoke.forwardResume(RI); 279 } 280 281 // Now that everything is happy, we have one final detail. The PHI nodes in 282 // the exception destination block still have entries due to the original 283 // invoke instruction. Eliminate these entries (which might even delete the 284 // PHI node) now. 285 InvokeDest->removePredecessor(II->getParent()); 286} 287 288/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 289/// into the caller, update the specified callgraph to reflect the changes we 290/// made. Note that it's possible that not all code was copied over, so only 291/// some edges of the callgraph may remain. 292static void UpdateCallGraphAfterInlining(CallSite CS, 293 Function::iterator FirstNewBlock, 294 ValueToValueMapTy &VMap, 295 InlineFunctionInfo &IFI) { 296 CallGraph &CG = *IFI.CG; 297 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 298 const Function *Callee = CS.getCalledFunction(); 299 CallGraphNode *CalleeNode = CG[Callee]; 300 CallGraphNode *CallerNode = CG[Caller]; 301 302 // Since we inlined some uninlined call sites in the callee into the caller, 303 // add edges from the caller to all of the callees of the callee. 304 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 305 306 // Consider the case where CalleeNode == CallerNode. 307 CallGraphNode::CalledFunctionsVector CallCache; 308 if (CalleeNode == CallerNode) { 309 CallCache.assign(I, E); 310 I = CallCache.begin(); 311 E = CallCache.end(); 312 } 313 314 for (; I != E; ++I) { 315 const Value *OrigCall = I->first; 316 317 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); 318 // Only copy the edge if the call was inlined! 319 if (VMI == VMap.end() || VMI->second == 0) 320 continue; 321 322 // If the call was inlined, but then constant folded, there is no edge to 323 // add. Check for this case. 324 Instruction *NewCall = dyn_cast<Instruction>(VMI->second); 325 if (NewCall == 0) continue; 326 327 // Remember that this call site got inlined for the client of 328 // InlineFunction. 329 IFI.InlinedCalls.push_back(NewCall); 330 331 // It's possible that inlining the callsite will cause it to go from an 332 // indirect to a direct call by resolving a function pointer. If this 333 // happens, set the callee of the new call site to a more precise 334 // destination. This can also happen if the call graph node of the caller 335 // was just unnecessarily imprecise. 336 if (I->second->getFunction() == 0) 337 if (Function *F = CallSite(NewCall).getCalledFunction()) { 338 // Indirect call site resolved to direct call. 339 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); 340 341 continue; 342 } 343 344 CallerNode->addCalledFunction(CallSite(NewCall), I->second); 345 } 346 347 // Update the call graph by deleting the edge from Callee to Caller. We must 348 // do this after the loop above in case Caller and Callee are the same. 349 CallerNode->removeCallEdgeFor(CS); 350} 351 352/// HandleByValArgument - When inlining a call site that has a byval argument, 353/// we have to make the implicit memcpy explicit by adding it. 354static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, 355 const Function *CalledFunc, 356 InlineFunctionInfo &IFI, 357 unsigned ByValAlignment) { 358 Type *AggTy = cast<PointerType>(Arg->getType())->getElementType(); 359 360 // If the called function is readonly, then it could not mutate the caller's 361 // copy of the byval'd memory. In this case, it is safe to elide the copy and 362 // temporary. 363 if (CalledFunc->onlyReadsMemory()) { 364 // If the byval argument has a specified alignment that is greater than the 365 // passed in pointer, then we either have to round up the input pointer or 366 // give up on this transformation. 367 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. 368 return Arg; 369 370 // If the pointer is already known to be sufficiently aligned, or if we can 371 // round it up to a larger alignment, then we don't need a temporary. 372 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, 373 IFI.TD) >= ByValAlignment) 374 return Arg; 375 376 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad 377 // for code quality, but rarely happens and is required for correctness. 378 } 379 380 LLVMContext &Context = Arg->getContext(); 381 382 Type *VoidPtrTy = Type::getInt8PtrTy(Context); 383 384 // Create the alloca. If we have TargetData, use nice alignment. 385 unsigned Align = 1; 386 if (IFI.TD) 387 Align = IFI.TD->getPrefTypeAlignment(AggTy); 388 389 // If the byval had an alignment specified, we *must* use at least that 390 // alignment, as it is required by the byval argument (and uses of the 391 // pointer inside the callee). 392 Align = std::max(Align, ByValAlignment); 393 394 Function *Caller = TheCall->getParent()->getParent(); 395 396 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(), 397 &*Caller->begin()->begin()); 398 // Emit a memcpy. 399 Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)}; 400 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 401 Intrinsic::memcpy, 402 Tys); 403 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 404 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall); 405 406 Value *Size; 407 if (IFI.TD == 0) 408 Size = ConstantExpr::getSizeOf(AggTy); 409 else 410 Size = ConstantInt::get(Type::getInt64Ty(Context), 411 IFI.TD->getTypeStoreSize(AggTy)); 412 413 // Always generate a memcpy of alignment 1 here because we don't know 414 // the alignment of the src pointer. Other optimizations can infer 415 // better alignment. 416 Value *CallArgs[] = { 417 DestCast, SrcCast, Size, 418 ConstantInt::get(Type::getInt32Ty(Context), 1), 419 ConstantInt::getFalse(Context) // isVolatile 420 }; 421 IRBuilder<>(TheCall).CreateCall(MemCpyFn, CallArgs); 422 423 // Uses of the argument in the function should use our new alloca 424 // instead. 425 return NewAlloca; 426} 427 428// isUsedByLifetimeMarker - Check whether this Value is used by a lifetime 429// intrinsic. 430static bool isUsedByLifetimeMarker(Value *V) { 431 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE; 432 ++UI) { 433 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI)) { 434 switch (II->getIntrinsicID()) { 435 default: break; 436 case Intrinsic::lifetime_start: 437 case Intrinsic::lifetime_end: 438 return true; 439 } 440 } 441 } 442 return false; 443} 444 445// hasLifetimeMarkers - Check whether the given alloca already has 446// lifetime.start or lifetime.end intrinsics. 447static bool hasLifetimeMarkers(AllocaInst *AI) { 448 Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext()); 449 if (AI->getType() == Int8PtrTy) 450 return isUsedByLifetimeMarker(AI); 451 452 // Do a scan to find all the casts to i8*. 453 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E; 454 ++I) { 455 if (I->getType() != Int8PtrTy) continue; 456 if (I->stripPointerCasts() != AI) continue; 457 if (isUsedByLifetimeMarker(*I)) 458 return true; 459 } 460 return false; 461} 462 463/// updateInlinedAtInfo - Helper function used by fixupLineNumbers to recursively 464/// update InlinedAtEntry of a DebugLoc. 465static DebugLoc updateInlinedAtInfo(const DebugLoc &DL, 466 const DebugLoc &InlinedAtDL, 467 LLVMContext &Ctx) { 468 if (MDNode *IA = DL.getInlinedAt(Ctx)) { 469 DebugLoc NewInlinedAtDL 470 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx); 471 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), 472 NewInlinedAtDL.getAsMDNode(Ctx)); 473 } 474 475 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), 476 InlinedAtDL.getAsMDNode(Ctx)); 477} 478 479/// fixupLineNumbers - Update inlined instructions' line numbers to 480/// to encode location where these instructions are inlined. 481static void fixupLineNumbers(Function *Fn, Function::iterator FI, 482 Instruction *TheCall) { 483 DebugLoc TheCallDL = TheCall->getDebugLoc(); 484 if (TheCallDL.isUnknown()) 485 return; 486 487 for (; FI != Fn->end(); ++FI) { 488 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end(); 489 BI != BE; ++BI) { 490 DebugLoc DL = BI->getDebugLoc(); 491 if (!DL.isUnknown()) { 492 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext())); 493 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) { 494 LLVMContext &Ctx = BI->getContext(); 495 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx); 496 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(), 497 InlinedAt, Ctx)); 498 } 499 } 500 } 501 } 502} 503 504/// InlineFunction - This function inlines the called function into the basic 505/// block of the caller. This returns false if it is not possible to inline 506/// this call. The program is still in a well defined state if this occurs 507/// though. 508/// 509/// Note that this only does one level of inlining. For example, if the 510/// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 511/// exists in the instruction stream. Similarly this will inline a recursive 512/// function by one level. 513bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) { 514 Instruction *TheCall = CS.getInstruction(); 515 LLVMContext &Context = TheCall->getContext(); 516 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 517 "Instruction not in function!"); 518 519 // If IFI has any state in it, zap it before we fill it in. 520 IFI.reset(); 521 522 const Function *CalledFunc = CS.getCalledFunction(); 523 if (CalledFunc == 0 || // Can't inline external function or indirect 524 CalledFunc->isDeclaration() || // call, or call to a vararg function! 525 CalledFunc->getFunctionType()->isVarArg()) return false; 526 527 // If the call to the callee is not a tail call, we must clear the 'tail' 528 // flags on any calls that we inline. 529 bool MustClearTailCallFlags = 530 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall()); 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 = 0; 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; 594 595 assert(CalledFunc->arg_size() == CS.arg_size() && 596 "No varargs calls can be inlined!"); 597 598 // Calculate the vector of arguments to pass into the function cloner, which 599 // matches up the formal to the actual argument values. 600 CallSite::arg_iterator AI = CS.arg_begin(); 601 unsigned ArgNo = 0; 602 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 603 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 604 Value *ActualArg = *AI; 605 606 // When byval arguments actually inlined, we need to make the copy implied 607 // by them explicit. However, we don't do this if the callee is readonly 608 // or readnone, because the copy would be unneeded: the callee doesn't 609 // modify the struct. 610 if (CS.isByValArgument(ArgNo)) { 611 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, 612 CalledFunc->getParamAlignment(ArgNo+1)); 613 614 // Calls that we inline may use the new alloca, so we need to clear 615 // their 'tail' flags if HandleByValArgument introduced a new alloca and 616 // the callee has calls. 617 MustClearTailCallFlags |= ActualArg != *AI; 618 } 619 620 VMap[I] = ActualArg; 621 } 622 623 // We want the inliner to prune the code as it copies. We would LOVE to 624 // have no dead or constant instructions leftover after inlining occurs 625 // (which can happen, e.g., because an argument was constant), but we'll be 626 // happy with whatever the cloner can do. 627 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 628 /*ModuleLevelChanges=*/false, Returns, ".i", 629 &InlinedFunctionInfo, IFI.TD, TheCall); 630 631 // Remember the first block that is newly cloned over. 632 FirstNewBlock = LastBlock; ++FirstNewBlock; 633 634 // Update the callgraph if requested. 635 if (IFI.CG) 636 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); 637 638 // Update inlined instructions' line number information. 639 fixupLineNumbers(Caller, FirstNewBlock, TheCall); 640 } 641 642 // If there are any alloca instructions in the block that used to be the entry 643 // block for the callee, move them to the entry block of the caller. First 644 // calculate which instruction they should be inserted before. We insert the 645 // instructions at the end of the current alloca list. 646 { 647 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 648 for (BasicBlock::iterator I = FirstNewBlock->begin(), 649 E = FirstNewBlock->end(); I != E; ) { 650 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 651 if (AI == 0) continue; 652 653 // If the alloca is now dead, remove it. This often occurs due to code 654 // specialization. 655 if (AI->use_empty()) { 656 AI->eraseFromParent(); 657 continue; 658 } 659 660 if (!isa<Constant>(AI->getArraySize())) 661 continue; 662 663 // Keep track of the static allocas that we inline into the caller. 664 IFI.StaticAllocas.push_back(AI); 665 666 // Scan for the block of allocas that we can move over, and move them 667 // all at once. 668 while (isa<AllocaInst>(I) && 669 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 670 IFI.StaticAllocas.push_back(cast<AllocaInst>(I)); 671 ++I; 672 } 673 674 // Transfer all of the allocas over in a block. Using splice means 675 // that the instructions aren't removed from the symbol table, then 676 // reinserted. 677 Caller->getEntryBlock().getInstList().splice(InsertPoint, 678 FirstNewBlock->getInstList(), 679 AI, I); 680 } 681 } 682 683 // Leave lifetime markers for the static alloca's, scoping them to the 684 // function we just inlined. 685 if (!IFI.StaticAllocas.empty()) { 686 IRBuilder<> builder(FirstNewBlock->begin()); 687 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) { 688 AllocaInst *AI = IFI.StaticAllocas[ai]; 689 690 // If the alloca is already scoped to something smaller than the whole 691 // function then there's no need to add redundant, less accurate markers. 692 if (hasLifetimeMarkers(AI)) 693 continue; 694 695 builder.CreateLifetimeStart(AI); 696 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) { 697 IRBuilder<> builder(Returns[ri]); 698 builder.CreateLifetimeEnd(AI); 699 } 700 } 701 } 702 703 // If the inlined code contained dynamic alloca instructions, wrap the inlined 704 // code with llvm.stacksave/llvm.stackrestore intrinsics. 705 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 706 Module *M = Caller->getParent(); 707 // Get the two intrinsics we care about. 708 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 709 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 710 711 // Insert the llvm.stacksave. 712 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin()) 713 .CreateCall(StackSave, "savedstack"); 714 715 // Insert a call to llvm.stackrestore before any return instructions in the 716 // inlined function. 717 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 718 IRBuilder<>(Returns[i]).CreateCall(StackRestore, SavedPtr); 719 } 720 721 // Count the number of StackRestore calls we insert. 722 unsigned NumStackRestores = Returns.size(); 723 724 // If we are inlining an invoke instruction, insert restores before each 725 // unwind. These unwinds will be rewritten into branches later. 726 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 727 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 728 BB != E; ++BB) 729 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 730 IRBuilder<>(UI).CreateCall(StackRestore, SavedPtr); 731 ++NumStackRestores; 732 } 733 } 734 } 735 736 // If we are inlining tail call instruction through a call site that isn't 737 // marked 'tail', we must remove the tail marker for any calls in the inlined 738 // code. Also, calls inlined through a 'nounwind' call site should be marked 739 // 'nounwind'. 740 if (InlinedFunctionInfo.ContainsCalls && 741 (MustClearTailCallFlags || MarkNoUnwind)) { 742 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 743 BB != E; ++BB) 744 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 745 if (CallInst *CI = dyn_cast<CallInst>(I)) { 746 if (MustClearTailCallFlags) 747 CI->setTailCall(false); 748 if (MarkNoUnwind) 749 CI->setDoesNotThrow(); 750 } 751 } 752 753 // If we are inlining through a 'nounwind' call site then any inlined 'unwind' 754 // instructions are unreachable. 755 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind) 756 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 757 BB != E; ++BB) { 758 TerminatorInst *Term = BB->getTerminator(); 759 if (isa<UnwindInst>(Term)) { 760 new UnreachableInst(Context, Term); 761 BB->getInstList().erase(Term); 762 } 763 } 764 765 // If we are inlining for an invoke instruction, we must make sure to rewrite 766 // any inlined 'unwind' instructions into branches to the invoke exception 767 // destination, and call instructions into invoke instructions. 768 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 769 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 770 771 // If we cloned in _exactly one_ basic block, and if that block ends in a 772 // return instruction, we splice the body of the inlined callee directly into 773 // the calling basic block. 774 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 775 // Move all of the instructions right before the call. 776 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 777 FirstNewBlock->begin(), FirstNewBlock->end()); 778 // Remove the cloned basic block. 779 Caller->getBasicBlockList().pop_back(); 780 781 // If the call site was an invoke instruction, add a branch to the normal 782 // destination. 783 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 784 BranchInst::Create(II->getNormalDest(), TheCall); 785 786 // If the return instruction returned a value, replace uses of the call with 787 // uses of the returned value. 788 if (!TheCall->use_empty()) { 789 ReturnInst *R = Returns[0]; 790 if (TheCall == R->getReturnValue()) 791 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 792 else 793 TheCall->replaceAllUsesWith(R->getReturnValue()); 794 } 795 // Since we are now done with the Call/Invoke, we can delete it. 796 TheCall->eraseFromParent(); 797 798 // Since we are now done with the return instruction, delete it also. 799 Returns[0]->eraseFromParent(); 800 801 // We are now done with the inlining. 802 return true; 803 } 804 805 // Otherwise, we have the normal case, of more than one block to inline or 806 // multiple return sites. 807 808 // We want to clone the entire callee function into the hole between the 809 // "starter" and "ender" blocks. How we accomplish this depends on whether 810 // this is an invoke instruction or a call instruction. 811 BasicBlock *AfterCallBB; 812 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 813 814 // Add an unconditional branch to make this look like the CallInst case... 815 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 816 817 // Split the basic block. This guarantees that no PHI nodes will have to be 818 // updated due to new incoming edges, and make the invoke case more 819 // symmetric to the call case. 820 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 821 CalledFunc->getName()+".exit"); 822 823 } else { // It's a call 824 // If this is a call instruction, we need to split the basic block that 825 // the call lives in. 826 // 827 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 828 CalledFunc->getName()+".exit"); 829 } 830 831 // Change the branch that used to go to AfterCallBB to branch to the first 832 // basic block of the inlined function. 833 // 834 TerminatorInst *Br = OrigBB->getTerminator(); 835 assert(Br && Br->getOpcode() == Instruction::Br && 836 "splitBasicBlock broken!"); 837 Br->setOperand(0, FirstNewBlock); 838 839 840 // Now that the function is correct, make it a little bit nicer. In 841 // particular, move the basic blocks inserted from the end of the function 842 // into the space made by splitting the source basic block. 843 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 844 FirstNewBlock, Caller->end()); 845 846 // Handle all of the return instructions that we just cloned in, and eliminate 847 // any users of the original call/invoke instruction. 848 Type *RTy = CalledFunc->getReturnType(); 849 850 PHINode *PHI = 0; 851 if (Returns.size() > 1) { 852 // The PHI node should go at the front of the new basic block to merge all 853 // possible incoming values. 854 if (!TheCall->use_empty()) { 855 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(), 856 AfterCallBB->begin()); 857 // Anything that used the result of the function call should now use the 858 // PHI node as their operand. 859 TheCall->replaceAllUsesWith(PHI); 860 } 861 862 // Loop over all of the return instructions adding entries to the PHI node 863 // as appropriate. 864 if (PHI) { 865 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 866 ReturnInst *RI = Returns[i]; 867 assert(RI->getReturnValue()->getType() == PHI->getType() && 868 "Ret value not consistent in function!"); 869 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 870 } 871 } 872 873 874 // Add a branch to the merge points and remove return instructions. 875 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 876 ReturnInst *RI = Returns[i]; 877 BranchInst::Create(AfterCallBB, RI); 878 RI->eraseFromParent(); 879 } 880 } else if (!Returns.empty()) { 881 // Otherwise, if there is exactly one return value, just replace anything 882 // using the return value of the call with the computed value. 883 if (!TheCall->use_empty()) { 884 if (TheCall == Returns[0]->getReturnValue()) 885 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 886 else 887 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 888 } 889 890 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 891 BasicBlock *ReturnBB = Returns[0]->getParent(); 892 ReturnBB->replaceAllUsesWith(AfterCallBB); 893 894 // Splice the code from the return block into the block that it will return 895 // to, which contains the code that was after the call. 896 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 897 ReturnBB->getInstList()); 898 899 // Delete the return instruction now and empty ReturnBB now. 900 Returns[0]->eraseFromParent(); 901 ReturnBB->eraseFromParent(); 902 } else if (!TheCall->use_empty()) { 903 // No returns, but something is using the return value of the call. Just 904 // nuke the result. 905 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 906 } 907 908 // Since we are now done with the Call/Invoke, we can delete it. 909 TheCall->eraseFromParent(); 910 911 // We should always be able to fold the entry block of the function into the 912 // single predecessor of the block... 913 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 914 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 915 916 // Splice the code entry block into calling block, right before the 917 // unconditional branch. 918 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 919 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 920 921 // Remove the unconditional branch. 922 OrigBB->getInstList().erase(Br); 923 924 // Now we can remove the CalleeEntry block, which is now empty. 925 Caller->getBasicBlockList().erase(CalleeEntry); 926 927 // If we inserted a phi node, check to see if it has a single value (e.g. all 928 // the entries are the same or undef). If so, remove the PHI so it doesn't 929 // block other optimizations. 930 if (PHI) { 931 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) { 932 PHI->replaceAllUsesWith(V); 933 PHI->eraseFromParent(); 934 } 935 } 936 937 return true; 938} 939