InlineFunction.cpp revision d9ff8c83d137586d8c06f98bdf8adbf0d1fa79ca
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/DebugInfo.h" 21#include "llvm/IR/Attributes.h" 22#include "llvm/IR/Constants.h" 23#include "llvm/IR/DataLayout.h" 24#include "llvm/IR/DerivedTypes.h" 25#include "llvm/IR/IRBuilder.h" 26#include "llvm/IR/Instructions.h" 27#include "llvm/IR/IntrinsicInst.h" 28#include "llvm/IR/Intrinsics.h" 29#include "llvm/IR/Module.h" 30#include "llvm/Support/CallSite.h" 31#include "llvm/Transforms/Utils/Local.h" 32using namespace llvm; 33 34bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI, 35 bool InsertLifetime) { 36 return InlineFunction(CallSite(CI), IFI, InsertLifetime); 37} 38bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI, 39 bool InsertLifetime) { 40 return InlineFunction(CallSite(II), IFI, InsertLifetime); 41} 42 43namespace { 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, BasicBlock *FirstNewBlock); 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. 105BasicBlock *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. 143void InvokeInliningInfo::forwardResume(ResumeInst *RI, 144 BasicBlock *FirstNewBlock) { 145 BasicBlock *Dest = getInnerResumeDest(); 146 LandingPadInst *OuterLPad = getLandingPadInst(); 147 BasicBlock *Src = RI->getParent(); 148 149 BranchInst::Create(Dest, Src); 150 151 // Update the PHIs in the destination. They were inserted in an order which 152 // makes this work. 153 addIncomingPHIValuesForInto(Src, Dest); 154 155 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src); 156 RI->eraseFromParent(); 157 158 // Get all of the inlined landing pad instructions. 159 SmallPtrSet<LandingPadInst*, 16> InlinedLPads; 160 Function *Caller = FirstNewBlock->getParent(); 161 for (Function::iterator I = FirstNewBlock, E = Caller->end(); I != E; ++I) 162 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) { 163 LandingPadInst *LPI = II->getLandingPadInst(); 164 if (!LPI->hasCatchAll()) 165 InlinedLPads.insert(LPI); 166 } 167 168 // Merge the catch clauses from the outer landing pad instruction into the 169 // inlined landing pad instructions. 170 for (SmallPtrSet<LandingPadInst*, 16>::iterator I = InlinedLPads.begin(), 171 E = InlinedLPads.end(); I != E; ++I) { 172 LandingPadInst *InlinedLPad = *I; 173 for (unsigned OuterIdx = 0, OuterNum = OuterLPad->getNumClauses(); 174 OuterIdx != OuterNum; ++OuterIdx) { 175 bool hasClause = false; 176 if (OuterLPad->isFilter(OuterIdx)) continue; 177 Value *OuterClause = OuterLPad->getClause(OuterIdx); 178 for (unsigned Idx = 0, N = InlinedLPad->getNumClauses(); Idx != N; ++Idx) 179 if (OuterClause == InlinedLPad->getClause(Idx)) { 180 hasClause = true; 181 break; 182 } 183 if (!hasClause) 184 InlinedLPad->addClause(OuterClause); 185 } 186 } 187} 188 189/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into 190/// an invoke, we have to turn all of the calls that can throw into 191/// invokes. This function analyze BB to see if there are any calls, and if so, 192/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI 193/// nodes in that block with the values specified in InvokeDestPHIValues. 194/// 195/// Returns true to indicate that the next block should be skipped. 196static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, 197 InvokeInliningInfo &Invoke) { 198 LandingPadInst *LPI = Invoke.getLandingPadInst(); 199 200 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 201 Instruction *I = BBI++; 202 203 if (LandingPadInst *L = dyn_cast<LandingPadInst>(I)) { 204 unsigned NumClauses = LPI->getNumClauses(); 205 L->reserveClauses(NumClauses); 206 for (unsigned i = 0; i != NumClauses; ++i) 207 L->addClause(LPI->getClause(i)); 208 } 209 210 // We only need to check for function calls: inlined invoke 211 // instructions require no special handling. 212 CallInst *CI = dyn_cast<CallInst>(I); 213 214 // If this call cannot unwind, don't convert it to an invoke. 215 if (!CI || CI->doesNotThrow()) 216 continue; 217 218 // Convert this function call into an invoke instruction. First, split the 219 // basic block. 220 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 221 222 // Delete the unconditional branch inserted by splitBasicBlock 223 BB->getInstList().pop_back(); 224 225 // Create the new invoke instruction. 226 ImmutableCallSite CS(CI); 227 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end()); 228 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split, 229 Invoke.getOuterResumeDest(), 230 InvokeArgs, CI->getName(), BB); 231 II->setCallingConv(CI->getCallingConv()); 232 II->setAttributes(CI->getAttributes()); 233 234 // Make sure that anything using the call now uses the invoke! This also 235 // updates the CallGraph if present, because it uses a WeakVH. 236 CI->replaceAllUsesWith(II); 237 238 // Delete the original call 239 Split->getInstList().pop_front(); 240 241 // Update any PHI nodes in the exceptional block to indicate that there is 242 // now a new entry in them. 243 Invoke.addIncomingPHIValuesFor(BB); 244 return false; 245 } 246 247 return false; 248} 249 250/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 251/// in the body of the inlined function into invokes. 252/// 253/// II is the invoke instruction being inlined. FirstNewBlock is the first 254/// block of the inlined code (the last block is the end of the function), 255/// and InlineCodeInfo is information about the code that got inlined. 256static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 257 ClonedCodeInfo &InlinedCodeInfo) { 258 BasicBlock *InvokeDest = II->getUnwindDest(); 259 260 Function *Caller = FirstNewBlock->getParent(); 261 262 // The inlined code is currently at the end of the function, scan from the 263 // start of the inlined code to its end, checking for stuff we need to 264 // rewrite. 265 InvokeInliningInfo Invoke(II); 266 267 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){ 268 if (InlinedCodeInfo.ContainsCalls) 269 if (HandleCallsInBlockInlinedThroughInvoke(BB, Invoke)) { 270 // Honor a request to skip the next block. 271 ++BB; 272 continue; 273 } 274 275 // Forward any resumes that are remaining here. 276 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) 277 Invoke.forwardResume(RI, FirstNewBlock); 278 } 279 280 // Now that everything is happy, we have one final detail. The PHI nodes in 281 // the exception destination block still have entries due to the original 282 // invoke instruction. Eliminate these entries (which might even delete the 283 // PHI node) now. 284 InvokeDest->removePredecessor(II->getParent()); 285} 286 287/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 288/// into the caller, update the specified callgraph to reflect the changes we 289/// made. Note that it's possible that not all code was copied over, so only 290/// some edges of the callgraph may remain. 291static void UpdateCallGraphAfterInlining(CallSite CS, 292 Function::iterator FirstNewBlock, 293 ValueToValueMapTy &VMap, 294 InlineFunctionInfo &IFI) { 295 CallGraph &CG = *IFI.CG; 296 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 297 const Function *Callee = CS.getCalledFunction(); 298 CallGraphNode *CalleeNode = CG[Callee]; 299 CallGraphNode *CallerNode = CG[Caller]; 300 301 // Since we inlined some uninlined call sites in the callee into the caller, 302 // add edges from the caller to all of the callees of the callee. 303 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 304 305 // Consider the case where CalleeNode == CallerNode. 306 CallGraphNode::CalledFunctionsVector CallCache; 307 if (CalleeNode == CallerNode) { 308 CallCache.assign(I, E); 309 I = CallCache.begin(); 310 E = CallCache.end(); 311 } 312 313 for (; I != E; ++I) { 314 const Value *OrigCall = I->first; 315 316 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); 317 // Only copy the edge if the call was inlined! 318 if (VMI == VMap.end() || VMI->second == 0) 319 continue; 320 321 // If the call was inlined, but then constant folded, there is no edge to 322 // add. Check for this case. 323 Instruction *NewCall = dyn_cast<Instruction>(VMI->second); 324 if (NewCall == 0) continue; 325 326 // Remember that this call site got inlined for the client of 327 // InlineFunction. 328 IFI.InlinedCalls.push_back(NewCall); 329 330 // It's possible that inlining the callsite will cause it to go from an 331 // indirect to a direct call by resolving a function pointer. If this 332 // happens, set the callee of the new call site to a more precise 333 // destination. This can also happen if the call graph node of the caller 334 // was just unnecessarily imprecise. 335 if (I->second->getFunction() == 0) 336 if (Function *F = CallSite(NewCall).getCalledFunction()) { 337 // Indirect call site resolved to direct call. 338 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); 339 340 continue; 341 } 342 343 CallerNode->addCalledFunction(CallSite(NewCall), I->second); 344 } 345 346 // Update the call graph by deleting the edge from Callee to Caller. We must 347 // do this after the loop above in case Caller and Callee are the same. 348 CallerNode->removeCallEdgeFor(CS); 349} 350 351/// HandleByValArgument - When inlining a call site that has a byval argument, 352/// we have to make the implicit memcpy explicit by adding it. 353static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, 354 const Function *CalledFunc, 355 InlineFunctionInfo &IFI, 356 unsigned ByValAlignment) { 357 Type *AggTy = cast<PointerType>(Arg->getType())->getElementType(); 358 359 // If the called function is readonly, then it could not mutate the caller's 360 // copy of the byval'd memory. In this case, it is safe to elide the copy and 361 // temporary. 362 if (CalledFunc->onlyReadsMemory()) { 363 // If the byval argument has a specified alignment that is greater than the 364 // passed in pointer, then we either have to round up the input pointer or 365 // give up on this transformation. 366 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. 367 return Arg; 368 369 // If the pointer is already known to be sufficiently aligned, or if we can 370 // round it up to a larger alignment, then we don't need a temporary. 371 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, 372 IFI.TD) >= ByValAlignment) 373 return Arg; 374 375 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad 376 // for code quality, but rarely happens and is required for correctness. 377 } 378 379 LLVMContext &Context = Arg->getContext(); 380 381 Type *VoidPtrTy = Type::getInt8PtrTy(Context); 382 383 // Create the alloca. If we have DataLayout, use nice alignment. 384 unsigned Align = 1; 385 if (IFI.TD) 386 Align = IFI.TD->getPrefTypeAlignment(AggTy); 387 388 // If the byval had an alignment specified, we *must* use at least that 389 // alignment, as it is required by the byval argument (and uses of the 390 // pointer inside the callee). 391 Align = std::max(Align, ByValAlignment); 392 393 Function *Caller = TheCall->getParent()->getParent(); 394 395 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(), 396 &*Caller->begin()->begin()); 397 // Emit a memcpy. 398 Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)}; 399 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 400 Intrinsic::memcpy, 401 Tys); 402 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 403 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall); 404 405 Value *Size; 406 if (IFI.TD == 0) 407 Size = ConstantExpr::getSizeOf(AggTy); 408 else 409 Size = ConstantInt::get(Type::getInt64Ty(Context), 410 IFI.TD->getTypeStoreSize(AggTy)); 411 412 // Always generate a memcpy of alignment 1 here because we don't know 413 // the alignment of the src pointer. Other optimizations can infer 414 // better alignment. 415 Value *CallArgs[] = { 416 DestCast, SrcCast, Size, 417 ConstantInt::get(Type::getInt32Ty(Context), 1), 418 ConstantInt::getFalse(Context) // isVolatile 419 }; 420 IRBuilder<>(TheCall).CreateCall(MemCpyFn, CallArgs); 421 422 // Uses of the argument in the function should use our new alloca 423 // instead. 424 return NewAlloca; 425} 426 427// isUsedByLifetimeMarker - Check whether this Value is used by a lifetime 428// intrinsic. 429static bool isUsedByLifetimeMarker(Value *V) { 430 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE; 431 ++UI) { 432 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI)) { 433 switch (II->getIntrinsicID()) { 434 default: break; 435 case Intrinsic::lifetime_start: 436 case Intrinsic::lifetime_end: 437 return true; 438 } 439 } 440 } 441 return false; 442} 443 444// hasLifetimeMarkers - Check whether the given alloca already has 445// lifetime.start or lifetime.end intrinsics. 446static bool hasLifetimeMarkers(AllocaInst *AI) { 447 Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext()); 448 if (AI->getType() == Int8PtrTy) 449 return isUsedByLifetimeMarker(AI); 450 451 // Do a scan to find all the casts to i8*. 452 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E; 453 ++I) { 454 if (I->getType() != Int8PtrTy) continue; 455 if (I->stripPointerCasts() != AI) continue; 456 if (isUsedByLifetimeMarker(*I)) 457 return true; 458 } 459 return false; 460} 461 462/// updateInlinedAtInfo - Helper function used by fixupLineNumbers to 463/// recursively update InlinedAtEntry of a DebugLoc. 464static DebugLoc updateInlinedAtInfo(const DebugLoc &DL, 465 const DebugLoc &InlinedAtDL, 466 LLVMContext &Ctx) { 467 if (MDNode *IA = DL.getInlinedAt(Ctx)) { 468 DebugLoc NewInlinedAtDL 469 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx); 470 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), 471 NewInlinedAtDL.getAsMDNode(Ctx)); 472 } 473 474 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), 475 InlinedAtDL.getAsMDNode(Ctx)); 476} 477 478/// fixupLineNumbers - Update inlined instructions' line numbers to 479/// to encode location where these instructions are inlined. 480static void fixupLineNumbers(Function *Fn, Function::iterator FI, 481 Instruction *TheCall) { 482 DebugLoc TheCallDL = TheCall->getDebugLoc(); 483 if (TheCallDL.isUnknown()) 484 return; 485 486 for (; FI != Fn->end(); ++FI) { 487 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end(); 488 BI != BE; ++BI) { 489 DebugLoc DL = BI->getDebugLoc(); 490 if (!DL.isUnknown()) { 491 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext())); 492 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) { 493 LLVMContext &Ctx = BI->getContext(); 494 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx); 495 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(), 496 InlinedAt, Ctx)); 497 } 498 } 499 } 500 } 501} 502 503/// InlineFunction - This function inlines the called function into the basic 504/// block of the caller. This returns false if it is not possible to inline 505/// this call. The program is still in a well defined state if this occurs 506/// though. 507/// 508/// Note that this only does one level of inlining. For example, if the 509/// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 510/// exists in the instruction stream. Similarly this will inline a recursive 511/// function by one level. 512bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI, 513 bool InsertLifetime) { 514 Instruction *TheCall = CS.getInstruction(); 515 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 516 "Instruction not in function!"); 517 518 // If IFI has any state in it, zap it before we fill it in. 519 IFI.reset(); 520 521 const Function *CalledFunc = CS.getCalledFunction(); 522 if (CalledFunc == 0 || // Can't inline external function or indirect 523 CalledFunc->isDeclaration() || // call, or call to a vararg function! 524 CalledFunc->getFunctionType()->isVarArg()) return false; 525 526 // If the call to the callee is not a tail call, we must clear the 'tail' 527 // flags on any calls that we inline. 528 bool MustClearTailCallFlags = 529 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall()); 530 531 // If the call to the callee cannot throw, set the 'nounwind' flag on any 532 // calls that we inline. 533 bool MarkNoUnwind = CS.doesNotThrow(); 534 535 BasicBlock *OrigBB = TheCall->getParent(); 536 Function *Caller = OrigBB->getParent(); 537 538 // GC poses two hazards to inlining, which only occur when the callee has GC: 539 // 1. If the caller has no GC, then the callee's GC must be propagated to the 540 // caller. 541 // 2. If the caller has a differing GC, it is invalid to inline. 542 if (CalledFunc->hasGC()) { 543 if (!Caller->hasGC()) 544 Caller->setGC(CalledFunc->getGC()); 545 else if (CalledFunc->getGC() != Caller->getGC()) 546 return false; 547 } 548 549 // Get the personality function from the callee if it contains a landing pad. 550 Value *CalleePersonality = 0; 551 for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end(); 552 I != E; ++I) 553 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) { 554 const BasicBlock *BB = II->getUnwindDest(); 555 const LandingPadInst *LP = BB->getLandingPadInst(); 556 CalleePersonality = LP->getPersonalityFn(); 557 break; 558 } 559 560 // Find the personality function used by the landing pads of the caller. If it 561 // exists, then check to see that it matches the personality function used in 562 // the callee. 563 if (CalleePersonality) { 564 for (Function::const_iterator I = Caller->begin(), E = Caller->end(); 565 I != E; ++I) 566 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) { 567 const BasicBlock *BB = II->getUnwindDest(); 568 const LandingPadInst *LP = BB->getLandingPadInst(); 569 570 // If the personality functions match, then we can perform the 571 // inlining. Otherwise, we can't inline. 572 // TODO: This isn't 100% true. Some personality functions are proper 573 // supersets of others and can be used in place of the other. 574 if (LP->getPersonalityFn() != CalleePersonality) 575 return false; 576 577 break; 578 } 579 } 580 581 // Get an iterator to the last basic block in the function, which will have 582 // the new function inlined after it. 583 Function::iterator LastBlock = &Caller->back(); 584 585 // Make sure to capture all of the return instructions from the cloned 586 // function. 587 SmallVector<ReturnInst*, 8> Returns; 588 ClonedCodeInfo InlinedFunctionInfo; 589 Function::iterator FirstNewBlock; 590 591 { // Scope to destroy VMap after cloning. 592 ValueToValueMapTy VMap; 593 594 assert(CalledFunc->arg_size() == CS.arg_size() && 595 "No varargs calls can be inlined!"); 596 597 // Calculate the vector of arguments to pass into the function cloner, which 598 // matches up the formal to the actual argument values. 599 CallSite::arg_iterator AI = CS.arg_begin(); 600 unsigned ArgNo = 0; 601 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 602 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 603 Value *ActualArg = *AI; 604 605 // When byval arguments actually inlined, we need to make the copy implied 606 // by them explicit. However, we don't do this if the callee is readonly 607 // or readnone, because the copy would be unneeded: the callee doesn't 608 // modify the struct. 609 if (CS.isByValArgument(ArgNo)) { 610 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, 611 CalledFunc->getParamAlignment(ArgNo+1)); 612 613 // Calls that we inline may use the new alloca, so we need to clear 614 // their 'tail' flags if HandleByValArgument introduced a new alloca and 615 // the callee has calls. 616 MustClearTailCallFlags |= ActualArg != *AI; 617 } 618 619 VMap[I] = ActualArg; 620 } 621 622 // We want the inliner to prune the code as it copies. We would LOVE to 623 // have no dead or constant instructions leftover after inlining occurs 624 // (which can happen, e.g., because an argument was constant), but we'll be 625 // happy with whatever the cloner can do. 626 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 627 /*ModuleLevelChanges=*/false, Returns, ".i", 628 &InlinedFunctionInfo, IFI.TD, TheCall); 629 630 // Remember the first block that is newly cloned over. 631 FirstNewBlock = LastBlock; ++FirstNewBlock; 632 633 // Update the callgraph if requested. 634 if (IFI.CG) 635 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); 636 637 // Update inlined instructions' line number information. 638 fixupLineNumbers(Caller, FirstNewBlock, TheCall); 639 } 640 641 // If there are any alloca instructions in the block that used to be the entry 642 // block for the callee, move them to the entry block of the caller. First 643 // calculate which instruction they should be inserted before. We insert the 644 // instructions at the end of the current alloca list. 645 { 646 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 647 for (BasicBlock::iterator I = FirstNewBlock->begin(), 648 E = FirstNewBlock->end(); I != E; ) { 649 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 650 if (AI == 0) continue; 651 652 // If the alloca is now dead, remove it. This often occurs due to code 653 // specialization. 654 if (AI->use_empty()) { 655 AI->eraseFromParent(); 656 continue; 657 } 658 659 if (!isa<Constant>(AI->getArraySize())) 660 continue; 661 662 // Keep track of the static allocas that we inline into the caller. 663 IFI.StaticAllocas.push_back(AI); 664 665 // Scan for the block of allocas that we can move over, and move them 666 // all at once. 667 while (isa<AllocaInst>(I) && 668 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 669 IFI.StaticAllocas.push_back(cast<AllocaInst>(I)); 670 ++I; 671 } 672 673 // Transfer all of the allocas over in a block. Using splice means 674 // that the instructions aren't removed from the symbol table, then 675 // reinserted. 676 Caller->getEntryBlock().getInstList().splice(InsertPoint, 677 FirstNewBlock->getInstList(), 678 AI, I); 679 } 680 } 681 682 // Leave lifetime markers for the static alloca's, scoping them to the 683 // function we just inlined. 684 if (InsertLifetime && !IFI.StaticAllocas.empty()) { 685 IRBuilder<> builder(FirstNewBlock->begin()); 686 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) { 687 AllocaInst *AI = IFI.StaticAllocas[ai]; 688 689 // If the alloca is already scoped to something smaller than the whole 690 // function then there's no need to add redundant, less accurate markers. 691 if (hasLifetimeMarkers(AI)) 692 continue; 693 694 // Try to determine the size of the allocation. 695 ConstantInt *AllocaSize = 0; 696 if (ConstantInt *AIArraySize = 697 dyn_cast<ConstantInt>(AI->getArraySize())) { 698 if (IFI.TD) { 699 Type *AllocaType = AI->getAllocatedType(); 700 uint64_t AllocaTypeSize = IFI.TD->getTypeAllocSize(AllocaType); 701 uint64_t AllocaArraySize = AIArraySize->getLimitedValue(); 702 assert(AllocaArraySize > 0 && "array size of AllocaInst is zero"); 703 // Check that array size doesn't saturate uint64_t and doesn't 704 // overflow when it's multiplied by type size. 705 if (AllocaArraySize != ~0ULL && 706 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) { 707 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()), 708 AllocaArraySize * AllocaTypeSize); 709 } 710 } 711 } 712 713 builder.CreateLifetimeStart(AI, AllocaSize); 714 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) { 715 IRBuilder<> builder(Returns[ri]); 716 builder.CreateLifetimeEnd(AI, AllocaSize); 717 } 718 } 719 } 720 721 // If the inlined code contained dynamic alloca instructions, wrap the inlined 722 // code with llvm.stacksave/llvm.stackrestore intrinsics. 723 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 724 Module *M = Caller->getParent(); 725 // Get the two intrinsics we care about. 726 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 727 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 728 729 // Insert the llvm.stacksave. 730 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin()) 731 .CreateCall(StackSave, "savedstack"); 732 733 // Insert a call to llvm.stackrestore before any return instructions in the 734 // inlined function. 735 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 736 IRBuilder<>(Returns[i]).CreateCall(StackRestore, SavedPtr); 737 } 738 } 739 740 // If we are inlining tail call instruction through a call site that isn't 741 // marked 'tail', we must remove the tail marker for any calls in the inlined 742 // code. Also, calls inlined through a 'nounwind' call site should be marked 743 // 'nounwind'. 744 if (InlinedFunctionInfo.ContainsCalls && 745 (MustClearTailCallFlags || MarkNoUnwind)) { 746 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 747 BB != E; ++BB) 748 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 749 if (CallInst *CI = dyn_cast<CallInst>(I)) { 750 if (MustClearTailCallFlags) 751 CI->setTailCall(false); 752 if (MarkNoUnwind) 753 CI->setDoesNotThrow(); 754 } 755 } 756 757 // If we are inlining for an invoke instruction, we must make sure to rewrite 758 // any call instructions into invoke instructions. 759 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 760 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 761 762 // If we cloned in _exactly one_ basic block, and if that block ends in a 763 // return instruction, we splice the body of the inlined callee directly into 764 // the calling basic block. 765 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 766 // Move all of the instructions right before the call. 767 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 768 FirstNewBlock->begin(), FirstNewBlock->end()); 769 // Remove the cloned basic block. 770 Caller->getBasicBlockList().pop_back(); 771 772 // If the call site was an invoke instruction, add a branch to the normal 773 // destination. 774 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 775 BranchInst::Create(II->getNormalDest(), TheCall); 776 777 // If the return instruction returned a value, replace uses of the call with 778 // uses of the returned value. 779 if (!TheCall->use_empty()) { 780 ReturnInst *R = Returns[0]; 781 if (TheCall == R->getReturnValue()) 782 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 783 else 784 TheCall->replaceAllUsesWith(R->getReturnValue()); 785 } 786 // Since we are now done with the Call/Invoke, we can delete it. 787 TheCall->eraseFromParent(); 788 789 // Since we are now done with the return instruction, delete it also. 790 Returns[0]->eraseFromParent(); 791 792 // We are now done with the inlining. 793 return true; 794 } 795 796 // Otherwise, we have the normal case, of more than one block to inline or 797 // multiple return sites. 798 799 // We want to clone the entire callee function into the hole between the 800 // "starter" and "ender" blocks. How we accomplish this depends on whether 801 // this is an invoke instruction or a call instruction. 802 BasicBlock *AfterCallBB; 803 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 804 805 // Add an unconditional branch to make this look like the CallInst case... 806 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 807 808 // Split the basic block. This guarantees that no PHI nodes will have to be 809 // updated due to new incoming edges, and make the invoke case more 810 // symmetric to the call case. 811 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 812 CalledFunc->getName()+".exit"); 813 814 } else { // It's a call 815 // If this is a call instruction, we need to split the basic block that 816 // the call lives in. 817 // 818 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 819 CalledFunc->getName()+".exit"); 820 } 821 822 // Change the branch that used to go to AfterCallBB to branch to the first 823 // basic block of the inlined function. 824 // 825 TerminatorInst *Br = OrigBB->getTerminator(); 826 assert(Br && Br->getOpcode() == Instruction::Br && 827 "splitBasicBlock broken!"); 828 Br->setOperand(0, FirstNewBlock); 829 830 831 // Now that the function is correct, make it a little bit nicer. In 832 // particular, move the basic blocks inserted from the end of the function 833 // into the space made by splitting the source basic block. 834 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 835 FirstNewBlock, Caller->end()); 836 837 // Handle all of the return instructions that we just cloned in, and eliminate 838 // any users of the original call/invoke instruction. 839 Type *RTy = CalledFunc->getReturnType(); 840 841 PHINode *PHI = 0; 842 if (Returns.size() > 1) { 843 // The PHI node should go at the front of the new basic block to merge all 844 // possible incoming values. 845 if (!TheCall->use_empty()) { 846 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(), 847 AfterCallBB->begin()); 848 // Anything that used the result of the function call should now use the 849 // PHI node as their operand. 850 TheCall->replaceAllUsesWith(PHI); 851 } 852 853 // Loop over all of the return instructions adding entries to the PHI node 854 // as appropriate. 855 if (PHI) { 856 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 857 ReturnInst *RI = Returns[i]; 858 assert(RI->getReturnValue()->getType() == PHI->getType() && 859 "Ret value not consistent in function!"); 860 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 861 } 862 } 863 864 865 // Add a branch to the merge points and remove return instructions. 866 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 867 ReturnInst *RI = Returns[i]; 868 BranchInst::Create(AfterCallBB, RI); 869 RI->eraseFromParent(); 870 } 871 } else if (!Returns.empty()) { 872 // Otherwise, if there is exactly one return value, just replace anything 873 // using the return value of the call with the computed value. 874 if (!TheCall->use_empty()) { 875 if (TheCall == Returns[0]->getReturnValue()) 876 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 877 else 878 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 879 } 880 881 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 882 BasicBlock *ReturnBB = Returns[0]->getParent(); 883 ReturnBB->replaceAllUsesWith(AfterCallBB); 884 885 // Splice the code from the return block into the block that it will return 886 // to, which contains the code that was after the call. 887 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 888 ReturnBB->getInstList()); 889 890 // Delete the return instruction now and empty ReturnBB now. 891 Returns[0]->eraseFromParent(); 892 ReturnBB->eraseFromParent(); 893 } else if (!TheCall->use_empty()) { 894 // No returns, but something is using the return value of the call. Just 895 // nuke the result. 896 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 897 } 898 899 // Since we are now done with the Call/Invoke, we can delete it. 900 TheCall->eraseFromParent(); 901 902 // We should always be able to fold the entry block of the function into the 903 // single predecessor of the block... 904 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 905 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 906 907 // Splice the code entry block into calling block, right before the 908 // unconditional branch. 909 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 910 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 911 912 // Remove the unconditional branch. 913 OrigBB->getInstList().erase(Br); 914 915 // Now we can remove the CalleeEntry block, which is now empty. 916 Caller->getBasicBlockList().erase(CalleeEntry); 917 918 // If we inserted a phi node, check to see if it has a single value (e.g. all 919 // the entries are the same or undef). If so, remove the PHI so it doesn't 920 // block other optimizations. 921 if (PHI) { 922 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) { 923 PHI->replaceAllUsesWith(V); 924 PHI->eraseFromParent(); 925 } 926 } 927 928 return true; 929} 930