InlineFunction.cpp revision a3de16bc8f36638d5444e3e7b0112998af54f826
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// The code in this file for handling inlines through invoke 14// instructions preserves semantics only under some assumptions about 15// the behavior of unwinders which correspond to gcc-style libUnwind 16// exception personality functions. Eventually the IR will be 17// improved to make this unnecessary, but until then, this code is 18// marked [LIBUNWIND]. 19// 20//===----------------------------------------------------------------------===// 21 22#include "llvm/Transforms/Utils/Cloning.h" 23#include "llvm/Constants.h" 24#include "llvm/DerivedTypes.h" 25#include "llvm/Module.h" 26#include "llvm/Instructions.h" 27#include "llvm/IntrinsicInst.h" 28#include "llvm/Intrinsics.h" 29#include "llvm/Attributes.h" 30#include "llvm/Analysis/CallGraph.h" 31#include "llvm/Analysis/DebugInfo.h" 32#include "llvm/Analysis/InstructionSimplify.h" 33#include "llvm/Target/TargetData.h" 34#include "llvm/Transforms/Utils/Local.h" 35#include "llvm/ADT/SmallVector.h" 36#include "llvm/ADT/StringExtras.h" 37#include "llvm/Support/CallSite.h" 38#include "llvm/Support/IRBuilder.h" 39using namespace llvm; 40 41bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI) { 42 return InlineFunction(CallSite(CI), IFI); 43} 44bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI) { 45 return InlineFunction(CallSite(II), IFI); 46} 47 48namespace { 49 /// A class for recording information about inlining through an invoke. 50 class InvokeInliningInfo { 51 BasicBlock *UnwindDest; 52 SmallVector<Value*, 8> UnwindDestPHIValues; 53 54 public: 55 InvokeInliningInfo(InvokeInst *II) : UnwindDest(II->getUnwindDest()) { 56 // If there are PHI nodes in the unwind destination block, we 57 // need to keep track of which values came into them from the 58 // invoke before removing the edge from this block. 59 llvm::BasicBlock *InvokeBlock = II->getParent(); 60 for (BasicBlock::iterator I = UnwindDest->begin(); isa<PHINode>(I); ++I) { 61 PHINode *PN = cast<PHINode>(I); 62 // Save the value to use for this edge. 63 llvm::Value *Incoming = PN->getIncomingValueForBlock(InvokeBlock); 64 UnwindDestPHIValues.push_back(Incoming); 65 } 66 } 67 68 BasicBlock *getUnwindDest() const { 69 return UnwindDest; 70 } 71 72 /// Add incoming-PHI values to the unwind destination block for 73 /// the given basic block, using the values for the original 74 /// invoke's source block. 75 void addIncomingPHIValuesFor(BasicBlock *BB) const { 76 BasicBlock::iterator I = UnwindDest->begin(); 77 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 78 PHINode *PN = cast<PHINode>(I); 79 PN->addIncoming(UnwindDestPHIValues[i], BB); 80 } 81 } 82 }; 83} 84 85/// [LIBUNWIND] Check whether the given value is the _Unwind_Resume 86/// function specified by the Itanium EH ABI. 87static bool isUnwindResume(Value *value) { 88 Function *fn = dyn_cast<Function>(value); 89 if (!fn) return false; 90 91 // declare void @_Unwind_Resume(i8*) 92 if (fn->getName() != "_Unwind_Resume") return false; 93 const FunctionType *fnType = fn->getFunctionType(); 94 if (!fnType->getReturnType()->isVoidTy()) return false; 95 if (fnType->isVarArg()) return false; 96 if (fnType->getNumParams() != 1) return false; 97 const PointerType *paramType = dyn_cast<PointerType>(fnType->getParamType(0)); 98 return (paramType && paramType->getElementType()->isIntegerTy(8)); 99} 100 101/// [LIBUNWIND] Find the (possibly absent) call to @llvm.eh.selector in 102/// the given landing pad. 103static EHSelectorInst *findSelectorForLandingPad(BasicBlock *lpad) { 104 for (BasicBlock::iterator i = lpad->begin(), e = lpad->end(); i != e; i++) 105 if (EHSelectorInst *selector = dyn_cast<EHSelectorInst>(i)) 106 return selector; 107 return 0; 108} 109 110/// [LIBUNWIND] Check whether this selector is "only cleanups": 111/// call i32 @llvm.eh.selector(blah, blah, i32 0) 112static bool isCleanupOnlySelector(EHSelectorInst *selector) { 113 if (selector->getNumArgOperands() != 3) return false; 114 ConstantInt *val = dyn_cast<ConstantInt>(selector->getArgOperand(2)); 115 return (val && val->isZero()); 116} 117 118/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into 119/// an invoke, we have to turn all of the calls that can throw into 120/// invokes. This function analyze BB to see if there are any calls, and if so, 121/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI 122/// nodes in that block with the values specified in InvokeDestPHIValues. 123/// 124/// Returns true to indicate that the next block should be skipped. 125static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, 126 InvokeInliningInfo &Invoke) { 127 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 128 Instruction *I = BBI++; 129 130 // We only need to check for function calls: inlined invoke 131 // instructions require no special handling. 132 CallInst *CI = dyn_cast<CallInst>(I); 133 if (CI == 0) continue; 134 135 // LIBUNWIND: merge selector instructions. 136 if (EHSelectorInst *Inner = dyn_cast<EHSelectorInst>(CI)) { 137 EHSelectorInst *Outer = findSelectorForLandingPad(Invoke.getUnwindDest()); 138 if (!Outer) continue; 139 140 bool innerIsOnlyCleanup = isCleanupOnlySelector(Inner); 141 bool outerIsOnlyCleanup = isCleanupOnlySelector(Outer); 142 143 // If both selectors contain only cleanups, we don't need to do 144 // anything. TODO: this is really just a very specific instance 145 // of a much more general optimization. 146 if (innerIsOnlyCleanup && outerIsOnlyCleanup) continue; 147 148 // Otherwise, we just append the outer selector to the inner selector. 149 SmallVector<Value*, 16> NewSelector; 150 for (unsigned i = 0, e = Inner->getNumArgOperands(); i != e; ++i) 151 NewSelector.push_back(Inner->getArgOperand(i)); 152 for (unsigned i = 2, e = Outer->getNumArgOperands(); i != e; ++i) 153 NewSelector.push_back(Outer->getArgOperand(i)); 154 155 CallInst *NewInner = CallInst::Create(Inner->getCalledValue(), 156 NewSelector.begin(), 157 NewSelector.end(), 158 "", 159 Inner); 160 // No need to copy attributes, calling convention, etc. 161 NewInner->takeName(Inner); 162 Inner->replaceAllUsesWith(NewInner); 163 Inner->eraseFromParent(); 164 continue; 165 } 166 167 // If this call cannot unwind, don't convert it to an invoke. 168 if (CI->doesNotThrow()) 169 continue; 170 171 // Convert this function call into an invoke instruction. 172 // First, split the basic block. 173 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 174 175 bool skipNextBlock = false; 176 177 // LIBUNWIND: If this is a call to @_Unwind_Resume, just branch 178 // directly to the new landing pad. 179 if (isUnwindResume(CI->getCalledValue())) { 180 BranchInst::Create(Invoke.getUnwindDest(), BB->getTerminator()); 181 182 // TODO: 'Split' is now unreachable; clean it up. 183 184 // We want to leave the original call intact so that the call 185 // graph and other structures won't get misled. We also have to 186 // avoid processing the next block, or we'll iterate here forever. 187 skipNextBlock = true; 188 189 // Otherwise, create the new invoke instruction. 190 } else { 191 ImmutableCallSite CS(CI); 192 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end()); 193 InvokeInst *II = 194 InvokeInst::Create(CI->getCalledValue(), Split, Invoke.getUnwindDest(), 195 InvokeArgs.begin(), InvokeArgs.end(), 196 CI->getName(), BB->getTerminator()); 197 II->setCallingConv(CI->getCallingConv()); 198 II->setAttributes(CI->getAttributes()); 199 200 // Make sure that anything using the call now uses the invoke! This also 201 // updates the CallGraph if present, because it uses a WeakVH. 202 CI->replaceAllUsesWith(II); 203 204 Split->getInstList().pop_front(); // Delete the original call 205 } 206 207 // Delete the unconditional branch inserted by splitBasicBlock 208 BB->getInstList().pop_back(); 209 210 // Update any PHI nodes in the exceptional block to indicate that 211 // there is now a new entry in them. 212 Invoke.addIncomingPHIValuesFor(BB); 213 214 // This basic block is now complete, the caller will continue scanning the 215 // next one. 216 return skipNextBlock; 217 } 218 219 return false; 220} 221 222 223/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 224/// in the body of the inlined function into invokes and turn unwind 225/// instructions into branches to the invoke unwind dest. 226/// 227/// II is the invoke instruction being inlined. FirstNewBlock is the first 228/// block of the inlined code (the last block is the end of the function), 229/// and InlineCodeInfo is information about the code that got inlined. 230static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 231 ClonedCodeInfo &InlinedCodeInfo) { 232 BasicBlock *InvokeDest = II->getUnwindDest(); 233 234 Function *Caller = FirstNewBlock->getParent(); 235 236 // The inlined code is currently at the end of the function, scan from the 237 // start of the inlined code to its end, checking for stuff we need to 238 // rewrite. If the code doesn't have calls or unwinds, we know there is 239 // nothing to rewrite. 240 if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) { 241 // Now that everything is happy, we have one final detail. The PHI nodes in 242 // the exception destination block still have entries due to the original 243 // invoke instruction. Eliminate these entries (which might even delete the 244 // PHI node) now. 245 InvokeDest->removePredecessor(II->getParent()); 246 return; 247 } 248 249 InvokeInliningInfo Invoke(II); 250 251 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){ 252 if (InlinedCodeInfo.ContainsCalls) 253 if (HandleCallsInBlockInlinedThroughInvoke(BB, Invoke)) { 254 // Honor a request to skip the next block. We don't need to 255 // consider UnwindInsts in this case either. 256 ++BB; 257 continue; 258 } 259 260 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 261 // An UnwindInst requires special handling when it gets inlined into an 262 // invoke site. Once this happens, we know that the unwind would cause 263 // a control transfer to the invoke exception destination, so we can 264 // transform it into a direct branch to the exception destination. 265 BranchInst::Create(InvokeDest, UI); 266 267 // Delete the unwind instruction! 268 UI->eraseFromParent(); 269 270 // Update any PHI nodes in the exceptional block to indicate that 271 // there is now a new entry in them. 272 Invoke.addIncomingPHIValuesFor(BB); 273 } 274 } 275 276 // Now that everything is happy, we have one final detail. The PHI nodes in 277 // the exception destination block still have entries due to the original 278 // invoke instruction. Eliminate these entries (which might even delete the 279 // PHI node) now. 280 InvokeDest->removePredecessor(II->getParent()); 281} 282 283/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 284/// into the caller, update the specified callgraph to reflect the changes we 285/// made. Note that it's possible that not all code was copied over, so only 286/// some edges of the callgraph may remain. 287static void UpdateCallGraphAfterInlining(CallSite CS, 288 Function::iterator FirstNewBlock, 289 ValueToValueMapTy &VMap, 290 InlineFunctionInfo &IFI) { 291 CallGraph &CG = *IFI.CG; 292 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 293 const Function *Callee = CS.getCalledFunction(); 294 CallGraphNode *CalleeNode = CG[Callee]; 295 CallGraphNode *CallerNode = CG[Caller]; 296 297 // Since we inlined some uninlined call sites in the callee into the caller, 298 // add edges from the caller to all of the callees of the callee. 299 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 300 301 // Consider the case where CalleeNode == CallerNode. 302 CallGraphNode::CalledFunctionsVector CallCache; 303 if (CalleeNode == CallerNode) { 304 CallCache.assign(I, E); 305 I = CallCache.begin(); 306 E = CallCache.end(); 307 } 308 309 for (; I != E; ++I) { 310 const Value *OrigCall = I->first; 311 312 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); 313 // Only copy the edge if the call was inlined! 314 if (VMI == VMap.end() || VMI->second == 0) 315 continue; 316 317 // If the call was inlined, but then constant folded, there is no edge to 318 // add. Check for this case. 319 Instruction *NewCall = dyn_cast<Instruction>(VMI->second); 320 if (NewCall == 0) continue; 321 322 // Remember that this call site got inlined for the client of 323 // InlineFunction. 324 IFI.InlinedCalls.push_back(NewCall); 325 326 // It's possible that inlining the callsite will cause it to go from an 327 // indirect to a direct call by resolving a function pointer. If this 328 // happens, set the callee of the new call site to a more precise 329 // destination. This can also happen if the call graph node of the caller 330 // was just unnecessarily imprecise. 331 if (I->second->getFunction() == 0) 332 if (Function *F = CallSite(NewCall).getCalledFunction()) { 333 // Indirect call site resolved to direct call. 334 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); 335 336 continue; 337 } 338 339 CallerNode->addCalledFunction(CallSite(NewCall), I->second); 340 } 341 342 // Update the call graph by deleting the edge from Callee to Caller. We must 343 // do this after the loop above in case Caller and Callee are the same. 344 CallerNode->removeCallEdgeFor(CS); 345} 346 347/// HandleByValArgument - When inlining a call site that has a byval argument, 348/// we have to make the implicit memcpy explicit by adding it. 349static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, 350 const Function *CalledFunc, 351 InlineFunctionInfo &IFI, 352 unsigned ByValAlignment) { 353 const Type *AggTy = cast<PointerType>(Arg->getType())->getElementType(); 354 355 // If the called function is readonly, then it could not mutate the caller's 356 // copy of the byval'd memory. In this case, it is safe to elide the copy and 357 // temporary. 358 if (CalledFunc->onlyReadsMemory()) { 359 // If the byval argument has a specified alignment that is greater than the 360 // passed in pointer, then we either have to round up the input pointer or 361 // give up on this transformation. 362 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. 363 return Arg; 364 365 // If the pointer is already known to be sufficiently aligned, or if we can 366 // round it up to a larger alignment, then we don't need a temporary. 367 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, 368 IFI.TD) >= ByValAlignment) 369 return Arg; 370 371 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad 372 // for code quality, but rarely happens and is required for correctness. 373 } 374 375 LLVMContext &Context = Arg->getContext(); 376 377 const Type *VoidPtrTy = Type::getInt8PtrTy(Context); 378 379 // Create the alloca. If we have TargetData, use nice alignment. 380 unsigned Align = 1; 381 if (IFI.TD) 382 Align = IFI.TD->getPrefTypeAlignment(AggTy); 383 384 // If the byval had an alignment specified, we *must* use at least that 385 // alignment, as it is required by the byval argument (and uses of the 386 // pointer inside the callee). 387 Align = std::max(Align, ByValAlignment); 388 389 Function *Caller = TheCall->getParent()->getParent(); 390 391 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(), 392 &*Caller->begin()->begin()); 393 // Emit a memcpy. 394 const Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)}; 395 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 396 Intrinsic::memcpy, 397 Tys, 3); 398 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 399 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall); 400 401 Value *Size; 402 if (IFI.TD == 0) 403 Size = ConstantExpr::getSizeOf(AggTy); 404 else 405 Size = ConstantInt::get(Type::getInt64Ty(Context), 406 IFI.TD->getTypeStoreSize(AggTy)); 407 408 // Always generate a memcpy of alignment 1 here because we don't know 409 // the alignment of the src pointer. Other optimizations can infer 410 // better alignment. 411 Value *CallArgs[] = { 412 DestCast, SrcCast, Size, 413 ConstantInt::get(Type::getInt32Ty(Context), 1), 414 ConstantInt::getFalse(Context) // isVolatile 415 }; 416 CallInst *TheMemCpy = 417 CallInst::Create(MemCpyFn, CallArgs, CallArgs+5, "", TheCall); 418 419 // If we have a call graph, update it. 420 if (CallGraph *CG = IFI.CG) { 421 CallGraphNode *MemCpyCGN = CG->getOrInsertFunction(MemCpyFn); 422 CallGraphNode *CallerNode = (*CG)[Caller]; 423 CallerNode->addCalledFunction(TheMemCpy, MemCpyCGN); 424 } 425 426 // Uses of the argument in the function should use our new alloca 427 // instead. 428 return NewAlloca; 429} 430 431// isUsedByLifetimeMarker - Check whether this Value is used by a lifetime 432// intrinsic. 433static bool isUsedByLifetimeMarker(Value *V) { 434 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE; 435 ++UI) { 436 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI)) { 437 switch (II->getIntrinsicID()) { 438 default: break; 439 case Intrinsic::lifetime_start: 440 case Intrinsic::lifetime_end: 441 return true; 442 } 443 } 444 } 445 return false; 446} 447 448// hasLifetimeMarkers - Check whether the given alloca already has 449// lifetime.start or lifetime.end intrinsics. 450static bool hasLifetimeMarkers(AllocaInst *AI) { 451 const Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext()); 452 if (AI->getType() == Int8PtrTy) 453 return isUsedByLifetimeMarker(AI); 454 455 // Do a scan to find all the bitcasts to i8*. 456 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E; 457 ++I) { 458 if (I->getType() != Int8PtrTy) continue; 459 if (!isa<BitCastInst>(*I)) continue; 460 if (isUsedByLifetimeMarker(*I)) 461 return true; 462 } 463 return false; 464} 465 466// InlineFunction - This function inlines the called function into the basic 467// block of the caller. This returns false if it is not possible to inline this 468// call. The program is still in a well defined state if this occurs though. 469// 470// Note that this only does one level of inlining. For example, if the 471// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 472// exists in the instruction stream. Similarly this will inline a recursive 473// function by one level. 474// 475bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) { 476 Instruction *TheCall = CS.getInstruction(); 477 LLVMContext &Context = TheCall->getContext(); 478 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 479 "Instruction not in function!"); 480 481 // If IFI has any state in it, zap it before we fill it in. 482 IFI.reset(); 483 484 const Function *CalledFunc = CS.getCalledFunction(); 485 if (CalledFunc == 0 || // Can't inline external function or indirect 486 CalledFunc->isDeclaration() || // call, or call to a vararg function! 487 CalledFunc->getFunctionType()->isVarArg()) return false; 488 489 // If the call to the callee is not a tail call, we must clear the 'tail' 490 // flags on any calls that we inline. 491 bool MustClearTailCallFlags = 492 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall()); 493 494 // If the call to the callee cannot throw, set the 'nounwind' flag on any 495 // calls that we inline. 496 bool MarkNoUnwind = CS.doesNotThrow(); 497 498 BasicBlock *OrigBB = TheCall->getParent(); 499 Function *Caller = OrigBB->getParent(); 500 501 // GC poses two hazards to inlining, which only occur when the callee has GC: 502 // 1. If the caller has no GC, then the callee's GC must be propagated to the 503 // caller. 504 // 2. If the caller has a differing GC, it is invalid to inline. 505 if (CalledFunc->hasGC()) { 506 if (!Caller->hasGC()) 507 Caller->setGC(CalledFunc->getGC()); 508 else if (CalledFunc->getGC() != Caller->getGC()) 509 return false; 510 } 511 512 // Get an iterator to the last basic block in the function, which will have 513 // the new function inlined after it. 514 // 515 Function::iterator LastBlock = &Caller->back(); 516 517 // Make sure to capture all of the return instructions from the cloned 518 // function. 519 SmallVector<ReturnInst*, 8> Returns; 520 ClonedCodeInfo InlinedFunctionInfo; 521 Function::iterator FirstNewBlock; 522 523 { // Scope to destroy VMap after cloning. 524 ValueToValueMapTy VMap; 525 526 assert(CalledFunc->arg_size() == CS.arg_size() && 527 "No varargs calls can be inlined!"); 528 529 // Calculate the vector of arguments to pass into the function cloner, which 530 // matches up the formal to the actual argument values. 531 CallSite::arg_iterator AI = CS.arg_begin(); 532 unsigned ArgNo = 0; 533 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 534 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 535 Value *ActualArg = *AI; 536 537 // When byval arguments actually inlined, we need to make the copy implied 538 // by them explicit. However, we don't do this if the callee is readonly 539 // or readnone, because the copy would be unneeded: the callee doesn't 540 // modify the struct. 541 if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal)) { 542 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, 543 CalledFunc->getParamAlignment(ArgNo+1)); 544 545 // Calls that we inline may use the new alloca, so we need to clear 546 // their 'tail' flags if HandleByValArgument introduced a new alloca and 547 // the callee has calls. 548 MustClearTailCallFlags |= ActualArg != *AI; 549 } 550 551 VMap[I] = ActualArg; 552 } 553 554 // We want the inliner to prune the code as it copies. We would LOVE to 555 // have no dead or constant instructions leftover after inlining occurs 556 // (which can happen, e.g., because an argument was constant), but we'll be 557 // happy with whatever the cloner can do. 558 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 559 /*ModuleLevelChanges=*/false, Returns, ".i", 560 &InlinedFunctionInfo, IFI.TD, TheCall); 561 562 // Remember the first block that is newly cloned over. 563 FirstNewBlock = LastBlock; ++FirstNewBlock; 564 565 // Update the callgraph if requested. 566 if (IFI.CG) 567 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); 568 } 569 570 // If there are any alloca instructions in the block that used to be the entry 571 // block for the callee, move them to the entry block of the caller. First 572 // calculate which instruction they should be inserted before. We insert the 573 // instructions at the end of the current alloca list. 574 // 575 { 576 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 577 for (BasicBlock::iterator I = FirstNewBlock->begin(), 578 E = FirstNewBlock->end(); I != E; ) { 579 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 580 if (AI == 0) continue; 581 582 // If the alloca is now dead, remove it. This often occurs due to code 583 // specialization. 584 if (AI->use_empty()) { 585 AI->eraseFromParent(); 586 continue; 587 } 588 589 if (!isa<Constant>(AI->getArraySize())) 590 continue; 591 592 // Keep track of the static allocas that we inline into the caller. 593 IFI.StaticAllocas.push_back(AI); 594 595 // Scan for the block of allocas that we can move over, and move them 596 // all at once. 597 while (isa<AllocaInst>(I) && 598 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 599 IFI.StaticAllocas.push_back(cast<AllocaInst>(I)); 600 ++I; 601 } 602 603 // Transfer all of the allocas over in a block. Using splice means 604 // that the instructions aren't removed from the symbol table, then 605 // reinserted. 606 Caller->getEntryBlock().getInstList().splice(InsertPoint, 607 FirstNewBlock->getInstList(), 608 AI, I); 609 } 610 } 611 612 // Leave lifetime markers for the static alloca's, scoping them to the 613 // function we just inlined. 614 if (!IFI.StaticAllocas.empty()) { 615 // Also preserve the call graph, if applicable. 616 CallGraphNode *StartCGN = 0, *EndCGN = 0, *CallerNode = 0; 617 if (CallGraph *CG = IFI.CG) { 618 Function *Start = Intrinsic::getDeclaration(Caller->getParent(), 619 Intrinsic::lifetime_start); 620 Function *End = Intrinsic::getDeclaration(Caller->getParent(), 621 Intrinsic::lifetime_end); 622 StartCGN = CG->getOrInsertFunction(Start); 623 EndCGN = CG->getOrInsertFunction(End); 624 CallerNode = (*CG)[Caller]; 625 } 626 627 IRBuilder<> builder(FirstNewBlock->begin()); 628 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) { 629 AllocaInst *AI = IFI.StaticAllocas[ai]; 630 631 // If the alloca is already scoped to something smaller than the whole 632 // function then there's no need to add redundant, less accurate markers. 633 if (hasLifetimeMarkers(AI)) 634 continue; 635 636 CallInst *StartCall = builder.CreateLifetimeStart(AI); 637 if (IFI.CG) CallerNode->addCalledFunction(StartCall, StartCGN); 638 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) { 639 IRBuilder<> builder(Returns[ri]); 640 CallInst *EndCall = builder.CreateLifetimeEnd(AI); 641 if (IFI.CG) CallerNode->addCalledFunction(EndCall, EndCGN); 642 } 643 } 644 } 645 646 // If the inlined code contained dynamic alloca instructions, wrap the inlined 647 // code with llvm.stacksave/llvm.stackrestore intrinsics. 648 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 649 Module *M = Caller->getParent(); 650 // Get the two intrinsics we care about. 651 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 652 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 653 654 // If we are preserving the callgraph, add edges to the stacksave/restore 655 // functions for the calls we insert. 656 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0; 657 if (CallGraph *CG = IFI.CG) { 658 StackSaveCGN = CG->getOrInsertFunction(StackSave); 659 StackRestoreCGN = CG->getOrInsertFunction(StackRestore); 660 CallerNode = (*CG)[Caller]; 661 } 662 663 // Insert the llvm.stacksave. 664 CallInst *SavedPtr = CallInst::Create(StackSave, "savedstack", 665 FirstNewBlock->begin()); 666 if (IFI.CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN); 667 668 // Insert a call to llvm.stackrestore before any return instructions in the 669 // inlined function. 670 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 671 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", Returns[i]); 672 if (IFI.CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 673 } 674 675 // Count the number of StackRestore calls we insert. 676 unsigned NumStackRestores = Returns.size(); 677 678 // If we are inlining an invoke instruction, insert restores before each 679 // unwind. These unwinds will be rewritten into branches later. 680 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 681 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 682 BB != E; ++BB) 683 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 684 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", UI); 685 if (IFI.CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 686 ++NumStackRestores; 687 } 688 } 689 } 690 691 // If we are inlining tail call instruction through a call site that isn't 692 // marked 'tail', we must remove the tail marker for any calls in the inlined 693 // code. Also, calls inlined through a 'nounwind' call site should be marked 694 // 'nounwind'. 695 if (InlinedFunctionInfo.ContainsCalls && 696 (MustClearTailCallFlags || MarkNoUnwind)) { 697 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 698 BB != E; ++BB) 699 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 700 if (CallInst *CI = dyn_cast<CallInst>(I)) { 701 if (MustClearTailCallFlags) 702 CI->setTailCall(false); 703 if (MarkNoUnwind) 704 CI->setDoesNotThrow(); 705 } 706 } 707 708 // If we are inlining through a 'nounwind' call site then any inlined 'unwind' 709 // instructions are unreachable. 710 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind) 711 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 712 BB != E; ++BB) { 713 TerminatorInst *Term = BB->getTerminator(); 714 if (isa<UnwindInst>(Term)) { 715 new UnreachableInst(Context, Term); 716 BB->getInstList().erase(Term); 717 } 718 } 719 720 // If we are inlining for an invoke instruction, we must make sure to rewrite 721 // any inlined 'unwind' instructions into branches to the invoke exception 722 // destination, and call instructions into invoke instructions. 723 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 724 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 725 726 // If we cloned in _exactly one_ basic block, and if that block ends in a 727 // return instruction, we splice the body of the inlined callee directly into 728 // the calling basic block. 729 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 730 // Move all of the instructions right before the call. 731 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 732 FirstNewBlock->begin(), FirstNewBlock->end()); 733 // Remove the cloned basic block. 734 Caller->getBasicBlockList().pop_back(); 735 736 // If the call site was an invoke instruction, add a branch to the normal 737 // destination. 738 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 739 BranchInst::Create(II->getNormalDest(), TheCall); 740 741 // If the return instruction returned a value, replace uses of the call with 742 // uses of the returned value. 743 if (!TheCall->use_empty()) { 744 ReturnInst *R = Returns[0]; 745 if (TheCall == R->getReturnValue()) 746 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 747 else 748 TheCall->replaceAllUsesWith(R->getReturnValue()); 749 } 750 // Since we are now done with the Call/Invoke, we can delete it. 751 TheCall->eraseFromParent(); 752 753 // Since we are now done with the return instruction, delete it also. 754 Returns[0]->eraseFromParent(); 755 756 // We are now done with the inlining. 757 return true; 758 } 759 760 // Otherwise, we have the normal case, of more than one block to inline or 761 // multiple return sites. 762 763 // We want to clone the entire callee function into the hole between the 764 // "starter" and "ender" blocks. How we accomplish this depends on whether 765 // this is an invoke instruction or a call instruction. 766 BasicBlock *AfterCallBB; 767 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 768 769 // Add an unconditional branch to make this look like the CallInst case... 770 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 771 772 // Split the basic block. This guarantees that no PHI nodes will have to be 773 // updated due to new incoming edges, and make the invoke case more 774 // symmetric to the call case. 775 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 776 CalledFunc->getName()+".exit"); 777 778 } else { // It's a call 779 // If this is a call instruction, we need to split the basic block that 780 // the call lives in. 781 // 782 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 783 CalledFunc->getName()+".exit"); 784 } 785 786 // Change the branch that used to go to AfterCallBB to branch to the first 787 // basic block of the inlined function. 788 // 789 TerminatorInst *Br = OrigBB->getTerminator(); 790 assert(Br && Br->getOpcode() == Instruction::Br && 791 "splitBasicBlock broken!"); 792 Br->setOperand(0, FirstNewBlock); 793 794 795 // Now that the function is correct, make it a little bit nicer. In 796 // particular, move the basic blocks inserted from the end of the function 797 // into the space made by splitting the source basic block. 798 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 799 FirstNewBlock, Caller->end()); 800 801 // Handle all of the return instructions that we just cloned in, and eliminate 802 // any users of the original call/invoke instruction. 803 const Type *RTy = CalledFunc->getReturnType(); 804 805 PHINode *PHI = 0; 806 if (Returns.size() > 1) { 807 // The PHI node should go at the front of the new basic block to merge all 808 // possible incoming values. 809 if (!TheCall->use_empty()) { 810 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(), 811 AfterCallBB->begin()); 812 // Anything that used the result of the function call should now use the 813 // PHI node as their operand. 814 TheCall->replaceAllUsesWith(PHI); 815 } 816 817 // Loop over all of the return instructions adding entries to the PHI node 818 // as appropriate. 819 if (PHI) { 820 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 821 ReturnInst *RI = Returns[i]; 822 assert(RI->getReturnValue()->getType() == PHI->getType() && 823 "Ret value not consistent in function!"); 824 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 825 } 826 } 827 828 829 // Add a branch to the merge points and remove return instructions. 830 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 831 ReturnInst *RI = Returns[i]; 832 BranchInst::Create(AfterCallBB, RI); 833 RI->eraseFromParent(); 834 } 835 } else if (!Returns.empty()) { 836 // Otherwise, if there is exactly one return value, just replace anything 837 // using the return value of the call with the computed value. 838 if (!TheCall->use_empty()) { 839 if (TheCall == Returns[0]->getReturnValue()) 840 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 841 else 842 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 843 } 844 845 // Splice the code from the return block into the block that it will return 846 // to, which contains the code that was after the call. 847 BasicBlock *ReturnBB = Returns[0]->getParent(); 848 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 849 ReturnBB->getInstList()); 850 851 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 852 ReturnBB->replaceAllUsesWith(AfterCallBB); 853 854 // Delete the return instruction now and empty ReturnBB now. 855 Returns[0]->eraseFromParent(); 856 ReturnBB->eraseFromParent(); 857 } else if (!TheCall->use_empty()) { 858 // No returns, but something is using the return value of the call. Just 859 // nuke the result. 860 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 861 } 862 863 // Since we are now done with the Call/Invoke, we can delete it. 864 TheCall->eraseFromParent(); 865 866 // We should always be able to fold the entry block of the function into the 867 // single predecessor of the block... 868 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 869 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 870 871 // Splice the code entry block into calling block, right before the 872 // unconditional branch. 873 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 874 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 875 876 // Remove the unconditional branch. 877 OrigBB->getInstList().erase(Br); 878 879 // Now we can remove the CalleeEntry block, which is now empty. 880 Caller->getBasicBlockList().erase(CalleeEntry); 881 882 // If we inserted a phi node, check to see if it has a single value (e.g. all 883 // the entries are the same or undef). If so, remove the PHI so it doesn't 884 // block other optimizations. 885 if (PHI) 886 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) { 887 PHI->replaceAllUsesWith(V); 888 PHI->eraseFromParent(); 889 } 890 891 return true; 892} 893