InlineFunction.cpp revision e669d83a21d7ebf01d3c9e37a20c7348b5a77c11
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 48/// [LIBUNWIND] Find the (possibly absent) call to @llvm.eh.selector in 49/// the given landing pad. 50static EHSelectorInst *findSelectorForLandingPad(BasicBlock *lpad) { 51 // The llvm.eh.exception call is required to be in the landing pad. 52 for (BasicBlock::iterator i = lpad->begin(), e = lpad->end(); i != e; i++) { 53 EHExceptionInst *exn = dyn_cast<EHExceptionInst>(i); 54 if (!exn) continue; 55 56 EHSelectorInst *selector = 0; 57 for (Instruction::use_iterator 58 ui = exn->use_begin(), ue = exn->use_end(); ui != ue; ++ui) { 59 EHSelectorInst *sel = dyn_cast<EHSelectorInst>(*ui); 60 if (!sel) continue; 61 62 // Immediately accept an eh.selector in the landing pad. 63 if (sel->getParent() == lpad) return sel; 64 65 // Otherwise, use the first selector we see. 66 if (!selector) selector = sel; 67 } 68 69 return selector; 70 } 71 72 return 0; 73} 74 75namespace { 76 /// A class for recording information about inlining through an invoke. 77 class InvokeInliningInfo { 78 BasicBlock *OuterUnwindDest; 79 EHSelectorInst *OuterSelector; 80 BasicBlock *InnerUnwindDest; 81 PHINode *InnerExceptionPHI; 82 PHINode *InnerSelectorPHI; 83 SmallVector<Value*, 8> UnwindDestPHIValues; 84 85 public: 86 InvokeInliningInfo(InvokeInst *II) : 87 OuterUnwindDest(II->getUnwindDest()), OuterSelector(0), 88 InnerUnwindDest(0), InnerExceptionPHI(0), InnerSelectorPHI(0) { 89 90 // If there are PHI nodes in the unwind destination block, we 91 // need to keep track of which values came into them from the 92 // invoke before removing the edge from this block. 93 llvm::BasicBlock *invokeBB = II->getParent(); 94 for (BasicBlock::iterator I = OuterUnwindDest->begin(); 95 isa<PHINode>(I); ++I) { 96 // Save the value to use for this edge. 97 PHINode *phi = cast<PHINode>(I); 98 UnwindDestPHIValues.push_back(phi->getIncomingValueForBlock(invokeBB)); 99 } 100 } 101 102 /// The outer unwind destination is the target of unwind edges 103 /// introduced for calls within the inlined function. 104 BasicBlock *getOuterUnwindDest() const { 105 return OuterUnwindDest; 106 } 107 108 EHSelectorInst *getOuterSelector() { 109 if (!OuterSelector) 110 OuterSelector = findSelectorForLandingPad(OuterUnwindDest); 111 return OuterSelector; 112 } 113 114 BasicBlock *getInnerUnwindDest(); 115 116 bool forwardEHResume(CallInst *call, BasicBlock *src); 117 118 /// Add incoming-PHI values to the unwind destination block for 119 /// the given basic block, using the values for the original 120 /// invoke's source block. 121 void addIncomingPHIValuesFor(BasicBlock *BB) const { 122 addIncomingPHIValuesForInto(BB, OuterUnwindDest); 123 } 124 125 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const { 126 BasicBlock::iterator I = dest->begin(); 127 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 128 PHINode *phi = cast<PHINode>(I); 129 phi->addIncoming(UnwindDestPHIValues[i], src); 130 } 131 } 132 }; 133} 134 135/// Get or create a target for the branch out of rewritten calls to 136/// llvm.eh.resume. 137BasicBlock *InvokeInliningInfo::getInnerUnwindDest() { 138 if (InnerUnwindDest) return InnerUnwindDest; 139 140 // Find and hoist the llvm.eh.exception and llvm.eh.selector calls 141 // in the outer landing pad to immediately following the phis. 142 EHSelectorInst *selector = getOuterSelector(); 143 if (!selector) return 0; 144 145 // The call to llvm.eh.exception *must* be in the landing pad. 146 Instruction *exn = cast<Instruction>(selector->getArgOperand(0)); 147 assert(exn->getParent() == OuterUnwindDest); 148 149 // TODO: recognize when we've already done this, so that we don't 150 // get a linear number of these when inlining calls into lots of 151 // invokes with the same landing pad. 152 153 // Do the hoisting. 154 Instruction *splitPoint = exn->getParent()->getFirstNonPHI(); 155 assert(splitPoint != selector && "selector-on-exception dominance broken!"); 156 if (splitPoint == exn) { 157 selector->removeFromParent(); 158 selector->insertAfter(exn); 159 splitPoint = selector->getNextNode(); 160 } else { 161 exn->moveBefore(splitPoint); 162 selector->moveBefore(splitPoint); 163 } 164 165 // Split the landing pad. 166 InnerUnwindDest = OuterUnwindDest->splitBasicBlock(splitPoint, 167 OuterUnwindDest->getName() + ".body"); 168 169 // The number of incoming edges we expect to the inner landing pad. 170 const unsigned phiCapacity = 2; 171 172 // Create corresponding new phis for all the phis in the outer landing pad. 173 BasicBlock::iterator insertPoint = InnerUnwindDest->begin(); 174 BasicBlock::iterator I = OuterUnwindDest->begin(); 175 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 176 PHINode *outerPhi = cast<PHINode>(I); 177 PHINode *innerPhi = PHINode::Create(outerPhi->getType(), phiCapacity, 178 outerPhi->getName() + ".lpad-body", 179 insertPoint); 180 outerPhi->replaceAllUsesWith(innerPhi); 181 innerPhi->addIncoming(outerPhi, OuterUnwindDest); 182 } 183 184 // Create a phi for the exception value... 185 InnerExceptionPHI = PHINode::Create(exn->getType(), phiCapacity, 186 "exn.lpad-body", insertPoint); 187 exn->replaceAllUsesWith(InnerExceptionPHI); 188 selector->setArgOperand(0, exn); // restore this use 189 InnerExceptionPHI->addIncoming(exn, OuterUnwindDest); 190 191 // ...and the selector. 192 InnerSelectorPHI = PHINode::Create(selector->getType(), phiCapacity, 193 "selector.lpad-body", insertPoint); 194 selector->replaceAllUsesWith(InnerSelectorPHI); 195 InnerSelectorPHI->addIncoming(selector, OuterUnwindDest); 196 197 // All done. 198 return InnerUnwindDest; 199} 200 201/// [LIBUNWIND] Try to forward the given call, which logically occurs 202/// at the end of the given block, as a branch to the inner unwind 203/// block. Returns true if the call was forwarded. 204bool InvokeInliningInfo::forwardEHResume(CallInst *call, BasicBlock *src) { 205 // First, check whether this is a call to the intrinsic. 206 Function *fn = dyn_cast<Function>(call->getCalledValue()); 207 if (!fn || fn->getName() != "llvm.eh.resume") 208 return false; 209 210 // At this point, we need to return true on all paths, because 211 // otherwise we'll construct an invoke of the intrinsic, which is 212 // not well-formed. 213 214 // Try to find or make an inner unwind dest, which will fail if we 215 // can't find a selector call for the outer unwind dest. 216 BasicBlock *dest = getInnerUnwindDest(); 217 bool hasSelector = (dest != 0); 218 219 // If we failed, just use the outer unwind dest, dropping the 220 // exception and selector on the floor. 221 if (!hasSelector) 222 dest = OuterUnwindDest; 223 224 // Make a branch. 225 BranchInst::Create(dest, src); 226 227 // Update the phis in the destination. They were inserted in an 228 // order which makes this work. 229 addIncomingPHIValuesForInto(src, dest); 230 231 if (hasSelector) { 232 InnerExceptionPHI->addIncoming(call->getArgOperand(0), src); 233 InnerSelectorPHI->addIncoming(call->getArgOperand(1), src); 234 } 235 236 return true; 237} 238 239/// [LIBUNWIND] Check whether this selector is "only cleanups": 240/// call i32 @llvm.eh.selector(blah, blah, i32 0) 241static bool isCleanupOnlySelector(EHSelectorInst *selector) { 242 if (selector->getNumArgOperands() != 3) return false; 243 ConstantInt *val = dyn_cast<ConstantInt>(selector->getArgOperand(2)); 244 return (val && val->isZero()); 245} 246 247/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into 248/// an invoke, we have to turn all of the calls that can throw into 249/// invokes. This function analyze BB to see if there are any calls, and if so, 250/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI 251/// nodes in that block with the values specified in InvokeDestPHIValues. 252/// 253/// Returns true to indicate that the next block should be skipped. 254static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, 255 InvokeInliningInfo &Invoke) { 256 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 257 Instruction *I = BBI++; 258 259 // We only need to check for function calls: inlined invoke 260 // instructions require no special handling. 261 CallInst *CI = dyn_cast<CallInst>(I); 262 if (CI == 0) continue; 263 264 // LIBUNWIND: merge selector instructions. 265 if (EHSelectorInst *Inner = dyn_cast<EHSelectorInst>(CI)) { 266 EHSelectorInst *Outer = Invoke.getOuterSelector(); 267 if (!Outer) continue; 268 269 bool innerIsOnlyCleanup = isCleanupOnlySelector(Inner); 270 bool outerIsOnlyCleanup = isCleanupOnlySelector(Outer); 271 272 // If both selectors contain only cleanups, we don't need to do 273 // anything. TODO: this is really just a very specific instance 274 // of a much more general optimization. 275 if (innerIsOnlyCleanup && outerIsOnlyCleanup) continue; 276 277 // Otherwise, we just append the outer selector to the inner selector. 278 SmallVector<Value*, 16> NewSelector; 279 for (unsigned i = 0, e = Inner->getNumArgOperands(); i != e; ++i) 280 NewSelector.push_back(Inner->getArgOperand(i)); 281 for (unsigned i = 2, e = Outer->getNumArgOperands(); i != e; ++i) 282 NewSelector.push_back(Outer->getArgOperand(i)); 283 284 CallInst *NewInner = CallInst::Create(Inner->getCalledValue(), 285 NewSelector.begin(), 286 NewSelector.end(), 287 "", 288 Inner); 289 // No need to copy attributes, calling convention, etc. 290 NewInner->takeName(Inner); 291 Inner->replaceAllUsesWith(NewInner); 292 Inner->eraseFromParent(); 293 continue; 294 } 295 296 // If this call cannot unwind, don't convert it to an invoke. 297 if (CI->doesNotThrow()) 298 continue; 299 300 // Convert this function call into an invoke instruction. 301 // First, split the basic block. 302 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 303 304 // Delete the unconditional branch inserted by splitBasicBlock 305 BB->getInstList().pop_back(); 306 307 // LIBUNWIND: If this is a call to @llvm.eh.resume, just branch 308 // directly to the new landing pad. 309 if (Invoke.forwardEHResume(CI, BB)) { 310 // TODO: 'Split' is now unreachable; clean it up. 311 312 // We want to leave the original call intact so that the call 313 // graph and other structures won't get misled. We also have to 314 // avoid processing the next block, or we'll iterate here forever. 315 return true; 316 } 317 318 // Otherwise, create the new invoke instruction. 319 ImmutableCallSite CS(CI); 320 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end()); 321 InvokeInst *II = 322 InvokeInst::Create(CI->getCalledValue(), Split, 323 Invoke.getOuterUnwindDest(), 324 InvokeArgs.begin(), InvokeArgs.end(), 325 CI->getName(), BB); 326 II->setCallingConv(CI->getCallingConv()); 327 II->setAttributes(CI->getAttributes()); 328 329 // Make sure that anything using the call now uses the invoke! This also 330 // updates the CallGraph if present, because it uses a WeakVH. 331 CI->replaceAllUsesWith(II); 332 333 Split->getInstList().pop_front(); // Delete the original call 334 335 // Update any PHI nodes in the exceptional block to indicate that 336 // there is now a new entry in them. 337 Invoke.addIncomingPHIValuesFor(BB); 338 return false; 339 } 340 341 return false; 342} 343 344 345/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 346/// in the body of the inlined function into invokes and turn unwind 347/// instructions into branches to the invoke unwind dest. 348/// 349/// II is the invoke instruction being inlined. FirstNewBlock is the first 350/// block of the inlined code (the last block is the end of the function), 351/// and InlineCodeInfo is information about the code that got inlined. 352static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 353 ClonedCodeInfo &InlinedCodeInfo) { 354 BasicBlock *InvokeDest = II->getUnwindDest(); 355 356 Function *Caller = FirstNewBlock->getParent(); 357 358 // The inlined code is currently at the end of the function, scan from the 359 // start of the inlined code to its end, checking for stuff we need to 360 // rewrite. If the code doesn't have calls or unwinds, we know there is 361 // nothing to rewrite. 362 if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) { 363 // Now that everything is happy, we have one final detail. The PHI nodes in 364 // the exception destination block still have entries due to the original 365 // invoke instruction. Eliminate these entries (which might even delete the 366 // PHI node) now. 367 InvokeDest->removePredecessor(II->getParent()); 368 return; 369 } 370 371 InvokeInliningInfo Invoke(II); 372 373 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){ 374 if (InlinedCodeInfo.ContainsCalls) 375 if (HandleCallsInBlockInlinedThroughInvoke(BB, Invoke)) { 376 // Honor a request to skip the next block. We don't need to 377 // consider UnwindInsts in this case either. 378 ++BB; 379 continue; 380 } 381 382 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 383 // An UnwindInst requires special handling when it gets inlined into an 384 // invoke site. Once this happens, we know that the unwind would cause 385 // a control transfer to the invoke exception destination, so we can 386 // transform it into a direct branch to the exception destination. 387 BranchInst::Create(InvokeDest, UI); 388 389 // Delete the unwind instruction! 390 UI->eraseFromParent(); 391 392 // Update any PHI nodes in the exceptional block to indicate that 393 // there is now a new entry in them. 394 Invoke.addIncomingPHIValuesFor(BB); 395 } 396 } 397 398 // Now that everything is happy, we have one final detail. The PHI nodes in 399 // the exception destination block still have entries due to the original 400 // invoke instruction. Eliminate these entries (which might even delete the 401 // PHI node) now. 402 InvokeDest->removePredecessor(II->getParent()); 403} 404 405/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 406/// into the caller, update the specified callgraph to reflect the changes we 407/// made. Note that it's possible that not all code was copied over, so only 408/// some edges of the callgraph may remain. 409static void UpdateCallGraphAfterInlining(CallSite CS, 410 Function::iterator FirstNewBlock, 411 ValueToValueMapTy &VMap, 412 InlineFunctionInfo &IFI) { 413 CallGraph &CG = *IFI.CG; 414 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 415 const Function *Callee = CS.getCalledFunction(); 416 CallGraphNode *CalleeNode = CG[Callee]; 417 CallGraphNode *CallerNode = CG[Caller]; 418 419 // Since we inlined some uninlined call sites in the callee into the caller, 420 // add edges from the caller to all of the callees of the callee. 421 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 422 423 // Consider the case where CalleeNode == CallerNode. 424 CallGraphNode::CalledFunctionsVector CallCache; 425 if (CalleeNode == CallerNode) { 426 CallCache.assign(I, E); 427 I = CallCache.begin(); 428 E = CallCache.end(); 429 } 430 431 for (; I != E; ++I) { 432 const Value *OrigCall = I->first; 433 434 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); 435 // Only copy the edge if the call was inlined! 436 if (VMI == VMap.end() || VMI->second == 0) 437 continue; 438 439 // If the call was inlined, but then constant folded, there is no edge to 440 // add. Check for this case. 441 Instruction *NewCall = dyn_cast<Instruction>(VMI->second); 442 if (NewCall == 0) continue; 443 444 // Remember that this call site got inlined for the client of 445 // InlineFunction. 446 IFI.InlinedCalls.push_back(NewCall); 447 448 // It's possible that inlining the callsite will cause it to go from an 449 // indirect to a direct call by resolving a function pointer. If this 450 // happens, set the callee of the new call site to a more precise 451 // destination. This can also happen if the call graph node of the caller 452 // was just unnecessarily imprecise. 453 if (I->second->getFunction() == 0) 454 if (Function *F = CallSite(NewCall).getCalledFunction()) { 455 // Indirect call site resolved to direct call. 456 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); 457 458 continue; 459 } 460 461 CallerNode->addCalledFunction(CallSite(NewCall), I->second); 462 } 463 464 // Update the call graph by deleting the edge from Callee to Caller. We must 465 // do this after the loop above in case Caller and Callee are the same. 466 CallerNode->removeCallEdgeFor(CS); 467} 468 469/// HandleByValArgument - When inlining a call site that has a byval argument, 470/// we have to make the implicit memcpy explicit by adding it. 471static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, 472 const Function *CalledFunc, 473 InlineFunctionInfo &IFI, 474 unsigned ByValAlignment) { 475 const Type *AggTy = cast<PointerType>(Arg->getType())->getElementType(); 476 477 // If the called function is readonly, then it could not mutate the caller's 478 // copy of the byval'd memory. In this case, it is safe to elide the copy and 479 // temporary. 480 if (CalledFunc->onlyReadsMemory()) { 481 // If the byval argument has a specified alignment that is greater than the 482 // passed in pointer, then we either have to round up the input pointer or 483 // give up on this transformation. 484 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. 485 return Arg; 486 487 // If the pointer is already known to be sufficiently aligned, or if we can 488 // round it up to a larger alignment, then we don't need a temporary. 489 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, 490 IFI.TD) >= ByValAlignment) 491 return Arg; 492 493 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad 494 // for code quality, but rarely happens and is required for correctness. 495 } 496 497 LLVMContext &Context = Arg->getContext(); 498 499 const Type *VoidPtrTy = Type::getInt8PtrTy(Context); 500 501 // Create the alloca. If we have TargetData, use nice alignment. 502 unsigned Align = 1; 503 if (IFI.TD) 504 Align = IFI.TD->getPrefTypeAlignment(AggTy); 505 506 // If the byval had an alignment specified, we *must* use at least that 507 // alignment, as it is required by the byval argument (and uses of the 508 // pointer inside the callee). 509 Align = std::max(Align, ByValAlignment); 510 511 Function *Caller = TheCall->getParent()->getParent(); 512 513 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(), 514 &*Caller->begin()->begin()); 515 // Emit a memcpy. 516 const Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)}; 517 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 518 Intrinsic::memcpy, 519 Tys, 3); 520 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 521 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall); 522 523 Value *Size; 524 if (IFI.TD == 0) 525 Size = ConstantExpr::getSizeOf(AggTy); 526 else 527 Size = ConstantInt::get(Type::getInt64Ty(Context), 528 IFI.TD->getTypeStoreSize(AggTy)); 529 530 // Always generate a memcpy of alignment 1 here because we don't know 531 // the alignment of the src pointer. Other optimizations can infer 532 // better alignment. 533 Value *CallArgs[] = { 534 DestCast, SrcCast, Size, 535 ConstantInt::get(Type::getInt32Ty(Context), 1), 536 ConstantInt::getFalse(Context) // isVolatile 537 }; 538 CallInst::Create(MemCpyFn, CallArgs, CallArgs+5, "", TheCall); 539 540 // Uses of the argument in the function should use our new alloca 541 // instead. 542 return NewAlloca; 543} 544 545// isUsedByLifetimeMarker - Check whether this Value is used by a lifetime 546// intrinsic. 547static bool isUsedByLifetimeMarker(Value *V) { 548 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE; 549 ++UI) { 550 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI)) { 551 switch (II->getIntrinsicID()) { 552 default: break; 553 case Intrinsic::lifetime_start: 554 case Intrinsic::lifetime_end: 555 return true; 556 } 557 } 558 } 559 return false; 560} 561 562// hasLifetimeMarkers - Check whether the given alloca already has 563// lifetime.start or lifetime.end intrinsics. 564static bool hasLifetimeMarkers(AllocaInst *AI) { 565 const Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext()); 566 if (AI->getType() == Int8PtrTy) 567 return isUsedByLifetimeMarker(AI); 568 569 // Do a scan to find all the bitcasts to i8*. 570 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E; 571 ++I) { 572 if (I->getType() != Int8PtrTy) continue; 573 if (!isa<BitCastInst>(*I)) continue; 574 if (isUsedByLifetimeMarker(*I)) 575 return true; 576 } 577 return false; 578} 579 580// InlineFunction - This function inlines the called function into the basic 581// block of the caller. This returns false if it is not possible to inline this 582// call. The program is still in a well defined state if this occurs though. 583// 584// Note that this only does one level of inlining. For example, if the 585// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 586// exists in the instruction stream. Similarly this will inline a recursive 587// function by one level. 588// 589bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) { 590 Instruction *TheCall = CS.getInstruction(); 591 LLVMContext &Context = TheCall->getContext(); 592 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 593 "Instruction not in function!"); 594 595 // If IFI has any state in it, zap it before we fill it in. 596 IFI.reset(); 597 598 const Function *CalledFunc = CS.getCalledFunction(); 599 if (CalledFunc == 0 || // Can't inline external function or indirect 600 CalledFunc->isDeclaration() || // call, or call to a vararg function! 601 CalledFunc->getFunctionType()->isVarArg()) return false; 602 603 // If the call to the callee is not a tail call, we must clear the 'tail' 604 // flags on any calls that we inline. 605 bool MustClearTailCallFlags = 606 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall()); 607 608 // If the call to the callee cannot throw, set the 'nounwind' flag on any 609 // calls that we inline. 610 bool MarkNoUnwind = CS.doesNotThrow(); 611 612 BasicBlock *OrigBB = TheCall->getParent(); 613 Function *Caller = OrigBB->getParent(); 614 615 // GC poses two hazards to inlining, which only occur when the callee has GC: 616 // 1. If the caller has no GC, then the callee's GC must be propagated to the 617 // caller. 618 // 2. If the caller has a differing GC, it is invalid to inline. 619 if (CalledFunc->hasGC()) { 620 if (!Caller->hasGC()) 621 Caller->setGC(CalledFunc->getGC()); 622 else if (CalledFunc->getGC() != Caller->getGC()) 623 return false; 624 } 625 626 // Get an iterator to the last basic block in the function, which will have 627 // the new function inlined after it. 628 // 629 Function::iterator LastBlock = &Caller->back(); 630 631 // Make sure to capture all of the return instructions from the cloned 632 // function. 633 SmallVector<ReturnInst*, 8> Returns; 634 ClonedCodeInfo InlinedFunctionInfo; 635 Function::iterator FirstNewBlock; 636 637 { // Scope to destroy VMap after cloning. 638 ValueToValueMapTy VMap; 639 640 assert(CalledFunc->arg_size() == CS.arg_size() && 641 "No varargs calls can be inlined!"); 642 643 // Calculate the vector of arguments to pass into the function cloner, which 644 // matches up the formal to the actual argument values. 645 CallSite::arg_iterator AI = CS.arg_begin(); 646 unsigned ArgNo = 0; 647 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 648 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 649 Value *ActualArg = *AI; 650 651 // When byval arguments actually inlined, we need to make the copy implied 652 // by them explicit. However, we don't do this if the callee is readonly 653 // or readnone, because the copy would be unneeded: the callee doesn't 654 // modify the struct. 655 if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal)) { 656 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, 657 CalledFunc->getParamAlignment(ArgNo+1)); 658 659 // Calls that we inline may use the new alloca, so we need to clear 660 // their 'tail' flags if HandleByValArgument introduced a new alloca and 661 // the callee has calls. 662 MustClearTailCallFlags |= ActualArg != *AI; 663 } 664 665 VMap[I] = ActualArg; 666 } 667 668 // We want the inliner to prune the code as it copies. We would LOVE to 669 // have no dead or constant instructions leftover after inlining occurs 670 // (which can happen, e.g., because an argument was constant), but we'll be 671 // happy with whatever the cloner can do. 672 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 673 /*ModuleLevelChanges=*/false, Returns, ".i", 674 &InlinedFunctionInfo, IFI.TD, TheCall); 675 676 // Remember the first block that is newly cloned over. 677 FirstNewBlock = LastBlock; ++FirstNewBlock; 678 679 // Update the callgraph if requested. 680 if (IFI.CG) 681 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); 682 } 683 684 // If there are any alloca instructions in the block that used to be the entry 685 // block for the callee, move them to the entry block of the caller. First 686 // calculate which instruction they should be inserted before. We insert the 687 // instructions at the end of the current alloca list. 688 // 689 { 690 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 691 for (BasicBlock::iterator I = FirstNewBlock->begin(), 692 E = FirstNewBlock->end(); I != E; ) { 693 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 694 if (AI == 0) continue; 695 696 // If the alloca is now dead, remove it. This often occurs due to code 697 // specialization. 698 if (AI->use_empty()) { 699 AI->eraseFromParent(); 700 continue; 701 } 702 703 if (!isa<Constant>(AI->getArraySize())) 704 continue; 705 706 // Keep track of the static allocas that we inline into the caller. 707 IFI.StaticAllocas.push_back(AI); 708 709 // Scan for the block of allocas that we can move over, and move them 710 // all at once. 711 while (isa<AllocaInst>(I) && 712 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 713 IFI.StaticAllocas.push_back(cast<AllocaInst>(I)); 714 ++I; 715 } 716 717 // Transfer all of the allocas over in a block. Using splice means 718 // that the instructions aren't removed from the symbol table, then 719 // reinserted. 720 Caller->getEntryBlock().getInstList().splice(InsertPoint, 721 FirstNewBlock->getInstList(), 722 AI, I); 723 } 724 } 725 726 // Leave lifetime markers for the static alloca's, scoping them to the 727 // function we just inlined. 728 if (!IFI.StaticAllocas.empty()) { 729 // Also preserve the call graph, if applicable. 730 CallGraphNode *StartCGN = 0, *EndCGN = 0, *CallerNode = 0; 731 if (CallGraph *CG = IFI.CG) { 732 Function *Start = Intrinsic::getDeclaration(Caller->getParent(), 733 Intrinsic::lifetime_start); 734 Function *End = Intrinsic::getDeclaration(Caller->getParent(), 735 Intrinsic::lifetime_end); 736 StartCGN = CG->getOrInsertFunction(Start); 737 EndCGN = CG->getOrInsertFunction(End); 738 CallerNode = (*CG)[Caller]; 739 } 740 741 IRBuilder<> builder(FirstNewBlock->begin()); 742 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) { 743 AllocaInst *AI = IFI.StaticAllocas[ai]; 744 745 // If the alloca is already scoped to something smaller than the whole 746 // function then there's no need to add redundant, less accurate markers. 747 if (hasLifetimeMarkers(AI)) 748 continue; 749 750 builder.CreateLifetimeStart(AI); 751 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) { 752 IRBuilder<> builder(Returns[ri]); 753 builder.CreateLifetimeEnd(AI); 754 } 755 } 756 } 757 758 // If the inlined code contained dynamic alloca instructions, wrap the inlined 759 // code with llvm.stacksave/llvm.stackrestore intrinsics. 760 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 761 Module *M = Caller->getParent(); 762 // Get the two intrinsics we care about. 763 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 764 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 765 766 // Insert the llvm.stacksave. 767 CallInst *SavedPtr = CallInst::Create(StackSave, "savedstack", 768 FirstNewBlock->begin()); 769 770 // Insert a call to llvm.stackrestore before any return instructions in the 771 // inlined function. 772 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 773 CallInst::Create(StackRestore, SavedPtr, "", Returns[i]); 774 } 775 776 // Count the number of StackRestore calls we insert. 777 unsigned NumStackRestores = Returns.size(); 778 779 // If we are inlining an invoke instruction, insert restores before each 780 // unwind. These unwinds will be rewritten into branches later. 781 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 782 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 783 BB != E; ++BB) 784 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 785 CallInst::Create(StackRestore, SavedPtr, "", UI); 786 ++NumStackRestores; 787 } 788 } 789 } 790 791 // If we are inlining tail call instruction through a call site that isn't 792 // marked 'tail', we must remove the tail marker for any calls in the inlined 793 // code. Also, calls inlined through a 'nounwind' call site should be marked 794 // 'nounwind'. 795 if (InlinedFunctionInfo.ContainsCalls && 796 (MustClearTailCallFlags || MarkNoUnwind)) { 797 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 798 BB != E; ++BB) 799 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 800 if (CallInst *CI = dyn_cast<CallInst>(I)) { 801 if (MustClearTailCallFlags) 802 CI->setTailCall(false); 803 if (MarkNoUnwind) 804 CI->setDoesNotThrow(); 805 } 806 } 807 808 // If we are inlining through a 'nounwind' call site then any inlined 'unwind' 809 // instructions are unreachable. 810 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind) 811 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 812 BB != E; ++BB) { 813 TerminatorInst *Term = BB->getTerminator(); 814 if (isa<UnwindInst>(Term)) { 815 new UnreachableInst(Context, Term); 816 BB->getInstList().erase(Term); 817 } 818 } 819 820 // If we are inlining for an invoke instruction, we must make sure to rewrite 821 // any inlined 'unwind' instructions into branches to the invoke exception 822 // destination, and call instructions into invoke instructions. 823 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 824 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 825 826 // If we cloned in _exactly one_ basic block, and if that block ends in a 827 // return instruction, we splice the body of the inlined callee directly into 828 // the calling basic block. 829 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 830 // Move all of the instructions right before the call. 831 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 832 FirstNewBlock->begin(), FirstNewBlock->end()); 833 // Remove the cloned basic block. 834 Caller->getBasicBlockList().pop_back(); 835 836 // If the call site was an invoke instruction, add a branch to the normal 837 // destination. 838 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 839 BranchInst::Create(II->getNormalDest(), TheCall); 840 841 // If the return instruction returned a value, replace uses of the call with 842 // uses of the returned value. 843 if (!TheCall->use_empty()) { 844 ReturnInst *R = Returns[0]; 845 if (TheCall == R->getReturnValue()) 846 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 847 else 848 TheCall->replaceAllUsesWith(R->getReturnValue()); 849 } 850 // Since we are now done with the Call/Invoke, we can delete it. 851 TheCall->eraseFromParent(); 852 853 // Since we are now done with the return instruction, delete it also. 854 Returns[0]->eraseFromParent(); 855 856 // We are now done with the inlining. 857 return true; 858 } 859 860 // Otherwise, we have the normal case, of more than one block to inline or 861 // multiple return sites. 862 863 // We want to clone the entire callee function into the hole between the 864 // "starter" and "ender" blocks. How we accomplish this depends on whether 865 // this is an invoke instruction or a call instruction. 866 BasicBlock *AfterCallBB; 867 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 868 869 // Add an unconditional branch to make this look like the CallInst case... 870 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 871 872 // Split the basic block. This guarantees that no PHI nodes will have to be 873 // updated due to new incoming edges, and make the invoke case more 874 // symmetric to the call case. 875 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 876 CalledFunc->getName()+".exit"); 877 878 } else { // It's a call 879 // If this is a call instruction, we need to split the basic block that 880 // the call lives in. 881 // 882 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 883 CalledFunc->getName()+".exit"); 884 } 885 886 // Change the branch that used to go to AfterCallBB to branch to the first 887 // basic block of the inlined function. 888 // 889 TerminatorInst *Br = OrigBB->getTerminator(); 890 assert(Br && Br->getOpcode() == Instruction::Br && 891 "splitBasicBlock broken!"); 892 Br->setOperand(0, FirstNewBlock); 893 894 895 // Now that the function is correct, make it a little bit nicer. In 896 // particular, move the basic blocks inserted from the end of the function 897 // into the space made by splitting the source basic block. 898 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 899 FirstNewBlock, Caller->end()); 900 901 // Handle all of the return instructions that we just cloned in, and eliminate 902 // any users of the original call/invoke instruction. 903 const Type *RTy = CalledFunc->getReturnType(); 904 905 PHINode *PHI = 0; 906 if (Returns.size() > 1) { 907 // The PHI node should go at the front of the new basic block to merge all 908 // possible incoming values. 909 if (!TheCall->use_empty()) { 910 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(), 911 AfterCallBB->begin()); 912 // Anything that used the result of the function call should now use the 913 // PHI node as their operand. 914 TheCall->replaceAllUsesWith(PHI); 915 } 916 917 // Loop over all of the return instructions adding entries to the PHI node 918 // as appropriate. 919 if (PHI) { 920 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 921 ReturnInst *RI = Returns[i]; 922 assert(RI->getReturnValue()->getType() == PHI->getType() && 923 "Ret value not consistent in function!"); 924 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 925 } 926 } 927 928 929 // Add a branch to the merge points and remove return instructions. 930 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 931 ReturnInst *RI = Returns[i]; 932 BranchInst::Create(AfterCallBB, RI); 933 RI->eraseFromParent(); 934 } 935 } else if (!Returns.empty()) { 936 // Otherwise, if there is exactly one return value, just replace anything 937 // using the return value of the call with the computed value. 938 if (!TheCall->use_empty()) { 939 if (TheCall == Returns[0]->getReturnValue()) 940 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 941 else 942 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 943 } 944 945 // Splice the code from the return block into the block that it will return 946 // to, which contains the code that was after the call. 947 BasicBlock *ReturnBB = Returns[0]->getParent(); 948 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 949 ReturnBB->getInstList()); 950 951 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 952 ReturnBB->replaceAllUsesWith(AfterCallBB); 953 954 // Delete the return instruction now and empty ReturnBB now. 955 Returns[0]->eraseFromParent(); 956 ReturnBB->eraseFromParent(); 957 } else if (!TheCall->use_empty()) { 958 // No returns, but something is using the return value of the call. Just 959 // nuke the result. 960 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 961 } 962 963 // Since we are now done with the Call/Invoke, we can delete it. 964 TheCall->eraseFromParent(); 965 966 // We should always be able to fold the entry block of the function into the 967 // single predecessor of the block... 968 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 969 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 970 971 // Splice the code entry block into calling block, right before the 972 // unconditional branch. 973 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 974 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 975 976 // Remove the unconditional branch. 977 OrigBB->getInstList().erase(Br); 978 979 // Now we can remove the CalleeEntry block, which is now empty. 980 Caller->getBasicBlockList().erase(CalleeEntry); 981 982 // If we inserted a phi node, check to see if it has a single value (e.g. all 983 // the entries are the same or undef). If so, remove the PHI so it doesn't 984 // block other optimizations. 985 if (PHI) 986 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) { 987 PHI->replaceAllUsesWith(V); 988 PHI->eraseFromParent(); 989 } 990 991 return true; 992} 993