InlineFunction.cpp revision 6d55f2269e20298a1d6a683be72d9552482156a9
1//===- InlineFunction.cpp - Code to perform function inlining -------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements inlining of a function into a call site, resolving 11// parameters and the return value as appropriate. 12// 13//===----------------------------------------------------------------------===// 14 15#include "llvm/Transforms/Utils/Cloning.h" 16#include "llvm/Constants.h" 17#include "llvm/DerivedTypes.h" 18#include "llvm/Module.h" 19#include "llvm/Instructions.h" 20#include "llvm/IntrinsicInst.h" 21#include "llvm/Intrinsics.h" 22#include "llvm/Attributes.h" 23#include "llvm/Analysis/CallGraph.h" 24#include "llvm/Analysis/DebugInfo.h" 25#include "llvm/Analysis/InstructionSimplify.h" 26#include "llvm/Target/TargetData.h" 27#include "llvm/Transforms/Utils/Local.h" 28#include "llvm/ADT/SmallVector.h" 29#include "llvm/ADT/StringExtras.h" 30#include "llvm/Support/CallSite.h" 31#include "llvm/Support/IRBuilder.h" 32using namespace llvm; 33 34bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI) { 35 return InlineFunction(CallSite(CI), IFI); 36} 37bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI) { 38 return InlineFunction(CallSite(II), IFI); 39} 40 41 42/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into 43/// an invoke, we have to turn all of the calls that can throw into 44/// invokes. This function analyze BB to see if there are any calls, and if so, 45/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI 46/// nodes in that block with the values specified in InvokeDestPHIValues. 47/// 48static void HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, 49 BasicBlock *InvokeDest, 50 const SmallVectorImpl<Value*> &InvokeDestPHIValues) { 51 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 52 Instruction *I = BBI++; 53 54 // We only need to check for function calls: inlined invoke 55 // instructions require no special handling. 56 CallInst *CI = dyn_cast<CallInst>(I); 57 if (CI == 0) continue; 58 59 // If this call cannot unwind, don't convert it to an invoke. 60 if (CI->doesNotThrow()) 61 continue; 62 63 // Convert this function call into an invoke instruction. 64 // First, split the basic block. 65 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 66 67 // Next, create the new invoke instruction, inserting it at the end 68 // of the old basic block. 69 ImmutableCallSite CS(CI); 70 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end()); 71 InvokeInst *II = 72 InvokeInst::Create(CI->getCalledValue(), Split, InvokeDest, 73 InvokeArgs.begin(), InvokeArgs.end(), 74 CI->getName(), BB->getTerminator()); 75 II->setCallingConv(CI->getCallingConv()); 76 II->setAttributes(CI->getAttributes()); 77 78 // Make sure that anything using the call now uses the invoke! This also 79 // updates the CallGraph if present, because it uses a WeakVH. 80 CI->replaceAllUsesWith(II); 81 82 // Delete the unconditional branch inserted by splitBasicBlock 83 BB->getInstList().pop_back(); 84 Split->getInstList().pop_front(); // Delete the original call 85 86 // Update any PHI nodes in the exceptional block to indicate that 87 // there is now a new entry in them. 88 unsigned i = 0; 89 for (BasicBlock::iterator I = InvokeDest->begin(); 90 isa<PHINode>(I); ++I, ++i) 91 cast<PHINode>(I)->addIncoming(InvokeDestPHIValues[i], BB); 92 93 // This basic block is now complete, the caller will continue scanning the 94 // next one. 95 return; 96 } 97} 98 99 100/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 101/// in the body of the inlined function into invokes and turn unwind 102/// instructions into branches to the invoke unwind dest. 103/// 104/// II is the invoke instruction being inlined. FirstNewBlock is the first 105/// block of the inlined code (the last block is the end of the function), 106/// and InlineCodeInfo is information about the code that got inlined. 107static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 108 ClonedCodeInfo &InlinedCodeInfo) { 109 BasicBlock *InvokeDest = II->getUnwindDest(); 110 SmallVector<Value*, 8> InvokeDestPHIValues; 111 112 // If there are PHI nodes in the unwind destination block, we need to 113 // keep track of which values came into them from this invoke, then remove 114 // the entry for this block. 115 BasicBlock *InvokeBlock = II->getParent(); 116 for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) { 117 PHINode *PN = cast<PHINode>(I); 118 // Save the value to use for this edge. 119 InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock)); 120 } 121 122 Function *Caller = FirstNewBlock->getParent(); 123 124 // The inlined code is currently at the end of the function, scan from the 125 // start of the inlined code to its end, checking for stuff we need to 126 // rewrite. If the code doesn't have calls or unwinds, we know there is 127 // nothing to rewrite. 128 if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) { 129 // Now that everything is happy, we have one final detail. The PHI nodes in 130 // the exception destination block still have entries due to the original 131 // invoke instruction. Eliminate these entries (which might even delete the 132 // PHI node) now. 133 InvokeDest->removePredecessor(II->getParent()); 134 return; 135 } 136 137 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){ 138 if (InlinedCodeInfo.ContainsCalls) 139 HandleCallsInBlockInlinedThroughInvoke(BB, InvokeDest, 140 InvokeDestPHIValues); 141 142 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 143 // An UnwindInst requires special handling when it gets inlined into an 144 // invoke site. Once this happens, we know that the unwind would cause 145 // a control transfer to the invoke exception destination, so we can 146 // transform it into a direct branch to the exception destination. 147 BranchInst::Create(InvokeDest, UI); 148 149 // Delete the unwind instruction! 150 UI->eraseFromParent(); 151 152 // Update any PHI nodes in the exceptional block to indicate that 153 // there is now a new entry in them. 154 unsigned i = 0; 155 for (BasicBlock::iterator I = InvokeDest->begin(); 156 isa<PHINode>(I); ++I, ++i) { 157 PHINode *PN = cast<PHINode>(I); 158 PN->addIncoming(InvokeDestPHIValues[i], BB); 159 } 160 } 161 } 162 163 // Now that everything is happy, we have one final detail. The PHI nodes in 164 // the exception destination block still have entries due to the original 165 // invoke instruction. Eliminate these entries (which might even delete the 166 // PHI node) now. 167 InvokeDest->removePredecessor(II->getParent()); 168} 169 170/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 171/// into the caller, update the specified callgraph to reflect the changes we 172/// made. Note that it's possible that not all code was copied over, so only 173/// some edges of the callgraph may remain. 174static void UpdateCallGraphAfterInlining(CallSite CS, 175 Function::iterator FirstNewBlock, 176 ValueToValueMapTy &VMap, 177 InlineFunctionInfo &IFI) { 178 CallGraph &CG = *IFI.CG; 179 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 180 const Function *Callee = CS.getCalledFunction(); 181 CallGraphNode *CalleeNode = CG[Callee]; 182 CallGraphNode *CallerNode = CG[Caller]; 183 184 // Since we inlined some uninlined call sites in the callee into the caller, 185 // add edges from the caller to all of the callees of the callee. 186 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 187 188 // Consider the case where CalleeNode == CallerNode. 189 CallGraphNode::CalledFunctionsVector CallCache; 190 if (CalleeNode == CallerNode) { 191 CallCache.assign(I, E); 192 I = CallCache.begin(); 193 E = CallCache.end(); 194 } 195 196 for (; I != E; ++I) { 197 const Value *OrigCall = I->first; 198 199 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); 200 // Only copy the edge if the call was inlined! 201 if (VMI == VMap.end() || VMI->second == 0) 202 continue; 203 204 // If the call was inlined, but then constant folded, there is no edge to 205 // add. Check for this case. 206 Instruction *NewCall = dyn_cast<Instruction>(VMI->second); 207 if (NewCall == 0) continue; 208 209 // Remember that this call site got inlined for the client of 210 // InlineFunction. 211 IFI.InlinedCalls.push_back(NewCall); 212 213 // It's possible that inlining the callsite will cause it to go from an 214 // indirect to a direct call by resolving a function pointer. If this 215 // happens, set the callee of the new call site to a more precise 216 // destination. This can also happen if the call graph node of the caller 217 // was just unnecessarily imprecise. 218 if (I->second->getFunction() == 0) 219 if (Function *F = CallSite(NewCall).getCalledFunction()) { 220 // Indirect call site resolved to direct call. 221 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); 222 223 continue; 224 } 225 226 CallerNode->addCalledFunction(CallSite(NewCall), I->second); 227 } 228 229 // Update the call graph by deleting the edge from Callee to Caller. We must 230 // do this after the loop above in case Caller and Callee are the same. 231 CallerNode->removeCallEdgeFor(CS); 232} 233 234/// HandleByValArgument - When inlining a call site that has a byval argument, 235/// we have to make the implicit memcpy explicit by adding it. 236static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, 237 const Function *CalledFunc, 238 InlineFunctionInfo &IFI, 239 unsigned ByValAlignment) { 240 const Type *AggTy = cast<PointerType>(Arg->getType())->getElementType(); 241 242 // If the called function is readonly, then it could not mutate the caller's 243 // copy of the byval'd memory. In this case, it is safe to elide the copy and 244 // temporary. 245 if (CalledFunc->onlyReadsMemory()) { 246 // If the byval argument has a specified alignment that is greater than the 247 // passed in pointer, then we either have to round up the input pointer or 248 // give up on this transformation. 249 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. 250 return Arg; 251 252 // If the pointer is already known to be sufficiently aligned, or if we can 253 // round it up to a larger alignment, then we don't need a temporary. 254 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, 255 IFI.TD) >= ByValAlignment) 256 return Arg; 257 258 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad 259 // for code quality, but rarely happens and is required for correctness. 260 } 261 262 LLVMContext &Context = Arg->getContext(); 263 264 const Type *VoidPtrTy = Type::getInt8PtrTy(Context); 265 266 // Create the alloca. If we have TargetData, use nice alignment. 267 unsigned Align = 1; 268 if (IFI.TD) 269 Align = IFI.TD->getPrefTypeAlignment(AggTy); 270 271 // If the byval had an alignment specified, we *must* use at least that 272 // alignment, as it is required by the byval argument (and uses of the 273 // pointer inside the callee). 274 Align = std::max(Align, ByValAlignment); 275 276 Function *Caller = TheCall->getParent()->getParent(); 277 278 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(), 279 &*Caller->begin()->begin()); 280 // Emit a memcpy. 281 const Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)}; 282 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 283 Intrinsic::memcpy, 284 Tys, 3); 285 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 286 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall); 287 288 Value *Size; 289 if (IFI.TD == 0) 290 Size = ConstantExpr::getSizeOf(AggTy); 291 else 292 Size = ConstantInt::get(Type::getInt64Ty(Context), 293 IFI.TD->getTypeStoreSize(AggTy)); 294 295 // Always generate a memcpy of alignment 1 here because we don't know 296 // the alignment of the src pointer. Other optimizations can infer 297 // better alignment. 298 Value *CallArgs[] = { 299 DestCast, SrcCast, Size, 300 ConstantInt::get(Type::getInt32Ty(Context), 1), 301 ConstantInt::getFalse(Context) // isVolatile 302 }; 303 CallInst *TheMemCpy = 304 CallInst::Create(MemCpyFn, CallArgs, CallArgs+5, "", TheCall); 305 306 // If we have a call graph, update it. 307 if (CallGraph *CG = IFI.CG) { 308 CallGraphNode *MemCpyCGN = CG->getOrInsertFunction(MemCpyFn); 309 CallGraphNode *CallerNode = (*CG)[Caller]; 310 CallerNode->addCalledFunction(TheMemCpy, MemCpyCGN); 311 } 312 313 // Uses of the argument in the function should use our new alloca 314 // instead. 315 return NewAlloca; 316} 317 318// isUsedByLifetimeMarker - Check whether this Value is used by a lifetime 319// intrinsic. 320static bool isUsedByLifetimeMarker(Value *V) { 321 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE; 322 ++UI) { 323 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI)) { 324 switch (II->getIntrinsicID()) { 325 default: break; 326 case Intrinsic::lifetime_start: 327 case Intrinsic::lifetime_end: 328 return true; 329 } 330 } 331 } 332 return false; 333} 334 335// hasLifetimeMarkers - Check whether the given alloca already has 336// lifetime.start or lifetime.end intrinsics. 337static bool hasLifetimeMarkers(AllocaInst *AI) { 338 const Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext()); 339 if (AI->getType() == Int8PtrTy) 340 return isUsedByLifetimeMarker(AI); 341 342 // Do a scan to find all the bitcasts to i8*. 343 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E; 344 ++I) { 345 if (I->getType() != Int8PtrTy) continue; 346 if (!isa<BitCastInst>(*I)) continue; 347 if (isUsedByLifetimeMarker(*I)) 348 return true; 349 } 350 return false; 351} 352 353// InlineFunction - This function inlines the called function into the basic 354// block of the caller. This returns false if it is not possible to inline this 355// call. The program is still in a well defined state if this occurs though. 356// 357// Note that this only does one level of inlining. For example, if the 358// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 359// exists in the instruction stream. Similarly this will inline a recursive 360// function by one level. 361// 362bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) { 363 Instruction *TheCall = CS.getInstruction(); 364 LLVMContext &Context = TheCall->getContext(); 365 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 366 "Instruction not in function!"); 367 368 // If IFI has any state in it, zap it before we fill it in. 369 IFI.reset(); 370 371 const Function *CalledFunc = CS.getCalledFunction(); 372 if (CalledFunc == 0 || // Can't inline external function or indirect 373 CalledFunc->isDeclaration() || // call, or call to a vararg function! 374 CalledFunc->getFunctionType()->isVarArg()) return false; 375 376 // If the call to the callee is not a tail call, we must clear the 'tail' 377 // flags on any calls that we inline. 378 bool MustClearTailCallFlags = 379 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall()); 380 381 // If the call to the callee cannot throw, set the 'nounwind' flag on any 382 // calls that we inline. 383 bool MarkNoUnwind = CS.doesNotThrow(); 384 385 BasicBlock *OrigBB = TheCall->getParent(); 386 Function *Caller = OrigBB->getParent(); 387 388 // GC poses two hazards to inlining, which only occur when the callee has GC: 389 // 1. If the caller has no GC, then the callee's GC must be propagated to the 390 // caller. 391 // 2. If the caller has a differing GC, it is invalid to inline. 392 if (CalledFunc->hasGC()) { 393 if (!Caller->hasGC()) 394 Caller->setGC(CalledFunc->getGC()); 395 else if (CalledFunc->getGC() != Caller->getGC()) 396 return false; 397 } 398 399 // Get an iterator to the last basic block in the function, which will have 400 // the new function inlined after it. 401 // 402 Function::iterator LastBlock = &Caller->back(); 403 404 // Make sure to capture all of the return instructions from the cloned 405 // function. 406 SmallVector<ReturnInst*, 8> Returns; 407 ClonedCodeInfo InlinedFunctionInfo; 408 Function::iterator FirstNewBlock; 409 410 { // Scope to destroy VMap after cloning. 411 ValueToValueMapTy VMap; 412 413 assert(CalledFunc->arg_size() == CS.arg_size() && 414 "No varargs calls can be inlined!"); 415 416 // Calculate the vector of arguments to pass into the function cloner, which 417 // matches up the formal to the actual argument values. 418 CallSite::arg_iterator AI = CS.arg_begin(); 419 unsigned ArgNo = 0; 420 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 421 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 422 Value *ActualArg = *AI; 423 424 // When byval arguments actually inlined, we need to make the copy implied 425 // by them explicit. However, we don't do this if the callee is readonly 426 // or readnone, because the copy would be unneeded: the callee doesn't 427 // modify the struct. 428 if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal)) { 429 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, 430 CalledFunc->getParamAlignment(ArgNo+1)); 431 432 // Calls that we inline may use the new alloca, so we need to clear 433 // their 'tail' flags if HandleByValArgument introduced a new alloca and 434 // the callee has calls. 435 MustClearTailCallFlags |= ActualArg != *AI; 436 } 437 438 VMap[I] = ActualArg; 439 } 440 441 // We want the inliner to prune the code as it copies. We would LOVE to 442 // have no dead or constant instructions leftover after inlining occurs 443 // (which can happen, e.g., because an argument was constant), but we'll be 444 // happy with whatever the cloner can do. 445 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 446 /*ModuleLevelChanges=*/false, Returns, ".i", 447 &InlinedFunctionInfo, IFI.TD, TheCall); 448 449 // Remember the first block that is newly cloned over. 450 FirstNewBlock = LastBlock; ++FirstNewBlock; 451 452 // Update the callgraph if requested. 453 if (IFI.CG) 454 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); 455 } 456 457 // If there are any alloca instructions in the block that used to be the entry 458 // block for the callee, move them to the entry block of the caller. First 459 // calculate which instruction they should be inserted before. We insert the 460 // instructions at the end of the current alloca list. 461 // 462 { 463 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 464 for (BasicBlock::iterator I = FirstNewBlock->begin(), 465 E = FirstNewBlock->end(); I != E; ) { 466 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 467 if (AI == 0) continue; 468 469 // If the alloca is now dead, remove it. This often occurs due to code 470 // specialization. 471 if (AI->use_empty()) { 472 AI->eraseFromParent(); 473 continue; 474 } 475 476 if (!isa<Constant>(AI->getArraySize())) 477 continue; 478 479 // Keep track of the static allocas that we inline into the caller. 480 IFI.StaticAllocas.push_back(AI); 481 482 // Scan for the block of allocas that we can move over, and move them 483 // all at once. 484 while (isa<AllocaInst>(I) && 485 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 486 IFI.StaticAllocas.push_back(cast<AllocaInst>(I)); 487 ++I; 488 } 489 490 // Transfer all of the allocas over in a block. Using splice means 491 // that the instructions aren't removed from the symbol table, then 492 // reinserted. 493 Caller->getEntryBlock().getInstList().splice(InsertPoint, 494 FirstNewBlock->getInstList(), 495 AI, I); 496 } 497 } 498 499 // Leave lifetime markers for the static alloca's, scoping them to the 500 // function we just inlined. 501 if (!IFI.StaticAllocas.empty()) { 502 // Also preserve the call graph, if applicable. 503 CallGraphNode *StartCGN = 0, *EndCGN = 0, *CallerNode = 0; 504 if (CallGraph *CG = IFI.CG) { 505 Function *Start = Intrinsic::getDeclaration(Caller->getParent(), 506 Intrinsic::lifetime_start); 507 Function *End = Intrinsic::getDeclaration(Caller->getParent(), 508 Intrinsic::lifetime_end); 509 StartCGN = CG->getOrInsertFunction(Start); 510 EndCGN = CG->getOrInsertFunction(End); 511 CallerNode = (*CG)[Caller]; 512 } 513 514 IRBuilder<> builder(FirstNewBlock->begin()); 515 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) { 516 AllocaInst *AI = IFI.StaticAllocas[ai]; 517 518 // If the alloca is already scoped to something smaller than the whole 519 // function then there's no need to add redundant, less accurate markers. 520 if (hasLifetimeMarkers(AI)) 521 continue; 522 523 CallInst *StartCall = builder.CreateLifetimeStart(AI); 524 if (IFI.CG) CallerNode->addCalledFunction(StartCall, StartCGN); 525 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) { 526 IRBuilder<> builder(Returns[ri]); 527 CallInst *EndCall = builder.CreateLifetimeEnd(AI); 528 if (IFI.CG) CallerNode->addCalledFunction(EndCall, EndCGN); 529 } 530 } 531 } 532 533 // If the inlined code contained dynamic alloca instructions, wrap the inlined 534 // code with llvm.stacksave/llvm.stackrestore intrinsics. 535 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 536 Module *M = Caller->getParent(); 537 // Get the two intrinsics we care about. 538 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 539 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 540 541 // If we are preserving the callgraph, add edges to the stacksave/restore 542 // functions for the calls we insert. 543 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0; 544 if (CallGraph *CG = IFI.CG) { 545 StackSaveCGN = CG->getOrInsertFunction(StackSave); 546 StackRestoreCGN = CG->getOrInsertFunction(StackRestore); 547 CallerNode = (*CG)[Caller]; 548 } 549 550 // Insert the llvm.stacksave. 551 CallInst *SavedPtr = CallInst::Create(StackSave, "savedstack", 552 FirstNewBlock->begin()); 553 if (IFI.CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN); 554 555 // Insert a call to llvm.stackrestore before any return instructions in the 556 // inlined function. 557 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 558 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", Returns[i]); 559 if (IFI.CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 560 } 561 562 // Count the number of StackRestore calls we insert. 563 unsigned NumStackRestores = Returns.size(); 564 565 // If we are inlining an invoke instruction, insert restores before each 566 // unwind. These unwinds will be rewritten into branches later. 567 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 568 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 569 BB != E; ++BB) 570 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 571 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", UI); 572 if (IFI.CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 573 ++NumStackRestores; 574 } 575 } 576 } 577 578 // If we are inlining tail call instruction through a call site that isn't 579 // marked 'tail', we must remove the tail marker for any calls in the inlined 580 // code. Also, calls inlined through a 'nounwind' call site should be marked 581 // 'nounwind'. 582 if (InlinedFunctionInfo.ContainsCalls && 583 (MustClearTailCallFlags || MarkNoUnwind)) { 584 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 585 BB != E; ++BB) 586 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 587 if (CallInst *CI = dyn_cast<CallInst>(I)) { 588 if (MustClearTailCallFlags) 589 CI->setTailCall(false); 590 if (MarkNoUnwind) 591 CI->setDoesNotThrow(); 592 } 593 } 594 595 // If we are inlining through a 'nounwind' call site then any inlined 'unwind' 596 // instructions are unreachable. 597 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind) 598 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 599 BB != E; ++BB) { 600 TerminatorInst *Term = BB->getTerminator(); 601 if (isa<UnwindInst>(Term)) { 602 new UnreachableInst(Context, Term); 603 BB->getInstList().erase(Term); 604 } 605 } 606 607 // If we are inlining for an invoke instruction, we must make sure to rewrite 608 // any inlined 'unwind' instructions into branches to the invoke exception 609 // destination, and call instructions into invoke instructions. 610 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 611 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 612 613 // If we cloned in _exactly one_ basic block, and if that block ends in a 614 // return instruction, we splice the body of the inlined callee directly into 615 // the calling basic block. 616 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 617 // Move all of the instructions right before the call. 618 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 619 FirstNewBlock->begin(), FirstNewBlock->end()); 620 // Remove the cloned basic block. 621 Caller->getBasicBlockList().pop_back(); 622 623 // If the call site was an invoke instruction, add a branch to the normal 624 // destination. 625 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 626 BranchInst::Create(II->getNormalDest(), TheCall); 627 628 // If the return instruction returned a value, replace uses of the call with 629 // uses of the returned value. 630 if (!TheCall->use_empty()) { 631 ReturnInst *R = Returns[0]; 632 if (TheCall == R->getReturnValue()) 633 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 634 else 635 TheCall->replaceAllUsesWith(R->getReturnValue()); 636 } 637 // Since we are now done with the Call/Invoke, we can delete it. 638 TheCall->eraseFromParent(); 639 640 // Since we are now done with the return instruction, delete it also. 641 Returns[0]->eraseFromParent(); 642 643 // We are now done with the inlining. 644 return true; 645 } 646 647 // Otherwise, we have the normal case, of more than one block to inline or 648 // multiple return sites. 649 650 // We want to clone the entire callee function into the hole between the 651 // "starter" and "ender" blocks. How we accomplish this depends on whether 652 // this is an invoke instruction or a call instruction. 653 BasicBlock *AfterCallBB; 654 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 655 656 // Add an unconditional branch to make this look like the CallInst case... 657 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 658 659 // Split the basic block. This guarantees that no PHI nodes will have to be 660 // updated due to new incoming edges, and make the invoke case more 661 // symmetric to the call case. 662 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 663 CalledFunc->getName()+".exit"); 664 665 } else { // It's a call 666 // If this is a call instruction, we need to split the basic block that 667 // the call lives in. 668 // 669 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 670 CalledFunc->getName()+".exit"); 671 } 672 673 // Change the branch that used to go to AfterCallBB to branch to the first 674 // basic block of the inlined function. 675 // 676 TerminatorInst *Br = OrigBB->getTerminator(); 677 assert(Br && Br->getOpcode() == Instruction::Br && 678 "splitBasicBlock broken!"); 679 Br->setOperand(0, FirstNewBlock); 680 681 682 // Now that the function is correct, make it a little bit nicer. In 683 // particular, move the basic blocks inserted from the end of the function 684 // into the space made by splitting the source basic block. 685 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 686 FirstNewBlock, Caller->end()); 687 688 // Handle all of the return instructions that we just cloned in, and eliminate 689 // any users of the original call/invoke instruction. 690 const Type *RTy = CalledFunc->getReturnType(); 691 692 PHINode *PHI = 0; 693 if (Returns.size() > 1) { 694 // The PHI node should go at the front of the new basic block to merge all 695 // possible incoming values. 696 if (!TheCall->use_empty()) { 697 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(), 698 AfterCallBB->begin()); 699 // Anything that used the result of the function call should now use the 700 // PHI node as their operand. 701 TheCall->replaceAllUsesWith(PHI); 702 } 703 704 // Loop over all of the return instructions adding entries to the PHI node 705 // as appropriate. 706 if (PHI) { 707 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 708 ReturnInst *RI = Returns[i]; 709 assert(RI->getReturnValue()->getType() == PHI->getType() && 710 "Ret value not consistent in function!"); 711 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 712 } 713 } 714 715 716 // Add a branch to the merge points and remove return instructions. 717 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 718 ReturnInst *RI = Returns[i]; 719 BranchInst::Create(AfterCallBB, RI); 720 RI->eraseFromParent(); 721 } 722 } else if (!Returns.empty()) { 723 // Otherwise, if there is exactly one return value, just replace anything 724 // using the return value of the call with the computed value. 725 if (!TheCall->use_empty()) { 726 if (TheCall == Returns[0]->getReturnValue()) 727 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 728 else 729 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 730 } 731 732 // Splice the code from the return block into the block that it will return 733 // to, which contains the code that was after the call. 734 BasicBlock *ReturnBB = Returns[0]->getParent(); 735 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 736 ReturnBB->getInstList()); 737 738 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 739 ReturnBB->replaceAllUsesWith(AfterCallBB); 740 741 // Delete the return instruction now and empty ReturnBB now. 742 Returns[0]->eraseFromParent(); 743 ReturnBB->eraseFromParent(); 744 } else if (!TheCall->use_empty()) { 745 // No returns, but something is using the return value of the call. Just 746 // nuke the result. 747 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 748 } 749 750 // Since we are now done with the Call/Invoke, we can delete it. 751 TheCall->eraseFromParent(); 752 753 // We should always be able to fold the entry block of the function into the 754 // single predecessor of the block... 755 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 756 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 757 758 // Splice the code entry block into calling block, right before the 759 // unconditional branch. 760 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 761 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 762 763 // Remove the unconditional branch. 764 OrigBB->getInstList().erase(Br); 765 766 // Now we can remove the CalleeEntry block, which is now empty. 767 Caller->getBasicBlockList().erase(CalleeEntry); 768 769 // If we inserted a phi node, check to see if it has a single value (e.g. all 770 // the entries are the same or undef). If so, remove the PHI so it doesn't 771 // block other optimizations. 772 if (PHI) 773 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) { 774 PHI->replaceAllUsesWith(V); 775 PHI->eraseFromParent(); 776 } 777 778 return true; 779} 780