InlineFunction.cpp revision 218893
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" 31using namespace llvm; 32 33bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI) { 34 return InlineFunction(CallSite(CI), IFI); 35} 36bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI) { 37 return InlineFunction(CallSite(II), IFI); 38} 39 40 41/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into 42/// an invoke, we have to turn all of the calls that can throw into 43/// invokes. This function analyze BB to see if there are any calls, and if so, 44/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI 45/// nodes in that block with the values specified in InvokeDestPHIValues. 46/// 47static void HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, 48 BasicBlock *InvokeDest, 49 const SmallVectorImpl<Value*> &InvokeDestPHIValues) { 50 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 51 Instruction *I = BBI++; 52 53 // We only need to check for function calls: inlined invoke 54 // instructions require no special handling. 55 CallInst *CI = dyn_cast<CallInst>(I); 56 if (CI == 0) continue; 57 58 // If this call cannot unwind, don't convert it to an invoke. 59 if (CI->doesNotThrow()) 60 continue; 61 62 // Convert this function call into an invoke instruction. 63 // First, split the basic block. 64 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 65 66 // Next, create the new invoke instruction, inserting it at the end 67 // of the old basic block. 68 ImmutableCallSite CS(CI); 69 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end()); 70 InvokeInst *II = 71 InvokeInst::Create(CI->getCalledValue(), Split, InvokeDest, 72 InvokeArgs.begin(), InvokeArgs.end(), 73 CI->getName(), BB->getTerminator()); 74 II->setCallingConv(CI->getCallingConv()); 75 II->setAttributes(CI->getAttributes()); 76 77 // Make sure that anything using the call now uses the invoke! This also 78 // updates the CallGraph if present, because it uses a WeakVH. 79 CI->replaceAllUsesWith(II); 80 81 // Delete the unconditional branch inserted by splitBasicBlock 82 BB->getInstList().pop_back(); 83 Split->getInstList().pop_front(); // Delete the original call 84 85 // Update any PHI nodes in the exceptional block to indicate that 86 // there is now a new entry in them. 87 unsigned i = 0; 88 for (BasicBlock::iterator I = InvokeDest->begin(); 89 isa<PHINode>(I); ++I, ++i) 90 cast<PHINode>(I)->addIncoming(InvokeDestPHIValues[i], BB); 91 92 // This basic block is now complete, the caller will continue scanning the 93 // next one. 94 return; 95 } 96} 97 98 99/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 100/// in the body of the inlined function into invokes and turn unwind 101/// instructions into branches to the invoke unwind dest. 102/// 103/// II is the invoke instruction being inlined. FirstNewBlock is the first 104/// block of the inlined code (the last block is the end of the function), 105/// and InlineCodeInfo is information about the code that got inlined. 106static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 107 ClonedCodeInfo &InlinedCodeInfo) { 108 BasicBlock *InvokeDest = II->getUnwindDest(); 109 SmallVector<Value*, 8> InvokeDestPHIValues; 110 111 // If there are PHI nodes in the unwind destination block, we need to 112 // keep track of which values came into them from this invoke, then remove 113 // the entry for this block. 114 BasicBlock *InvokeBlock = II->getParent(); 115 for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) { 116 PHINode *PN = cast<PHINode>(I); 117 // Save the value to use for this edge. 118 InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock)); 119 } 120 121 Function *Caller = FirstNewBlock->getParent(); 122 123 // The inlined code is currently at the end of the function, scan from the 124 // start of the inlined code to its end, checking for stuff we need to 125 // rewrite. If the code doesn't have calls or unwinds, we know there is 126 // nothing to rewrite. 127 if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) { 128 // Now that everything is happy, we have one final detail. The PHI nodes in 129 // the exception destination block still have entries due to the original 130 // invoke instruction. Eliminate these entries (which might even delete the 131 // PHI node) now. 132 InvokeDest->removePredecessor(II->getParent()); 133 return; 134 } 135 136 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){ 137 if (InlinedCodeInfo.ContainsCalls) 138 HandleCallsInBlockInlinedThroughInvoke(BB, InvokeDest, 139 InvokeDestPHIValues); 140 141 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 142 // An UnwindInst requires special handling when it gets inlined into an 143 // invoke site. Once this happens, we know that the unwind would cause 144 // a control transfer to the invoke exception destination, so we can 145 // transform it into a direct branch to the exception destination. 146 BranchInst::Create(InvokeDest, UI); 147 148 // Delete the unwind instruction! 149 UI->eraseFromParent(); 150 151 // Update any PHI nodes in the exceptional block to indicate that 152 // there is now a new entry in them. 153 unsigned i = 0; 154 for (BasicBlock::iterator I = InvokeDest->begin(); 155 isa<PHINode>(I); ++I, ++i) { 156 PHINode *PN = cast<PHINode>(I); 157 PN->addIncoming(InvokeDestPHIValues[i], BB); 158 } 159 } 160 } 161 162 // Now that everything is happy, we have one final detail. The PHI nodes in 163 // the exception destination block still have entries due to the original 164 // invoke instruction. Eliminate these entries (which might even delete the 165 // PHI node) now. 166 InvokeDest->removePredecessor(II->getParent()); 167} 168 169/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 170/// into the caller, update the specified callgraph to reflect the changes we 171/// made. Note that it's possible that not all code was copied over, so only 172/// some edges of the callgraph may remain. 173static void UpdateCallGraphAfterInlining(CallSite CS, 174 Function::iterator FirstNewBlock, 175 ValueToValueMapTy &VMap, 176 InlineFunctionInfo &IFI) { 177 CallGraph &CG = *IFI.CG; 178 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 179 const Function *Callee = CS.getCalledFunction(); 180 CallGraphNode *CalleeNode = CG[Callee]; 181 CallGraphNode *CallerNode = CG[Caller]; 182 183 // Since we inlined some uninlined call sites in the callee into the caller, 184 // add edges from the caller to all of the callees of the callee. 185 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 186 187 // Consider the case where CalleeNode == CallerNode. 188 CallGraphNode::CalledFunctionsVector CallCache; 189 if (CalleeNode == CallerNode) { 190 CallCache.assign(I, E); 191 I = CallCache.begin(); 192 E = CallCache.end(); 193 } 194 195 for (; I != E; ++I) { 196 const Value *OrigCall = I->first; 197 198 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); 199 // Only copy the edge if the call was inlined! 200 if (VMI == VMap.end() || VMI->second == 0) 201 continue; 202 203 // If the call was inlined, but then constant folded, there is no edge to 204 // add. Check for this case. 205 Instruction *NewCall = dyn_cast<Instruction>(VMI->second); 206 if (NewCall == 0) continue; 207 208 // Remember that this call site got inlined for the client of 209 // InlineFunction. 210 IFI.InlinedCalls.push_back(NewCall); 211 212 // It's possible that inlining the callsite will cause it to go from an 213 // indirect to a direct call by resolving a function pointer. If this 214 // happens, set the callee of the new call site to a more precise 215 // destination. This can also happen if the call graph node of the caller 216 // was just unnecessarily imprecise. 217 if (I->second->getFunction() == 0) 218 if (Function *F = CallSite(NewCall).getCalledFunction()) { 219 // Indirect call site resolved to direct call. 220 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); 221 222 continue; 223 } 224 225 CallerNode->addCalledFunction(CallSite(NewCall), I->second); 226 } 227 228 // Update the call graph by deleting the edge from Callee to Caller. We must 229 // do this after the loop above in case Caller and Callee are the same. 230 CallerNode->removeCallEdgeFor(CS); 231} 232 233/// HandleByValArgument - When inlining a call site that has a byval argument, 234/// we have to make the implicit memcpy explicit by adding it. 235static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, 236 const Function *CalledFunc, 237 InlineFunctionInfo &IFI, 238 unsigned ByValAlignment) { 239 const Type *AggTy = cast<PointerType>(Arg->getType())->getElementType(); 240 241 // If the called function is readonly, then it could not mutate the caller's 242 // copy of the byval'd memory. In this case, it is safe to elide the copy and 243 // temporary. 244 if (CalledFunc->onlyReadsMemory()) { 245 // If the byval argument has a specified alignment that is greater than the 246 // passed in pointer, then we either have to round up the input pointer or 247 // give up on this transformation. 248 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. 249 return Arg; 250 251 // If the pointer is already known to be sufficiently aligned, or if we can 252 // round it up to a larger alignment, then we don't need a temporary. 253 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, 254 IFI.TD) >= ByValAlignment) 255 return Arg; 256 257 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad 258 // for code quality, but rarely happens and is required for correctness. 259 } 260 261 LLVMContext &Context = Arg->getContext(); 262 263 const Type *VoidPtrTy = Type::getInt8PtrTy(Context); 264 265 // Create the alloca. If we have TargetData, use nice alignment. 266 unsigned Align = 1; 267 if (IFI.TD) 268 Align = IFI.TD->getPrefTypeAlignment(AggTy); 269 270 // If the byval had an alignment specified, we *must* use at least that 271 // alignment, as it is required by the byval argument (and uses of the 272 // pointer inside the callee). 273 Align = std::max(Align, ByValAlignment); 274 275 Function *Caller = TheCall->getParent()->getParent(); 276 277 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(), 278 &*Caller->begin()->begin()); 279 // Emit a memcpy. 280 const Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)}; 281 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 282 Intrinsic::memcpy, 283 Tys, 3); 284 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 285 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall); 286 287 Value *Size; 288 if (IFI.TD == 0) 289 Size = ConstantExpr::getSizeOf(AggTy); 290 else 291 Size = ConstantInt::get(Type::getInt64Ty(Context), 292 IFI.TD->getTypeStoreSize(AggTy)); 293 294 // Always generate a memcpy of alignment 1 here because we don't know 295 // the alignment of the src pointer. Other optimizations can infer 296 // better alignment. 297 Value *CallArgs[] = { 298 DestCast, SrcCast, Size, 299 ConstantInt::get(Type::getInt32Ty(Context), 1), 300 ConstantInt::getFalse(Context) // isVolatile 301 }; 302 CallInst *TheMemCpy = 303 CallInst::Create(MemCpyFn, CallArgs, CallArgs+5, "", TheCall); 304 305 // If we have a call graph, update it. 306 if (CallGraph *CG = IFI.CG) { 307 CallGraphNode *MemCpyCGN = CG->getOrInsertFunction(MemCpyFn); 308 CallGraphNode *CallerNode = (*CG)[Caller]; 309 CallerNode->addCalledFunction(TheMemCpy, MemCpyCGN); 310 } 311 312 // Uses of the argument in the function should use our new alloca 313 // instead. 314 return NewAlloca; 315} 316 317// InlineFunction - This function inlines the called function into the basic 318// block of the caller. This returns false if it is not possible to inline this 319// call. The program is still in a well defined state if this occurs though. 320// 321// Note that this only does one level of inlining. For example, if the 322// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 323// exists in the instruction stream. Similiarly this will inline a recursive 324// function by one level. 325// 326bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) { 327 Instruction *TheCall = CS.getInstruction(); 328 LLVMContext &Context = TheCall->getContext(); 329 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 330 "Instruction not in function!"); 331 332 // If IFI has any state in it, zap it before we fill it in. 333 IFI.reset(); 334 335 const Function *CalledFunc = CS.getCalledFunction(); 336 if (CalledFunc == 0 || // Can't inline external function or indirect 337 CalledFunc->isDeclaration() || // call, or call to a vararg function! 338 CalledFunc->getFunctionType()->isVarArg()) return false; 339 340 // If the call to the callee is not a tail call, we must clear the 'tail' 341 // flags on any calls that we inline. 342 bool MustClearTailCallFlags = 343 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall()); 344 345 // If the call to the callee cannot throw, set the 'nounwind' flag on any 346 // calls that we inline. 347 bool MarkNoUnwind = CS.doesNotThrow(); 348 349 BasicBlock *OrigBB = TheCall->getParent(); 350 Function *Caller = OrigBB->getParent(); 351 352 // GC poses two hazards to inlining, which only occur when the callee has GC: 353 // 1. If the caller has no GC, then the callee's GC must be propagated to the 354 // caller. 355 // 2. If the caller has a differing GC, it is invalid to inline. 356 if (CalledFunc->hasGC()) { 357 if (!Caller->hasGC()) 358 Caller->setGC(CalledFunc->getGC()); 359 else if (CalledFunc->getGC() != Caller->getGC()) 360 return false; 361 } 362 363 // Get an iterator to the last basic block in the function, which will have 364 // the new function inlined after it. 365 // 366 Function::iterator LastBlock = &Caller->back(); 367 368 // Make sure to capture all of the return instructions from the cloned 369 // function. 370 SmallVector<ReturnInst*, 8> Returns; 371 ClonedCodeInfo InlinedFunctionInfo; 372 Function::iterator FirstNewBlock; 373 374 { // Scope to destroy VMap after cloning. 375 ValueToValueMapTy VMap; 376 377 assert(CalledFunc->arg_size() == CS.arg_size() && 378 "No varargs calls can be inlined!"); 379 380 // Calculate the vector of arguments to pass into the function cloner, which 381 // matches up the formal to the actual argument values. 382 CallSite::arg_iterator AI = CS.arg_begin(); 383 unsigned ArgNo = 0; 384 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 385 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 386 Value *ActualArg = *AI; 387 388 // When byval arguments actually inlined, we need to make the copy implied 389 // by them explicit. However, we don't do this if the callee is readonly 390 // or readnone, because the copy would be unneeded: the callee doesn't 391 // modify the struct. 392 if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal)) { 393 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, 394 CalledFunc->getParamAlignment(ArgNo+1)); 395 396 // Calls that we inline may use the new alloca, so we need to clear 397 // their 'tail' flags if HandleByValArgument introduced a new alloca and 398 // the callee has calls. 399 MustClearTailCallFlags |= ActualArg != *AI; 400 } 401 402 VMap[I] = ActualArg; 403 } 404 405 // We want the inliner to prune the code as it copies. We would LOVE to 406 // have no dead or constant instructions leftover after inlining occurs 407 // (which can happen, e.g., because an argument was constant), but we'll be 408 // happy with whatever the cloner can do. 409 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 410 /*ModuleLevelChanges=*/false, Returns, ".i", 411 &InlinedFunctionInfo, IFI.TD, TheCall); 412 413 // Remember the first block that is newly cloned over. 414 FirstNewBlock = LastBlock; ++FirstNewBlock; 415 416 // Update the callgraph if requested. 417 if (IFI.CG) 418 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); 419 } 420 421 // If there are any alloca instructions in the block that used to be the entry 422 // block for the callee, move them to the entry block of the caller. First 423 // calculate which instruction they should be inserted before. We insert the 424 // instructions at the end of the current alloca list. 425 // 426 { 427 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 428 for (BasicBlock::iterator I = FirstNewBlock->begin(), 429 E = FirstNewBlock->end(); I != E; ) { 430 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 431 if (AI == 0) continue; 432 433 // If the alloca is now dead, remove it. This often occurs due to code 434 // specialization. 435 if (AI->use_empty()) { 436 AI->eraseFromParent(); 437 continue; 438 } 439 440 if (!isa<Constant>(AI->getArraySize())) 441 continue; 442 443 // Keep track of the static allocas that we inline into the caller. 444 IFI.StaticAllocas.push_back(AI); 445 446 // Scan for the block of allocas that we can move over, and move them 447 // all at once. 448 while (isa<AllocaInst>(I) && 449 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 450 IFI.StaticAllocas.push_back(cast<AllocaInst>(I)); 451 ++I; 452 } 453 454 // Transfer all of the allocas over in a block. Using splice means 455 // that the instructions aren't removed from the symbol table, then 456 // reinserted. 457 Caller->getEntryBlock().getInstList().splice(InsertPoint, 458 FirstNewBlock->getInstList(), 459 AI, I); 460 } 461 } 462 463 // If the inlined code contained dynamic alloca instructions, wrap the inlined 464 // code with llvm.stacksave/llvm.stackrestore intrinsics. 465 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 466 Module *M = Caller->getParent(); 467 // Get the two intrinsics we care about. 468 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 469 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 470 471 // If we are preserving the callgraph, add edges to the stacksave/restore 472 // functions for the calls we insert. 473 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0; 474 if (CallGraph *CG = IFI.CG) { 475 StackSaveCGN = CG->getOrInsertFunction(StackSave); 476 StackRestoreCGN = CG->getOrInsertFunction(StackRestore); 477 CallerNode = (*CG)[Caller]; 478 } 479 480 // Insert the llvm.stacksave. 481 CallInst *SavedPtr = CallInst::Create(StackSave, "savedstack", 482 FirstNewBlock->begin()); 483 if (IFI.CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN); 484 485 // Insert a call to llvm.stackrestore before any return instructions in the 486 // inlined function. 487 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 488 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", Returns[i]); 489 if (IFI.CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 490 } 491 492 // Count the number of StackRestore calls we insert. 493 unsigned NumStackRestores = Returns.size(); 494 495 // If we are inlining an invoke instruction, insert restores before each 496 // unwind. These unwinds will be rewritten into branches later. 497 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 498 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 499 BB != E; ++BB) 500 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 501 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", UI); 502 if (IFI.CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 503 ++NumStackRestores; 504 } 505 } 506 } 507 508 // If we are inlining tail call instruction through a call site that isn't 509 // marked 'tail', we must remove the tail marker for any calls in the inlined 510 // code. Also, calls inlined through a 'nounwind' call site should be marked 511 // 'nounwind'. 512 if (InlinedFunctionInfo.ContainsCalls && 513 (MustClearTailCallFlags || MarkNoUnwind)) { 514 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 515 BB != E; ++BB) 516 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 517 if (CallInst *CI = dyn_cast<CallInst>(I)) { 518 if (MustClearTailCallFlags) 519 CI->setTailCall(false); 520 if (MarkNoUnwind) 521 CI->setDoesNotThrow(); 522 } 523 } 524 525 // If we are inlining through a 'nounwind' call site then any inlined 'unwind' 526 // instructions are unreachable. 527 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind) 528 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 529 BB != E; ++BB) { 530 TerminatorInst *Term = BB->getTerminator(); 531 if (isa<UnwindInst>(Term)) { 532 new UnreachableInst(Context, Term); 533 BB->getInstList().erase(Term); 534 } 535 } 536 537 // If we are inlining for an invoke instruction, we must make sure to rewrite 538 // any inlined 'unwind' instructions into branches to the invoke exception 539 // destination, and call instructions into invoke instructions. 540 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 541 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 542 543 // If we cloned in _exactly one_ basic block, and if that block ends in a 544 // return instruction, we splice the body of the inlined callee directly into 545 // the calling basic block. 546 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 547 // Move all of the instructions right before the call. 548 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 549 FirstNewBlock->begin(), FirstNewBlock->end()); 550 // Remove the cloned basic block. 551 Caller->getBasicBlockList().pop_back(); 552 553 // If the call site was an invoke instruction, add a branch to the normal 554 // destination. 555 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 556 BranchInst::Create(II->getNormalDest(), TheCall); 557 558 // If the return instruction returned a value, replace uses of the call with 559 // uses of the returned value. 560 if (!TheCall->use_empty()) { 561 ReturnInst *R = Returns[0]; 562 if (TheCall == R->getReturnValue()) 563 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 564 else 565 TheCall->replaceAllUsesWith(R->getReturnValue()); 566 } 567 // Since we are now done with the Call/Invoke, we can delete it. 568 TheCall->eraseFromParent(); 569 570 // Since we are now done with the return instruction, delete it also. 571 Returns[0]->eraseFromParent(); 572 573 // We are now done with the inlining. 574 return true; 575 } 576 577 // Otherwise, we have the normal case, of more than one block to inline or 578 // multiple return sites. 579 580 // We want to clone the entire callee function into the hole between the 581 // "starter" and "ender" blocks. How we accomplish this depends on whether 582 // this is an invoke instruction or a call instruction. 583 BasicBlock *AfterCallBB; 584 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 585 586 // Add an unconditional branch to make this look like the CallInst case... 587 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 588 589 // Split the basic block. This guarantees that no PHI nodes will have to be 590 // updated due to new incoming edges, and make the invoke case more 591 // symmetric to the call case. 592 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 593 CalledFunc->getName()+".exit"); 594 595 } else { // It's a call 596 // If this is a call instruction, we need to split the basic block that 597 // the call lives in. 598 // 599 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 600 CalledFunc->getName()+".exit"); 601 } 602 603 // Change the branch that used to go to AfterCallBB to branch to the first 604 // basic block of the inlined function. 605 // 606 TerminatorInst *Br = OrigBB->getTerminator(); 607 assert(Br && Br->getOpcode() == Instruction::Br && 608 "splitBasicBlock broken!"); 609 Br->setOperand(0, FirstNewBlock); 610 611 612 // Now that the function is correct, make it a little bit nicer. In 613 // particular, move the basic blocks inserted from the end of the function 614 // into the space made by splitting the source basic block. 615 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 616 FirstNewBlock, Caller->end()); 617 618 // Handle all of the return instructions that we just cloned in, and eliminate 619 // any users of the original call/invoke instruction. 620 const Type *RTy = CalledFunc->getReturnType(); 621 622 PHINode *PHI = 0; 623 if (Returns.size() > 1) { 624 // The PHI node should go at the front of the new basic block to merge all 625 // possible incoming values. 626 if (!TheCall->use_empty()) { 627 PHI = PHINode::Create(RTy, TheCall->getName(), 628 AfterCallBB->begin()); 629 // Anything that used the result of the function call should now use the 630 // PHI node as their operand. 631 TheCall->replaceAllUsesWith(PHI); 632 } 633 634 // Loop over all of the return instructions adding entries to the PHI node 635 // as appropriate. 636 if (PHI) { 637 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 638 ReturnInst *RI = Returns[i]; 639 assert(RI->getReturnValue()->getType() == PHI->getType() && 640 "Ret value not consistent in function!"); 641 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 642 } 643 } 644 645 646 // Add a branch to the merge points and remove return instructions. 647 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 648 ReturnInst *RI = Returns[i]; 649 BranchInst::Create(AfterCallBB, RI); 650 RI->eraseFromParent(); 651 } 652 } else if (!Returns.empty()) { 653 // Otherwise, if there is exactly one return value, just replace anything 654 // using the return value of the call with the computed value. 655 if (!TheCall->use_empty()) { 656 if (TheCall == Returns[0]->getReturnValue()) 657 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 658 else 659 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 660 } 661 662 // Splice the code from the return block into the block that it will return 663 // to, which contains the code that was after the call. 664 BasicBlock *ReturnBB = Returns[0]->getParent(); 665 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 666 ReturnBB->getInstList()); 667 668 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 669 ReturnBB->replaceAllUsesWith(AfterCallBB); 670 671 // Delete the return instruction now and empty ReturnBB now. 672 Returns[0]->eraseFromParent(); 673 ReturnBB->eraseFromParent(); 674 } else if (!TheCall->use_empty()) { 675 // No returns, but something is using the return value of the call. Just 676 // nuke the result. 677 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 678 } 679 680 // Since we are now done with the Call/Invoke, we can delete it. 681 TheCall->eraseFromParent(); 682 683 // We should always be able to fold the entry block of the function into the 684 // single predecessor of the block... 685 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 686 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 687 688 // Splice the code entry block into calling block, right before the 689 // unconditional branch. 690 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 691 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 692 693 // Remove the unconditional branch. 694 OrigBB->getInstList().erase(Br); 695 696 // Now we can remove the CalleeEntry block, which is now empty. 697 Caller->getBasicBlockList().erase(CalleeEntry); 698 699 // If we inserted a phi node, check to see if it has a single value (e.g. all 700 // the entries are the same or undef). If so, remove the PHI so it doesn't 701 // block other optimizations. 702 if (PHI) 703 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) { 704 PHI->replaceAllUsesWith(V); 705 PHI->eraseFromParent(); 706 } 707 708 return true; 709} 710