InlineFunction.cpp revision 198892
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/LLVMContext.h" 19#include "llvm/Module.h" 20#include "llvm/Instructions.h" 21#include "llvm/IntrinsicInst.h" 22#include "llvm/Intrinsics.h" 23#include "llvm/Attributes.h" 24#include "llvm/Analysis/CallGraph.h" 25#include "llvm/Analysis/DebugInfo.h" 26#include "llvm/Target/TargetData.h" 27#include "llvm/ADT/SmallVector.h" 28#include "llvm/ADT/StringExtras.h" 29#include "llvm/Support/CallSite.h" 30using namespace llvm; 31 32bool llvm::InlineFunction(CallInst *CI, CallGraph *CG, const TargetData *TD, 33 SmallVectorImpl<AllocaInst*> *StaticAllocas) { 34 return InlineFunction(CallSite(CI), CG, TD, StaticAllocas); 35} 36bool llvm::InlineFunction(InvokeInst *II, CallGraph *CG, const TargetData *TD, 37 SmallVectorImpl<AllocaInst*> *StaticAllocas) { 38 return InlineFunction(CallSite(II), CG, TD, StaticAllocas); 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 SmallVector<Value*, 8> InvokeArgs(CI->op_begin()+1, CI->op_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. 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 DenseMap<const Value*, Value*> &ValueMap, 176 CallGraph &CG) { 177 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 178 const Function *Callee = CS.getCalledFunction(); 179 CallGraphNode *CalleeNode = CG[Callee]; 180 CallGraphNode *CallerNode = CG[Caller]; 181 182 // Since we inlined some uninlined call sites in the callee into the caller, 183 // add edges from the caller to all of the callees of the callee. 184 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 185 186 // Consider the case where CalleeNode == CallerNode. 187 CallGraphNode::CalledFunctionsVector CallCache; 188 if (CalleeNode == CallerNode) { 189 CallCache.assign(I, E); 190 I = CallCache.begin(); 191 E = CallCache.end(); 192 } 193 194 for (; I != E; ++I) { 195 const Value *OrigCall = I->first; 196 197 DenseMap<const Value*, Value*>::iterator VMI = ValueMap.find(OrigCall); 198 // Only copy the edge if the call was inlined! 199 if (VMI == ValueMap.end() || VMI->second == 0) 200 continue; 201 202 // If the call was inlined, but then constant folded, there is no edge to 203 // add. Check for this case. 204 if (Instruction *NewCall = dyn_cast<Instruction>(VMI->second)) 205 CallerNode->addCalledFunction(CallSite::get(NewCall), I->second); 206 } 207 208 // Update the call graph by deleting the edge from Callee to Caller. We must 209 // do this after the loop above in case Caller and Callee are the same. 210 CallerNode->removeCallEdgeFor(CS); 211} 212 213/// findFnRegionEndMarker - This is a utility routine that is used by 214/// InlineFunction. Return llvm.dbg.region.end intrinsic that corresponds 215/// to the llvm.dbg.func.start of the function F. Otherwise return NULL. 216/// 217static const DbgRegionEndInst *findFnRegionEndMarker(const Function *F) { 218 219 MDNode *FnStart = NULL; 220 const DbgRegionEndInst *FnEnd = NULL; 221 for (Function::const_iterator FI = F->begin(), FE =F->end(); FI != FE; ++FI) 222 for (BasicBlock::const_iterator BI = FI->begin(), BE = FI->end(); BI != BE; 223 ++BI) { 224 if (FnStart == NULL) { 225 if (const DbgFuncStartInst *FSI = dyn_cast<DbgFuncStartInst>(BI)) { 226 DISubprogram SP(FSI->getSubprogram()); 227 assert (SP.isNull() == false && "Invalid llvm.dbg.func.start"); 228 if (SP.describes(F)) 229 FnStart = SP.getNode(); 230 } 231 continue; 232 } 233 234 if (const DbgRegionEndInst *REI = dyn_cast<DbgRegionEndInst>(BI)) 235 if (REI->getContext() == FnStart) 236 FnEnd = REI; 237 } 238 return FnEnd; 239} 240 241// InlineFunction - This function inlines the called function into the basic 242// block of the caller. This returns false if it is not possible to inline this 243// call. The program is still in a well defined state if this occurs though. 244// 245// Note that this only does one level of inlining. For example, if the 246// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 247// exists in the instruction stream. Similiarly this will inline a recursive 248// function by one level. 249// 250bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD, 251 SmallVectorImpl<AllocaInst*> *StaticAllocas) { 252 Instruction *TheCall = CS.getInstruction(); 253 LLVMContext &Context = TheCall->getContext(); 254 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 255 "Instruction not in function!"); 256 257 const Function *CalledFunc = CS.getCalledFunction(); 258 if (CalledFunc == 0 || // Can't inline external function or indirect 259 CalledFunc->isDeclaration() || // call, or call to a vararg function! 260 CalledFunc->getFunctionType()->isVarArg()) return false; 261 262 263 // If the call to the callee is not a tail call, we must clear the 'tail' 264 // flags on any calls that we inline. 265 bool MustClearTailCallFlags = 266 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall()); 267 268 // If the call to the callee cannot throw, set the 'nounwind' flag on any 269 // calls that we inline. 270 bool MarkNoUnwind = CS.doesNotThrow(); 271 272 BasicBlock *OrigBB = TheCall->getParent(); 273 Function *Caller = OrigBB->getParent(); 274 275 // GC poses two hazards to inlining, which only occur when the callee has GC: 276 // 1. If the caller has no GC, then the callee's GC must be propagated to the 277 // caller. 278 // 2. If the caller has a differing GC, it is invalid to inline. 279 if (CalledFunc->hasGC()) { 280 if (!Caller->hasGC()) 281 Caller->setGC(CalledFunc->getGC()); 282 else if (CalledFunc->getGC() != Caller->getGC()) 283 return false; 284 } 285 286 // Get an iterator to the last basic block in the function, which will have 287 // the new function inlined after it. 288 // 289 Function::iterator LastBlock = &Caller->back(); 290 291 // Make sure to capture all of the return instructions from the cloned 292 // function. 293 SmallVector<ReturnInst*, 8> Returns; 294 ClonedCodeInfo InlinedFunctionInfo; 295 Function::iterator FirstNewBlock; 296 297 { // Scope to destroy ValueMap after cloning. 298 DenseMap<const Value*, Value*> ValueMap; 299 300 assert(CalledFunc->arg_size() == CS.arg_size() && 301 "No varargs calls can be inlined!"); 302 303 // Calculate the vector of arguments to pass into the function cloner, which 304 // matches up the formal to the actual argument values. 305 CallSite::arg_iterator AI = CS.arg_begin(); 306 unsigned ArgNo = 0; 307 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 308 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 309 Value *ActualArg = *AI; 310 311 // When byval arguments actually inlined, we need to make the copy implied 312 // by them explicit. However, we don't do this if the callee is readonly 313 // or readnone, because the copy would be unneeded: the callee doesn't 314 // modify the struct. 315 if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal) && 316 !CalledFunc->onlyReadsMemory()) { 317 const Type *AggTy = cast<PointerType>(I->getType())->getElementType(); 318 const Type *VoidPtrTy = 319 Type::getInt8PtrTy(Context); 320 321 // Create the alloca. If we have TargetData, use nice alignment. 322 unsigned Align = 1; 323 if (TD) Align = TD->getPrefTypeAlignment(AggTy); 324 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, 325 I->getName(), 326 &*Caller->begin()->begin()); 327 // Emit a memcpy. 328 const Type *Tys[] = { Type::getInt64Ty(Context) }; 329 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 330 Intrinsic::memcpy, 331 Tys, 1); 332 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 333 Value *SrcCast = new BitCastInst(*AI, VoidPtrTy, "tmp", TheCall); 334 335 Value *Size; 336 if (TD == 0) 337 Size = ConstantExpr::getSizeOf(AggTy); 338 else 339 Size = ConstantInt::get(Type::getInt64Ty(Context), 340 TD->getTypeStoreSize(AggTy)); 341 342 // Always generate a memcpy of alignment 1 here because we don't know 343 // the alignment of the src pointer. Other optimizations can infer 344 // better alignment. 345 Value *CallArgs[] = { 346 DestCast, SrcCast, Size, 347 ConstantInt::get(Type::getInt32Ty(Context), 1) 348 }; 349 CallInst *TheMemCpy = 350 CallInst::Create(MemCpyFn, CallArgs, CallArgs+4, "", TheCall); 351 352 // If we have a call graph, update it. 353 if (CG) { 354 CallGraphNode *MemCpyCGN = CG->getOrInsertFunction(MemCpyFn); 355 CallGraphNode *CallerNode = (*CG)[Caller]; 356 CallerNode->addCalledFunction(TheMemCpy, MemCpyCGN); 357 } 358 359 // Uses of the argument in the function should use our new alloca 360 // instead. 361 ActualArg = NewAlloca; 362 } 363 364 ValueMap[I] = ActualArg; 365 } 366 367 // Adjust llvm.dbg.region.end. If the CalledFunc has region end 368 // marker then clone that marker after next stop point at the 369 // call site. The function body cloner does not clone original 370 // region end marker from the CalledFunc. This will ensure that 371 // inlined function's scope ends at the right place. 372 if (const DbgRegionEndInst *DREI = findFnRegionEndMarker(CalledFunc)) { 373 for (BasicBlock::iterator BI = TheCall, BE = TheCall->getParent()->end(); 374 BI != BE; ++BI) { 375 if (DbgStopPointInst *DSPI = dyn_cast<DbgStopPointInst>(BI)) { 376 if (DbgRegionEndInst *NewDREI = 377 dyn_cast<DbgRegionEndInst>(DREI->clone())) 378 NewDREI->insertAfter(DSPI); 379 break; 380 } 381 } 382 } 383 384 // We want the inliner to prune the code as it copies. We would LOVE to 385 // have no dead or constant instructions leftover after inlining occurs 386 // (which can happen, e.g., because an argument was constant), but we'll be 387 // happy with whatever the cloner can do. 388 CloneAndPruneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i", 389 &InlinedFunctionInfo, TD); 390 391 // Remember the first block that is newly cloned over. 392 FirstNewBlock = LastBlock; ++FirstNewBlock; 393 394 // Update the callgraph if requested. 395 if (CG) 396 UpdateCallGraphAfterInlining(CS, FirstNewBlock, ValueMap, *CG); 397 } 398 399 // If there are any alloca instructions in the block that used to be the entry 400 // block for the callee, move them to the entry block of the caller. First 401 // calculate which instruction they should be inserted before. We insert the 402 // instructions at the end of the current alloca list. 403 // 404 { 405 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 406 for (BasicBlock::iterator I = FirstNewBlock->begin(), 407 E = FirstNewBlock->end(); I != E; ) { 408 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 409 if (AI == 0) continue; 410 411 // If the alloca is now dead, remove it. This often occurs due to code 412 // specialization. 413 if (AI->use_empty()) { 414 AI->eraseFromParent(); 415 continue; 416 } 417 418 if (!isa<Constant>(AI->getArraySize())) 419 continue; 420 421 // Keep track of the static allocas that we inline into the caller if the 422 // StaticAllocas pointer is non-null. 423 if (StaticAllocas) StaticAllocas->push_back(AI); 424 425 // Scan for the block of allocas that we can move over, and move them 426 // all at once. 427 while (isa<AllocaInst>(I) && 428 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 429 if (StaticAllocas) StaticAllocas->push_back(cast<AllocaInst>(I)); 430 ++I; 431 } 432 433 // Transfer all of the allocas over in a block. Using splice means 434 // that the instructions aren't removed from the symbol table, then 435 // reinserted. 436 Caller->getEntryBlock().getInstList().splice(InsertPoint, 437 FirstNewBlock->getInstList(), 438 AI, I); 439 } 440 } 441 442 // If the inlined code contained dynamic alloca instructions, wrap the inlined 443 // code with llvm.stacksave/llvm.stackrestore intrinsics. 444 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 445 Module *M = Caller->getParent(); 446 // Get the two intrinsics we care about. 447 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 448 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 449 450 // If we are preserving the callgraph, add edges to the stacksave/restore 451 // functions for the calls we insert. 452 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0; 453 if (CG) { 454 StackSaveCGN = CG->getOrInsertFunction(StackSave); 455 StackRestoreCGN = CG->getOrInsertFunction(StackRestore); 456 CallerNode = (*CG)[Caller]; 457 } 458 459 // Insert the llvm.stacksave. 460 CallInst *SavedPtr = CallInst::Create(StackSave, "savedstack", 461 FirstNewBlock->begin()); 462 if (CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN); 463 464 // Insert a call to llvm.stackrestore before any return instructions in the 465 // inlined function. 466 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 467 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", Returns[i]); 468 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 469 } 470 471 // Count the number of StackRestore calls we insert. 472 unsigned NumStackRestores = Returns.size(); 473 474 // If we are inlining an invoke instruction, insert restores before each 475 // unwind. These unwinds will be rewritten into branches later. 476 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 477 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 478 BB != E; ++BB) 479 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 480 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", UI); 481 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 482 ++NumStackRestores; 483 } 484 } 485 } 486 487 // If we are inlining tail call instruction through a call site that isn't 488 // marked 'tail', we must remove the tail marker for any calls in the inlined 489 // code. Also, calls inlined through a 'nounwind' call site should be marked 490 // 'nounwind'. 491 if (InlinedFunctionInfo.ContainsCalls && 492 (MustClearTailCallFlags || MarkNoUnwind)) { 493 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 494 BB != E; ++BB) 495 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 496 if (CallInst *CI = dyn_cast<CallInst>(I)) { 497 if (MustClearTailCallFlags) 498 CI->setTailCall(false); 499 if (MarkNoUnwind) 500 CI->setDoesNotThrow(); 501 } 502 } 503 504 // If we are inlining through a 'nounwind' call site then any inlined 'unwind' 505 // instructions are unreachable. 506 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind) 507 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 508 BB != E; ++BB) { 509 TerminatorInst *Term = BB->getTerminator(); 510 if (isa<UnwindInst>(Term)) { 511 new UnreachableInst(Context, Term); 512 BB->getInstList().erase(Term); 513 } 514 } 515 516 // If we are inlining for an invoke instruction, we must make sure to rewrite 517 // any inlined 'unwind' instructions into branches to the invoke exception 518 // destination, and call instructions into invoke instructions. 519 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 520 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 521 522 // If we cloned in _exactly one_ basic block, and if that block ends in a 523 // return instruction, we splice the body of the inlined callee directly into 524 // the calling basic block. 525 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 526 // Move all of the instructions right before the call. 527 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 528 FirstNewBlock->begin(), FirstNewBlock->end()); 529 // Remove the cloned basic block. 530 Caller->getBasicBlockList().pop_back(); 531 532 // If the call site was an invoke instruction, add a branch to the normal 533 // destination. 534 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 535 BranchInst::Create(II->getNormalDest(), TheCall); 536 537 // If the return instruction returned a value, replace uses of the call with 538 // uses of the returned value. 539 if (!TheCall->use_empty()) { 540 ReturnInst *R = Returns[0]; 541 if (TheCall == R->getReturnValue()) 542 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 543 else 544 TheCall->replaceAllUsesWith(R->getReturnValue()); 545 } 546 // Since we are now done with the Call/Invoke, we can delete it. 547 TheCall->eraseFromParent(); 548 549 // Since we are now done with the return instruction, delete it also. 550 Returns[0]->eraseFromParent(); 551 552 // We are now done with the inlining. 553 return true; 554 } 555 556 // Otherwise, we have the normal case, of more than one block to inline or 557 // multiple return sites. 558 559 // We want to clone the entire callee function into the hole between the 560 // "starter" and "ender" blocks. How we accomplish this depends on whether 561 // this is an invoke instruction or a call instruction. 562 BasicBlock *AfterCallBB; 563 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 564 565 // Add an unconditional branch to make this look like the CallInst case... 566 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 567 568 // Split the basic block. This guarantees that no PHI nodes will have to be 569 // updated due to new incoming edges, and make the invoke case more 570 // symmetric to the call case. 571 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 572 CalledFunc->getName()+".exit"); 573 574 } else { // It's a call 575 // If this is a call instruction, we need to split the basic block that 576 // the call lives in. 577 // 578 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 579 CalledFunc->getName()+".exit"); 580 } 581 582 // Change the branch that used to go to AfterCallBB to branch to the first 583 // basic block of the inlined function. 584 // 585 TerminatorInst *Br = OrigBB->getTerminator(); 586 assert(Br && Br->getOpcode() == Instruction::Br && 587 "splitBasicBlock broken!"); 588 Br->setOperand(0, FirstNewBlock); 589 590 591 // Now that the function is correct, make it a little bit nicer. In 592 // particular, move the basic blocks inserted from the end of the function 593 // into the space made by splitting the source basic block. 594 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 595 FirstNewBlock, Caller->end()); 596 597 // Handle all of the return instructions that we just cloned in, and eliminate 598 // any users of the original call/invoke instruction. 599 const Type *RTy = CalledFunc->getReturnType(); 600 601 if (Returns.size() > 1) { 602 // The PHI node should go at the front of the new basic block to merge all 603 // possible incoming values. 604 PHINode *PHI = 0; 605 if (!TheCall->use_empty()) { 606 PHI = PHINode::Create(RTy, TheCall->getName(), 607 AfterCallBB->begin()); 608 // Anything that used the result of the function call should now use the 609 // PHI node as their operand. 610 TheCall->replaceAllUsesWith(PHI); 611 } 612 613 // Loop over all of the return instructions adding entries to the PHI node 614 // as appropriate. 615 if (PHI) { 616 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 617 ReturnInst *RI = Returns[i]; 618 assert(RI->getReturnValue()->getType() == PHI->getType() && 619 "Ret value not consistent in function!"); 620 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 621 } 622 623 // Now that we inserted the PHI, check to see if it has a single value 624 // (e.g. all the entries are the same or undef). If so, remove the PHI so 625 // it doesn't block other optimizations. 626 if (Value *V = PHI->hasConstantValue()) { 627 PHI->replaceAllUsesWith(V); 628 PHI->eraseFromParent(); 629 } 630 } 631 632 633 // Add a branch to the merge points and remove return instructions. 634 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 635 ReturnInst *RI = Returns[i]; 636 BranchInst::Create(AfterCallBB, RI); 637 RI->eraseFromParent(); 638 } 639 } else if (!Returns.empty()) { 640 // Otherwise, if there is exactly one return value, just replace anything 641 // using the return value of the call with the computed value. 642 if (!TheCall->use_empty()) { 643 if (TheCall == Returns[0]->getReturnValue()) 644 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 645 else 646 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 647 } 648 649 // Splice the code from the return block into the block that it will return 650 // to, which contains the code that was after the call. 651 BasicBlock *ReturnBB = Returns[0]->getParent(); 652 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 653 ReturnBB->getInstList()); 654 655 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 656 ReturnBB->replaceAllUsesWith(AfterCallBB); 657 658 // Delete the return instruction now and empty ReturnBB now. 659 Returns[0]->eraseFromParent(); 660 ReturnBB->eraseFromParent(); 661 } else if (!TheCall->use_empty()) { 662 // No returns, but something is using the return value of the call. Just 663 // nuke the result. 664 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 665 } 666 667 // Since we are now done with the Call/Invoke, we can delete it. 668 TheCall->eraseFromParent(); 669 670 // We should always be able to fold the entry block of the function into the 671 // single predecessor of the block... 672 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 673 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 674 675 // Splice the code entry block into calling block, right before the 676 // unconditional branch. 677 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 678 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 679 680 // Remove the unconditional branch. 681 OrigBB->getInstList().erase(Br); 682 683 // Now we can remove the CalleeEntry block, which is now empty. 684 Caller->getBasicBlockList().erase(CalleeEntry); 685 686 return true; 687} 688