InlineFunction.cpp revision 6af31aab63583e61b7c7b51bc285541750bd834f
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/Intrinsics.h" 21#include "llvm/ParameterAttributes.h" 22#include "llvm/Analysis/CallGraph.h" 23#include "llvm/Target/TargetData.h" 24#include "llvm/ADT/SmallVector.h" 25#include "llvm/Support/CallSite.h" 26using namespace llvm; 27 28bool llvm::InlineFunction(CallInst *CI, CallGraph *CG, const TargetData *TD) { 29 return InlineFunction(CallSite(CI), CG, TD); 30} 31bool llvm::InlineFunction(InvokeInst *II, CallGraph *CG, const TargetData *TD) { 32 return InlineFunction(CallSite(II), CG, TD); 33} 34 35/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 36/// in the body of the inlined function into invokes and turn unwind 37/// instructions into branches to the invoke unwind dest. 38/// 39/// II is the invoke instruction begin inlined. FirstNewBlock is the first 40/// block of the inlined code (the last block is the end of the function), 41/// and InlineCodeInfo is information about the code that got inlined. 42static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 43 ClonedCodeInfo &InlinedCodeInfo) { 44 BasicBlock *InvokeDest = II->getUnwindDest(); 45 std::vector<Value*> InvokeDestPHIValues; 46 47 // If there are PHI nodes in the unwind destination block, we need to 48 // keep track of which values came into them from this invoke, then remove 49 // the entry for this block. 50 BasicBlock *InvokeBlock = II->getParent(); 51 for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) { 52 PHINode *PN = cast<PHINode>(I); 53 // Save the value to use for this edge. 54 InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock)); 55 } 56 57 Function *Caller = FirstNewBlock->getParent(); 58 59 // The inlined code is currently at the end of the function, scan from the 60 // start of the inlined code to its end, checking for stuff we need to 61 // rewrite. 62 if (InlinedCodeInfo.ContainsCalls || InlinedCodeInfo.ContainsUnwinds) { 63 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 64 BB != E; ++BB) { 65 if (InlinedCodeInfo.ContainsCalls) { 66 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ){ 67 Instruction *I = BBI++; 68 69 // We only need to check for function calls: inlined invoke 70 // instructions require no special handling. 71 if (!isa<CallInst>(I)) continue; 72 CallInst *CI = cast<CallInst>(I); 73 74 // If this call cannot unwind, don't convert it to an invoke. 75 if (CI->doesNotThrow()) 76 continue; 77 78 // Convert this function call into an invoke instruction. 79 // First, split the basic block. 80 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 81 82 // Next, create the new invoke instruction, inserting it at the end 83 // of the old basic block. 84 SmallVector<Value*, 8> InvokeArgs(CI->op_begin()+1, CI->op_end()); 85 InvokeInst *II = 86 new InvokeInst(CI->getCalledValue(), Split, InvokeDest, 87 InvokeArgs.begin(), InvokeArgs.end(), 88 CI->getName(), BB->getTerminator()); 89 II->setCallingConv(CI->getCallingConv()); 90 II->setParamAttrs(CI->getParamAttrs()); 91 92 // Make sure that anything using the call now uses the invoke! 93 CI->replaceAllUsesWith(II); 94 95 // Delete the unconditional branch inserted by splitBasicBlock 96 BB->getInstList().pop_back(); 97 Split->getInstList().pop_front(); // Delete the original call 98 99 // Update any PHI nodes in the exceptional block to indicate that 100 // there is now a new entry in them. 101 unsigned i = 0; 102 for (BasicBlock::iterator I = InvokeDest->begin(); 103 isa<PHINode>(I); ++I, ++i) { 104 PHINode *PN = cast<PHINode>(I); 105 PN->addIncoming(InvokeDestPHIValues[i], BB); 106 } 107 108 // This basic block is now complete, start scanning the next one. 109 break; 110 } 111 } 112 113 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 114 // An UnwindInst requires special handling when it gets inlined into an 115 // invoke site. Once this happens, we know that the unwind would cause 116 // a control transfer to the invoke exception destination, so we can 117 // transform it into a direct branch to the exception destination. 118 new BranchInst(InvokeDest, UI); 119 120 // Delete the unwind instruction! 121 UI->getParent()->getInstList().pop_back(); 122 123 // Update any PHI nodes in the exceptional block to indicate that 124 // there is now a new entry in them. 125 unsigned i = 0; 126 for (BasicBlock::iterator I = InvokeDest->begin(); 127 isa<PHINode>(I); ++I, ++i) { 128 PHINode *PN = cast<PHINode>(I); 129 PN->addIncoming(InvokeDestPHIValues[i], BB); 130 } 131 } 132 } 133 } 134 135 // Now that everything is happy, we have one final detail. The PHI nodes in 136 // the exception destination block still have entries due to the original 137 // invoke instruction. Eliminate these entries (which might even delete the 138 // PHI node) now. 139 InvokeDest->removePredecessor(II->getParent()); 140} 141 142/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 143/// into the caller, update the specified callgraph to reflect the changes we 144/// made. Note that it's possible that not all code was copied over, so only 145/// some edges of the callgraph will be remain. 146static void UpdateCallGraphAfterInlining(const Function *Caller, 147 const Function *Callee, 148 Function::iterator FirstNewBlock, 149 DenseMap<const Value*, Value*> &ValueMap, 150 CallGraph &CG) { 151 // Update the call graph by deleting the edge from Callee to Caller 152 CallGraphNode *CalleeNode = CG[Callee]; 153 CallGraphNode *CallerNode = CG[Caller]; 154 CallerNode->removeCallEdgeTo(CalleeNode); 155 156 // Since we inlined some uninlined call sites in the callee into the caller, 157 // add edges from the caller to all of the callees of the callee. 158 for (CallGraphNode::iterator I = CalleeNode->begin(), 159 E = CalleeNode->end(); I != E; ++I) { 160 const Instruction *OrigCall = I->first.getInstruction(); 161 162 DenseMap<const Value*, Value*>::iterator VMI = ValueMap.find(OrigCall); 163 // Only copy the edge if the call was inlined! 164 if (VMI != ValueMap.end() && VMI->second) { 165 // If the call was inlined, but then constant folded, there is no edge to 166 // add. Check for this case. 167 if (Instruction *NewCall = dyn_cast<Instruction>(VMI->second)) 168 CallerNode->addCalledFunction(CallSite::get(NewCall), I->second); 169 } 170 } 171} 172 173 174// InlineFunction - This function inlines the called function into the basic 175// block of the caller. This returns false if it is not possible to inline this 176// call. The program is still in a well defined state if this occurs though. 177// 178// Note that this only does one level of inlining. For example, if the 179// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 180// exists in the instruction stream. Similiarly this will inline a recursive 181// function by one level. 182// 183bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { 184 Instruction *TheCall = CS.getInstruction(); 185 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 186 "Instruction not in function!"); 187 188 const Function *CalledFunc = CS.getCalledFunction(); 189 if (CalledFunc == 0 || // Can't inline external function or indirect 190 CalledFunc->isDeclaration() || // call, or call to a vararg function! 191 CalledFunc->getFunctionType()->isVarArg()) return false; 192 193 194 // If the call to the callee is a non-tail call, we must clear the 'tail' 195 // flags on any calls that we inline. 196 bool MustClearTailCallFlags = 197 isa<CallInst>(TheCall) && !cast<CallInst>(TheCall)->isTailCall(); 198 199 // If the call to the callee cannot throw, set the 'nounwind' flag on any 200 // calls that we inline. 201 bool MarkNoUnwind = CS.doesNotThrow(); 202 203 BasicBlock *OrigBB = TheCall->getParent(); 204 Function *Caller = OrigBB->getParent(); 205 BasicBlock *UnwindBB = OrigBB->getUnwindDest(); 206 207 // GC poses two hazards to inlining, which only occur when the callee has GC: 208 // 1. If the caller has no GC, then the callee's GC must be propagated to the 209 // caller. 210 // 2. If the caller has a differing GC, it is invalid to inline. 211 if (CalledFunc->hasCollector()) { 212 if (!Caller->hasCollector()) 213 Caller->setCollector(CalledFunc->getCollector()); 214 else if (CalledFunc->getCollector() != Caller->getCollector()) 215 return false; 216 } 217 218 // Get an iterator to the last basic block in the function, which will have 219 // the new function inlined after it. 220 // 221 Function::iterator LastBlock = &Caller->back(); 222 223 // Make sure to capture all of the return instructions from the cloned 224 // function. 225 std::vector<ReturnInst*> Returns; 226 ClonedCodeInfo InlinedFunctionInfo; 227 Function::iterator FirstNewBlock; 228 229 { // Scope to destroy ValueMap after cloning. 230 DenseMap<const Value*, Value*> ValueMap; 231 232 assert(std::distance(CalledFunc->arg_begin(), CalledFunc->arg_end()) == 233 std::distance(CS.arg_begin(), CS.arg_end()) && 234 "No varargs calls can be inlined!"); 235 236 // Calculate the vector of arguments to pass into the function cloner, which 237 // matches up the formal to the actual argument values. 238 CallSite::arg_iterator AI = CS.arg_begin(); 239 unsigned ArgNo = 0; 240 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 241 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 242 Value *ActualArg = *AI; 243 244 // When byval arguments actually inlined, we need to make the copy implied 245 // by them explicit. However, we don't do this if the callee is readonly 246 // or readnone, because the copy would be unneeded: the callee doesn't 247 // modify the struct. 248 if (CalledFunc->paramHasAttr(ArgNo+1, ParamAttr::ByVal) && 249 !CalledFunc->onlyReadsMemory()) { 250 const Type *AggTy = cast<PointerType>(I->getType())->getElementType(); 251 const Type *VoidPtrTy = PointerType::getUnqual(Type::Int8Ty); 252 253 // Create the alloca. If we have TargetData, use nice alignment. 254 unsigned Align = 1; 255 if (TD) Align = TD->getPrefTypeAlignment(AggTy); 256 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, I->getName(), 257 Caller->begin()->begin()); 258 // Emit a memcpy. 259 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 260 Intrinsic::memcpy_i64); 261 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 262 Value *SrcCast = new BitCastInst(*AI, VoidPtrTy, "tmp", TheCall); 263 264 Value *Size; 265 if (TD == 0) 266 Size = ConstantExpr::getSizeOf(AggTy); 267 else 268 Size = ConstantInt::get(Type::Int64Ty, TD->getTypeStoreSize(AggTy)); 269 270 // Always generate a memcpy of alignment 1 here because we don't know 271 // the alignment of the src pointer. Other optimizations can infer 272 // better alignment. 273 Value *CallArgs[] = { 274 DestCast, SrcCast, Size, ConstantInt::get(Type::Int32Ty, 1) 275 }; 276 CallInst *TheMemCpy = 277 new CallInst(MemCpyFn, CallArgs, CallArgs+4, "", TheCall); 278 279 // If we have a call graph, update it. 280 if (CG) { 281 CallGraphNode *MemCpyCGN = CG->getOrInsertFunction(MemCpyFn); 282 CallGraphNode *CallerNode = (*CG)[Caller]; 283 CallerNode->addCalledFunction(TheMemCpy, MemCpyCGN); 284 } 285 286 // Uses of the argument in the function should use our new alloca 287 // instead. 288 ActualArg = NewAlloca; 289 } 290 291 ValueMap[I] = ActualArg; 292 } 293 294 // We want the inliner to prune the code as it copies. We would LOVE to 295 // have no dead or constant instructions leftover after inlining occurs 296 // (which can happen, e.g., because an argument was constant), but we'll be 297 // happy with whatever the cloner can do. 298 CloneAndPruneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i", 299 &InlinedFunctionInfo, TD); 300 301 // Remember the first block that is newly cloned over. 302 FirstNewBlock = LastBlock; ++FirstNewBlock; 303 304 // Update the callgraph if requested. 305 if (CG) 306 UpdateCallGraphAfterInlining(Caller, CalledFunc, FirstNewBlock, ValueMap, 307 *CG); 308 } 309 310 // If there are any alloca instructions in the block that used to be the entry 311 // block for the callee, move them to the entry block of the caller. First 312 // calculate which instruction they should be inserted before. We insert the 313 // instructions at the end of the current alloca list. 314 // 315 { 316 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 317 for (BasicBlock::iterator I = FirstNewBlock->begin(), 318 E = FirstNewBlock->end(); I != E; ) 319 if (AllocaInst *AI = dyn_cast<AllocaInst>(I++)) { 320 // If the alloca is now dead, remove it. This often occurs due to code 321 // specialization. 322 if (AI->use_empty()) { 323 AI->eraseFromParent(); 324 continue; 325 } 326 327 if (isa<Constant>(AI->getArraySize())) { 328 // Scan for the block of allocas that we can move over, and move them 329 // all at once. 330 while (isa<AllocaInst>(I) && 331 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) 332 ++I; 333 334 // Transfer all of the allocas over in a block. Using splice means 335 // that the instructions aren't removed from the symbol table, then 336 // reinserted. 337 Caller->getEntryBlock().getInstList().splice( 338 InsertPoint, 339 FirstNewBlock->getInstList(), 340 AI, I); 341 } 342 } 343 } 344 345 // If the inlined code contained dynamic alloca instructions, wrap the inlined 346 // code with llvm.stacksave/llvm.stackrestore intrinsics. 347 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 348 Module *M = Caller->getParent(); 349 const Type *BytePtr = PointerType::getUnqual(Type::Int8Ty); 350 // Get the two intrinsics we care about. 351 Constant *StackSave, *StackRestore; 352 StackSave = M->getOrInsertFunction("llvm.stacksave", BytePtr, NULL); 353 StackRestore = M->getOrInsertFunction("llvm.stackrestore", Type::VoidTy, 354 BytePtr, NULL); 355 356 // If we are preserving the callgraph, add edges to the stacksave/restore 357 // functions for the calls we insert. 358 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0; 359 if (CG) { 360 // We know that StackSave/StackRestore are Function*'s, because they are 361 // intrinsics which must have the right types. 362 StackSaveCGN = CG->getOrInsertFunction(cast<Function>(StackSave)); 363 StackRestoreCGN = CG->getOrInsertFunction(cast<Function>(StackRestore)); 364 CallerNode = (*CG)[Caller]; 365 } 366 367 // Insert the llvm.stacksave. 368 CallInst *SavedPtr = new CallInst(StackSave, "savedstack", 369 FirstNewBlock->begin()); 370 if (CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN); 371 372 // Insert a call to llvm.stackrestore before any return instructions in the 373 // inlined function. 374 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 375 CallInst *CI = new CallInst(StackRestore, SavedPtr, "", Returns[i]); 376 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 377 } 378 379 // Count the number of StackRestore calls we insert. 380 unsigned NumStackRestores = Returns.size(); 381 382 // If we are inlining an invoke instruction, insert restores before each 383 // unwind. These unwinds will be rewritten into branches later. 384 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 385 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 386 BB != E; ++BB) 387 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 388 new CallInst(StackRestore, SavedPtr, "", UI); 389 ++NumStackRestores; 390 } 391 } 392 } 393 394 // If we are inlining tail call instruction through a call site that isn't 395 // marked 'tail', we must remove the tail marker for any calls in the inlined 396 // code. Also, calls inlined through a 'nounwind' call site should be marked 397 // 'nounwind'. 398 if (InlinedFunctionInfo.ContainsCalls && 399 (MustClearTailCallFlags || MarkNoUnwind)) { 400 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 401 BB != E; ++BB) 402 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 403 if (CallInst *CI = dyn_cast<CallInst>(I)) { 404 if (MustClearTailCallFlags) 405 CI->setTailCall(false); 406 if (MarkNoUnwind) 407 CI->setDoesNotThrow(); 408 } 409 } 410 411 // If we are inlining through a 'nounwind' call site then any inlined 'unwind' 412 // instructions are unreachable. 413 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind) 414 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 415 BB != E; ++BB) { 416 TerminatorInst *Term = BB->getTerminator(); 417 if (isa<UnwindInst>(Term)) { 418 new UnreachableInst(Term); 419 BB->getInstList().erase(Term); 420 } 421 } 422 423 // If we are inlining a function that unwinds into a BB with an unwind dest, 424 // turn the inlined unwinds into branches to the unwind dest. 425 if (InlinedFunctionInfo.ContainsUnwinds && UnwindBB && isa<CallInst>(TheCall)) 426 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 427 BB != E; ++BB) { 428 TerminatorInst *Term = BB->getTerminator(); 429 if (isa<UnwindInst>(Term)) { 430 new BranchInst(UnwindBB, Term); 431 BB->getInstList().erase(Term); 432 } 433 } 434 435 // If we are inlining for an invoke instruction, we must make sure to rewrite 436 // any inlined 'unwind' instructions into branches to the invoke exception 437 // destination, and call instructions into invoke instructions. 438 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 439 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 440 441 // If we cloned in _exactly one_ basic block, and if that block ends in a 442 // return instruction, we splice the body of the inlined callee directly into 443 // the calling basic block. 444 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 445 // Move all of the instructions right before the call. 446 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 447 FirstNewBlock->begin(), FirstNewBlock->end()); 448 // Remove the cloned basic block. 449 Caller->getBasicBlockList().pop_back(); 450 451 // If the call site was an invoke instruction, add a branch to the normal 452 // destination. 453 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 454 new BranchInst(II->getNormalDest(), TheCall); 455 456 // If the return instruction returned a value, replace uses of the call with 457 // uses of the returned value. 458 if (!TheCall->use_empty()) { 459 ReturnInst *R = Returns[0]; 460 if (R->getNumOperands() > 1) { 461 // Multiple return values. 462 while (!TheCall->use_empty()) { 463 GetResultInst *GR = cast<GetResultInst>(TheCall->use_back()); 464 Value *RV = R->getOperand(GR->getIndex()); 465 GR->replaceAllUsesWith(RV); 466 GR->eraseFromParent(); 467 } 468 } else 469 TheCall->replaceAllUsesWith(R->getReturnValue()); 470 } 471 // Since we are now done with the Call/Invoke, we can delete it. 472 TheCall->getParent()->getInstList().erase(TheCall); 473 474 // Since we are now done with the return instruction, delete it also. 475 Returns[0]->getParent()->getInstList().erase(Returns[0]); 476 477 // We are now done with the inlining. 478 return true; 479 } 480 481 // Otherwise, we have the normal case, of more than one block to inline or 482 // multiple return sites. 483 484 // We want to clone the entire callee function into the hole between the 485 // "starter" and "ender" blocks. How we accomplish this depends on whether 486 // this is an invoke instruction or a call instruction. 487 BasicBlock *AfterCallBB; 488 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 489 490 // Add an unconditional branch to make this look like the CallInst case... 491 BranchInst *NewBr = new BranchInst(II->getNormalDest(), TheCall); 492 493 // Split the basic block. This guarantees that no PHI nodes will have to be 494 // updated due to new incoming edges, and make the invoke case more 495 // symmetric to the call case. 496 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 497 CalledFunc->getName()+".exit"); 498 499 } else { // It's a call 500 // If this is a call instruction, we need to split the basic block that 501 // the call lives in. 502 // 503 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 504 CalledFunc->getName()+".exit"); 505 } 506 507 // Change the branch that used to go to AfterCallBB to branch to the first 508 // basic block of the inlined function. 509 // 510 TerminatorInst *Br = OrigBB->getTerminator(); 511 assert(Br && Br->getOpcode() == Instruction::Br && 512 "splitBasicBlock broken!"); 513 Br->setOperand(0, FirstNewBlock); 514 515 516 // Now that the function is correct, make it a little bit nicer. In 517 // particular, move the basic blocks inserted from the end of the function 518 // into the space made by splitting the source basic block. 519 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 520 FirstNewBlock, Caller->end()); 521 522 // Handle all of the return instructions that we just cloned in, and eliminate 523 // any users of the original call/invoke instruction. 524 if (!Returns.empty()) { 525 // The PHI node should go at the front of the new basic block to merge all 526 // possible incoming values. 527 SmallVector<PHINode *, 4> PHIs; 528 if (!TheCall->use_empty()) { 529 const Type *RTy = CalledFunc->getReturnType(); 530 if (const StructType *STy = dyn_cast<StructType>(RTy)) { 531 unsigned NumRetVals = STy->getNumElements(); 532 // Create new phi nodes such that phi node number in the PHIs vector 533 // match corresponding return value operand number. 534 for (unsigned i = 0; i < NumRetVals; ++i) { 535 PHINode *PHI = new PHINode(STy->getElementType(i), 536 TheCall->getName(), AfterCallBB->begin()); 537 PHIs.push_back(PHI); 538 } 539 // TheCall results are used by GetResult instructions. 540 while (!TheCall->use_empty()) { 541 GetResultInst *GR = cast<GetResultInst>(TheCall->use_back()); 542 GR->replaceAllUsesWith(PHIs[GR->getIndex()]); 543 GR->eraseFromParent(); 544 } 545 } else { 546 PHINode *PHI = new PHINode(RTy, TheCall->getName(), AfterCallBB->begin()); 547 PHIs.push_back(PHI); 548 // Anything that used the result of the function call should now use the 549 // PHI node as their operand. 550 TheCall->replaceAllUsesWith(PHI); 551 } 552 } 553 554 // Loop over all of the return instructions adding entries to the PHI node as 555 // appropriate. 556 if (!PHIs.empty()) { 557 const Type *RTy = CalledFunc->getReturnType(); 558 if (const StructType *STy = dyn_cast<StructType>(RTy)) { 559 unsigned NumRetVals = STy->getNumElements(); 560 for (unsigned j = 0; j < NumRetVals; ++j) { 561 PHINode *PHI = PHIs[j]; 562 // Each PHI node will receive one value from each return instruction. 563 for(unsigned i = 0, e = Returns.size(); i != e; ++i) { 564 ReturnInst *RI = Returns[i]; 565 PHI->addIncoming(RI->getReturnValue(j /*PHI number matches operand number*/), 566 RI->getParent()); 567 } 568 } 569 } else { 570 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 571 ReturnInst *RI = Returns[i]; 572 assert(PHIs.size() == 1 && "Invalid number of PHI nodes"); 573 assert(RI->getReturnValue() && "Ret should have value!"); 574 assert(RI->getReturnValue()->getType() == PHIs[0]->getType() && 575 "Ret value not consistent in function!"); 576 PHIs[0]->addIncoming(RI->getReturnValue(), RI->getParent()); 577 } 578 } 579 } 580 581 // Add a branch to the merge points and remove retrun instructions. 582 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 583 ReturnInst *RI = Returns[i]; 584 new BranchInst(AfterCallBB, RI); 585 RI->getParent()->getInstList().erase(RI); 586 } 587 } else if (!TheCall->use_empty()) { 588 // No returns, but something is using the return value of the call. Just 589 // nuke the result. 590 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 591 } 592 593 // Since we are now done with the Call/Invoke, we can delete it. 594 TheCall->eraseFromParent(); 595 596 // We should always be able to fold the entry block of the function into the 597 // single predecessor of the block... 598 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 599 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 600 601 // Splice the code entry block into calling block, right before the 602 // unconditional branch. 603 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 604 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 605 606 // Remove the unconditional branch. 607 OrigBB->getInstList().erase(Br); 608 609 // Now we can remove the CalleeEntry block, which is now empty. 610 Caller->getBasicBlockList().erase(CalleeEntry); 611 612 return true; 613} 614