InlineFunction.cpp revision f8b99cb03bd2278328e995c4b8400a09f77ef19b
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/Target/TargetData.h" 26#include "llvm/ADT/SmallVector.h" 27#include "llvm/ADT/StringExtras.h" 28#include "llvm/Support/CallSite.h" 29using namespace llvm; 30 31bool llvm::InlineFunction(CallInst *CI, CallGraph *CG, const TargetData *TD, 32 SmallVectorImpl<AllocaInst*> *StaticAllocas) { 33 return InlineFunction(CallSite(CI), CG, TD, StaticAllocas); 34} 35bool llvm::InlineFunction(InvokeInst *II, CallGraph *CG, const TargetData *TD, 36 SmallVectorImpl<AllocaInst*> *StaticAllocas) { 37 return InlineFunction(CallSite(II), CG, TD, StaticAllocas); 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 SmallVector<Value*, 8> InvokeArgs(CI->op_begin()+1, CI->op_end()); 69 InvokeInst *II = 70 InvokeInst::Create(CI->getCalledValue(), Split, InvokeDest, 71 InvokeArgs.begin(), InvokeArgs.end(), 72 CI->getName(), BB->getTerminator()); 73 II->setCallingConv(CI->getCallingConv()); 74 II->setAttributes(CI->getAttributes()); 75 76 // Make sure that anything using the call now uses the invoke! This also 77 // updates the CallGraph if present. 78 CI->replaceAllUsesWith(II); 79 80 // Delete the unconditional branch inserted by splitBasicBlock 81 BB->getInstList().pop_back(); 82 Split->getInstList().pop_front(); // Delete the original call 83 84 // Update any PHI nodes in the exceptional block to indicate that 85 // there is now a new entry in them. 86 unsigned i = 0; 87 for (BasicBlock::iterator I = InvokeDest->begin(); 88 isa<PHINode>(I); ++I, ++i) 89 cast<PHINode>(I)->addIncoming(InvokeDestPHIValues[i], BB); 90 91 // This basic block is now complete, the caller will continue scanning the 92 // next one. 93 return; 94 } 95} 96 97 98/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 99/// in the body of the inlined function into invokes and turn unwind 100/// instructions into branches to the invoke unwind dest. 101/// 102/// II is the invoke instruction being inlined. FirstNewBlock is the first 103/// block of the inlined code (the last block is the end of the function), 104/// and InlineCodeInfo is information about the code that got inlined. 105static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 106 ClonedCodeInfo &InlinedCodeInfo) { 107 BasicBlock *InvokeDest = II->getUnwindDest(); 108 SmallVector<Value*, 8> InvokeDestPHIValues; 109 110 // If there are PHI nodes in the unwind destination block, we need to 111 // keep track of which values came into them from this invoke, then remove 112 // the entry for this block. 113 BasicBlock *InvokeBlock = II->getParent(); 114 for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) { 115 PHINode *PN = cast<PHINode>(I); 116 // Save the value to use for this edge. 117 InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock)); 118 } 119 120 Function *Caller = FirstNewBlock->getParent(); 121 122 // The inlined code is currently at the end of the function, scan from the 123 // start of the inlined code to its end, checking for stuff we need to 124 // rewrite. If the code doesn't have calls or unwinds, we know there is 125 // nothing to rewrite. 126 if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) { 127 // Now that everything is happy, we have one final detail. The PHI nodes in 128 // the exception destination block still have entries due to the original 129 // invoke instruction. Eliminate these entries (which might even delete the 130 // PHI node) now. 131 InvokeDest->removePredecessor(II->getParent()); 132 return; 133 } 134 135 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){ 136 if (InlinedCodeInfo.ContainsCalls) 137 HandleCallsInBlockInlinedThroughInvoke(BB, InvokeDest, 138 InvokeDestPHIValues); 139 140 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 141 // An UnwindInst requires special handling when it gets inlined into an 142 // invoke site. Once this happens, we know that the unwind would cause 143 // a control transfer to the invoke exception destination, so we can 144 // transform it into a direct branch to the exception destination. 145 BranchInst::Create(InvokeDest, UI); 146 147 // Delete the unwind instruction! 148 UI->eraseFromParent(); 149 150 // Update any PHI nodes in the exceptional block to indicate that 151 // there is now a new entry in them. 152 unsigned i = 0; 153 for (BasicBlock::iterator I = InvokeDest->begin(); 154 isa<PHINode>(I); ++I, ++i) { 155 PHINode *PN = cast<PHINode>(I); 156 PN->addIncoming(InvokeDestPHIValues[i], BB); 157 } 158 } 159 } 160 161 // Now that everything is happy, we have one final detail. The PHI nodes in 162 // the exception destination block still have entries due to the original 163 // invoke instruction. Eliminate these entries (which might even delete the 164 // PHI node) now. 165 InvokeDest->removePredecessor(II->getParent()); 166} 167 168/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 169/// into the caller, update the specified callgraph to reflect the changes we 170/// made. Note that it's possible that not all code was copied over, so only 171/// some edges of the callgraph may remain. 172static void UpdateCallGraphAfterInlining(CallSite CS, 173 Function::iterator FirstNewBlock, 174 DenseMap<const Value*, Value*> &ValueMap, 175 CallGraph &CG) { 176 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 177 const Function *Callee = CS.getCalledFunction(); 178 CallGraphNode *CalleeNode = CG[Callee]; 179 CallGraphNode *CallerNode = CG[Caller]; 180 181 // Since we inlined some uninlined call sites in the callee into the caller, 182 // add edges from the caller to all of the callees of the callee. 183 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 184 185 // Consider the case where CalleeNode == CallerNode. 186 CallGraphNode::CalledFunctionsVector CallCache; 187 if (CalleeNode == CallerNode) { 188 CallCache.assign(I, E); 189 I = CallCache.begin(); 190 E = CallCache.end(); 191 } 192 193 for (; I != E; ++I) { 194 const Value *OrigCall = I->first; 195 196 DenseMap<const Value*, Value*>::iterator VMI = ValueMap.find(OrigCall); 197 // Only copy the edge if the call was inlined! 198 if (VMI == ValueMap.end() || VMI->second == 0) 199 continue; 200 201 // If the call was inlined, but then constant folded, there is no edge to 202 // add. Check for this case. 203 if (Instruction *NewCall = dyn_cast<Instruction>(VMI->second)) 204 CallerNode->addCalledFunction(CallSite::get(NewCall), I->second); 205 } 206 207 // Update the call graph by deleting the edge from Callee to Caller. We must 208 // do this after the loop above in case Caller and Callee are the same. 209 CallerNode->removeCallEdgeFor(CS); 210} 211 212// InlineFunction - This function inlines the called function into the basic 213// block of the caller. This returns false if it is not possible to inline this 214// call. The program is still in a well defined state if this occurs though. 215// 216// Note that this only does one level of inlining. For example, if the 217// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 218// exists in the instruction stream. Similiarly this will inline a recursive 219// function by one level. 220// 221bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD, 222 SmallVectorImpl<AllocaInst*> *StaticAllocas) { 223 Instruction *TheCall = CS.getInstruction(); 224 LLVMContext &Context = TheCall->getContext(); 225 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 226 "Instruction not in function!"); 227 228 const Function *CalledFunc = CS.getCalledFunction(); 229 if (CalledFunc == 0 || // Can't inline external function or indirect 230 CalledFunc->isDeclaration() || // call, or call to a vararg function! 231 CalledFunc->getFunctionType()->isVarArg()) return false; 232 233 234 // If the call to the callee is not a tail call, we must clear the 'tail' 235 // flags on any calls that we inline. 236 bool MustClearTailCallFlags = 237 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall()); 238 239 // If the call to the callee cannot throw, set the 'nounwind' flag on any 240 // calls that we inline. 241 bool MarkNoUnwind = CS.doesNotThrow(); 242 243 BasicBlock *OrigBB = TheCall->getParent(); 244 Function *Caller = OrigBB->getParent(); 245 246 // GC poses two hazards to inlining, which only occur when the callee has GC: 247 // 1. If the caller has no GC, then the callee's GC must be propagated to the 248 // caller. 249 // 2. If the caller has a differing GC, it is invalid to inline. 250 if (CalledFunc->hasGC()) { 251 if (!Caller->hasGC()) 252 Caller->setGC(CalledFunc->getGC()); 253 else if (CalledFunc->getGC() != Caller->getGC()) 254 return false; 255 } 256 257 // Get an iterator to the last basic block in the function, which will have 258 // the new function inlined after it. 259 // 260 Function::iterator LastBlock = &Caller->back(); 261 262 // Make sure to capture all of the return instructions from the cloned 263 // function. 264 SmallVector<ReturnInst*, 8> Returns; 265 ClonedCodeInfo InlinedFunctionInfo; 266 Function::iterator FirstNewBlock; 267 268 { // Scope to destroy ValueMap after cloning. 269 DenseMap<const Value*, Value*> ValueMap; 270 271 assert(CalledFunc->arg_size() == CS.arg_size() && 272 "No varargs calls can be inlined!"); 273 274 // Calculate the vector of arguments to pass into the function cloner, which 275 // matches up the formal to the actual argument values. 276 CallSite::arg_iterator AI = CS.arg_begin(); 277 unsigned ArgNo = 0; 278 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 279 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 280 Value *ActualArg = *AI; 281 282 // When byval arguments actually inlined, we need to make the copy implied 283 // by them explicit. However, we don't do this if the callee is readonly 284 // or readnone, because the copy would be unneeded: the callee doesn't 285 // modify the struct. 286 if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal) && 287 !CalledFunc->onlyReadsMemory()) { 288 const Type *AggTy = cast<PointerType>(I->getType())->getElementType(); 289 const Type *VoidPtrTy = 290 Type::getInt8PtrTy(Context); 291 292 // Create the alloca. If we have TargetData, use nice alignment. 293 unsigned Align = 1; 294 if (TD) Align = TD->getPrefTypeAlignment(AggTy); 295 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, 296 I->getName(), 297 &*Caller->begin()->begin()); 298 // Emit a memcpy. 299 const Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)}; 300 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 301 Intrinsic::memcpy, 302 Tys, 3); 303 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 304 Value *SrcCast = new BitCastInst(*AI, VoidPtrTy, "tmp", TheCall); 305 306 Value *Size; 307 if (TD == 0) 308 Size = ConstantExpr::getSizeOf(AggTy); 309 else 310 Size = ConstantInt::get(Type::getInt64Ty(Context), 311 TD->getTypeStoreSize(AggTy)); 312 313 // Always generate a memcpy of alignment 1 here because we don't know 314 // the alignment of the src pointer. Other optimizations can infer 315 // better alignment. 316 Value *CallArgs[] = { 317 DestCast, SrcCast, Size, 318 ConstantInt::get(Type::getInt32Ty(Context), 1), 319 ConstantInt::get(Type::getInt1Ty(Context), 0) 320 }; 321 CallInst *TheMemCpy = 322 CallInst::Create(MemCpyFn, CallArgs, CallArgs+5, "", TheCall); 323 324 // If we have a call graph, update it. 325 if (CG) { 326 CallGraphNode *MemCpyCGN = CG->getOrInsertFunction(MemCpyFn); 327 CallGraphNode *CallerNode = (*CG)[Caller]; 328 CallerNode->addCalledFunction(TheMemCpy, MemCpyCGN); 329 } 330 331 // Uses of the argument in the function should use our new alloca 332 // instead. 333 ActualArg = NewAlloca; 334 } 335 336 ValueMap[I] = ActualArg; 337 } 338 339 // We want the inliner to prune the code as it copies. We would LOVE to 340 // have no dead or constant instructions leftover after inlining occurs 341 // (which can happen, e.g., because an argument was constant), but we'll be 342 // happy with whatever the cloner can do. 343 CloneAndPruneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i", 344 &InlinedFunctionInfo, TD, TheCall); 345 346 // Remember the first block that is newly cloned over. 347 FirstNewBlock = LastBlock; ++FirstNewBlock; 348 349 // Update the callgraph if requested. 350 if (CG) 351 UpdateCallGraphAfterInlining(CS, FirstNewBlock, ValueMap, *CG); 352 } 353 354 // If there are any alloca instructions in the block that used to be the entry 355 // block for the callee, move them to the entry block of the caller. First 356 // calculate which instruction they should be inserted before. We insert the 357 // instructions at the end of the current alloca list. 358 // 359 { 360 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 361 for (BasicBlock::iterator I = FirstNewBlock->begin(), 362 E = FirstNewBlock->end(); I != E; ) { 363 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 364 if (AI == 0) continue; 365 366 // If the alloca is now dead, remove it. This often occurs due to code 367 // specialization. 368 if (AI->use_empty()) { 369 AI->eraseFromParent(); 370 continue; 371 } 372 373 if (!isa<Constant>(AI->getArraySize())) 374 continue; 375 376 // Keep track of the static allocas that we inline into the caller if the 377 // StaticAllocas pointer is non-null. 378 if (StaticAllocas) StaticAllocas->push_back(AI); 379 380 // Scan for the block of allocas that we can move over, and move them 381 // all at once. 382 while (isa<AllocaInst>(I) && 383 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 384 if (StaticAllocas) StaticAllocas->push_back(cast<AllocaInst>(I)); 385 ++I; 386 } 387 388 // Transfer all of the allocas over in a block. Using splice means 389 // that the instructions aren't removed from the symbol table, then 390 // reinserted. 391 Caller->getEntryBlock().getInstList().splice(InsertPoint, 392 FirstNewBlock->getInstList(), 393 AI, I); 394 } 395 } 396 397 // If the inlined code contained dynamic alloca instructions, wrap the inlined 398 // code with llvm.stacksave/llvm.stackrestore intrinsics. 399 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 400 Module *M = Caller->getParent(); 401 // Get the two intrinsics we care about. 402 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 403 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 404 405 // If we are preserving the callgraph, add edges to the stacksave/restore 406 // functions for the calls we insert. 407 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0; 408 if (CG) { 409 StackSaveCGN = CG->getOrInsertFunction(StackSave); 410 StackRestoreCGN = CG->getOrInsertFunction(StackRestore); 411 CallerNode = (*CG)[Caller]; 412 } 413 414 // Insert the llvm.stacksave. 415 CallInst *SavedPtr = CallInst::Create(StackSave, "savedstack", 416 FirstNewBlock->begin()); 417 if (CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN); 418 419 // Insert a call to llvm.stackrestore before any return instructions in the 420 // inlined function. 421 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 422 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", Returns[i]); 423 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 424 } 425 426 // Count the number of StackRestore calls we insert. 427 unsigned NumStackRestores = Returns.size(); 428 429 // If we are inlining an invoke instruction, insert restores before each 430 // unwind. These unwinds will be rewritten into branches later. 431 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 432 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 433 BB != E; ++BB) 434 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 435 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", UI); 436 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 437 ++NumStackRestores; 438 } 439 } 440 } 441 442 // If we are inlining tail call instruction through a call site that isn't 443 // marked 'tail', we must remove the tail marker for any calls in the inlined 444 // code. Also, calls inlined through a 'nounwind' call site should be marked 445 // 'nounwind'. 446 if (InlinedFunctionInfo.ContainsCalls && 447 (MustClearTailCallFlags || MarkNoUnwind)) { 448 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 449 BB != E; ++BB) 450 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 451 if (CallInst *CI = dyn_cast<CallInst>(I)) { 452 if (MustClearTailCallFlags) 453 CI->setTailCall(false); 454 if (MarkNoUnwind) 455 CI->setDoesNotThrow(); 456 } 457 } 458 459 // If we are inlining through a 'nounwind' call site then any inlined 'unwind' 460 // instructions are unreachable. 461 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind) 462 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 463 BB != E; ++BB) { 464 TerminatorInst *Term = BB->getTerminator(); 465 if (isa<UnwindInst>(Term)) { 466 new UnreachableInst(Context, Term); 467 BB->getInstList().erase(Term); 468 } 469 } 470 471 // If we are inlining for an invoke instruction, we must make sure to rewrite 472 // any inlined 'unwind' instructions into branches to the invoke exception 473 // destination, and call instructions into invoke instructions. 474 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 475 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 476 477 // If we cloned in _exactly one_ basic block, and if that block ends in a 478 // return instruction, we splice the body of the inlined callee directly into 479 // the calling basic block. 480 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 481 // Move all of the instructions right before the call. 482 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 483 FirstNewBlock->begin(), FirstNewBlock->end()); 484 // Remove the cloned basic block. 485 Caller->getBasicBlockList().pop_back(); 486 487 // If the call site was an invoke instruction, add a branch to the normal 488 // destination. 489 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 490 BranchInst::Create(II->getNormalDest(), TheCall); 491 492 // If the return instruction returned a value, replace uses of the call with 493 // uses of the returned value. 494 if (!TheCall->use_empty()) { 495 ReturnInst *R = Returns[0]; 496 if (TheCall == R->getReturnValue()) 497 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 498 else 499 TheCall->replaceAllUsesWith(R->getReturnValue()); 500 } 501 // Since we are now done with the Call/Invoke, we can delete it. 502 TheCall->eraseFromParent(); 503 504 // Since we are now done with the return instruction, delete it also. 505 Returns[0]->eraseFromParent(); 506 507 // We are now done with the inlining. 508 return true; 509 } 510 511 // Otherwise, we have the normal case, of more than one block to inline or 512 // multiple return sites. 513 514 // We want to clone the entire callee function into the hole between the 515 // "starter" and "ender" blocks. How we accomplish this depends on whether 516 // this is an invoke instruction or a call instruction. 517 BasicBlock *AfterCallBB; 518 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 519 520 // Add an unconditional branch to make this look like the CallInst case... 521 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 522 523 // Split the basic block. This guarantees that no PHI nodes will have to be 524 // updated due to new incoming edges, and make the invoke case more 525 // symmetric to the call case. 526 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 527 CalledFunc->getName()+".exit"); 528 529 } else { // It's a call 530 // If this is a call instruction, we need to split the basic block that 531 // the call lives in. 532 // 533 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 534 CalledFunc->getName()+".exit"); 535 } 536 537 // Change the branch that used to go to AfterCallBB to branch to the first 538 // basic block of the inlined function. 539 // 540 TerminatorInst *Br = OrigBB->getTerminator(); 541 assert(Br && Br->getOpcode() == Instruction::Br && 542 "splitBasicBlock broken!"); 543 Br->setOperand(0, FirstNewBlock); 544 545 546 // Now that the function is correct, make it a little bit nicer. In 547 // particular, move the basic blocks inserted from the end of the function 548 // into the space made by splitting the source basic block. 549 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 550 FirstNewBlock, Caller->end()); 551 552 // Handle all of the return instructions that we just cloned in, and eliminate 553 // any users of the original call/invoke instruction. 554 const Type *RTy = CalledFunc->getReturnType(); 555 556 if (Returns.size() > 1) { 557 // The PHI node should go at the front of the new basic block to merge all 558 // possible incoming values. 559 PHINode *PHI = 0; 560 if (!TheCall->use_empty()) { 561 PHI = PHINode::Create(RTy, TheCall->getName(), 562 AfterCallBB->begin()); 563 // Anything that used the result of the function call should now use the 564 // PHI node as their operand. 565 TheCall->replaceAllUsesWith(PHI); 566 } 567 568 // Loop over all of the return instructions adding entries to the PHI node 569 // as appropriate. 570 if (PHI) { 571 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 572 ReturnInst *RI = Returns[i]; 573 assert(RI->getReturnValue()->getType() == PHI->getType() && 574 "Ret value not consistent in function!"); 575 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 576 } 577 578 // Now that we inserted the PHI, check to see if it has a single value 579 // (e.g. all the entries are the same or undef). If so, remove the PHI so 580 // it doesn't block other optimizations. 581 if (Value *V = PHI->hasConstantValue()) { 582 PHI->replaceAllUsesWith(V); 583 PHI->eraseFromParent(); 584 } 585 } 586 587 588 // Add a branch to the merge points and remove return instructions. 589 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 590 ReturnInst *RI = Returns[i]; 591 BranchInst::Create(AfterCallBB, RI); 592 RI->eraseFromParent(); 593 } 594 } else if (!Returns.empty()) { 595 // Otherwise, if there is exactly one return value, just replace anything 596 // using the return value of the call with the computed value. 597 if (!TheCall->use_empty()) { 598 if (TheCall == Returns[0]->getReturnValue()) 599 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 600 else 601 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 602 } 603 604 // Splice the code from the return block into the block that it will return 605 // to, which contains the code that was after the call. 606 BasicBlock *ReturnBB = Returns[0]->getParent(); 607 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 608 ReturnBB->getInstList()); 609 610 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 611 ReturnBB->replaceAllUsesWith(AfterCallBB); 612 613 // Delete the return instruction now and empty ReturnBB now. 614 Returns[0]->eraseFromParent(); 615 ReturnBB->eraseFromParent(); 616 } else if (!TheCall->use_empty()) { 617 // No returns, but something is using the return value of the call. Just 618 // nuke the result. 619 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 620 } 621 622 // Since we are now done with the Call/Invoke, we can delete it. 623 TheCall->eraseFromParent(); 624 625 // We should always be able to fold the entry block of the function into the 626 // single predecessor of the block... 627 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 628 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 629 630 // Splice the code entry block into calling block, right before the 631 // unconditional branch. 632 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 633 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 634 635 // Remove the unconditional branch. 636 OrigBB->getInstList().erase(Br); 637 638 // Now we can remove the CalleeEntry block, which is now empty. 639 Caller->getBasicBlockList().erase(CalleeEntry); 640 641 return true; 642} 643