InlineFunction.cpp revision 607a7ab3da72a2eb53553a520507cbb8068dd1d8
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// InlineFunction - This function inlines the called function into the basic 214// block of the caller. This returns false if it is not possible to inline this 215// call. The program is still in a well defined state if this occurs though. 216// 217// Note that this only does one level of inlining. For example, if the 218// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 219// exists in the instruction stream. Similiarly this will inline a recursive 220// function by one level. 221// 222bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD, 223 SmallVectorImpl<AllocaInst*> *StaticAllocas) { 224 Instruction *TheCall = CS.getInstruction(); 225 LLVMContext &Context = TheCall->getContext(); 226 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 227 "Instruction not in function!"); 228 229 const Function *CalledFunc = CS.getCalledFunction(); 230 if (CalledFunc == 0 || // Can't inline external function or indirect 231 CalledFunc->isDeclaration() || // call, or call to a vararg function! 232 CalledFunc->getFunctionType()->isVarArg()) return false; 233 234 235 // If the call to the callee is not a tail call, we must clear the 'tail' 236 // flags on any calls that we inline. 237 bool MustClearTailCallFlags = 238 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall()); 239 240 // If the call to the callee cannot throw, set the 'nounwind' flag on any 241 // calls that we inline. 242 bool MarkNoUnwind = CS.doesNotThrow(); 243 244 BasicBlock *OrigBB = TheCall->getParent(); 245 Function *Caller = OrigBB->getParent(); 246 247 // GC poses two hazards to inlining, which only occur when the callee has GC: 248 // 1. If the caller has no GC, then the callee's GC must be propagated to the 249 // caller. 250 // 2. If the caller has a differing GC, it is invalid to inline. 251 if (CalledFunc->hasGC()) { 252 if (!Caller->hasGC()) 253 Caller->setGC(CalledFunc->getGC()); 254 else if (CalledFunc->getGC() != Caller->getGC()) 255 return false; 256 } 257 258 // Get an iterator to the last basic block in the function, which will have 259 // the new function inlined after it. 260 // 261 Function::iterator LastBlock = &Caller->back(); 262 263 // Make sure to capture all of the return instructions from the cloned 264 // function. 265 SmallVector<ReturnInst*, 8> Returns; 266 ClonedCodeInfo InlinedFunctionInfo; 267 Function::iterator FirstNewBlock; 268 269 { // Scope to destroy ValueMap after cloning. 270 DenseMap<const Value*, Value*> ValueMap; 271 272 assert(CalledFunc->arg_size() == CS.arg_size() && 273 "No varargs calls can be inlined!"); 274 275 // Calculate the vector of arguments to pass into the function cloner, which 276 // matches up the formal to the actual argument values. 277 CallSite::arg_iterator AI = CS.arg_begin(); 278 unsigned ArgNo = 0; 279 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 280 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 281 Value *ActualArg = *AI; 282 283 // When byval arguments actually inlined, we need to make the copy implied 284 // by them explicit. However, we don't do this if the callee is readonly 285 // or readnone, because the copy would be unneeded: the callee doesn't 286 // modify the struct. 287 if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal) && 288 !CalledFunc->onlyReadsMemory()) { 289 const Type *AggTy = cast<PointerType>(I->getType())->getElementType(); 290 const Type *VoidPtrTy = 291 Type::getInt8PtrTy(Context); 292 293 // Create the alloca. If we have TargetData, use nice alignment. 294 unsigned Align = 1; 295 if (TD) Align = TD->getPrefTypeAlignment(AggTy); 296 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, 297 I->getName(), 298 &*Caller->begin()->begin()); 299 // Emit a memcpy. 300 const Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)}; 301 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 302 Intrinsic::memcpy, 303 Tys, 3); 304 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 305 Value *SrcCast = new BitCastInst(*AI, VoidPtrTy, "tmp", TheCall); 306 307 Value *Size; 308 if (TD == 0) 309 Size = ConstantExpr::getSizeOf(AggTy); 310 else 311 Size = ConstantInt::get(Type::getInt64Ty(Context), 312 TD->getTypeStoreSize(AggTy)); 313 314 // Always generate a memcpy of alignment 1 here because we don't know 315 // the alignment of the src pointer. Other optimizations can infer 316 // better alignment. 317 Value *CallArgs[] = { 318 DestCast, SrcCast, Size, 319 ConstantInt::get(Type::getInt32Ty(Context), 1), 320 ConstantInt::get(Type::getInt1Ty(Context), 0) 321 }; 322 CallInst *TheMemCpy = 323 CallInst::Create(MemCpyFn, CallArgs, CallArgs+5, "", TheCall); 324 325 // If we have a call graph, update it. 326 if (CG) { 327 CallGraphNode *MemCpyCGN = CG->getOrInsertFunction(MemCpyFn); 328 CallGraphNode *CallerNode = (*CG)[Caller]; 329 CallerNode->addCalledFunction(TheMemCpy, MemCpyCGN); 330 } 331 332 // Uses of the argument in the function should use our new alloca 333 // instead. 334 ActualArg = NewAlloca; 335 } 336 337 ValueMap[I] = ActualArg; 338 } 339 340 // We want the inliner to prune the code as it copies. We would LOVE to 341 // have no dead or constant instructions leftover after inlining occurs 342 // (which can happen, e.g., because an argument was constant), but we'll be 343 // happy with whatever the cloner can do. 344 CloneAndPruneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i", 345 &InlinedFunctionInfo, TD, TheCall); 346 347 // Remember the first block that is newly cloned over. 348 FirstNewBlock = LastBlock; ++FirstNewBlock; 349 350 // Update the callgraph if requested. 351 if (CG) 352 UpdateCallGraphAfterInlining(CS, FirstNewBlock, ValueMap, *CG); 353 } 354 355 // If there are any alloca instructions in the block that used to be the entry 356 // block for the callee, move them to the entry block of the caller. First 357 // calculate which instruction they should be inserted before. We insert the 358 // instructions at the end of the current alloca list. 359 // 360 { 361 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 362 for (BasicBlock::iterator I = FirstNewBlock->begin(), 363 E = FirstNewBlock->end(); I != E; ) { 364 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 365 if (AI == 0) continue; 366 367 // If the alloca is now dead, remove it. This often occurs due to code 368 // specialization. 369 if (AI->use_empty()) { 370 AI->eraseFromParent(); 371 continue; 372 } 373 374 if (!isa<Constant>(AI->getArraySize())) 375 continue; 376 377 // Keep track of the static allocas that we inline into the caller if the 378 // StaticAllocas pointer is non-null. 379 if (StaticAllocas) StaticAllocas->push_back(AI); 380 381 // Scan for the block of allocas that we can move over, and move them 382 // all at once. 383 while (isa<AllocaInst>(I) && 384 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 385 if (StaticAllocas) StaticAllocas->push_back(cast<AllocaInst>(I)); 386 ++I; 387 } 388 389 // Transfer all of the allocas over in a block. Using splice means 390 // that the instructions aren't removed from the symbol table, then 391 // reinserted. 392 Caller->getEntryBlock().getInstList().splice(InsertPoint, 393 FirstNewBlock->getInstList(), 394 AI, I); 395 } 396 } 397 398 // If the inlined code contained dynamic alloca instructions, wrap the inlined 399 // code with llvm.stacksave/llvm.stackrestore intrinsics. 400 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 401 Module *M = Caller->getParent(); 402 // Get the two intrinsics we care about. 403 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 404 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 405 406 // If we are preserving the callgraph, add edges to the stacksave/restore 407 // functions for the calls we insert. 408 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0; 409 if (CG) { 410 StackSaveCGN = CG->getOrInsertFunction(StackSave); 411 StackRestoreCGN = CG->getOrInsertFunction(StackRestore); 412 CallerNode = (*CG)[Caller]; 413 } 414 415 // Insert the llvm.stacksave. 416 CallInst *SavedPtr = CallInst::Create(StackSave, "savedstack", 417 FirstNewBlock->begin()); 418 if (CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN); 419 420 // Insert a call to llvm.stackrestore before any return instructions in the 421 // inlined function. 422 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 423 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", Returns[i]); 424 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 425 } 426 427 // Count the number of StackRestore calls we insert. 428 unsigned NumStackRestores = Returns.size(); 429 430 // If we are inlining an invoke instruction, insert restores before each 431 // unwind. These unwinds will be rewritten into branches later. 432 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 433 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 434 BB != E; ++BB) 435 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 436 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", UI); 437 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 438 ++NumStackRestores; 439 } 440 } 441 } 442 443 // If we are inlining tail call instruction through a call site that isn't 444 // marked 'tail', we must remove the tail marker for any calls in the inlined 445 // code. Also, calls inlined through a 'nounwind' call site should be marked 446 // 'nounwind'. 447 if (InlinedFunctionInfo.ContainsCalls && 448 (MustClearTailCallFlags || MarkNoUnwind)) { 449 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 450 BB != E; ++BB) 451 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 452 if (CallInst *CI = dyn_cast<CallInst>(I)) { 453 if (MustClearTailCallFlags) 454 CI->setTailCall(false); 455 if (MarkNoUnwind) 456 CI->setDoesNotThrow(); 457 } 458 } 459 460 // If we are inlining through a 'nounwind' call site then any inlined 'unwind' 461 // instructions are unreachable. 462 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind) 463 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 464 BB != E; ++BB) { 465 TerminatorInst *Term = BB->getTerminator(); 466 if (isa<UnwindInst>(Term)) { 467 new UnreachableInst(Context, Term); 468 BB->getInstList().erase(Term); 469 } 470 } 471 472 // If we are inlining for an invoke instruction, we must make sure to rewrite 473 // any inlined 'unwind' instructions into branches to the invoke exception 474 // destination, and call instructions into invoke instructions. 475 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 476 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 477 478 // If we cloned in _exactly one_ basic block, and if that block ends in a 479 // return instruction, we splice the body of the inlined callee directly into 480 // the calling basic block. 481 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 482 // Move all of the instructions right before the call. 483 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 484 FirstNewBlock->begin(), FirstNewBlock->end()); 485 // Remove the cloned basic block. 486 Caller->getBasicBlockList().pop_back(); 487 488 // If the call site was an invoke instruction, add a branch to the normal 489 // destination. 490 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 491 BranchInst::Create(II->getNormalDest(), TheCall); 492 493 // If the return instruction returned a value, replace uses of the call with 494 // uses of the returned value. 495 if (!TheCall->use_empty()) { 496 ReturnInst *R = Returns[0]; 497 if (TheCall == R->getReturnValue()) 498 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 499 else 500 TheCall->replaceAllUsesWith(R->getReturnValue()); 501 } 502 // Since we are now done with the Call/Invoke, we can delete it. 503 TheCall->eraseFromParent(); 504 505 // Since we are now done with the return instruction, delete it also. 506 Returns[0]->eraseFromParent(); 507 508 // We are now done with the inlining. 509 return true; 510 } 511 512 // Otherwise, we have the normal case, of more than one block to inline or 513 // multiple return sites. 514 515 // We want to clone the entire callee function into the hole between the 516 // "starter" and "ender" blocks. How we accomplish this depends on whether 517 // this is an invoke instruction or a call instruction. 518 BasicBlock *AfterCallBB; 519 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 520 521 // Add an unconditional branch to make this look like the CallInst case... 522 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 523 524 // Split the basic block. This guarantees that no PHI nodes will have to be 525 // updated due to new incoming edges, and make the invoke case more 526 // symmetric to the call case. 527 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 528 CalledFunc->getName()+".exit"); 529 530 } else { // It's a call 531 // If this is a call instruction, we need to split the basic block that 532 // the call lives in. 533 // 534 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 535 CalledFunc->getName()+".exit"); 536 } 537 538 // Change the branch that used to go to AfterCallBB to branch to the first 539 // basic block of the inlined function. 540 // 541 TerminatorInst *Br = OrigBB->getTerminator(); 542 assert(Br && Br->getOpcode() == Instruction::Br && 543 "splitBasicBlock broken!"); 544 Br->setOperand(0, FirstNewBlock); 545 546 547 // Now that the function is correct, make it a little bit nicer. In 548 // particular, move the basic blocks inserted from the end of the function 549 // into the space made by splitting the source basic block. 550 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 551 FirstNewBlock, Caller->end()); 552 553 // Handle all of the return instructions that we just cloned in, and eliminate 554 // any users of the original call/invoke instruction. 555 const Type *RTy = CalledFunc->getReturnType(); 556 557 if (Returns.size() > 1) { 558 // The PHI node should go at the front of the new basic block to merge all 559 // possible incoming values. 560 PHINode *PHI = 0; 561 if (!TheCall->use_empty()) { 562 PHI = PHINode::Create(RTy, TheCall->getName(), 563 AfterCallBB->begin()); 564 // Anything that used the result of the function call should now use the 565 // PHI node as their operand. 566 TheCall->replaceAllUsesWith(PHI); 567 } 568 569 // Loop over all of the return instructions adding entries to the PHI node 570 // as appropriate. 571 if (PHI) { 572 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 573 ReturnInst *RI = Returns[i]; 574 assert(RI->getReturnValue()->getType() == PHI->getType() && 575 "Ret value not consistent in function!"); 576 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 577 } 578 579 // Now that we inserted the PHI, check to see if it has a single value 580 // (e.g. all the entries are the same or undef). If so, remove the PHI so 581 // it doesn't block other optimizations. 582 if (Value *V = PHI->hasConstantValue()) { 583 PHI->replaceAllUsesWith(V); 584 PHI->eraseFromParent(); 585 } 586 } 587 588 589 // Add a branch to the merge points and remove return instructions. 590 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 591 ReturnInst *RI = Returns[i]; 592 BranchInst::Create(AfterCallBB, RI); 593 RI->eraseFromParent(); 594 } 595 } else if (!Returns.empty()) { 596 // Otherwise, if there is exactly one return value, just replace anything 597 // using the return value of the call with the computed value. 598 if (!TheCall->use_empty()) { 599 if (TheCall == Returns[0]->getReturnValue()) 600 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 601 else 602 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 603 } 604 605 // Splice the code from the return block into the block that it will return 606 // to, which contains the code that was after the call. 607 BasicBlock *ReturnBB = Returns[0]->getParent(); 608 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 609 ReturnBB->getInstList()); 610 611 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 612 ReturnBB->replaceAllUsesWith(AfterCallBB); 613 614 // Delete the return instruction now and empty ReturnBB now. 615 Returns[0]->eraseFromParent(); 616 ReturnBB->eraseFromParent(); 617 } else if (!TheCall->use_empty()) { 618 // No returns, but something is using the return value of the call. Just 619 // nuke the result. 620 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 621 } 622 623 // Since we are now done with the Call/Invoke, we can delete it. 624 TheCall->eraseFromParent(); 625 626 // We should always be able to fold the entry block of the function into the 627 // single predecessor of the block... 628 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 629 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 630 631 // Splice the code entry block into calling block, right before the 632 // unconditional branch. 633 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 634 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 635 636 // Remove the unconditional branch. 637 OrigBB->getInstList().erase(Br); 638 639 // Now we can remove the CalleeEntry block, which is now empty. 640 Caller->getBasicBlockList().erase(CalleeEntry); 641 642 return true; 643} 644