InlineFunction.cpp revision baf3c404409d5e47b13984a7f95bfbd6d1f2e79e
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 return InlineFunction(CallSite(CI), CG, TD); 34} 35bool llvm::InlineFunction(InvokeInst *II, CallGraph *CG, const TargetData *TD) { 36 return InlineFunction(CallSite(II), CG, TD); 37} 38 39/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 40/// in the body of the inlined function into invokes and turn unwind 41/// instructions into branches to the invoke unwind dest. 42/// 43/// II is the invoke instruction being inlined. FirstNewBlock is the first 44/// block of the inlined code (the last block is the end of the function), 45/// and InlineCodeInfo is information about the code that got inlined. 46static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 47 ClonedCodeInfo &InlinedCodeInfo, 48 CallGraph *CG) { 49 BasicBlock *InvokeDest = II->getUnwindDest(); 50 std::vector<Value*> InvokeDestPHIValues; 51 52 // If there are PHI nodes in the unwind destination block, we need to 53 // keep track of which values came into them from this invoke, then remove 54 // the entry for this block. 55 BasicBlock *InvokeBlock = II->getParent(); 56 for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) { 57 PHINode *PN = cast<PHINode>(I); 58 // Save the value to use for this edge. 59 InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock)); 60 } 61 62 Function *Caller = FirstNewBlock->getParent(); 63 64 // The inlined code is currently at the end of the function, scan from the 65 // start of the inlined code to its end, checking for stuff we need to 66 // rewrite. 67 if (InlinedCodeInfo.ContainsCalls || InlinedCodeInfo.ContainsUnwinds) { 68 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 69 BB != E; ++BB) { 70 if (InlinedCodeInfo.ContainsCalls) { 71 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ){ 72 Instruction *I = BBI++; 73 74 // We only need to check for function calls: inlined invoke 75 // instructions require no special handling. 76 if (!isa<CallInst>(I)) continue; 77 CallInst *CI = cast<CallInst>(I); 78 79 // If this call cannot unwind, don't convert it to an invoke. 80 if (CI->doesNotThrow()) 81 continue; 82 83 // Convert this function call into an invoke instruction. 84 // First, split the basic block. 85 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 86 87 // Next, create the new invoke instruction, inserting it at the end 88 // of the old basic block. 89 SmallVector<Value*, 8> InvokeArgs(CI->op_begin()+1, CI->op_end()); 90 InvokeInst *II = 91 InvokeInst::Create(CI->getCalledValue(), Split, InvokeDest, 92 InvokeArgs.begin(), InvokeArgs.end(), 93 CI->getName(), BB->getTerminator()); 94 II->setCallingConv(CI->getCallingConv()); 95 II->setAttributes(CI->getAttributes()); 96 97 // Make sure that anything using the call now uses the invoke! 98 CI->replaceAllUsesWith(II); 99 100 // Update the callgraph. 101 if (CG) { 102 // We should be able to do this: 103 // (*CG)[Caller]->replaceCallSite(CI, II); 104 // but that fails if the old call site isn't in the call graph, 105 // which, because of LLVM bug 3601, it sometimes isn't. 106 CallGraphNode *CGN = (*CG)[Caller]; 107 for (CallGraphNode::iterator NI = CGN->begin(), NE = CGN->end(); 108 NI != NE; ++NI) { 109 if (NI->first == CI) { 110 NI->first = II; 111 break; 112 } 113 } 114 } 115 116 // Delete the unconditional branch inserted by splitBasicBlock 117 BB->getInstList().pop_back(); 118 Split->getInstList().pop_front(); // Delete the original call 119 120 // Update any PHI nodes in the exceptional block to indicate that 121 // there is now a new entry in them. 122 unsigned i = 0; 123 for (BasicBlock::iterator I = InvokeDest->begin(); 124 isa<PHINode>(I); ++I, ++i) { 125 PHINode *PN = cast<PHINode>(I); 126 PN->addIncoming(InvokeDestPHIValues[i], BB); 127 } 128 129 // This basic block is now complete, start scanning the next one. 130 break; 131 } 132 } 133 134 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 135 // An UnwindInst requires special handling when it gets inlined into an 136 // invoke site. Once this happens, we know that the unwind would cause 137 // a control transfer to the invoke exception destination, so we can 138 // transform it into a direct branch to the exception destination. 139 BranchInst::Create(InvokeDest, UI); 140 141 // Delete the unwind instruction! 142 UI->eraseFromParent(); 143 144 // Update any PHI nodes in the exceptional block to indicate that 145 // there is now a new entry in them. 146 unsigned i = 0; 147 for (BasicBlock::iterator I = InvokeDest->begin(); 148 isa<PHINode>(I); ++I, ++i) { 149 PHINode *PN = cast<PHINode>(I); 150 PN->addIncoming(InvokeDestPHIValues[i], BB); 151 } 152 } 153 } 154 } 155 156 // Now that everything is happy, we have one final detail. The PHI nodes in 157 // the exception destination block still have entries due to the original 158 // invoke instruction. Eliminate these entries (which might even delete the 159 // PHI node) now. 160 InvokeDest->removePredecessor(II->getParent()); 161} 162 163/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 164/// into the caller, update the specified callgraph to reflect the changes we 165/// made. Note that it's possible that not all code was copied over, so only 166/// some edges of the callgraph may remain. 167static void UpdateCallGraphAfterInlining(CallSite CS, 168 Function::iterator FirstNewBlock, 169 DenseMap<const Value*, Value*> &ValueMap, 170 CallGraph &CG) { 171 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 172 const Function *Callee = CS.getCalledFunction(); 173 CallGraphNode *CalleeNode = CG[Callee]; 174 CallGraphNode *CallerNode = CG[Caller]; 175 176 // Since we inlined some uninlined call sites in the callee into the caller, 177 // add edges from the caller to all of the callees of the callee. 178 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 179 180 // Consider the case where CalleeNode == CallerNode. 181 CallGraphNode::CalledFunctionsVector CallCache; 182 if (CalleeNode == CallerNode) { 183 CallCache.assign(I, E); 184 I = CallCache.begin(); 185 E = CallCache.end(); 186 } 187 188 for (; I != E; ++I) { 189 const Instruction *OrigCall = I->first.getInstruction(); 190 191 DenseMap<const Value*, Value*>::iterator VMI = ValueMap.find(OrigCall); 192 // Only copy the edge if the call was inlined! 193 if (VMI != ValueMap.end() && VMI->second) { 194 // If the call was inlined, but then constant folded, there is no edge to 195 // add. Check for this case. 196 if (Instruction *NewCall = dyn_cast<Instruction>(VMI->second)) 197 CallerNode->addCalledFunction(CallSite::get(NewCall), I->second); 198 } 199 } 200 // Update the call graph by deleting the edge from Callee to Caller. We must 201 // do this after the loop above in case Caller and Callee are the same. 202 CallerNode->removeCallEdgeFor(CS); 203} 204 205/// findFnRegionEndMarker - This is a utility routine that is used by 206/// InlineFunction. Return llvm.dbg.region.end intrinsic that corresponds 207/// to the llvm.dbg.func.start of the function F. Otherwise return NULL. 208static const DbgRegionEndInst *findFnRegionEndMarker(const Function *F) { 209 210 GlobalVariable *FnStart = NULL; 211 const DbgRegionEndInst *FnEnd = NULL; 212 for (Function::const_iterator FI = F->begin(), FE =F->end(); FI != FE; ++FI) 213 for (BasicBlock::const_iterator BI = FI->begin(), BE = FI->end(); BI != BE; 214 ++BI) { 215 if (FnStart == NULL) { 216 if (const DbgFuncStartInst *FSI = dyn_cast<DbgFuncStartInst>(BI)) { 217 DISubprogram SP(cast<GlobalVariable>(FSI->getSubprogram())); 218 assert (SP.isNull() == false && "Invalid llvm.dbg.func.start"); 219 if (SP.describes(F)) 220 FnStart = SP.getGV(); 221 } 222 } else { 223 if (const DbgRegionEndInst *REI = dyn_cast<DbgRegionEndInst>(BI)) 224 if (REI->getContext() == FnStart) 225 FnEnd = REI; 226 } 227 } 228 return FnEnd; 229} 230 231// InlineFunction - This function inlines the called function into the basic 232// block of the caller. This returns false if it is not possible to inline this 233// call. The program is still in a well defined state if this occurs though. 234// 235// Note that this only does one level of inlining. For example, if the 236// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 237// exists in the instruction stream. Similiarly this will inline a recursive 238// function by one level. 239// 240bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { 241 Instruction *TheCall = CS.getInstruction(); 242 LLVMContext &Context = TheCall->getContext(); 243 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 244 "Instruction not in function!"); 245 246 const Function *CalledFunc = CS.getCalledFunction(); 247 if (CalledFunc == 0 || // Can't inline external function or indirect 248 CalledFunc->isDeclaration() || // call, or call to a vararg function! 249 CalledFunc->getFunctionType()->isVarArg()) return false; 250 251 252 // If the call to the callee is not a tail call, we must clear the 'tail' 253 // flags on any calls that we inline. 254 bool MustClearTailCallFlags = 255 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall()); 256 257 // If the call to the callee cannot throw, set the 'nounwind' flag on any 258 // calls that we inline. 259 bool MarkNoUnwind = CS.doesNotThrow(); 260 261 BasicBlock *OrigBB = TheCall->getParent(); 262 Function *Caller = OrigBB->getParent(); 263 264 // GC poses two hazards to inlining, which only occur when the callee has GC: 265 // 1. If the caller has no GC, then the callee's GC must be propagated to the 266 // caller. 267 // 2. If the caller has a differing GC, it is invalid to inline. 268 if (CalledFunc->hasGC()) { 269 if (!Caller->hasGC()) 270 Caller->setGC(CalledFunc->getGC()); 271 else if (CalledFunc->getGC() != Caller->getGC()) 272 return false; 273 } 274 275 // Get an iterator to the last basic block in the function, which will have 276 // the new function inlined after it. 277 // 278 Function::iterator LastBlock = &Caller->back(); 279 280 // Make sure to capture all of the return instructions from the cloned 281 // function. 282 std::vector<ReturnInst*> Returns; 283 ClonedCodeInfo InlinedFunctionInfo; 284 Function::iterator FirstNewBlock; 285 286 { // Scope to destroy ValueMap after cloning. 287 DenseMap<const Value*, Value*> ValueMap; 288 289 assert(CalledFunc->arg_size() == CS.arg_size() && 290 "No varargs calls can be inlined!"); 291 292 // Calculate the vector of arguments to pass into the function cloner, which 293 // matches up the formal to the actual argument values. 294 CallSite::arg_iterator AI = CS.arg_begin(); 295 unsigned ArgNo = 0; 296 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 297 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 298 Value *ActualArg = *AI; 299 300 // When byval arguments actually inlined, we need to make the copy implied 301 // by them explicit. However, we don't do this if the callee is readonly 302 // or readnone, because the copy would be unneeded: the callee doesn't 303 // modify the struct. 304 if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal) && 305 !CalledFunc->onlyReadsMemory()) { 306 const Type *AggTy = cast<PointerType>(I->getType())->getElementType(); 307 const Type *VoidPtrTy = Context.getPointerTypeUnqual(Type::Int8Ty); 308 309 // Create the alloca. If we have TargetData, use nice alignment. 310 unsigned Align = 1; 311 if (TD) Align = TD->getPrefTypeAlignment(AggTy); 312 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, 313 I->getName(), 314 &*Caller->begin()->begin()); 315 // Emit a memcpy. 316 const Type *Tys[] = { Type::Int64Ty }; 317 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 318 Intrinsic::memcpy, 319 Tys, 1); 320 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 321 Value *SrcCast = new BitCastInst(*AI, VoidPtrTy, "tmp", TheCall); 322 323 Value *Size; 324 if (TD == 0) 325 Size = ConstantExpr::getSizeOf(AggTy); 326 else 327 Size = ConstantInt::get(Type::Int64Ty, 328 TD->getTypeStoreSize(AggTy)); 329 330 // Always generate a memcpy of alignment 1 here because we don't know 331 // the alignment of the src pointer. Other optimizations can infer 332 // better alignment. 333 Value *CallArgs[] = { 334 DestCast, SrcCast, Size, ConstantInt::get(Type::Int32Ty, 1) 335 }; 336 CallInst *TheMemCpy = 337 CallInst::Create(MemCpyFn, CallArgs, CallArgs+4, "", TheCall); 338 339 // If we have a call graph, update it. 340 if (CG) { 341 CallGraphNode *MemCpyCGN = CG->getOrInsertFunction(MemCpyFn); 342 CallGraphNode *CallerNode = (*CG)[Caller]; 343 CallerNode->addCalledFunction(TheMemCpy, MemCpyCGN); 344 } 345 346 // Uses of the argument in the function should use our new alloca 347 // instead. 348 ActualArg = NewAlloca; 349 } 350 351 ValueMap[I] = ActualArg; 352 } 353 354 // Adjust llvm.dbg.region.end. If the CalledFunc has region end 355 // marker then clone that marker after next stop point at the 356 // call site. The function body cloner does not clone original 357 // region end marker from the CalledFunc. This will ensure that 358 // inlined function's scope ends at the right place. 359 const DbgRegionEndInst *DREI = findFnRegionEndMarker(CalledFunc); 360 if (DREI) { 361 for (BasicBlock::iterator BI = TheCall, 362 BE = TheCall->getParent()->end(); BI != BE; ++BI) { 363 if (DbgStopPointInst *DSPI = dyn_cast<DbgStopPointInst>(BI)) { 364 if (DbgRegionEndInst *NewDREI = 365 dyn_cast<DbgRegionEndInst>(DREI->clone(Context))) 366 NewDREI->insertAfter(DSPI); 367 break; 368 } 369 } 370 } 371 372 // We want the inliner to prune the code as it copies. We would LOVE to 373 // have no dead or constant instructions leftover after inlining occurs 374 // (which can happen, e.g., because an argument was constant), but we'll be 375 // happy with whatever the cloner can do. 376 CloneAndPruneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i", 377 &InlinedFunctionInfo, TD); 378 379 // Remember the first block that is newly cloned over. 380 FirstNewBlock = LastBlock; ++FirstNewBlock; 381 382 // Update the callgraph if requested. 383 if (CG) 384 UpdateCallGraphAfterInlining(CS, FirstNewBlock, ValueMap, *CG); 385 } 386 387 // If there are any alloca instructions in the block that used to be the entry 388 // block for the callee, move them to the entry block of the caller. First 389 // calculate which instruction they should be inserted before. We insert the 390 // instructions at the end of the current alloca list. 391 // 392 { 393 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 394 for (BasicBlock::iterator I = FirstNewBlock->begin(), 395 E = FirstNewBlock->end(); I != E; ) 396 if (AllocaInst *AI = dyn_cast<AllocaInst>(I++)) { 397 // If the alloca is now dead, remove it. This often occurs due to code 398 // specialization. 399 if (AI->use_empty()) { 400 AI->eraseFromParent(); 401 continue; 402 } 403 404 if (isa<Constant>(AI->getArraySize())) { 405 // Scan for the block of allocas that we can move over, and move them 406 // all at once. 407 while (isa<AllocaInst>(I) && 408 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) 409 ++I; 410 411 // Transfer all of the allocas over in a block. Using splice means 412 // that the instructions aren't removed from the symbol table, then 413 // reinserted. 414 Caller->getEntryBlock().getInstList().splice( 415 InsertPoint, 416 FirstNewBlock->getInstList(), 417 AI, I); 418 } 419 } 420 } 421 422 // If the inlined code contained dynamic alloca instructions, wrap the inlined 423 // code with llvm.stacksave/llvm.stackrestore intrinsics. 424 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 425 Module *M = Caller->getParent(); 426 // Get the two intrinsics we care about. 427 Constant *StackSave, *StackRestore; 428 StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 429 StackRestore = Intrinsic::getDeclaration(M, Intrinsic::stackrestore); 430 431 // If we are preserving the callgraph, add edges to the stacksave/restore 432 // functions for the calls we insert. 433 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0; 434 if (CG) { 435 // We know that StackSave/StackRestore are Function*'s, because they are 436 // intrinsics which must have the right types. 437 StackSaveCGN = CG->getOrInsertFunction(cast<Function>(StackSave)); 438 StackRestoreCGN = CG->getOrInsertFunction(cast<Function>(StackRestore)); 439 CallerNode = (*CG)[Caller]; 440 } 441 442 // Insert the llvm.stacksave. 443 CallInst *SavedPtr = CallInst::Create(StackSave, "savedstack", 444 FirstNewBlock->begin()); 445 if (CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN); 446 447 // Insert a call to llvm.stackrestore before any return instructions in the 448 // inlined function. 449 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 450 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", Returns[i]); 451 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 452 } 453 454 // Count the number of StackRestore calls we insert. 455 unsigned NumStackRestores = Returns.size(); 456 457 // If we are inlining an invoke instruction, insert restores before each 458 // unwind. These unwinds will be rewritten into branches later. 459 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 460 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 461 BB != E; ++BB) 462 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 463 CallInst::Create(StackRestore, SavedPtr, "", UI); 464 ++NumStackRestores; 465 } 466 } 467 } 468 469 // If we are inlining tail call instruction through a call site that isn't 470 // marked 'tail', we must remove the tail marker for any calls in the inlined 471 // code. Also, calls inlined through a 'nounwind' call site should be marked 472 // 'nounwind'. 473 if (InlinedFunctionInfo.ContainsCalls && 474 (MustClearTailCallFlags || MarkNoUnwind)) { 475 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 476 BB != E; ++BB) 477 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 478 if (CallInst *CI = dyn_cast<CallInst>(I)) { 479 if (MustClearTailCallFlags) 480 CI->setTailCall(false); 481 if (MarkNoUnwind) 482 CI->setDoesNotThrow(); 483 } 484 } 485 486 // If we are inlining through a 'nounwind' call site then any inlined 'unwind' 487 // instructions are unreachable. 488 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind) 489 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 490 BB != E; ++BB) { 491 TerminatorInst *Term = BB->getTerminator(); 492 if (isa<UnwindInst>(Term)) { 493 new UnreachableInst(Term); 494 BB->getInstList().erase(Term); 495 } 496 } 497 498 // If we are inlining for an invoke instruction, we must make sure to rewrite 499 // any inlined 'unwind' instructions into branches to the invoke exception 500 // destination, and call instructions into invoke instructions. 501 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 502 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo, CG); 503 504 // If we cloned in _exactly one_ basic block, and if that block ends in a 505 // return instruction, we splice the body of the inlined callee directly into 506 // the calling basic block. 507 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 508 // Move all of the instructions right before the call. 509 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 510 FirstNewBlock->begin(), FirstNewBlock->end()); 511 // Remove the cloned basic block. 512 Caller->getBasicBlockList().pop_back(); 513 514 // If the call site was an invoke instruction, add a branch to the normal 515 // destination. 516 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 517 BranchInst::Create(II->getNormalDest(), TheCall); 518 519 // If the return instruction returned a value, replace uses of the call with 520 // uses of the returned value. 521 if (!TheCall->use_empty()) { 522 ReturnInst *R = Returns[0]; 523 if (TheCall == R->getReturnValue()) 524 TheCall->replaceAllUsesWith(Context.getUndef(TheCall->getType())); 525 else 526 TheCall->replaceAllUsesWith(R->getReturnValue()); 527 } 528 // Since we are now done with the Call/Invoke, we can delete it. 529 TheCall->eraseFromParent(); 530 531 // Since we are now done with the return instruction, delete it also. 532 Returns[0]->eraseFromParent(); 533 534 // We are now done with the inlining. 535 return true; 536 } 537 538 // Otherwise, we have the normal case, of more than one block to inline or 539 // multiple return sites. 540 541 // We want to clone the entire callee function into the hole between the 542 // "starter" and "ender" blocks. How we accomplish this depends on whether 543 // this is an invoke instruction or a call instruction. 544 BasicBlock *AfterCallBB; 545 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 546 547 // Add an unconditional branch to make this look like the CallInst case... 548 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 549 550 // Split the basic block. This guarantees that no PHI nodes will have to be 551 // updated due to new incoming edges, and make the invoke case more 552 // symmetric to the call case. 553 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 554 CalledFunc->getName()+".exit"); 555 556 } else { // It's a call 557 // If this is a call instruction, we need to split the basic block that 558 // the call lives in. 559 // 560 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 561 CalledFunc->getName()+".exit"); 562 } 563 564 // Change the branch that used to go to AfterCallBB to branch to the first 565 // basic block of the inlined function. 566 // 567 TerminatorInst *Br = OrigBB->getTerminator(); 568 assert(Br && Br->getOpcode() == Instruction::Br && 569 "splitBasicBlock broken!"); 570 Br->setOperand(0, FirstNewBlock); 571 572 573 // Now that the function is correct, make it a little bit nicer. In 574 // particular, move the basic blocks inserted from the end of the function 575 // into the space made by splitting the source basic block. 576 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 577 FirstNewBlock, Caller->end()); 578 579 // Handle all of the return instructions that we just cloned in, and eliminate 580 // any users of the original call/invoke instruction. 581 const Type *RTy = CalledFunc->getReturnType(); 582 583 if (Returns.size() > 1) { 584 // The PHI node should go at the front of the new basic block to merge all 585 // possible incoming values. 586 PHINode *PHI = 0; 587 if (!TheCall->use_empty()) { 588 PHI = PHINode::Create(RTy, TheCall->getName(), 589 AfterCallBB->begin()); 590 // Anything that used the result of the function call should now use the 591 // PHI node as their operand. 592 TheCall->replaceAllUsesWith(PHI); 593 } 594 595 // Loop over all of the return instructions adding entries to the PHI node 596 // as appropriate. 597 if (PHI) { 598 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 599 ReturnInst *RI = Returns[i]; 600 assert(RI->getReturnValue()->getType() == PHI->getType() && 601 "Ret value not consistent in function!"); 602 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 603 } 604 } 605 606 // Add a branch to the merge points and remove return instructions. 607 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 608 ReturnInst *RI = Returns[i]; 609 BranchInst::Create(AfterCallBB, RI); 610 RI->eraseFromParent(); 611 } 612 } else if (!Returns.empty()) { 613 // Otherwise, if there is exactly one return value, just replace anything 614 // using the return value of the call with the computed value. 615 if (!TheCall->use_empty()) { 616 if (TheCall == Returns[0]->getReturnValue()) 617 TheCall->replaceAllUsesWith(Context.getUndef(TheCall->getType())); 618 else 619 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 620 } 621 622 // Splice the code from the return block into the block that it will return 623 // to, which contains the code that was after the call. 624 BasicBlock *ReturnBB = Returns[0]->getParent(); 625 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 626 ReturnBB->getInstList()); 627 628 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 629 ReturnBB->replaceAllUsesWith(AfterCallBB); 630 631 // Delete the return instruction now and empty ReturnBB now. 632 Returns[0]->eraseFromParent(); 633 ReturnBB->eraseFromParent(); 634 } else if (!TheCall->use_empty()) { 635 // No returns, but something is using the return value of the call. Just 636 // nuke the result. 637 TheCall->replaceAllUsesWith(Context.getUndef(TheCall->getType())); 638 } 639 640 // Since we are now done with the Call/Invoke, we can delete it. 641 TheCall->eraseFromParent(); 642 643 // We should always be able to fold the entry block of the function into the 644 // single predecessor of the block... 645 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 646 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 647 648 // Splice the code entry block into calling block, right before the 649 // unconditional branch. 650 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 651 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 652 653 // Remove the unconditional branch. 654 OrigBB->getInstList().erase(Br); 655 656 // Now we can remove the CalleeEntry block, which is now empty. 657 Caller->getBasicBlockList().erase(CalleeEntry); 658 659 return true; 660} 661