LLVM API Documentation
00001 //===- InlineFunction.cpp - Code to perform function inlining -------------===// 00002 // 00003 // The LLVM Compiler Infrastructure 00004 // 00005 // This file was developed by the LLVM research group and is distributed under 00006 // the University of Illinois Open Source License. See LICENSE.TXT for details. 00007 // 00008 //===----------------------------------------------------------------------===// 00009 // 00010 // This file implements inlining of a function into a call site, resolving 00011 // parameters and the return value as appropriate. 00012 // 00013 //===----------------------------------------------------------------------===// 00014 00015 #include "llvm/Transforms/Utils/Cloning.h" 00016 #include "llvm/Constants.h" 00017 #include "llvm/DerivedTypes.h" 00018 #include "llvm/Module.h" 00019 #include "llvm/Instructions.h" 00020 #include "llvm/Intrinsics.h" 00021 #include "llvm/Analysis/CallGraph.h" 00022 #include "llvm/Support/CallSite.h" 00023 using namespace llvm; 00024 00025 bool llvm::InlineFunction(CallInst *CI, CallGraph *CG) { 00026 return InlineFunction(CallSite(CI), CG); 00027 } 00028 bool llvm::InlineFunction(InvokeInst *II, CallGraph *CG) { 00029 return InlineFunction(CallSite(II), CG); 00030 } 00031 00032 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 00033 /// in the body of the inlined function into invokes and turn unwind 00034 /// instructions into branches to the invoke unwind dest. 00035 /// 00036 /// II is the invoke instruction begin inlined. FirstNewBlock is the first 00037 /// block of the inlined code (the last block is the end of the function), 00038 /// and InlineCodeInfo is information about the code that got inlined. 00039 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 00040 ClonedCodeInfo &InlinedCodeInfo) { 00041 BasicBlock *InvokeDest = II->getUnwindDest(); 00042 std::vector<Value*> InvokeDestPHIValues; 00043 00044 // If there are PHI nodes in the unwind destination block, we need to 00045 // keep track of which values came into them from this invoke, then remove 00046 // the entry for this block. 00047 BasicBlock *InvokeBlock = II->getParent(); 00048 for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) { 00049 PHINode *PN = cast<PHINode>(I); 00050 // Save the value to use for this edge. 00051 InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock)); 00052 } 00053 00054 Function *Caller = FirstNewBlock->getParent(); 00055 00056 // The inlined code is currently at the end of the function, scan from the 00057 // start of the inlined code to its end, checking for stuff we need to 00058 // rewrite. 00059 if (InlinedCodeInfo.ContainsCalls || InlinedCodeInfo.ContainsUnwinds) { 00060 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 00061 BB != E; ++BB) { 00062 if (InlinedCodeInfo.ContainsCalls) { 00063 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ){ 00064 Instruction *I = BBI++; 00065 00066 // We only need to check for function calls: inlined invoke 00067 // instructions require no special handling. 00068 if (!isa<CallInst>(I)) continue; 00069 CallInst *CI = cast<CallInst>(I); 00070 00071 // If this is an intrinsic function call, don't convert it to an 00072 // invoke. 00073 if (CI->getCalledFunction() && 00074 CI->getCalledFunction()->getIntrinsicID()) 00075 continue; 00076 00077 // Convert this function call into an invoke instruction. 00078 // First, split the basic block. 00079 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 00080 00081 // Next, create the new invoke instruction, inserting it at the end 00082 // of the old basic block. 00083 InvokeInst *II = 00084 new InvokeInst(CI->getCalledValue(), Split, InvokeDest, 00085 std::vector<Value*>(CI->op_begin()+1, CI->op_end()), 00086 CI->getName(), BB->getTerminator()); 00087 II->setCallingConv(CI->getCallingConv()); 00088 00089 // Make sure that anything using the call now uses the invoke! 00090 CI->replaceAllUsesWith(II); 00091 00092 // Delete the unconditional branch inserted by splitBasicBlock 00093 BB->getInstList().pop_back(); 00094 Split->getInstList().pop_front(); // Delete the original call 00095 00096 // Update any PHI nodes in the exceptional block to indicate that 00097 // there is now a new entry in them. 00098 unsigned i = 0; 00099 for (BasicBlock::iterator I = InvokeDest->begin(); 00100 isa<PHINode>(I); ++I, ++i) { 00101 PHINode *PN = cast<PHINode>(I); 00102 PN->addIncoming(InvokeDestPHIValues[i], BB); 00103 } 00104 00105 // This basic block is now complete, start scanning the next one. 00106 break; 00107 } 00108 } 00109 00110 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 00111 // An UnwindInst requires special handling when it gets inlined into an 00112 // invoke site. Once this happens, we know that the unwind would cause 00113 // a control transfer to the invoke exception destination, so we can 00114 // transform it into a direct branch to the exception destination. 00115 new BranchInst(InvokeDest, UI); 00116 00117 // Delete the unwind instruction! 00118 UI->getParent()->getInstList().pop_back(); 00119 00120 // Update any PHI nodes in the exceptional block to indicate that 00121 // there is now a new entry in them. 00122 unsigned i = 0; 00123 for (BasicBlock::iterator I = InvokeDest->begin(); 00124 isa<PHINode>(I); ++I, ++i) { 00125 PHINode *PN = cast<PHINode>(I); 00126 PN->addIncoming(InvokeDestPHIValues[i], BB); 00127 } 00128 } 00129 } 00130 } 00131 00132 // Now that everything is happy, we have one final detail. The PHI nodes in 00133 // the exception destination block still have entries due to the original 00134 // invoke instruction. Eliminate these entries (which might even delete the 00135 // PHI node) now. 00136 InvokeDest->removePredecessor(II->getParent()); 00137 } 00138 00139 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 00140 /// into the caller, update the specified callgraph to reflect the changes we 00141 /// made. Note that it's possible that not all code was copied over, so only 00142 /// some edges of the callgraph will be remain. 00143 static void UpdateCallGraphAfterInlining(const Function *Caller, 00144 const Function *Callee, 00145 Function::iterator FirstNewBlock, 00146 std::map<const Value*, Value*> &ValueMap, 00147 CallGraph &CG) { 00148 // Update the call graph by deleting the edge from Callee to Caller 00149 CallGraphNode *CalleeNode = CG[Callee]; 00150 CallGraphNode *CallerNode = CG[Caller]; 00151 CallerNode->removeCallEdgeTo(CalleeNode); 00152 00153 // Since we inlined some uninlined call sites in the callee into the caller, 00154 // add edges from the caller to all of the callees of the callee. 00155 for (CallGraphNode::iterator I = CalleeNode->begin(), 00156 E = CalleeNode->end(); I != E; ++I) { 00157 const Instruction *OrigCall = I->first.getInstruction(); 00158 00159 std::map<const Value*, Value*>::iterator VMI = ValueMap.find(OrigCall); 00160 // Only copy the edge if the call was inlined! 00161 if (VMI != ValueMap.end() && VMI->second) { 00162 // If the call was inlined, but then constant folded, there is no edge to 00163 // add. Check for this case. 00164 if (Instruction *NewCall = dyn_cast<Instruction>(VMI->second)) 00165 CallerNode->addCalledFunction(CallSite::get(NewCall), I->second); 00166 } 00167 } 00168 } 00169 00170 00171 // InlineFunction - This function inlines the called function into the basic 00172 // block of the caller. This returns false if it is not possible to inline this 00173 // call. The program is still in a well defined state if this occurs though. 00174 // 00175 // Note that this only does one level of inlining. For example, if the 00176 // instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 00177 // exists in the instruction stream. Similiarly this will inline a recursive 00178 // function by one level. 00179 // 00180 bool llvm::InlineFunction(CallSite CS, CallGraph *CG) { 00181 Instruction *TheCall = CS.getInstruction(); 00182 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 00183 "Instruction not in function!"); 00184 00185 const Function *CalledFunc = CS.getCalledFunction(); 00186 if (CalledFunc == 0 || // Can't inline external function or indirect 00187 CalledFunc->isExternal() || // call, or call to a vararg function! 00188 CalledFunc->getFunctionType()->isVarArg()) return false; 00189 00190 00191 // If the call to the callee is a non-tail call, we must clear the 'tail' 00192 // flags on any calls that we inline. 00193 bool MustClearTailCallFlags = 00194 isa<CallInst>(TheCall) && !cast<CallInst>(TheCall)->isTailCall(); 00195 00196 BasicBlock *OrigBB = TheCall->getParent(); 00197 Function *Caller = OrigBB->getParent(); 00198 00199 // Get an iterator to the last basic block in the function, which will have 00200 // the new function inlined after it. 00201 // 00202 Function::iterator LastBlock = &Caller->back(); 00203 00204 // Make sure to capture all of the return instructions from the cloned 00205 // function. 00206 std::vector<ReturnInst*> Returns; 00207 ClonedCodeInfo InlinedFunctionInfo; 00208 Function::iterator FirstNewBlock; 00209 00210 { // Scope to destroy ValueMap after cloning. 00211 std::map<const Value*, Value*> ValueMap; 00212 00213 // Calculate the vector of arguments to pass into the function cloner, which 00214 // matches up the formal to the actual argument values. 00215 assert(std::distance(CalledFunc->arg_begin(), CalledFunc->arg_end()) == 00216 std::distance(CS.arg_begin(), CS.arg_end()) && 00217 "No varargs calls can be inlined!"); 00218 CallSite::arg_iterator AI = CS.arg_begin(); 00219 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 00220 E = CalledFunc->arg_end(); I != E; ++I, ++AI) 00221 ValueMap[I] = *AI; 00222 00223 // We want the inliner to prune the code as it copies. We would LOVE to 00224 // have no dead or constant instructions leftover after inlining occurs 00225 // (which can happen, e.g., because an argument was constant), but we'll be 00226 // happy with whatever the cloner can do. 00227 CloneAndPruneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i", 00228 &InlinedFunctionInfo); 00229 00230 // Remember the first block that is newly cloned over. 00231 FirstNewBlock = LastBlock; ++FirstNewBlock; 00232 00233 // Update the callgraph if requested. 00234 if (CG) 00235 UpdateCallGraphAfterInlining(Caller, CalledFunc, FirstNewBlock, ValueMap, 00236 *CG); 00237 } 00238 00239 // If there are any alloca instructions in the block that used to be the entry 00240 // block for the callee, move them to the entry block of the caller. First 00241 // calculate which instruction they should be inserted before. We insert the 00242 // instructions at the end of the current alloca list. 00243 // 00244 { 00245 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 00246 for (BasicBlock::iterator I = FirstNewBlock->begin(), 00247 E = FirstNewBlock->end(); I != E; ) 00248 if (AllocaInst *AI = dyn_cast<AllocaInst>(I++)) 00249 if (isa<Constant>(AI->getArraySize())) { 00250 // Scan for the block of allocas that we can move over, and move them 00251 // all at once. 00252 while (isa<AllocaInst>(I) && 00253 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) 00254 ++I; 00255 00256 // Transfer all of the allocas over in a block. Using splice means 00257 // that they instructions aren't removed from the symbol table, then 00258 // reinserted. 00259 Caller->front().getInstList().splice(InsertPoint, 00260 FirstNewBlock->getInstList(), 00261 AI, I); 00262 } 00263 } 00264 00265 // If the inlined code contained dynamic alloca instructions, wrap the inlined 00266 // code with llvm.stacksave/llvm.stackrestore intrinsics. 00267 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 00268 Module *M = Caller->getParent(); 00269 const Type *SBytePtr = PointerType::get(Type::SByteTy); 00270 // Get the two intrinsics we care about. 00271 Function *StackSave, *StackRestore; 00272 StackSave = M->getOrInsertFunction("llvm.stacksave", SBytePtr, NULL); 00273 StackRestore = M->getOrInsertFunction("llvm.stackrestore", Type::VoidTy, 00274 SBytePtr, NULL); 00275 00276 // If we are preserving the callgraph, add edges to the stacksave/restore 00277 // functions for the calls we insert. 00278 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0; 00279 if (CG) { 00280 StackSaveCGN = CG->getOrInsertFunction(StackSave); 00281 StackRestoreCGN = CG->getOrInsertFunction(StackRestore); 00282 CallerNode = (*CG)[Caller]; 00283 } 00284 00285 // Insert the llvm.stacksave. 00286 CallInst *SavedPtr = new CallInst(StackSave, "savedstack", 00287 FirstNewBlock->begin()); 00288 if (CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN); 00289 00290 // Insert a call to llvm.stackrestore before any return instructions in the 00291 // inlined function. 00292 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 00293 CallInst *CI = new CallInst(StackRestore, SavedPtr, "", Returns[i]); 00294 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 00295 } 00296 00297 // Count the number of StackRestore calls we insert. 00298 unsigned NumStackRestores = Returns.size(); 00299 00300 // If we are inlining an invoke instruction, insert restores before each 00301 // unwind. These unwinds will be rewritten into branches later. 00302 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 00303 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 00304 BB != E; ++BB) 00305 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 00306 new CallInst(StackRestore, SavedPtr, "", UI); 00307 ++NumStackRestores; 00308 } 00309 } 00310 } 00311 00312 // If we are inlining tail call instruction through a call site that isn't 00313 // marked 'tail', we must remove the tail marker for any calls in the inlined 00314 // code. 00315 if (MustClearTailCallFlags && InlinedFunctionInfo.ContainsCalls) { 00316 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 00317 BB != E; ++BB) 00318 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 00319 if (CallInst *CI = dyn_cast<CallInst>(I)) 00320 CI->setTailCall(false); 00321 } 00322 00323 // If we are inlining for an invoke instruction, we must make sure to rewrite 00324 // any inlined 'unwind' instructions into branches to the invoke exception 00325 // destination, and call instructions into invoke instructions. 00326 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 00327 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 00328 00329 // If we cloned in _exactly one_ basic block, and if that block ends in a 00330 // return instruction, we splice the body of the inlined callee directly into 00331 // the calling basic block. 00332 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 00333 // Move all of the instructions right before the call. 00334 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 00335 FirstNewBlock->begin(), FirstNewBlock->end()); 00336 // Remove the cloned basic block. 00337 Caller->getBasicBlockList().pop_back(); 00338 00339 // If the call site was an invoke instruction, add a branch to the normal 00340 // destination. 00341 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 00342 new BranchInst(II->getNormalDest(), TheCall); 00343 00344 // If the return instruction returned a value, replace uses of the call with 00345 // uses of the returned value. 00346 if (!TheCall->use_empty()) 00347 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 00348 00349 // Since we are now done with the Call/Invoke, we can delete it. 00350 TheCall->getParent()->getInstList().erase(TheCall); 00351 00352 // Since we are now done with the return instruction, delete it also. 00353 Returns[0]->getParent()->getInstList().erase(Returns[0]); 00354 00355 // We are now done with the inlining. 00356 return true; 00357 } 00358 00359 // Otherwise, we have the normal case, of more than one block to inline or 00360 // multiple return sites. 00361 00362 // We want to clone the entire callee function into the hole between the 00363 // "starter" and "ender" blocks. How we accomplish this depends on whether 00364 // this is an invoke instruction or a call instruction. 00365 BasicBlock *AfterCallBB; 00366 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 00367 00368 // Add an unconditional branch to make this look like the CallInst case... 00369 BranchInst *NewBr = new BranchInst(II->getNormalDest(), TheCall); 00370 00371 // Split the basic block. This guarantees that no PHI nodes will have to be 00372 // updated due to new incoming edges, and make the invoke case more 00373 // symmetric to the call case. 00374 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 00375 CalledFunc->getName()+".exit"); 00376 00377 } else { // It's a call 00378 // If this is a call instruction, we need to split the basic block that 00379 // the call lives in. 00380 // 00381 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 00382 CalledFunc->getName()+".exit"); 00383 } 00384 00385 // Change the branch that used to go to AfterCallBB to branch to the first 00386 // basic block of the inlined function. 00387 // 00388 TerminatorInst *Br = OrigBB->getTerminator(); 00389 assert(Br && Br->getOpcode() == Instruction::Br && 00390 "splitBasicBlock broken!"); 00391 Br->setOperand(0, FirstNewBlock); 00392 00393 00394 // Now that the function is correct, make it a little bit nicer. In 00395 // particular, move the basic blocks inserted from the end of the function 00396 // into the space made by splitting the source basic block. 00397 // 00398 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 00399 FirstNewBlock, Caller->end()); 00400 00401 // Handle all of the return instructions that we just cloned in, and eliminate 00402 // any users of the original call/invoke instruction. 00403 if (Returns.size() > 1) { 00404 // The PHI node should go at the front of the new basic block to merge all 00405 // possible incoming values. 00406 // 00407 PHINode *PHI = 0; 00408 if (!TheCall->use_empty()) { 00409 PHI = new PHINode(CalledFunc->getReturnType(), 00410 TheCall->getName(), AfterCallBB->begin()); 00411 00412 // Anything that used the result of the function call should now use the 00413 // PHI node as their operand. 00414 // 00415 TheCall->replaceAllUsesWith(PHI); 00416 } 00417 00418 // Loop over all of the return instructions, turning them into unconditional 00419 // branches to the merge point now, and adding entries to the PHI node as 00420 // appropriate. 00421 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 00422 ReturnInst *RI = Returns[i]; 00423 00424 if (PHI) { 00425 assert(RI->getReturnValue() && "Ret should have value!"); 00426 assert(RI->getReturnValue()->getType() == PHI->getType() && 00427 "Ret value not consistent in function!"); 00428 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 00429 } 00430 00431 // Add a branch to the merge point where the PHI node lives if it exists. 00432 new BranchInst(AfterCallBB, RI); 00433 00434 // Delete the return instruction now 00435 RI->getParent()->getInstList().erase(RI); 00436 } 00437 00438 } else if (!Returns.empty()) { 00439 // Otherwise, if there is exactly one return value, just replace anything 00440 // using the return value of the call with the computed value. 00441 if (!TheCall->use_empty()) 00442 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 00443 00444 // Splice the code from the return block into the block that it will return 00445 // to, which contains the code that was after the call. 00446 BasicBlock *ReturnBB = Returns[0]->getParent(); 00447 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 00448 ReturnBB->getInstList()); 00449 00450 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 00451 ReturnBB->replaceAllUsesWith(AfterCallBB); 00452 00453 // Delete the return instruction now and empty ReturnBB now. 00454 Returns[0]->eraseFromParent(); 00455 ReturnBB->eraseFromParent(); 00456 } else if (!TheCall->use_empty()) { 00457 // No returns, but something is using the return value of the call. Just 00458 // nuke the result. 00459 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 00460 } 00461 00462 // Since we are now done with the Call/Invoke, we can delete it. 00463 TheCall->eraseFromParent(); 00464 00465 // We should always be able to fold the entry block of the function into the 00466 // single predecessor of the block... 00467 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 00468 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 00469 00470 // Splice the code entry block into calling block, right before the 00471 // unconditional branch. 00472 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 00473 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 00474 00475 // Remove the unconditional branch. 00476 OrigBB->getInstList().erase(Br); 00477 00478 // Now we can remove the CalleeEntry block, which is now empty. 00479 Caller->getBasicBlockList().erase(CalleeEntry); 00480 00481 return true; 00482 }