InlineFunction.cpp revision e4d5c441e04bdc00ccf1804744af670655123b07
1//===- InlineFunction.cpp - Code to perform function inlining -------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source 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// FIXME: This pass should transform alloca instructions in the called function 14// into alloca/dealloca pairs! Or perhaps it should refuse to inline them! 15// 16//===----------------------------------------------------------------------===// 17 18#include "llvm/Transforms/Utils/Cloning.h" 19#include "llvm/Constants.h" 20#include "llvm/DerivedTypes.h" 21#include "llvm/Module.h" 22#include "llvm/Instructions.h" 23#include "llvm/Intrinsics.h" 24#include "llvm/Support/CallSite.h" 25using namespace llvm; 26 27bool llvm::InlineFunction(CallInst *CI) { return InlineFunction(CallSite(CI)); } 28bool llvm::InlineFunction(InvokeInst *II) {return InlineFunction(CallSite(II));} 29 30// InlineFunction - This function inlines the called function into the basic 31// block of the caller. This returns false if it is not possible to inline this 32// call. The program is still in a well defined state if this occurs though. 33// 34// Note that this only does one level of inlining. For example, if the 35// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 36// exists in the instruction stream. Similiarly this will inline a recursive 37// function by one level. 38// 39bool llvm::InlineFunction(CallSite CS) { 40 Instruction *TheCall = CS.getInstruction(); 41 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 42 "Instruction not in function!"); 43 44 const Function *CalledFunc = CS.getCalledFunction(); 45 if (CalledFunc == 0 || // Can't inline external function or indirect 46 CalledFunc->isExternal() || // call, or call to a vararg function! 47 CalledFunc->getFunctionType()->isVarArg()) return false; 48 49 BasicBlock *OrigBB = TheCall->getParent(); 50 Function *Caller = OrigBB->getParent(); 51 52 // Get an iterator to the last basic block in the function, which will have 53 // the new function inlined after it. 54 // 55 Function::iterator LastBlock = &Caller->back(); 56 57 // Make sure to capture all of the return instructions from the cloned 58 // function. 59 std::vector<ReturnInst*> Returns; 60 { // Scope to destroy ValueMap after cloning. 61 // Calculate the vector of arguments to pass into the function cloner... 62 std::map<const Value*, Value*> ValueMap; 63 assert(std::distance(CalledFunc->arg_begin(), CalledFunc->arg_end()) == 64 std::distance(CS.arg_begin(), CS.arg_end()) && 65 "No varargs calls can be inlined!"); 66 67 CallSite::arg_iterator AI = CS.arg_begin(); 68 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 69 E = CalledFunc->arg_end(); I != E; ++I, ++AI) 70 ValueMap[I] = *AI; 71 72 // Clone the entire body of the callee into the caller. 73 CloneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i"); 74 } 75 76 // Remember the first block that is newly cloned over. 77 Function::iterator FirstNewBlock = LastBlock; ++FirstNewBlock; 78 79 // If there are any alloca instructions in the block that used to be the entry 80 // block for the callee, move them to the entry block of the caller. First 81 // calculate which instruction they should be inserted before. We insert the 82 // instructions at the end of the current alloca list. 83 // 84 if (isa<AllocaInst>(FirstNewBlock->begin())) { 85 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 86 for (BasicBlock::iterator I = FirstNewBlock->begin(), 87 E = FirstNewBlock->end(); I != E; ) 88 if (AllocaInst *AI = dyn_cast<AllocaInst>(I++)) 89 if (isa<Constant>(AI->getArraySize())) { 90 // Scan for the block of allocas that we can move over. 91 while (isa<AllocaInst>(I) && 92 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) 93 ++I; 94 95 // Transfer all of the allocas over in a block. Using splice means 96 // that they instructions aren't removed from the symbol table, then 97 // reinserted. 98 Caller->front().getInstList().splice(InsertPoint, 99 FirstNewBlock->getInstList(), 100 AI, I); 101 } 102 } 103 104 // If we are inlining for an invoke instruction, we must make sure to rewrite 105 // any inlined 'unwind' instructions into branches to the invoke exception 106 // destination, and call instructions into invoke instructions. 107 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 108 BasicBlock *InvokeDest = II->getUnwindDest(); 109 std::vector<Value*> InvokeDestPHIValues; 110 111 // If there are PHI nodes in the exceptional 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 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(OrigBB)); 118 } 119 120 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 121 BB != E; ++BB) { 122 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) { 123 // We only need to check for function calls: inlined invoke instructions 124 // require no special handling... 125 if (CallInst *CI = dyn_cast<CallInst>(I)) { 126 // Convert this function call into an invoke instruction... if it's 127 // not an intrinsic function call (which are known to not throw). 128 if (CI->getCalledFunction() && 129 CI->getCalledFunction()->getIntrinsicID()) { 130 ++I; 131 } else { 132 // First, split the basic block... 133 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 134 135 // Next, create the new invoke instruction, inserting it at the end 136 // of the old basic block. 137 InvokeInst *II = 138 new InvokeInst(CI->getCalledValue(), Split, InvokeDest, 139 std::vector<Value*>(CI->op_begin()+1, CI->op_end()), 140 CI->getName(), BB->getTerminator()); 141 142 // Make sure that anything using the call now uses the invoke! 143 CI->replaceAllUsesWith(II); 144 145 // Delete the unconditional branch inserted by splitBasicBlock 146 BB->getInstList().pop_back(); 147 Split->getInstList().pop_front(); // Delete the original call 148 149 // Update any PHI nodes in the exceptional block to indicate that 150 // there is now a new entry in them. 151 unsigned i = 0; 152 for (BasicBlock::iterator I = InvokeDest->begin(); 153 isa<PHINode>(I); ++I, ++i) { 154 PHINode *PN = cast<PHINode>(I); 155 PN->addIncoming(InvokeDestPHIValues[i], BB); 156 } 157 158 // This basic block is now complete, start scanning the next one. 159 break; 160 } 161 } else { 162 ++I; 163 } 164 } 165 166 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 167 // An UnwindInst requires special handling when it gets inlined into an 168 // invoke site. Once this happens, we know that the unwind would cause 169 // a control transfer to the invoke exception destination, so we can 170 // transform it into a direct branch to the exception destination. 171 new BranchInst(InvokeDest, UI); 172 173 // Delete the unwind instruction! 174 UI->getParent()->getInstList().pop_back(); 175 176 // Update any PHI nodes in the exceptional block to indicate that 177 // there is now a new entry in them. 178 unsigned i = 0; 179 for (BasicBlock::iterator I = InvokeDest->begin(); 180 isa<PHINode>(I); ++I, ++i) { 181 PHINode *PN = cast<PHINode>(I); 182 PN->addIncoming(InvokeDestPHIValues[i], BB); 183 } 184 } 185 } 186 187 // Now that everything is happy, we have one final detail. The PHI nodes in 188 // the exception destination block still have entries due to the original 189 // invoke instruction. Eliminate these entries (which might even delete the 190 // PHI node) now. 191 InvokeDest->removePredecessor(II->getParent()); 192 } 193 194 // If we cloned in _exactly one_ basic block, and if that block ends in a 195 // return instruction, we splice the body of the inlined callee directly into 196 // the calling basic block. 197 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 198 // Move all of the instructions right before the call. 199 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 200 FirstNewBlock->begin(), FirstNewBlock->end()); 201 // Remove the cloned basic block. 202 Caller->getBasicBlockList().pop_back(); 203 204 // If the call site was an invoke instruction, add a branch to the normal 205 // destination. 206 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 207 new BranchInst(II->getNormalDest(), TheCall); 208 209 // If the return instruction returned a value, replace uses of the call with 210 // uses of the returned value. 211 if (!TheCall->use_empty()) 212 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 213 214 // Since we are now done with the Call/Invoke, we can delete it. 215 TheCall->getParent()->getInstList().erase(TheCall); 216 217 // Since we are now done with the return instruction, delete it also. 218 Returns[0]->getParent()->getInstList().erase(Returns[0]); 219 220 // We are now done with the inlining. 221 return true; 222 } 223 224 // Otherwise, we have the normal case, of more than one block to inline or 225 // multiple return sites. 226 227 // We want to clone the entire callee function into the hole between the 228 // "starter" and "ender" blocks. How we accomplish this depends on whether 229 // this is an invoke instruction or a call instruction. 230 BasicBlock *AfterCallBB; 231 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 232 233 // Add an unconditional branch to make this look like the CallInst case... 234 BranchInst *NewBr = new BranchInst(II->getNormalDest(), TheCall); 235 236 // Split the basic block. This guarantees that no PHI nodes will have to be 237 // updated due to new incoming edges, and make the invoke case more 238 // symmetric to the call case. 239 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 240 CalledFunc->getName()+".exit"); 241 242 } else { // It's a call 243 // If this is a call instruction, we need to split the basic block that 244 // the call lives in. 245 // 246 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 247 CalledFunc->getName()+".exit"); 248 } 249 250 // Change the branch that used to go to AfterCallBB to branch to the first 251 // basic block of the inlined function. 252 // 253 TerminatorInst *Br = OrigBB->getTerminator(); 254 assert(Br && Br->getOpcode() == Instruction::Br && 255 "splitBasicBlock broken!"); 256 Br->setOperand(0, FirstNewBlock); 257 258 259 // Now that the function is correct, make it a little bit nicer. In 260 // particular, move the basic blocks inserted from the end of the function 261 // into the space made by splitting the source basic block. 262 // 263 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 264 FirstNewBlock, Caller->end()); 265 266 // Handle all of the return instructions that we just cloned in, and eliminate 267 // any users of the original call/invoke instruction. 268 if (Returns.size() > 1) { 269 // The PHI node should go at the front of the new basic block to merge all 270 // possible incoming values. 271 // 272 PHINode *PHI = 0; 273 if (!TheCall->use_empty()) { 274 PHI = new PHINode(CalledFunc->getReturnType(), 275 TheCall->getName(), AfterCallBB->begin()); 276 277 // Anything that used the result of the function call should now use the 278 // PHI node as their operand. 279 // 280 TheCall->replaceAllUsesWith(PHI); 281 } 282 283 // Loop over all of the return instructions, turning them into unconditional 284 // branches to the merge point now, and adding entries to the PHI node as 285 // appropriate. 286 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 287 ReturnInst *RI = Returns[i]; 288 289 if (PHI) { 290 assert(RI->getReturnValue() && "Ret should have value!"); 291 assert(RI->getReturnValue()->getType() == PHI->getType() && 292 "Ret value not consistent in function!"); 293 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 294 } 295 296 // Add a branch to the merge point where the PHI node lives if it exists. 297 new BranchInst(AfterCallBB, RI); 298 299 // Delete the return instruction now 300 RI->getParent()->getInstList().erase(RI); 301 } 302 303 } else if (!Returns.empty()) { 304 // Otherwise, if there is exactly one return value, just replace anything 305 // using the return value of the call with the computed value. 306 if (!TheCall->use_empty()) 307 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 308 309 // Splice the code from the return block into the block that it will return 310 // to, which contains the code that was after the call. 311 BasicBlock *ReturnBB = Returns[0]->getParent(); 312 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 313 ReturnBB->getInstList()); 314 315 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 316 ReturnBB->replaceAllUsesWith(AfterCallBB); 317 318 // Delete the return instruction now and empty ReturnBB now. 319 Returns[0]->eraseFromParent(); 320 ReturnBB->eraseFromParent(); 321 } else if (!TheCall->use_empty()) { 322 // No returns, but something is using the return value of the call. Just 323 // nuke the result. 324 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 325 } 326 327 // Since we are now done with the Call/Invoke, we can delete it. 328 TheCall->eraseFromParent(); 329 330 // We should always be able to fold the entry block of the function into the 331 // single predecessor of the block... 332 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 333 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 334 335 // Splice the code entry block into calling block, right before the 336 // unconditional branch. 337 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 338 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 339 340 // Remove the unconditional branch. 341 OrigBB->getInstList().erase(Br); 342 343 // Now we can remove the CalleeEntry block, which is now empty. 344 Caller->getBasicBlockList().erase(CalleeEntry); 345 return true; 346} 347