InlineFunction.cpp revision 981418bf1562d0b5b470ddc7d0034c9f3297b893
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//===----------------------------------------------------------------------===// 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/Intrinsics.h" 21#include "llvm/Analysis/CallGraph.h" 22#include "llvm/Support/CallSite.h" 23using namespace llvm; 24 25bool llvm::InlineFunction(CallInst *CI, CallGraph *CG) { 26 return InlineFunction(CallSite(CI), CG); 27} 28bool llvm::InlineFunction(InvokeInst *II, CallGraph *CG) { 29 return InlineFunction(CallSite(II), CG); 30} 31 32/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 33/// in the body of the inlined function into invokes and turn unwind 34/// instructions into branches to the invoke unwind dest. 35/// 36/// II is the invoke instruction begin inlined. FirstNewBlock is the first 37/// block of the inlined code (the last block is the end of the function), 38/// and InlineCodeInfo is information about the code that got inlined. 39static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 40 ClonedCodeInfo &InlinedCodeInfo) { 41 BasicBlock *InvokeDest = II->getUnwindDest(); 42 std::vector<Value*> InvokeDestPHIValues; 43 44 // If there are PHI nodes in the unwind destination block, we need to 45 // keep track of which values came into them from this invoke, then remove 46 // the entry for this block. 47 BasicBlock *InvokeBlock = II->getParent(); 48 for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) { 49 PHINode *PN = cast<PHINode>(I); 50 // Save the value to use for this edge. 51 InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock)); 52 } 53 54 Function *Caller = FirstNewBlock->getParent(); 55 56 // The inlined code is currently at the end of the function, scan from the 57 // start of the inlined code to its end, checking for stuff we need to 58 // rewrite. 59 if (InlinedCodeInfo.ContainsCalls || InlinedCodeInfo.ContainsUnwinds) { 60 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 61 BB != E; ++BB) { 62 if (InlinedCodeInfo.ContainsCalls) { 63 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ){ 64 Instruction *I = BBI++; 65 66 // We only need to check for function calls: inlined invoke 67 // instructions require no special handling. 68 if (!isa<CallInst>(I)) continue; 69 CallInst *CI = cast<CallInst>(I); 70 71 // If this is an intrinsic function call, don't convert it to an 72 // invoke. 73 if (CI->getCalledFunction() && 74 CI->getCalledFunction()->getIntrinsicID()) 75 continue; 76 77 // Convert this function call into an invoke instruction. 78 // First, split the basic block. 79 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 80 81 // Next, create the new invoke instruction, inserting it at the end 82 // of the old basic block. 83 InvokeInst *II = 84 new InvokeInst(CI->getCalledValue(), Split, InvokeDest, 85 std::vector<Value*>(CI->op_begin()+1, CI->op_end()), 86 CI->getName(), BB->getTerminator()); 87 II->setCallingConv(CI->getCallingConv()); 88 89 // Make sure that anything using the call now uses the invoke! 90 CI->replaceAllUsesWith(II); 91 92 // Delete the unconditional branch inserted by splitBasicBlock 93 BB->getInstList().pop_back(); 94 Split->getInstList().pop_front(); // Delete the original call 95 96 // Update any PHI nodes in the exceptional block to indicate that 97 // there is now a new entry in them. 98 unsigned i = 0; 99 for (BasicBlock::iterator I = InvokeDest->begin(); 100 isa<PHINode>(I); ++I, ++i) { 101 PHINode *PN = cast<PHINode>(I); 102 PN->addIncoming(InvokeDestPHIValues[i], BB); 103 } 104 105 // This basic block is now complete, start scanning the next one. 106 break; 107 } 108 } 109 110 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 111 // An UnwindInst requires special handling when it gets inlined into an 112 // invoke site. Once this happens, we know that the unwind would cause 113 // a control transfer to the invoke exception destination, so we can 114 // transform it into a direct branch to the exception destination. 115 new BranchInst(InvokeDest, UI); 116 117 // Delete the unwind instruction! 118 UI->getParent()->getInstList().pop_back(); 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 } 130 } 131 132 // Now that everything is happy, we have one final detail. The PHI nodes in 133 // the exception destination block still have entries due to the original 134 // invoke instruction. Eliminate these entries (which might even delete the 135 // PHI node) now. 136 InvokeDest->removePredecessor(II->getParent()); 137} 138 139/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 140/// into the caller, update the specified callgraph to reflect the changes we 141/// made. Note that it's possible that not all code was copied over, so only 142/// some edges of the callgraph will be remain. 143static void UpdateCallGraphAfterInlining(const Function *Caller, 144 const Function *Callee, 145 Function::iterator FirstNewBlock, 146 std::map<const Value*, Value*> &ValueMap, 147 CallGraph &CG) { 148 // Update the call graph by deleting the edge from Callee to Caller 149 CallGraphNode *CalleeNode = CG[Callee]; 150 CallGraphNode *CallerNode = CG[Caller]; 151 CallerNode->removeCallEdgeTo(CalleeNode); 152 153 // Since we inlined some uninlined call sites in the callee into the caller, 154 // add edges from the caller to all of the callees of the callee. 155 for (CallGraphNode::iterator I = CalleeNode->begin(), 156 E = CalleeNode->end(); I != E; ++I) { 157 const Instruction *OrigCall = I->first.getInstruction(); 158 159 std::map<const Value*, Value*>::iterator VMI = ValueMap.find(OrigCall); 160 // Only copy the edge if the call was inlined! 161 if (VMI != ValueMap.end() && VMI->second) { 162 // If the call was inlined, but then constant folded, there is no edge to 163 // add. Check for this case. 164 if (Instruction *NewCall = dyn_cast<Instruction>(VMI->second)) 165 CallerNode->addCalledFunction(CallSite::get(NewCall), I->second); 166 } 167 } 168} 169 170 171// InlineFunction - This function inlines the called function into the basic 172// block of the caller. This returns false if it is not possible to inline this 173// call. The program is still in a well defined state if this occurs though. 174// 175// Note that this only does one level of inlining. For example, if the 176// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 177// exists in the instruction stream. Similiarly this will inline a recursive 178// function by one level. 179// 180bool llvm::InlineFunction(CallSite CS, CallGraph *CG) { 181 Instruction *TheCall = CS.getInstruction(); 182 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 183 "Instruction not in function!"); 184 185 const Function *CalledFunc = CS.getCalledFunction(); 186 if (CalledFunc == 0 || // Can't inline external function or indirect 187 CalledFunc->isExternal() || // call, or call to a vararg function! 188 CalledFunc->getFunctionType()->isVarArg()) return false; 189 190 191 // If the call to the callee is a non-tail call, we must clear the 'tail' 192 // flags on any calls that we inline. 193 bool MustClearTailCallFlags = 194 isa<CallInst>(TheCall) && !cast<CallInst>(TheCall)->isTailCall(); 195 196 BasicBlock *OrigBB = TheCall->getParent(); 197 Function *Caller = OrigBB->getParent(); 198 199 // Get an iterator to the last basic block in the function, which will have 200 // the new function inlined after it. 201 // 202 Function::iterator LastBlock = &Caller->back(); 203 204 // Make sure to capture all of the return instructions from the cloned 205 // function. 206 std::vector<ReturnInst*> Returns; 207 ClonedCodeInfo InlinedFunctionInfo; 208 Function::iterator FirstNewBlock; 209 210 { // Scope to destroy ValueMap after cloning. 211 std::map<const Value*, Value*> ValueMap; 212 213 // Calculate the vector of arguments to pass into the function cloner, which 214 // matches up the formal to the actual argument values. 215 assert(std::distance(CalledFunc->arg_begin(), CalledFunc->arg_end()) == 216 std::distance(CS.arg_begin(), CS.arg_end()) && 217 "No varargs calls can be inlined!"); 218 CallSite::arg_iterator AI = CS.arg_begin(); 219 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 220 E = CalledFunc->arg_end(); I != E; ++I, ++AI) 221 ValueMap[I] = *AI; 222 223 // We want the inliner to prune the code as it copies. We would LOVE to 224 // have no dead or constant instructions leftover after inlining occurs 225 // (which can happen, e.g., because an argument was constant), but we'll be 226 // happy with whatever the cloner can do. 227 CloneAndPruneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i", 228 &InlinedFunctionInfo); 229 230 // Remember the first block that is newly cloned over. 231 FirstNewBlock = LastBlock; ++FirstNewBlock; 232 233 // Update the callgraph if requested. 234 if (CG) 235 UpdateCallGraphAfterInlining(Caller, CalledFunc, FirstNewBlock, ValueMap, 236 *CG); 237 } 238 239 // If there are any alloca instructions in the block that used to be the entry 240 // block for the callee, move them to the entry block of the caller. First 241 // calculate which instruction they should be inserted before. We insert the 242 // instructions at the end of the current alloca list. 243 // 244 { 245 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 246 for (BasicBlock::iterator I = FirstNewBlock->begin(), 247 E = FirstNewBlock->end(); I != E; ) 248 if (AllocaInst *AI = dyn_cast<AllocaInst>(I++)) 249 if (isa<Constant>(AI->getArraySize())) { 250 // Scan for the block of allocas that we can move over, and move them 251 // all at once. 252 while (isa<AllocaInst>(I) && 253 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) 254 ++I; 255 256 // Transfer all of the allocas over in a block. Using splice means 257 // that they instructions aren't removed from the symbol table, then 258 // reinserted. 259 Caller->front().getInstList().splice(InsertPoint, 260 FirstNewBlock->getInstList(), 261 AI, I); 262 } 263 } 264 265 // If the inlined code contained dynamic alloca instructions, wrap the inlined 266 // code with llvm.stacksave/llvm.stackrestore intrinsics. 267 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 268 Module *M = Caller->getParent(); 269 const Type *SBytePtr = PointerType::get(Type::SByteTy); 270 // Get the two intrinsics we care about. 271 Function *StackSave, *StackRestore; 272 StackSave = M->getOrInsertFunction("llvm.stacksave", SBytePtr, NULL); 273 StackRestore = M->getOrInsertFunction("llvm.stackrestore", Type::VoidTy, 274 SBytePtr, NULL); 275 276 // If we are preserving the callgraph, add edges to the stacksave/restore 277 // functions for the calls we insert. 278 CallGraphNode *StackSaveCGN, *StackRestoreCGN, *CallerNode; 279 if (CG) { 280 StackSaveCGN = CG->getOrInsertFunction(StackSave); 281 StackRestoreCGN = CG->getOrInsertFunction(StackRestore); 282 CallerNode = (*CG)[Caller]; 283 } 284 285 // Insert the llvm.stacksave. 286 CallInst *SavedPtr = new CallInst(StackSave, "savedstack", 287 FirstNewBlock->begin()); 288 if (CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN); 289 290 // Insert a call to llvm.stackrestore before any return instructions in the 291 // inlined function. 292 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 293 CallInst *CI = new CallInst(StackRestore, SavedPtr, "", Returns[i]); 294 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); 295 } 296 297 // Count the number of StackRestore calls we insert. 298 unsigned NumStackRestores = Returns.size(); 299 300 // If we are inlining an invoke instruction, insert restores before each 301 // unwind. These unwinds will be rewritten into branches later. 302 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 303 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 304 BB != E; ++BB) 305 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 306 new CallInst(StackRestore, SavedPtr, "", UI); 307 ++NumStackRestores; 308 } 309 } 310 } 311 312 // If we are inlining tail call instruction through a call site that isn't 313 // marked 'tail', we must remove the tail marker for any calls in the inlined 314 // code. 315 if (MustClearTailCallFlags && InlinedFunctionInfo.ContainsCalls) { 316 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 317 BB != E; ++BB) 318 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 319 if (CallInst *CI = dyn_cast<CallInst>(I)) 320 CI->setTailCall(false); 321 } 322 323 // If we are inlining for an invoke instruction, we must make sure to rewrite 324 // any inlined 'unwind' instructions into branches to the invoke exception 325 // destination, and call instructions into invoke instructions. 326 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 327 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 328 329 // If we cloned in _exactly one_ basic block, and if that block ends in a 330 // return instruction, we splice the body of the inlined callee directly into 331 // the calling basic block. 332 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 333 // Move all of the instructions right before the call. 334 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 335 FirstNewBlock->begin(), FirstNewBlock->end()); 336 // Remove the cloned basic block. 337 Caller->getBasicBlockList().pop_back(); 338 339 // If the call site was an invoke instruction, add a branch to the normal 340 // destination. 341 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 342 new BranchInst(II->getNormalDest(), TheCall); 343 344 // If the return instruction returned a value, replace uses of the call with 345 // uses of the returned value. 346 if (!TheCall->use_empty()) 347 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 348 349 // Since we are now done with the Call/Invoke, we can delete it. 350 TheCall->getParent()->getInstList().erase(TheCall); 351 352 // Since we are now done with the return instruction, delete it also. 353 Returns[0]->getParent()->getInstList().erase(Returns[0]); 354 355 // We are now done with the inlining. 356 return true; 357 } 358 359 // Otherwise, we have the normal case, of more than one block to inline or 360 // multiple return sites. 361 362 // We want to clone the entire callee function into the hole between the 363 // "starter" and "ender" blocks. How we accomplish this depends on whether 364 // this is an invoke instruction or a call instruction. 365 BasicBlock *AfterCallBB; 366 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 367 368 // Add an unconditional branch to make this look like the CallInst case... 369 BranchInst *NewBr = new BranchInst(II->getNormalDest(), TheCall); 370 371 // Split the basic block. This guarantees that no PHI nodes will have to be 372 // updated due to new incoming edges, and make the invoke case more 373 // symmetric to the call case. 374 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 375 CalledFunc->getName()+".exit"); 376 377 } else { // It's a call 378 // If this is a call instruction, we need to split the basic block that 379 // the call lives in. 380 // 381 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 382 CalledFunc->getName()+".exit"); 383 } 384 385 // Change the branch that used to go to AfterCallBB to branch to the first 386 // basic block of the inlined function. 387 // 388 TerminatorInst *Br = OrigBB->getTerminator(); 389 assert(Br && Br->getOpcode() == Instruction::Br && 390 "splitBasicBlock broken!"); 391 Br->setOperand(0, FirstNewBlock); 392 393 394 // Now that the function is correct, make it a little bit nicer. In 395 // particular, move the basic blocks inserted from the end of the function 396 // into the space made by splitting the source basic block. 397 // 398 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 399 FirstNewBlock, Caller->end()); 400 401 // Handle all of the return instructions that we just cloned in, and eliminate 402 // any users of the original call/invoke instruction. 403 if (Returns.size() > 1) { 404 // The PHI node should go at the front of the new basic block to merge all 405 // possible incoming values. 406 // 407 PHINode *PHI = 0; 408 if (!TheCall->use_empty()) { 409 PHI = new PHINode(CalledFunc->getReturnType(), 410 TheCall->getName(), AfterCallBB->begin()); 411 412 // Anything that used the result of the function call should now use the 413 // PHI node as their operand. 414 // 415 TheCall->replaceAllUsesWith(PHI); 416 } 417 418 // Loop over all of the return instructions, turning them into unconditional 419 // branches to the merge point now, and adding entries to the PHI node as 420 // appropriate. 421 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 422 ReturnInst *RI = Returns[i]; 423 424 if (PHI) { 425 assert(RI->getReturnValue() && "Ret should have value!"); 426 assert(RI->getReturnValue()->getType() == PHI->getType() && 427 "Ret value not consistent in function!"); 428 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 429 } 430 431 // Add a branch to the merge point where the PHI node lives if it exists. 432 new BranchInst(AfterCallBB, RI); 433 434 // Delete the return instruction now 435 RI->getParent()->getInstList().erase(RI); 436 } 437 438 } else if (!Returns.empty()) { 439 // Otherwise, if there is exactly one return value, just replace anything 440 // using the return value of the call with the computed value. 441 if (!TheCall->use_empty()) 442 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 443 444 // Splice the code from the return block into the block that it will return 445 // to, which contains the code that was after the call. 446 BasicBlock *ReturnBB = Returns[0]->getParent(); 447 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 448 ReturnBB->getInstList()); 449 450 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 451 ReturnBB->replaceAllUsesWith(AfterCallBB); 452 453 // Delete the return instruction now and empty ReturnBB now. 454 Returns[0]->eraseFromParent(); 455 ReturnBB->eraseFromParent(); 456 } else if (!TheCall->use_empty()) { 457 // No returns, but something is using the return value of the call. Just 458 // nuke the result. 459 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 460 } 461 462 // Since we are now done with the Call/Invoke, we can delete it. 463 TheCall->eraseFromParent(); 464 465 // We should always be able to fold the entry block of the function into the 466 // single predecessor of the block... 467 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 468 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 469 470 // Splice the code entry block into calling block, right before the 471 // unconditional branch. 472 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 473 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 474 475 // Remove the unconditional branch. 476 OrigBB->getInstList().erase(Br); 477 478 // Now we can remove the CalleeEntry block, which is now empty. 479 Caller->getBasicBlockList().erase(CalleeEntry); 480 481 return true; 482} 483