TailRecursionElimination.cpp revision ce869ee05bfcb9d4750a3d0919a8260a727841c3
1//===- TailRecursionElimination.cpp - Eliminate Tail Calls ----------------===// 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 transforms calls of the current function (self recursion) followed 11// by a return instruction with a branch to the entry of the function, creating 12// a loop. This pass also implements the following extensions to the basic 13// algorithm: 14// 15// 1. Trivial instructions between the call and return do not prevent the 16// transformation from taking place, though currently the analysis cannot 17// support moving any really useful instructions (only dead ones). 18// 2. This pass transforms functions that are prevented from being tail 19// recursive by an associative expression to use an accumulator variable, 20// thus compiling the typical naive factorial or 'fib' implementation into 21// efficient code. 22// 3. TRE is performed if the function returns void, if the return 23// returns the result returned by the call, or if the function returns a 24// run-time constant on all exits from the function. It is possible, though 25// unlikely, that the return returns something else (like constant 0), and 26// can still be TRE'd. It can be TRE'd if ALL OTHER return instructions in 27// the function return the exact same value. 28// 4. If it can prove that callees do not access theier caller stack frame, 29// they are marked as eligible for tail call elimination (by the code 30// generator). 31// 32// There are several improvements that could be made: 33// 34// 1. If the function has any alloca instructions, these instructions will be 35// moved out of the entry block of the function, causing them to be 36// evaluated each time through the tail recursion. Safely keeping allocas 37// in the entry block requires analysis to proves that the tail-called 38// function does not read or write the stack object. 39// 2. Tail recursion is only performed if the call immediately preceeds the 40// return instruction. It's possible that there could be a jump between 41// the call and the return. 42// 3. There can be intervening operations between the call and the return that 43// prevent the TRE from occurring. For example, there could be GEP's and 44// stores to memory that will not be read or written by the call. This 45// requires some substantial analysis (such as with DSA) to prove safe to 46// move ahead of the call, but doing so could allow many more TREs to be 47// performed, for example in TreeAdd/TreeAlloc from the treeadd benchmark. 48// 4. The algorithm we use to detect if callees access their caller stack 49// frames is very primitive. 50// 51//===----------------------------------------------------------------------===// 52 53#include "llvm/Transforms/Scalar.h" 54#include "llvm/Constants.h" 55#include "llvm/DerivedTypes.h" 56#include "llvm/Function.h" 57#include "llvm/Instructions.h" 58#include "llvm/Pass.h" 59#include "llvm/Support/CFG.h" 60#include "llvm/ADT/Statistic.h" 61using namespace llvm; 62 63namespace { 64 Statistic<> NumEliminated("tailcallelim", "Number of tail calls removed"); 65 Statistic<> NumAccumAdded("tailcallelim","Number of accumulators introduced"); 66 67 struct TailCallElim : public FunctionPass { 68 virtual bool runOnFunction(Function &F); 69 70 private: 71 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry, 72 bool &TailCallsAreMarkedTail, 73 std::vector<PHINode*> &ArgumentPHIs, 74 bool CannotTailCallElimCallsMarkedTail); 75 bool CanMoveAboveCall(Instruction *I, CallInst *CI); 76 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI); 77 }; 78 RegisterOpt<TailCallElim> X("tailcallelim", "Tail Call Elimination"); 79} 80 81// Public interface to the TailCallElimination pass 82FunctionPass *llvm::createTailCallEliminationPass() { 83 return new TailCallElim(); 84} 85 86 87/// AllocaMightEscapeToCalls - Return true if this alloca may be accessed by 88/// callees of this function. We only do very simple analysis right now, this 89/// could be expanded in the future to use mod/ref information for particular 90/// call sites if desired. 91static bool AllocaMightEscapeToCalls(AllocaInst *AI) { 92 // FIXME: do simple 'address taken' analysis. 93 return true; 94} 95 96/// FunctionContainsAllocas - Scan the specified basic block for alloca 97/// instructions. If it contains any that might be accessed by calls, return 98/// true. 99static bool CheckForEscapingAllocas(BasicBlock *BB, 100 bool &CannotTCETailMarkedCall) { 101 bool RetVal = false; 102 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 103 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) { 104 RetVal |= AllocaMightEscapeToCalls(AI); 105 106 // If this alloca is in the body of the function, or if it is a variable 107 // sized allocation, we cannot tail call eliminate calls marked 'tail' 108 // with this mechanism. 109 if (BB != &BB->getParent()->front() || 110 !isa<ConstantInt>(AI->getArraySize())) 111 CannotTCETailMarkedCall = true; 112 } 113 return RetVal; 114} 115 116bool TailCallElim::runOnFunction(Function &F) { 117 // If this function is a varargs function, we won't be able to PHI the args 118 // right, so don't even try to convert it... 119 if (F.getFunctionType()->isVarArg()) return false; 120 121 BasicBlock *OldEntry = 0; 122 bool TailCallsAreMarkedTail = false; 123 std::vector<PHINode*> ArgumentPHIs; 124 bool MadeChange = false; 125 126 bool FunctionContainsEscapingAllocas = false; 127 128 // CannotTCETailMarkedCall - If true, we cannot perform TCE on tail calls 129 // marked with the 'tail' attribute, because doing so would cause the stack 130 // size to increase (real TCE would deallocate variable sized allocas, TCE 131 // doesn't). 132 bool CannotTCETailMarkedCall = false; 133 134 // Loop over the function, looking for any returning blocks, and keeping track 135 // of whether this function has any non-trivially used allocas. 136 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 137 if (FunctionContainsEscapingAllocas && CannotTCETailMarkedCall) 138 break; 139 140 FunctionContainsEscapingAllocas = 141 CheckForEscapingAllocas(BB, CannotTCETailMarkedCall); 142 } 143 144 // Second pass, change any tail calls to loops. 145 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 146 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) 147 MadeChange |= ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail, 148 ArgumentPHIs,CannotTCETailMarkedCall); 149 150 // If we eliminated any tail recursions, it's possible that we inserted some 151 // silly PHI nodes which just merge an initial value (the incoming operand) 152 // with themselves. Check to see if we did and clean up our mess if so. This 153 // occurs when a function passes an argument straight through to its tail 154 // call. 155 if (!ArgumentPHIs.empty()) { 156 unsigned NumIncoming = ArgumentPHIs[0]->getNumIncomingValues(); 157 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) { 158 PHINode *PN = ArgumentPHIs[i]; 159 160 // If the PHI Node is a dynamic constant, replace it with the value it is. 161 if (Value *PNV = PN->hasConstantValue()) { 162 PN->replaceAllUsesWith(PNV); 163 PN->eraseFromParent(); 164 } 165 } 166 } 167 168 // Finally, if this function contains no non-escaping allocas, mark all calls 169 // in the function as eligible for tail calls (there is no stack memory for 170 // them to access). 171 if (!FunctionContainsEscapingAllocas) 172 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 173 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 174 if (CallInst *CI = dyn_cast<CallInst>(I)) 175 CI->setTailCall(); 176 177 return MadeChange; 178} 179 180 181/// CanMoveAboveCall - Return true if it is safe to move the specified 182/// instruction from after the call to before the call, assuming that all 183/// instructions between the call and this instruction are movable. 184/// 185bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) { 186 // FIXME: We can move load/store/call/free instructions above the call if the 187 // call does not mod/ref the memory location being processed. 188 if (I->mayWriteToMemory() || isa<LoadInst>(I)) 189 return false; 190 191 // Otherwise, if this is a side-effect free instruction, check to make sure 192 // that it does not use the return value of the call. If it doesn't use the 193 // return value of the call, it must only use things that are defined before 194 // the call, or movable instructions between the call and the instruction 195 // itself. 196 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 197 if (I->getOperand(i) == CI) 198 return false; 199 return true; 200} 201 202// isDynamicConstant - Return true if the specified value is the same when the 203// return would exit as it was when the initial iteration of the recursive 204// function was executed. 205// 206// We currently handle static constants and arguments that are not modified as 207// part of the recursion. 208// 209static bool isDynamicConstant(Value *V, CallInst *CI) { 210 if (isa<Constant>(V)) return true; // Static constants are always dyn consts 211 212 // Check to see if this is an immutable argument, if so, the value 213 // will be available to initialize the accumulator. 214 if (Argument *Arg = dyn_cast<Argument>(V)) { 215 // Figure out which argument number this is... 216 unsigned ArgNo = 0; 217 Function *F = CI->getParent()->getParent(); 218 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI) 219 ++ArgNo; 220 221 // If we are passing this argument into call as the corresponding 222 // argument operand, then the argument is dynamically constant. 223 // Otherwise, we cannot transform this function safely. 224 if (CI->getOperand(ArgNo+1) == Arg) 225 return true; 226 } 227 // Not a constant or immutable argument, we can't safely transform. 228 return false; 229} 230 231// getCommonReturnValue - Check to see if the function containing the specified 232// return instruction and tail call consistently returns the same 233// runtime-constant value at all exit points. If so, return the returned value. 234// 235static Value *getCommonReturnValue(ReturnInst *TheRI, CallInst *CI) { 236 Function *F = TheRI->getParent()->getParent(); 237 Value *ReturnedValue = 0; 238 239 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) 240 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator())) 241 if (RI != TheRI) { 242 Value *RetOp = RI->getOperand(0); 243 244 // We can only perform this transformation if the value returned is 245 // evaluatable at the start of the initial invocation of the function, 246 // instead of at the end of the evaluation. 247 // 248 if (!isDynamicConstant(RetOp, CI)) 249 return 0; 250 251 if (ReturnedValue && RetOp != ReturnedValue) 252 return 0; // Cannot transform if differing values are returned. 253 ReturnedValue = RetOp; 254 } 255 return ReturnedValue; 256} 257 258/// CanTransformAccumulatorRecursion - If the specified instruction can be 259/// transformed using accumulator recursion elimination, return the constant 260/// which is the start of the accumulator value. Otherwise return null. 261/// 262Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I, 263 CallInst *CI) { 264 if (!I->isAssociative()) return 0; 265 assert(I->getNumOperands() == 2 && 266 "Associative operations should have 2 args!"); 267 268 // Exactly one operand should be the result of the call instruction... 269 if (I->getOperand(0) == CI && I->getOperand(1) == CI || 270 I->getOperand(0) != CI && I->getOperand(1) != CI) 271 return 0; 272 273 // The only user of this instruction we allow is a single return instruction. 274 if (!I->hasOneUse() || !isa<ReturnInst>(I->use_back())) 275 return 0; 276 277 // Ok, now we have to check all of the other return instructions in this 278 // function. If they return non-constants or differing values, then we cannot 279 // transform the function safely. 280 return getCommonReturnValue(cast<ReturnInst>(I->use_back()), CI); 281} 282 283bool TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry, 284 bool &TailCallsAreMarkedTail, 285 std::vector<PHINode*> &ArgumentPHIs, 286 bool CannotTailCallElimCallsMarkedTail) { 287 BasicBlock *BB = Ret->getParent(); 288 Function *F = BB->getParent(); 289 290 if (&BB->front() == Ret) // Make sure there is something before the ret... 291 return false; 292 293 // Scan backwards from the return, checking to see if there is a tail call in 294 // this block. If so, set CI to it. 295 CallInst *CI; 296 BasicBlock::iterator BBI = Ret; 297 while (1) { 298 CI = dyn_cast<CallInst>(BBI); 299 if (CI && CI->getCalledFunction() == F) 300 break; 301 302 if (BBI == BB->begin()) 303 return false; // Didn't find a potential tail call. 304 --BBI; 305 } 306 307 // If this call is marked as a tail call, and if there are dynamic allocas in 308 // the function, we cannot perform this optimization. 309 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail) 310 return false; 311 312 // If we are introducing accumulator recursion to eliminate associative 313 // operations after the call instruction, this variable contains the initial 314 // value for the accumulator. If this value is set, we actually perform 315 // accumulator recursion elimination instead of simple tail recursion 316 // elimination. 317 Value *AccumulatorRecursionEliminationInitVal = 0; 318 Instruction *AccumulatorRecursionInstr = 0; 319 320 // Ok, we found a potential tail call. We can currently only transform the 321 // tail call if all of the instructions between the call and the return are 322 // movable to above the call itself, leaving the call next to the return. 323 // Check that this is the case now. 324 for (BBI = CI, ++BBI; &*BBI != Ret; ++BBI) 325 if (!CanMoveAboveCall(BBI, CI)) { 326 // If we can't move the instruction above the call, it might be because it 327 // is an associative operation that could be tranformed using accumulator 328 // recursion elimination. Check to see if this is the case, and if so, 329 // remember the initial accumulator value for later. 330 if ((AccumulatorRecursionEliminationInitVal = 331 CanTransformAccumulatorRecursion(BBI, CI))) { 332 // Yes, this is accumulator recursion. Remember which instruction 333 // accumulates. 334 AccumulatorRecursionInstr = BBI; 335 } else { 336 return false; // Otherwise, we cannot eliminate the tail recursion! 337 } 338 } 339 340 // We can only transform call/return pairs that either ignore the return value 341 // of the call and return void, ignore the value of the call and return a 342 // constant, return the value returned by the tail call, or that are being 343 // accumulator recursion variable eliminated. 344 if (Ret->getNumOperands() != 0 && Ret->getReturnValue() != CI && 345 AccumulatorRecursionEliminationInitVal == 0 && 346 !getCommonReturnValue(Ret, CI)) 347 return false; 348 349 // OK! We can transform this tail call. If this is the first one found, 350 // create the new entry block, allowing us to branch back to the old entry. 351 if (OldEntry == 0) { 352 OldEntry = &F->getEntryBlock(); 353 std::string OldName = OldEntry->getName(); OldEntry->setName("tailrecurse"); 354 BasicBlock *NewEntry = new BasicBlock(OldName, F, OldEntry); 355 new BranchInst(OldEntry, NewEntry); 356 357 // If this tail call is marked 'tail' and if there are any allocas in the 358 // entry block, move them up to the new entry block. 359 TailCallsAreMarkedTail = CI->isTailCall(); 360 if (TailCallsAreMarkedTail) 361 // Move all fixed sized allocas from OldEntry to NewEntry. 362 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(), 363 NEBI = NewEntry->begin(); OEBI != E; ) 364 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++)) 365 if (isa<ConstantInt>(AI->getArraySize())) 366 NewEntry->getInstList().splice(NEBI, OldEntry->getInstList(), AI); 367 368 // Now that we have created a new block, which jumps to the entry 369 // block, insert a PHI node for each argument of the function. 370 // For now, we initialize each PHI to only have the real arguments 371 // which are passed in. 372 Instruction *InsertPos = OldEntry->begin(); 373 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 374 I != E; ++I) { 375 PHINode *PN = new PHINode(I->getType(), I->getName()+".tr", InsertPos); 376 I->replaceAllUsesWith(PN); // Everyone use the PHI node now! 377 PN->addIncoming(I, NewEntry); 378 ArgumentPHIs.push_back(PN); 379 } 380 } 381 382 // If this function has self recursive calls in the tail position where some 383 // are marked tail and some are not, only transform one flavor or another. We 384 // have to choose whether we move allocas in the entry block to the new entry 385 // block or not, so we can't make a good choice for both. NOTE: We could do 386 // slightly better here in the case that the function has no entry block 387 // allocas. 388 if (TailCallsAreMarkedTail && !CI->isTailCall()) 389 return false; 390 391 // Ok, now that we know we have a pseudo-entry block WITH all of the 392 // required PHI nodes, add entries into the PHI node for the actual 393 // parameters passed into the tail-recursive call. 394 for (unsigned i = 0, e = CI->getNumOperands()-1; i != e; ++i) 395 ArgumentPHIs[i]->addIncoming(CI->getOperand(i+1), BB); 396 397 // If we are introducing an accumulator variable to eliminate the recursion, 398 // do so now. Note that we _know_ that no subsequent tail recursion 399 // eliminations will happen on this function because of the way the 400 // accumulator recursion predicate is set up. 401 // 402 if (AccumulatorRecursionEliminationInitVal) { 403 Instruction *AccRecInstr = AccumulatorRecursionInstr; 404 // Start by inserting a new PHI node for the accumulator. 405 PHINode *AccPN = new PHINode(AccRecInstr->getType(), "accumulator.tr", 406 OldEntry->begin()); 407 408 // Loop over all of the predecessors of the tail recursion block. For the 409 // real entry into the function we seed the PHI with the initial value, 410 // computed earlier. For any other existing branches to this block (due to 411 // other tail recursions eliminated) the accumulator is not modified. 412 // Because we haven't added the branch in the current block to OldEntry yet, 413 // it will not show up as a predecessor. 414 for (pred_iterator PI = pred_begin(OldEntry), PE = pred_end(OldEntry); 415 PI != PE; ++PI) { 416 if (*PI == &F->getEntryBlock()) 417 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, *PI); 418 else 419 AccPN->addIncoming(AccPN, *PI); 420 } 421 422 // Add an incoming argument for the current block, which is computed by our 423 // associative accumulator instruction. 424 AccPN->addIncoming(AccRecInstr, BB); 425 426 // Next, rewrite the accumulator recursion instruction so that it does not 427 // use the result of the call anymore, instead, use the PHI node we just 428 // inserted. 429 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN); 430 431 // Finally, rewrite any return instructions in the program to return the PHI 432 // node instead of the "initval" that they do currently. This loop will 433 // actually rewrite the return value we are destroying, but that's ok. 434 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) 435 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator())) 436 RI->setOperand(0, AccPN); 437 ++NumAccumAdded; 438 } 439 440 // Now that all of the PHI nodes are in place, remove the call and 441 // ret instructions, replacing them with an unconditional branch. 442 new BranchInst(OldEntry, Ret); 443 BB->getInstList().erase(Ret); // Remove return. 444 BB->getInstList().erase(CI); // Remove call. 445 ++NumEliminated; 446 return true; 447} 448