TailRecursionElimination.cpp revision 3e8b6631e67e01e4960a7ba4668a50c596607473
1//===- TailRecursionElimination.cpp - Eliminate Tail Calls ----------------===// 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 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#define DEBUG_TYPE "tailcallelim" 54#include "llvm/Transforms/Scalar.h" 55#include "llvm/Transforms/Utils/Local.h" 56#include "llvm/Constants.h" 57#include "llvm/DerivedTypes.h" 58#include "llvm/Function.h" 59#include "llvm/Instructions.h" 60#include "llvm/Pass.h" 61#include "llvm/Support/CFG.h" 62#include "llvm/ADT/Statistic.h" 63using namespace llvm; 64 65STATISTIC(NumEliminated, "Number of tail calls removed"); 66STATISTIC(NumAccumAdded, "Number of accumulators introduced"); 67 68namespace { 69 struct TailCallElim : public FunctionPass { 70 static char ID; // Pass identification, replacement for typeid 71 TailCallElim() : FunctionPass(&ID) {} 72 73 virtual bool runOnFunction(Function &F); 74 75 private: 76 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry, 77 bool &TailCallsAreMarkedTail, 78 std::vector<PHINode*> &ArgumentPHIs, 79 bool CannotTailCallElimCallsMarkedTail); 80 bool CanMoveAboveCall(Instruction *I, CallInst *CI); 81 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI); 82 }; 83} 84 85char TailCallElim::ID = 0; 86static RegisterPass<TailCallElim> X("tailcallelim", "Tail Call Elimination"); 87 88// Public interface to the TailCallElimination pass 89FunctionPass *llvm::createTailCallEliminationPass() { 90 return new TailCallElim(); 91} 92 93 94/// AllocaMightEscapeToCalls - Return true if this alloca may be accessed by 95/// callees of this function. We only do very simple analysis right now, this 96/// could be expanded in the future to use mod/ref information for particular 97/// call sites if desired. 98static bool AllocaMightEscapeToCalls(AllocaInst *AI) { 99 // FIXME: do simple 'address taken' analysis. 100 return true; 101} 102 103/// FunctionContainsAllocas - Scan the specified basic block for alloca 104/// instructions. If it contains any that might be accessed by calls, return 105/// true. 106static bool CheckForEscapingAllocas(BasicBlock *BB, 107 bool &CannotTCETailMarkedCall) { 108 bool RetVal = false; 109 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 110 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) { 111 RetVal |= AllocaMightEscapeToCalls(AI); 112 113 // If this alloca is in the body of the function, or if it is a variable 114 // sized allocation, we cannot tail call eliminate calls marked 'tail' 115 // with this mechanism. 116 if (BB != &BB->getParent()->getEntryBlock() || 117 !isa<ConstantInt>(AI->getArraySize())) 118 CannotTCETailMarkedCall = true; 119 } 120 return RetVal; 121} 122 123bool TailCallElim::runOnFunction(Function &F) { 124 // If this function is a varargs function, we won't be able to PHI the args 125 // right, so don't even try to convert it... 126 if (F.getFunctionType()->isVarArg()) return false; 127 128 BasicBlock *OldEntry = 0; 129 bool TailCallsAreMarkedTail = false; 130 std::vector<PHINode*> ArgumentPHIs; 131 bool MadeChange = false; 132 133 bool FunctionContainsEscapingAllocas = false; 134 135 // CannotTCETailMarkedCall - If true, we cannot perform TCE on tail calls 136 // marked with the 'tail' attribute, because doing so would cause the stack 137 // size to increase (real TCE would deallocate variable sized allocas, TCE 138 // doesn't). 139 bool CannotTCETailMarkedCall = false; 140 141 // Loop over the function, looking for any returning blocks, and keeping track 142 // of whether this function has any non-trivially used allocas. 143 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 144 if (FunctionContainsEscapingAllocas && CannotTCETailMarkedCall) 145 break; 146 147 FunctionContainsEscapingAllocas |= 148 CheckForEscapingAllocas(BB, CannotTCETailMarkedCall); 149 } 150 151 /// FIXME: The code generator produces really bad code when an 'escaping 152 /// alloca' is changed from being a static alloca to being a dynamic alloca. 153 /// Until this is resolved, disable this transformation if that would ever 154 /// happen. This bug is PR962. 155 if (FunctionContainsEscapingAllocas) 156 return false; 157 158 159 // Second pass, change any tail calls to loops. 160 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 161 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) 162 MadeChange |= ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail, 163 ArgumentPHIs,CannotTCETailMarkedCall); 164 165 // If we eliminated any tail recursions, it's possible that we inserted some 166 // silly PHI nodes which just merge an initial value (the incoming operand) 167 // with themselves. Check to see if we did and clean up our mess if so. This 168 // occurs when a function passes an argument straight through to its tail 169 // call. 170 if (!ArgumentPHIs.empty()) { 171 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) { 172 PHINode *PN = ArgumentPHIs[i]; 173 174 // If the PHI Node is a dynamic constant, replace it with the value it is. 175 if (Value *PNV = PN->hasConstantValue()) { 176 PN->replaceAllUsesWith(PNV); 177 PN->eraseFromParent(); 178 } 179 } 180 } 181 182 // Finally, if this function contains no non-escaping allocas, mark all calls 183 // in the function as eligible for tail calls (there is no stack memory for 184 // them to access). 185 if (!FunctionContainsEscapingAllocas) 186 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 187 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 188 if (CallInst *CI = dyn_cast<CallInst>(I)) { 189 CI->setTailCall(); 190 MadeChange = true; 191 } 192 193 return MadeChange; 194} 195 196 197/// CanMoveAboveCall - Return true if it is safe to move the specified 198/// instruction from after the call to before the call, assuming that all 199/// instructions between the call and this instruction are movable. 200/// 201bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) { 202 // FIXME: We can move load/store/call/free instructions above the call if the 203 // call does not mod/ref the memory location being processed. 204 if (I->mayHaveSideEffects()) // This also handles volatile loads. 205 return false; 206 207 if (LoadInst* L = dyn_cast<LoadInst>(I)) { 208 // Loads may always be moved above calls without side effects. 209 if (CI->mayHaveSideEffects()) { 210 // Non-volatile loads may be moved above a call with side effects if it 211 // does not write to memory and the load provably won't trap. 212 // FIXME: Writes to memory only matter if they may alias the pointer 213 // being loaded from. 214 if (CI->mayWriteToMemory() || 215 !isSafeToLoadUnconditionally(L->getPointerOperand(), L)) 216 return false; 217 } 218 } 219 220 // Otherwise, if this is a side-effect free instruction, check to make sure 221 // that it does not use the return value of the call. If it doesn't use the 222 // return value of the call, it must only use things that are defined before 223 // the call, or movable instructions between the call and the instruction 224 // itself. 225 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 226 if (I->getOperand(i) == CI) 227 return false; 228 return true; 229} 230 231// isDynamicConstant - Return true if the specified value is the same when the 232// return would exit as it was when the initial iteration of the recursive 233// function was executed. 234// 235// We currently handle static constants and arguments that are not modified as 236// part of the recursion. 237// 238static bool isDynamicConstant(Value *V, CallInst *CI) { 239 if (isa<Constant>(V)) return true; // Static constants are always dyn consts 240 241 // Check to see if this is an immutable argument, if so, the value 242 // will be available to initialize the accumulator. 243 if (Argument *Arg = dyn_cast<Argument>(V)) { 244 // Figure out which argument number this is... 245 unsigned ArgNo = 0; 246 Function *F = CI->getParent()->getParent(); 247 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI) 248 ++ArgNo; 249 250 // If we are passing this argument into call as the corresponding 251 // argument operand, then the argument is dynamically constant. 252 // Otherwise, we cannot transform this function safely. 253 if (CI->getOperand(ArgNo+1) == Arg) 254 return true; 255 } 256 // Not a constant or immutable argument, we can't safely transform. 257 return false; 258} 259 260// getCommonReturnValue - Check to see if the function containing the specified 261// return instruction and tail call consistently returns the same 262// runtime-constant value at all exit points. If so, return the returned value. 263// 264static Value *getCommonReturnValue(ReturnInst *TheRI, CallInst *CI) { 265 Function *F = TheRI->getParent()->getParent(); 266 Value *ReturnedValue = 0; 267 268 // TODO: Handle multiple value ret instructions; 269 if (isa<StructType>(F->getReturnType())) 270 return 0; 271 272 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) 273 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator())) 274 if (RI != TheRI) { 275 Value *RetOp = RI->getOperand(0); 276 277 // We can only perform this transformation if the value returned is 278 // evaluatable at the start of the initial invocation of the function, 279 // instead of at the end of the evaluation. 280 // 281 if (!isDynamicConstant(RetOp, CI)) 282 return 0; 283 284 if (ReturnedValue && RetOp != ReturnedValue) 285 return 0; // Cannot transform if differing values are returned. 286 ReturnedValue = RetOp; 287 } 288 return ReturnedValue; 289} 290 291/// CanTransformAccumulatorRecursion - If the specified instruction can be 292/// transformed using accumulator recursion elimination, return the constant 293/// which is the start of the accumulator value. Otherwise return null. 294/// 295Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I, 296 CallInst *CI) { 297 if (!I->isAssociative()) return 0; 298 assert(I->getNumOperands() == 2 && 299 "Associative operations should have 2 args!"); 300 301 // Exactly one operand should be the result of the call instruction... 302 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) || 303 (I->getOperand(0) != CI && I->getOperand(1) != CI)) 304 return 0; 305 306 // The only user of this instruction we allow is a single return instruction. 307 if (!I->hasOneUse() || !isa<ReturnInst>(I->use_back())) 308 return 0; 309 310 // Ok, now we have to check all of the other return instructions in this 311 // function. If they return non-constants or differing values, then we cannot 312 // transform the function safely. 313 return getCommonReturnValue(cast<ReturnInst>(I->use_back()), CI); 314} 315 316bool TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry, 317 bool &TailCallsAreMarkedTail, 318 std::vector<PHINode*> &ArgumentPHIs, 319 bool CannotTailCallElimCallsMarkedTail) { 320 BasicBlock *BB = Ret->getParent(); 321 Function *F = BB->getParent(); 322 323 if (&BB->front() == Ret) // Make sure there is something before the ret... 324 return false; 325 326 // If the return is in the entry block, then making this transformation would 327 // turn infinite recursion into an infinite loop. This transformation is ok 328 // in theory, but breaks some code like: 329 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call 330 // disable this xform in this case, because the code generator will lower the 331 // call to fabs into inline code. 332 if (BB == &F->getEntryBlock()) 333 return false; 334 335 // Scan backwards from the return, checking to see if there is a tail call in 336 // this block. If so, set CI to it. 337 CallInst *CI; 338 BasicBlock::iterator BBI = Ret; 339 while (1) { 340 CI = dyn_cast<CallInst>(BBI); 341 if (CI && CI->getCalledFunction() == F) 342 break; 343 344 if (BBI == BB->begin()) 345 return false; // Didn't find a potential tail call. 346 --BBI; 347 } 348 349 // If this call is marked as a tail call, and if there are dynamic allocas in 350 // the function, we cannot perform this optimization. 351 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail) 352 return false; 353 354 // If we are introducing accumulator recursion to eliminate associative 355 // operations after the call instruction, this variable contains the initial 356 // value for the accumulator. If this value is set, we actually perform 357 // accumulator recursion elimination instead of simple tail recursion 358 // elimination. 359 Value *AccumulatorRecursionEliminationInitVal = 0; 360 Instruction *AccumulatorRecursionInstr = 0; 361 362 // Ok, we found a potential tail call. We can currently only transform the 363 // tail call if all of the instructions between the call and the return are 364 // movable to above the call itself, leaving the call next to the return. 365 // Check that this is the case now. 366 for (BBI = CI, ++BBI; &*BBI != Ret; ++BBI) 367 if (!CanMoveAboveCall(BBI, CI)) { 368 // If we can't move the instruction above the call, it might be because it 369 // is an associative operation that could be tranformed using accumulator 370 // recursion elimination. Check to see if this is the case, and if so, 371 // remember the initial accumulator value for later. 372 if ((AccumulatorRecursionEliminationInitVal = 373 CanTransformAccumulatorRecursion(BBI, CI))) { 374 // Yes, this is accumulator recursion. Remember which instruction 375 // accumulates. 376 AccumulatorRecursionInstr = BBI; 377 } else { 378 return false; // Otherwise, we cannot eliminate the tail recursion! 379 } 380 } 381 382 // We can only transform call/return pairs that either ignore the return value 383 // of the call and return void, ignore the value of the call and return a 384 // constant, return the value returned by the tail call, or that are being 385 // accumulator recursion variable eliminated. 386 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI && 387 !isa<UndefValue>(Ret->getReturnValue()) && 388 AccumulatorRecursionEliminationInitVal == 0 && 389 !getCommonReturnValue(Ret, CI)) 390 return false; 391 392 // OK! We can transform this tail call. If this is the first one found, 393 // create the new entry block, allowing us to branch back to the old entry. 394 if (OldEntry == 0) { 395 OldEntry = &F->getEntryBlock(); 396 BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry); 397 NewEntry->takeName(OldEntry); 398 OldEntry->setName("tailrecurse"); 399 BranchInst::Create(OldEntry, NewEntry); 400 401 // If this tail call is marked 'tail' and if there are any allocas in the 402 // entry block, move them up to the new entry block. 403 TailCallsAreMarkedTail = CI->isTailCall(); 404 if (TailCallsAreMarkedTail) 405 // Move all fixed sized allocas from OldEntry to NewEntry. 406 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(), 407 NEBI = NewEntry->begin(); OEBI != E; ) 408 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++)) 409 if (isa<ConstantInt>(AI->getArraySize())) 410 AI->moveBefore(NEBI); 411 412 // Now that we have created a new block, which jumps to the entry 413 // block, insert a PHI node for each argument of the function. 414 // For now, we initialize each PHI to only have the real arguments 415 // which are passed in. 416 Instruction *InsertPos = OldEntry->begin(); 417 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 418 I != E; ++I) { 419 PHINode *PN = PHINode::Create(I->getType(), 420 I->getName() + ".tr", InsertPos); 421 I->replaceAllUsesWith(PN); // Everyone use the PHI node now! 422 PN->addIncoming(I, NewEntry); 423 ArgumentPHIs.push_back(PN); 424 } 425 } 426 427 // If this function has self recursive calls in the tail position where some 428 // are marked tail and some are not, only transform one flavor or another. We 429 // have to choose whether we move allocas in the entry block to the new entry 430 // block or not, so we can't make a good choice for both. NOTE: We could do 431 // slightly better here in the case that the function has no entry block 432 // allocas. 433 if (TailCallsAreMarkedTail && !CI->isTailCall()) 434 return false; 435 436 // Ok, now that we know we have a pseudo-entry block WITH all of the 437 // required PHI nodes, add entries into the PHI node for the actual 438 // parameters passed into the tail-recursive call. 439 for (unsigned i = 0, e = CI->getNumOperands()-1; i != e; ++i) 440 ArgumentPHIs[i]->addIncoming(CI->getOperand(i+1), BB); 441 442 // If we are introducing an accumulator variable to eliminate the recursion, 443 // do so now. Note that we _know_ that no subsequent tail recursion 444 // eliminations will happen on this function because of the way the 445 // accumulator recursion predicate is set up. 446 // 447 if (AccumulatorRecursionEliminationInitVal) { 448 Instruction *AccRecInstr = AccumulatorRecursionInstr; 449 // Start by inserting a new PHI node for the accumulator. 450 PHINode *AccPN = PHINode::Create(AccRecInstr->getType(), "accumulator.tr", 451 OldEntry->begin()); 452 453 // Loop over all of the predecessors of the tail recursion block. For the 454 // real entry into the function we seed the PHI with the initial value, 455 // computed earlier. For any other existing branches to this block (due to 456 // other tail recursions eliminated) the accumulator is not modified. 457 // Because we haven't added the branch in the current block to OldEntry yet, 458 // it will not show up as a predecessor. 459 for (pred_iterator PI = pred_begin(OldEntry), PE = pred_end(OldEntry); 460 PI != PE; ++PI) { 461 if (*PI == &F->getEntryBlock()) 462 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, *PI); 463 else 464 AccPN->addIncoming(AccPN, *PI); 465 } 466 467 // Add an incoming argument for the current block, which is computed by our 468 // associative accumulator instruction. 469 AccPN->addIncoming(AccRecInstr, BB); 470 471 // Next, rewrite the accumulator recursion instruction so that it does not 472 // use the result of the call anymore, instead, use the PHI node we just 473 // inserted. 474 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN); 475 476 // Finally, rewrite any return instructions in the program to return the PHI 477 // node instead of the "initval" that they do currently. This loop will 478 // actually rewrite the return value we are destroying, but that's ok. 479 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) 480 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator())) 481 RI->setOperand(0, AccPN); 482 ++NumAccumAdded; 483 } 484 485 // Now that all of the PHI nodes are in place, remove the call and 486 // ret instructions, replacing them with an unconditional branch. 487 BranchInst::Create(OldEntry, Ret); 488 BB->getInstList().erase(Ret); // Remove return. 489 BB->getInstList().erase(CI); // Remove call. 490 ++NumEliminated; 491 return true; 492} 493