TailRecursionElimination.cpp revision f80fcd00b3590626a6110fe5fbe9b5fd2a8be1ae
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 their 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/Analysis/CaptureTracking.h" 62#include "llvm/Support/CFG.h" 63#include "llvm/ADT/Statistic.h" 64using namespace llvm; 65 66STATISTIC(NumEliminated, "Number of tail calls removed"); 67STATISTIC(NumAccumAdded, "Number of accumulators introduced"); 68 69namespace { 70 struct TailCallElim : public FunctionPass { 71 static char ID; // Pass identification, replacement for typeid 72 TailCallElim() : FunctionPass(&ID) {} 73 74 virtual bool runOnFunction(Function &F); 75 76 private: 77 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry, 78 bool &TailCallsAreMarkedTail, 79 SmallVector<PHINode*, 8> &ArgumentPHIs, 80 bool CannotTailCallElimCallsMarkedTail); 81 bool CanMoveAboveCall(Instruction *I, CallInst *CI); 82 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI); 83 }; 84} 85 86char TailCallElim::ID = 0; 87static RegisterPass<TailCallElim> X("tailcallelim", "Tail Call Elimination"); 88 89// Public interface to the TailCallElimination pass 90FunctionPass *llvm::createTailCallEliminationPass() { 91 return new TailCallElim(); 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/// CheckForEscapingAllocas - 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 SmallVector<PHINode*, 8> 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 // Second pass, change any tail calls to loops. 159 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 160 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) 161 MadeChange |= ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail, 162 ArgumentPHIs,CannotTCETailMarkedCall); 163 164 // If we eliminated any tail recursions, it's possible that we inserted some 165 // silly PHI nodes which just merge an initial value (the incoming operand) 166 // with themselves. Check to see if we did and clean up our mess if so. This 167 // occurs when a function passes an argument straight through to its tail 168 // call. 169 if (!ArgumentPHIs.empty()) { 170 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) { 171 PHINode *PN = ArgumentPHIs[i]; 172 173 // If the PHI Node is a dynamic constant, replace it with the value it is. 174 if (Value *PNV = PN->hasConstantValue()) { 175 PN->replaceAllUsesWith(PNV); 176 PN->eraseFromParent(); 177 } 178 } 179 } 180 181 // Finally, if this function contains no non-escaping allocas, mark all calls 182 // in the function as eligible for tail calls (there is no stack memory for 183 // them to access). 184 if (!FunctionContainsEscapingAllocas) 185 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 186 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 187 if (CallInst *CI = dyn_cast<CallInst>(I)) { 188 CI->setTailCall(); 189 MadeChange = true; 190 } 191 192 return MadeChange; 193} 194 195 196/// CanMoveAboveCall - Return true if it is safe to move the specified 197/// instruction from after the call to before the call, assuming that all 198/// instructions between the call and this instruction are movable. 199/// 200bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) { 201 // FIXME: We can move load/store/call/free instructions above the call if the 202 // call does not mod/ref the memory location being processed. 203 if (I->mayHaveSideEffects()) // This also handles volatile loads. 204 return false; 205 206 if (LoadInst *L = dyn_cast<LoadInst>(I)) { 207 // Loads may always be moved above calls without side effects. 208 if (CI->mayHaveSideEffects()) { 209 // Non-volatile loads may be moved above a call with side effects if it 210 // does not write to memory and the load provably won't trap. 211 // FIXME: Writes to memory only matter if they may alias the pointer 212 // being loaded from. 213 if (CI->mayWriteToMemory() || 214 !isSafeToLoadUnconditionally(L->getPointerOperand(), L)) 215 return false; 216 } 217 } 218 219 // Otherwise, if this is a side-effect free instruction, check to make sure 220 // that it does not use the return value of the call. If it doesn't use the 221 // return value of the call, it must only use things that are defined before 222 // the call, or movable instructions between the call and the instruction 223 // itself. 224 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 225 if (I->getOperand(i) == CI) 226 return false; 227 return true; 228} 229 230// isDynamicConstant - Return true if the specified value is the same when the 231// return would exit as it was when the initial iteration of the recursive 232// function was executed. 233// 234// We currently handle static constants and arguments that are not modified as 235// part of the recursion. 236// 237static bool isDynamicConstant(Value *V, CallInst *CI, ReturnInst *RI) { 238 if (isa<Constant>(V)) return true; // Static constants are always dyn consts 239 240 // Check to see if this is an immutable argument, if so, the value 241 // will be available to initialize the accumulator. 242 if (Argument *Arg = dyn_cast<Argument>(V)) { 243 // Figure out which argument number this is... 244 unsigned ArgNo = 0; 245 Function *F = CI->getParent()->getParent(); 246 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI) 247 ++ArgNo; 248 249 // If we are passing this argument into call as the corresponding 250 // argument operand, then the argument is dynamically constant. 251 // Otherwise, we cannot transform this function safely. 252 if (CI->getOperand(ArgNo+1) == Arg) 253 return true; 254 } 255 256 // Switch cases are always constant integers. If the value is being switched 257 // on and the return is only reachable from one of its cases, it's 258 // effectively constant. 259 if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor()) 260 if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator())) 261 if (SI->getCondition() == V) 262 return SI->getDefaultDest() != RI->getParent(); 263 264 // Not a constant or immutable argument, we can't safely transform. 265 return false; 266} 267 268// getCommonReturnValue - Check to see if the function containing the specified 269// return instruction and tail call consistently returns the same 270// runtime-constant value at all exit points. If so, return the returned value. 271// 272static Value *getCommonReturnValue(ReturnInst *TheRI, CallInst *CI) { 273 Function *F = TheRI->getParent()->getParent(); 274 Value *ReturnedValue = 0; 275 276 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) 277 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator())) 278 if (RI != TheRI) { 279 Value *RetOp = RI->getOperand(0); 280 281 // We can only perform this transformation if the value returned is 282 // evaluatable at the start of the initial invocation of the function, 283 // instead of at the end of the evaluation. 284 // 285 if (!isDynamicConstant(RetOp, CI, RI)) 286 return 0; 287 288 if (ReturnedValue && RetOp != ReturnedValue) 289 return 0; // Cannot transform if differing values are returned. 290 ReturnedValue = RetOp; 291 } 292 return ReturnedValue; 293} 294 295/// CanTransformAccumulatorRecursion - If the specified instruction can be 296/// transformed using accumulator recursion elimination, return the constant 297/// which is the start of the accumulator value. Otherwise return null. 298/// 299Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I, 300 CallInst *CI) { 301 if (!I->isAssociative()) return 0; 302 assert(I->getNumOperands() == 2 && 303 "Associative operations should have 2 args!"); 304 305 // Exactly one operand should be the result of the call instruction... 306 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) || 307 (I->getOperand(0) != CI && I->getOperand(1) != CI)) 308 return 0; 309 310 // The only user of this instruction we allow is a single return instruction. 311 if (!I->hasOneUse() || !isa<ReturnInst>(I->use_back())) 312 return 0; 313 314 // Ok, now we have to check all of the other return instructions in this 315 // function. If they return non-constants or differing values, then we cannot 316 // transform the function safely. 317 return getCommonReturnValue(cast<ReturnInst>(I->use_back()), CI); 318} 319 320bool TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry, 321 bool &TailCallsAreMarkedTail, 322 SmallVector<PHINode*, 8> &ArgumentPHIs, 323 bool CannotTailCallElimCallsMarkedTail) { 324 BasicBlock *BB = Ret->getParent(); 325 Function *F = BB->getParent(); 326 327 if (&BB->front() == Ret) // Make sure there is something before the ret... 328 return false; 329 330 // If the return is in the entry block, then making this transformation would 331 // turn infinite recursion into an infinite loop. This transformation is ok 332 // in theory, but breaks some code like: 333 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call 334 // disable this xform in this case, because the code generator will lower the 335 // call to fabs into inline code. 336 if (BB == &F->getEntryBlock()) 337 return false; 338 339 // Scan backwards from the return, checking to see if there is a tail call in 340 // this block. If so, set CI to it. 341 CallInst *CI; 342 BasicBlock::iterator BBI = Ret; 343 while (1) { 344 CI = dyn_cast<CallInst>(BBI); 345 if (CI && CI->getCalledFunction() == F) 346 break; 347 348 if (BBI == BB->begin()) 349 return false; // Didn't find a potential tail call. 350 --BBI; 351 } 352 353 // If this call is marked as a tail call, and if there are dynamic allocas in 354 // the function, we cannot perform this optimization. 355 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail) 356 return false; 357 358 // If we are introducing accumulator recursion to eliminate associative 359 // operations after the call instruction, this variable contains the initial 360 // value for the accumulator. If this value is set, we actually perform 361 // accumulator recursion elimination instead of simple tail recursion 362 // elimination. 363 Value *AccumulatorRecursionEliminationInitVal = 0; 364 Instruction *AccumulatorRecursionInstr = 0; 365 366 // Ok, we found a potential tail call. We can currently only transform the 367 // tail call if all of the instructions between the call and the return are 368 // movable to above the call itself, leaving the call next to the return. 369 // Check that this is the case now. 370 for (BBI = CI, ++BBI; &*BBI != Ret; ++BBI) 371 if (!CanMoveAboveCall(BBI, CI)) { 372 // If we can't move the instruction above the call, it might be because it 373 // is an associative operation that could be tranformed using accumulator 374 // recursion elimination. Check to see if this is the case, and if so, 375 // remember the initial accumulator value for later. 376 if ((AccumulatorRecursionEliminationInitVal = 377 CanTransformAccumulatorRecursion(BBI, CI))) { 378 // Yes, this is accumulator recursion. Remember which instruction 379 // accumulates. 380 AccumulatorRecursionInstr = BBI; 381 } else { 382 return false; // Otherwise, we cannot eliminate the tail recursion! 383 } 384 } 385 386 // We can only transform call/return pairs that either ignore the return value 387 // of the call and return void, ignore the value of the call and return a 388 // constant, return the value returned by the tail call, or that are being 389 // accumulator recursion variable eliminated. 390 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI && 391 !isa<UndefValue>(Ret->getReturnValue()) && 392 AccumulatorRecursionEliminationInitVal == 0 && 393 !getCommonReturnValue(Ret, CI)) 394 return false; 395 396 // OK! We can transform this tail call. If this is the first one found, 397 // create the new entry block, allowing us to branch back to the old entry. 398 if (OldEntry == 0) { 399 OldEntry = &F->getEntryBlock(); 400 BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry); 401 NewEntry->takeName(OldEntry); 402 OldEntry->setName("tailrecurse"); 403 BranchInst::Create(OldEntry, NewEntry); 404 405 // If this tail call is marked 'tail' and if there are any allocas in the 406 // entry block, move them up to the new entry block. 407 TailCallsAreMarkedTail = CI->isTailCall(); 408 if (TailCallsAreMarkedTail) 409 // Move all fixed sized allocas from OldEntry to NewEntry. 410 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(), 411 NEBI = NewEntry->begin(); OEBI != E; ) 412 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++)) 413 if (isa<ConstantInt>(AI->getArraySize())) 414 AI->moveBefore(NEBI); 415 416 // Now that we have created a new block, which jumps to the entry 417 // block, insert a PHI node for each argument of the function. 418 // For now, we initialize each PHI to only have the real arguments 419 // which are passed in. 420 Instruction *InsertPos = OldEntry->begin(); 421 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 422 I != E; ++I) { 423 PHINode *PN = PHINode::Create(I->getType(), 424 I->getName() + ".tr", InsertPos); 425 I->replaceAllUsesWith(PN); // Everyone use the PHI node now! 426 PN->addIncoming(I, NewEntry); 427 ArgumentPHIs.push_back(PN); 428 } 429 } 430 431 // If this function has self recursive calls in the tail position where some 432 // are marked tail and some are not, only transform one flavor or another. We 433 // have to choose whether we move allocas in the entry block to the new entry 434 // block or not, so we can't make a good choice for both. NOTE: We could do 435 // slightly better here in the case that the function has no entry block 436 // allocas. 437 if (TailCallsAreMarkedTail && !CI->isTailCall()) 438 return false; 439 440 // Ok, now that we know we have a pseudo-entry block WITH all of the 441 // required PHI nodes, add entries into the PHI node for the actual 442 // parameters passed into the tail-recursive call. 443 for (unsigned i = 0, e = CI->getNumOperands()-1; i != e; ++i) 444 ArgumentPHIs[i]->addIncoming(CI->getOperand(i+1), BB); 445 446 // If we are introducing an accumulator variable to eliminate the recursion, 447 // do so now. Note that we _know_ that no subsequent tail recursion 448 // eliminations will happen on this function because of the way the 449 // accumulator recursion predicate is set up. 450 // 451 if (AccumulatorRecursionEliminationInitVal) { 452 Instruction *AccRecInstr = AccumulatorRecursionInstr; 453 // Start by inserting a new PHI node for the accumulator. 454 PHINode *AccPN = PHINode::Create(AccRecInstr->getType(), "accumulator.tr", 455 OldEntry->begin()); 456 457 // Loop over all of the predecessors of the tail recursion block. For the 458 // real entry into the function we seed the PHI with the initial value, 459 // computed earlier. For any other existing branches to this block (due to 460 // other tail recursions eliminated) the accumulator is not modified. 461 // Because we haven't added the branch in the current block to OldEntry yet, 462 // it will not show up as a predecessor. 463 for (pred_iterator PI = pred_begin(OldEntry), PE = pred_end(OldEntry); 464 PI != PE; ++PI) { 465 if (*PI == &F->getEntryBlock()) 466 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, *PI); 467 else 468 AccPN->addIncoming(AccPN, *PI); 469 } 470 471 // Add an incoming argument for the current block, which is computed by our 472 // associative accumulator instruction. 473 AccPN->addIncoming(AccRecInstr, BB); 474 475 // Next, rewrite the accumulator recursion instruction so that it does not 476 // use the result of the call anymore, instead, use the PHI node we just 477 // inserted. 478 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN); 479 480 // Finally, rewrite any return instructions in the program to return the PHI 481 // node instead of the "initval" that they do currently. This loop will 482 // actually rewrite the return value we are destroying, but that's ok. 483 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) 484 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator())) 485 RI->setOperand(0, AccPN); 486 ++NumAccumAdded; 487 } 488 489 // Now that all of the PHI nodes are in place, remove the call and 490 // ret instructions, replacing them with an unconditional branch. 491 BranchInst::Create(OldEntry, Ret); 492 BB->getInstList().erase(Ret); // Remove return. 493 BB->getInstList().erase(CI); // Remove call. 494 ++NumEliminated; 495 return true; 496} 497