TailRecursionElimination.cpp revision cdbd99262286e96729007ac535cd430ecb3d38ac
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 and commutative expression to use an 20// accumulator variable, thus compiling the typical naive factorial or 21// 'fib' implementation into 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/Analysis/InlineCost.h" 63#include "llvm/Analysis/InstructionSimplify.h" 64#include "llvm/Analysis/Loads.h" 65#include "llvm/Support/CallSite.h" 66#include "llvm/Support/CFG.h" 67#include "llvm/ADT/Statistic.h" 68using namespace llvm; 69 70STATISTIC(NumEliminated, "Number of tail calls removed"); 71STATISTIC(NumAccumAdded, "Number of accumulators introduced"); 72 73namespace { 74 struct TailCallElim : public FunctionPass { 75 static char ID; // Pass identification, replacement for typeid 76 TailCallElim() : FunctionPass(ID) { 77 initializeTailCallElimPass(*PassRegistry::getPassRegistry()); 78 } 79 80 virtual bool runOnFunction(Function &F); 81 82 private: 83 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry, 84 bool &TailCallsAreMarkedTail, 85 SmallVector<PHINode*, 8> &ArgumentPHIs, 86 bool CannotTailCallElimCallsMarkedTail); 87 bool CanMoveAboveCall(Instruction *I, CallInst *CI); 88 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI); 89 }; 90} 91 92char TailCallElim::ID = 0; 93INITIALIZE_PASS(TailCallElim, "tailcallelim", 94 "Tail Call Elimination", false, false) 95 96// Public interface to the TailCallElimination pass 97FunctionPass *llvm::createTailCallEliminationPass() { 98 return new TailCallElim(); 99} 100 101/// AllocaMightEscapeToCalls - Return true if this alloca may be accessed by 102/// callees of this function. We only do very simple analysis right now, this 103/// could be expanded in the future to use mod/ref information for particular 104/// call sites if desired. 105static bool AllocaMightEscapeToCalls(AllocaInst *AI) { 106 // FIXME: do simple 'address taken' analysis. 107 return true; 108} 109 110/// CheckForEscapingAllocas - Scan the specified basic block for alloca 111/// instructions. If it contains any that might be accessed by calls, return 112/// true. 113static bool CheckForEscapingAllocas(BasicBlock *BB, 114 bool &CannotTCETailMarkedCall) { 115 bool RetVal = false; 116 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 117 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) { 118 RetVal |= AllocaMightEscapeToCalls(AI); 119 120 // If this alloca is in the body of the function, or if it is a variable 121 // sized allocation, we cannot tail call eliminate calls marked 'tail' 122 // with this mechanism. 123 if (BB != &BB->getParent()->getEntryBlock() || 124 !isa<ConstantInt>(AI->getArraySize())) 125 CannotTCETailMarkedCall = true; 126 } 127 return RetVal; 128} 129 130bool TailCallElim::runOnFunction(Function &F) { 131 // If this function is a varargs function, we won't be able to PHI the args 132 // right, so don't even try to convert it... 133 if (F.getFunctionType()->isVarArg()) return false; 134 135 BasicBlock *OldEntry = 0; 136 bool TailCallsAreMarkedTail = false; 137 SmallVector<PHINode*, 8> ArgumentPHIs; 138 bool MadeChange = false; 139 140 bool FunctionContainsEscapingAllocas = false; 141 142 // CannotTCETailMarkedCall - If true, we cannot perform TCE on tail calls 143 // marked with the 'tail' attribute, because doing so would cause the stack 144 // size to increase (real TCE would deallocate variable sized allocas, TCE 145 // doesn't). 146 bool CannotTCETailMarkedCall = false; 147 148 // Loop over the function, looking for any returning blocks, and keeping track 149 // of whether this function has any non-trivially used allocas. 150 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 151 if (FunctionContainsEscapingAllocas && CannotTCETailMarkedCall) 152 break; 153 154 FunctionContainsEscapingAllocas |= 155 CheckForEscapingAllocas(BB, CannotTCETailMarkedCall); 156 } 157 158 /// FIXME: The code generator produces really bad code when an 'escaping 159 /// alloca' is changed from being a static alloca to being a dynamic alloca. 160 /// Until this is resolved, disable this transformation if that would ever 161 /// happen. This bug is PR962. 162 if (FunctionContainsEscapingAllocas) 163 return false; 164 165 // Second pass, change any tail calls to loops. 166 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 167 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) 168 MadeChange |= ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail, 169 ArgumentPHIs,CannotTCETailMarkedCall); 170 171 // If we eliminated any tail recursions, it's possible that we inserted some 172 // silly PHI nodes which just merge an initial value (the incoming operand) 173 // with themselves. Check to see if we did and clean up our mess if so. This 174 // occurs when a function passes an argument straight through to its tail 175 // call. 176 if (!ArgumentPHIs.empty()) { 177 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) { 178 PHINode *PN = ArgumentPHIs[i]; 179 180 // If the PHI Node is a dynamic constant, replace it with the value it is. 181 if (Value *PNV = SimplifyInstruction(PN)) { 182 PN->replaceAllUsesWith(PNV); 183 PN->eraseFromParent(); 184 } 185 } 186 } 187 188 // Finally, if this function contains no non-escaping allocas, mark all calls 189 // in the function as eligible for tail calls (there is no stack memory for 190 // them to access). 191 if (!FunctionContainsEscapingAllocas) 192 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 193 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 194 if (CallInst *CI = dyn_cast<CallInst>(I)) { 195 CI->setTailCall(); 196 MadeChange = true; 197 } 198 199 return MadeChange; 200} 201 202 203/// CanMoveAboveCall - Return true if it is safe to move the specified 204/// instruction from after the call to before the call, assuming that all 205/// instructions between the call and this instruction are movable. 206/// 207bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) { 208 // FIXME: We can move load/store/call/free instructions above the call if the 209 // call does not mod/ref the memory location being processed. 210 if (I->mayHaveSideEffects()) // This also handles volatile loads. 211 return false; 212 213 if (LoadInst *L = dyn_cast<LoadInst>(I)) { 214 // Loads may always be moved above calls without side effects. 215 if (CI->mayHaveSideEffects()) { 216 // Non-volatile loads may be moved above a call with side effects if it 217 // does not write to memory and the load provably won't trap. 218 // FIXME: Writes to memory only matter if they may alias the pointer 219 // being loaded from. 220 if (CI->mayWriteToMemory() || 221 !isSafeToLoadUnconditionally(L->getPointerOperand(), L, 222 L->getAlignment())) 223 return false; 224 } 225 } 226 227 // Otherwise, if this is a side-effect free instruction, check to make sure 228 // that it does not use the return value of the call. If it doesn't use the 229 // return value of the call, it must only use things that are defined before 230 // the call, or movable instructions between the call and the instruction 231 // itself. 232 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 233 if (I->getOperand(i) == CI) 234 return false; 235 return true; 236} 237 238// isDynamicConstant - Return true if the specified value is the same when the 239// return would exit as it was when the initial iteration of the recursive 240// function was executed. 241// 242// We currently handle static constants and arguments that are not modified as 243// part of the recursion. 244// 245static bool isDynamicConstant(Value *V, CallInst *CI, ReturnInst *RI) { 246 if (isa<Constant>(V)) return true; // Static constants are always dyn consts 247 248 // Check to see if this is an immutable argument, if so, the value 249 // will be available to initialize the accumulator. 250 if (Argument *Arg = dyn_cast<Argument>(V)) { 251 // Figure out which argument number this is... 252 unsigned ArgNo = 0; 253 Function *F = CI->getParent()->getParent(); 254 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI) 255 ++ArgNo; 256 257 // If we are passing this argument into call as the corresponding 258 // argument operand, then the argument is dynamically constant. 259 // Otherwise, we cannot transform this function safely. 260 if (CI->getArgOperand(ArgNo) == Arg) 261 return true; 262 } 263 264 // Switch cases are always constant integers. If the value is being switched 265 // on and the return is only reachable from one of its cases, it's 266 // effectively constant. 267 if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor()) 268 if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator())) 269 if (SI->getCondition() == V) 270 return SI->getDefaultDest() != RI->getParent(); 271 272 // Not a constant or immutable argument, we can't safely transform. 273 return false; 274} 275 276// getCommonReturnValue - Check to see if the function containing the specified 277// tail call consistently returns the same runtime-constant value at all exit 278// points except for IgnoreRI. If so, return the returned value. 279// 280static Value *getCommonReturnValue(ReturnInst *IgnoreRI, CallInst *CI) { 281 Function *F = CI->getParent()->getParent(); 282 Value *ReturnedValue = 0; 283 284 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) { 285 ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()); 286 if (RI == 0 || RI == IgnoreRI) continue; 287 288 // We can only perform this transformation if the value returned is 289 // evaluatable at the start of the initial invocation of the function, 290 // instead of at the end of the evaluation. 291 // 292 Value *RetOp = RI->getOperand(0); 293 if (!isDynamicConstant(RetOp, CI, RI)) 294 return 0; 295 296 if (ReturnedValue && RetOp != ReturnedValue) 297 return 0; // Cannot transform if differing values are returned. 298 ReturnedValue = RetOp; 299 } 300 return ReturnedValue; 301} 302 303/// CanTransformAccumulatorRecursion - If the specified instruction can be 304/// transformed using accumulator recursion elimination, return the constant 305/// which is the start of the accumulator value. Otherwise return null. 306/// 307Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I, 308 CallInst *CI) { 309 if (!I->isAssociative() || !I->isCommutative()) return 0; 310 assert(I->getNumOperands() == 2 && 311 "Associative/commutative operations should have 2 args!"); 312 313 // Exactly one operand should be the result of the call instruction. 314 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) || 315 (I->getOperand(0) != CI && I->getOperand(1) != CI)) 316 return 0; 317 318 // The only user of this instruction we allow is a single return instruction. 319 if (!I->hasOneUse() || !isa<ReturnInst>(I->use_back())) 320 return 0; 321 322 // Ok, now we have to check all of the other return instructions in this 323 // function. If they return non-constants or differing values, then we cannot 324 // transform the function safely. 325 return getCommonReturnValue(cast<ReturnInst>(I->use_back()), CI); 326} 327 328bool TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry, 329 bool &TailCallsAreMarkedTail, 330 SmallVector<PHINode*, 8> &ArgumentPHIs, 331 bool CannotTailCallElimCallsMarkedTail) { 332 BasicBlock *BB = Ret->getParent(); 333 Function *F = BB->getParent(); 334 335 if (&BB->front() == Ret) // Make sure there is something before the ret... 336 return false; 337 338 // Scan backwards from the return, checking to see if there is a tail call in 339 // this block. If so, set CI to it. 340 CallInst *CI; 341 BasicBlock::iterator BBI = Ret; 342 while (1) { 343 CI = dyn_cast<CallInst>(BBI); 344 if (CI && CI->getCalledFunction() == F) 345 break; 346 347 if (BBI == BB->begin()) 348 return false; // Didn't find a potential tail call. 349 --BBI; 350 } 351 352 // If this call is marked as a tail call, and if there are dynamic allocas in 353 // the function, we cannot perform this optimization. 354 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail) 355 return false; 356 357 // As a special case, detect code like this: 358 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call 359 // and disable this xform in this case, because the code generator will 360 // lower the call to fabs into inline code. 361 if (BB == &F->getEntryBlock() && 362 &BB->front() == CI && &*++BB->begin() == Ret && 363 callIsSmall(F)) { 364 // A single-block function with just a call and a return. Check that 365 // the arguments match. 366 CallSite::arg_iterator I = CallSite(CI).arg_begin(), 367 E = CallSite(CI).arg_end(); 368 Function::arg_iterator FI = F->arg_begin(), 369 FE = F->arg_end(); 370 for (; I != E && FI != FE; ++I, ++FI) 371 if (*I != &*FI) break; 372 if (I == E && FI == FE) 373 return false; 374 } 375 376 // If we are introducing accumulator recursion to eliminate operations after 377 // the call instruction that are both associative and commutative, the initial 378 // value for the accumulator is placed in this variable. If this value is set 379 // then we actually perform accumulator recursion elimination instead of 380 // simple tail recursion elimination. If the operation is an LLVM instruction 381 // (eg: "add") then it is recorded in AccumulatorRecursionInstr. If not, then 382 // we are handling the case when the return instruction returns a constant C 383 // which is different to the constant returned by other return instructions 384 // (which is recorded in AccumulatorRecursionEliminationInitVal). This is a 385 // special case of accumulator recursion, the operation being "return C". 386 Value *AccumulatorRecursionEliminationInitVal = 0; 387 Instruction *AccumulatorRecursionInstr = 0; 388 389 // Ok, we found a potential tail call. We can currently only transform the 390 // tail call if all of the instructions between the call and the return are 391 // movable to above the call itself, leaving the call next to the return. 392 // Check that this is the case now. 393 for (BBI = CI, ++BBI; &*BBI != Ret; ++BBI) { 394 if (CanMoveAboveCall(BBI, CI)) continue; 395 396 // If we can't move the instruction above the call, it might be because it 397 // is an associative and commutative operation that could be tranformed 398 // using accumulator recursion elimination. Check to see if this is the 399 // case, and if so, remember the initial accumulator value for later. 400 if ((AccumulatorRecursionEliminationInitVal = 401 CanTransformAccumulatorRecursion(BBI, CI))) { 402 // Yes, this is accumulator recursion. Remember which instruction 403 // accumulates. 404 AccumulatorRecursionInstr = BBI; 405 } else { 406 return false; // Otherwise, we cannot eliminate the tail recursion! 407 } 408 } 409 410 // We can only transform call/return pairs that either ignore the return value 411 // of the call and return void, ignore the value of the call and return a 412 // constant, return the value returned by the tail call, or that are being 413 // accumulator recursion variable eliminated. 414 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI && 415 !isa<UndefValue>(Ret->getReturnValue()) && 416 AccumulatorRecursionEliminationInitVal == 0 && 417 !getCommonReturnValue(0, CI)) { 418 // One case remains that we are able to handle: the current return 419 // instruction returns a constant, and all other return instructions 420 // return a different constant. 421 if (!isDynamicConstant(Ret->getReturnValue(), CI, Ret)) 422 return false; // Current return instruction does not return a constant. 423 // Check that all other return instructions return a common constant. If 424 // so, record it in AccumulatorRecursionEliminationInitVal. 425 AccumulatorRecursionEliminationInitVal = getCommonReturnValue(Ret, CI); 426 if (!AccumulatorRecursionEliminationInitVal) 427 return false; 428 } 429 430 // OK! We can transform this tail call. If this is the first one found, 431 // create the new entry block, allowing us to branch back to the old entry. 432 if (OldEntry == 0) { 433 OldEntry = &F->getEntryBlock(); 434 BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry); 435 NewEntry->takeName(OldEntry); 436 OldEntry->setName("tailrecurse"); 437 BranchInst::Create(OldEntry, NewEntry); 438 439 // If this tail call is marked 'tail' and if there are any allocas in the 440 // entry block, move them up to the new entry block. 441 TailCallsAreMarkedTail = CI->isTailCall(); 442 if (TailCallsAreMarkedTail) 443 // Move all fixed sized allocas from OldEntry to NewEntry. 444 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(), 445 NEBI = NewEntry->begin(); OEBI != E; ) 446 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++)) 447 if (isa<ConstantInt>(AI->getArraySize())) 448 AI->moveBefore(NEBI); 449 450 // Now that we have created a new block, which jumps to the entry 451 // block, insert a PHI node for each argument of the function. 452 // For now, we initialize each PHI to only have the real arguments 453 // which are passed in. 454 Instruction *InsertPos = OldEntry->begin(); 455 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 456 I != E; ++I) { 457 PHINode *PN = PHINode::Create(I->getType(), 458 I->getName() + ".tr", InsertPos); 459 I->replaceAllUsesWith(PN); // Everyone use the PHI node now! 460 PN->addIncoming(I, NewEntry); 461 ArgumentPHIs.push_back(PN); 462 } 463 } 464 465 // If this function has self recursive calls in the tail position where some 466 // are marked tail and some are not, only transform one flavor or another. We 467 // have to choose whether we move allocas in the entry block to the new entry 468 // block or not, so we can't make a good choice for both. NOTE: We could do 469 // slightly better here in the case that the function has no entry block 470 // allocas. 471 if (TailCallsAreMarkedTail && !CI->isTailCall()) 472 return false; 473 474 // Ok, now that we know we have a pseudo-entry block WITH all of the 475 // required PHI nodes, add entries into the PHI node for the actual 476 // parameters passed into the tail-recursive call. 477 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) 478 ArgumentPHIs[i]->addIncoming(CI->getArgOperand(i), BB); 479 480 // If we are introducing an accumulator variable to eliminate the recursion, 481 // do so now. Note that we _know_ that no subsequent tail recursion 482 // eliminations will happen on this function because of the way the 483 // accumulator recursion predicate is set up. 484 // 485 if (AccumulatorRecursionEliminationInitVal) { 486 Instruction *AccRecInstr = AccumulatorRecursionInstr; 487 // Start by inserting a new PHI node for the accumulator. 488 PHINode *AccPN = 489 PHINode::Create(AccumulatorRecursionEliminationInitVal->getType(), 490 "accumulator.tr", OldEntry->begin()); 491 492 // Loop over all of the predecessors of the tail recursion block. For the 493 // real entry into the function we seed the PHI with the initial value, 494 // computed earlier. For any other existing branches to this block (due to 495 // other tail recursions eliminated) the accumulator is not modified. 496 // Because we haven't added the branch in the current block to OldEntry yet, 497 // it will not show up as a predecessor. 498 for (pred_iterator PI = pred_begin(OldEntry), PE = pred_end(OldEntry); 499 PI != PE; ++PI) { 500 BasicBlock *P = *PI; 501 if (P == &F->getEntryBlock()) 502 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, P); 503 else 504 AccPN->addIncoming(AccPN, P); 505 } 506 507 if (AccRecInstr) { 508 // Add an incoming argument for the current block, which is computed by 509 // our associative and commutative accumulator instruction. 510 AccPN->addIncoming(AccRecInstr, BB); 511 512 // Next, rewrite the accumulator recursion instruction so that it does not 513 // use the result of the call anymore, instead, use the PHI node we just 514 // inserted. 515 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN); 516 } else { 517 // Add an incoming argument for the current block, which is just the 518 // constant returned by the current return instruction. 519 AccPN->addIncoming(Ret->getReturnValue(), BB); 520 } 521 522 // Finally, rewrite any return instructions in the program to return the PHI 523 // node instead of the "initval" that they do currently. This loop will 524 // actually rewrite the return value we are destroying, but that's ok. 525 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) 526 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator())) 527 RI->setOperand(0, AccPN); 528 ++NumAccumAdded; 529 } 530 531 // Now that all of the PHI nodes are in place, remove the call and 532 // ret instructions, replacing them with an unconditional branch. 533 BranchInst::Create(OldEntry, Ret); 534 BB->getInstList().erase(Ret); // Remove return. 535 BB->getInstList().erase(CI); // Remove call. 536 ++NumEliminated; 537 return true; 538} 539