TailRecursionElimination.cpp revision 9ee17208115482441953127615231c59a2f4d052
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 L->getAlignment())) 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, ReturnInst *RI) { 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 } 256 257 // Switch cases are always constant integers. If the value is being switched 258 // on and the return is only reachable from one of its cases, it's 259 // effectively constant. 260 if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor()) 261 if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator())) 262 if (SI->getCondition() == V) 263 return SI->getDefaultDest() != RI->getParent(); 264 265 // Not a constant or immutable argument, we can't safely transform. 266 return false; 267} 268 269// getCommonReturnValue - Check to see if the function containing the specified 270// return instruction and tail call consistently returns the same 271// runtime-constant value at all exit points. If so, return the returned value. 272// 273static Value *getCommonReturnValue(ReturnInst *TheRI, CallInst *CI) { 274 Function *F = TheRI->getParent()->getParent(); 275 Value *ReturnedValue = 0; 276 277 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) 278 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator())) 279 if (RI != TheRI) { 280 Value *RetOp = RI->getOperand(0); 281 282 // We can only perform this transformation if the value returned is 283 // evaluatable at the start of the initial invocation of the function, 284 // instead of at the end of the evaluation. 285 // 286 if (!isDynamicConstant(RetOp, CI, RI)) 287 return 0; 288 289 if (ReturnedValue && RetOp != ReturnedValue) 290 return 0; // Cannot transform if differing values are returned. 291 ReturnedValue = RetOp; 292 } 293 return ReturnedValue; 294} 295 296/// CanTransformAccumulatorRecursion - If the specified instruction can be 297/// transformed using accumulator recursion elimination, return the constant 298/// which is the start of the accumulator value. Otherwise return null. 299/// 300Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I, 301 CallInst *CI) { 302 if (!I->isAssociative()) return 0; 303 assert(I->getNumOperands() == 2 && 304 "Associative operations should have 2 args!"); 305 306 // Exactly one operand should be the result of the call instruction... 307 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) || 308 (I->getOperand(0) != CI && I->getOperand(1) != CI)) 309 return 0; 310 311 // The only user of this instruction we allow is a single return instruction. 312 if (!I->hasOneUse() || !isa<ReturnInst>(I->use_back())) 313 return 0; 314 315 // Ok, now we have to check all of the other return instructions in this 316 // function. If they return non-constants or differing values, then we cannot 317 // transform the function safely. 318 return getCommonReturnValue(cast<ReturnInst>(I->use_back()), CI); 319} 320 321bool TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry, 322 bool &TailCallsAreMarkedTail, 323 SmallVector<PHINode*, 8> &ArgumentPHIs, 324 bool CannotTailCallElimCallsMarkedTail) { 325 BasicBlock *BB = Ret->getParent(); 326 Function *F = BB->getParent(); 327 328 if (&BB->front() == Ret) // Make sure there is something before the ret... 329 return false; 330 331 // If the return is in the entry block, then making this transformation would 332 // turn infinite recursion into an infinite loop. This transformation is ok 333 // in theory, but breaks some code like: 334 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call 335 // disable this xform in this case, because the code generator will lower the 336 // call to fabs into inline code. 337 if (BB == &F->getEntryBlock()) 338 return false; 339 340 // Scan backwards from the return, checking to see if there is a tail call in 341 // this block. If so, set CI to it. 342 CallInst *CI; 343 BasicBlock::iterator BBI = Ret; 344 while (1) { 345 CI = dyn_cast<CallInst>(BBI); 346 if (CI && CI->getCalledFunction() == F) 347 break; 348 349 if (BBI == BB->begin()) 350 return false; // Didn't find a potential tail call. 351 --BBI; 352 } 353 354 // If this call is marked as a tail call, and if there are dynamic allocas in 355 // the function, we cannot perform this optimization. 356 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail) 357 return false; 358 359 // If we are introducing accumulator recursion to eliminate associative 360 // operations after the call instruction, this variable contains the initial 361 // value for the accumulator. If this value is set, we actually perform 362 // accumulator recursion elimination instead of simple tail recursion 363 // elimination. 364 Value *AccumulatorRecursionEliminationInitVal = 0; 365 Instruction *AccumulatorRecursionInstr = 0; 366 367 // Ok, we found a potential tail call. We can currently only transform the 368 // tail call if all of the instructions between the call and the return are 369 // movable to above the call itself, leaving the call next to the return. 370 // Check that this is the case now. 371 for (BBI = CI, ++BBI; &*BBI != Ret; ++BBI) 372 if (!CanMoveAboveCall(BBI, CI)) { 373 // If we can't move the instruction above the call, it might be because it 374 // is an associative operation that could be tranformed using accumulator 375 // recursion elimination. Check to see if this is the case, and if so, 376 // remember the initial accumulator value for later. 377 if ((AccumulatorRecursionEliminationInitVal = 378 CanTransformAccumulatorRecursion(BBI, CI))) { 379 // Yes, this is accumulator recursion. Remember which instruction 380 // accumulates. 381 AccumulatorRecursionInstr = BBI; 382 } else { 383 return false; // Otherwise, we cannot eliminate the tail recursion! 384 } 385 } 386 387 // We can only transform call/return pairs that either ignore the return value 388 // of the call and return void, ignore the value of the call and return a 389 // constant, return the value returned by the tail call, or that are being 390 // accumulator recursion variable eliminated. 391 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI && 392 !isa<UndefValue>(Ret->getReturnValue()) && 393 AccumulatorRecursionEliminationInitVal == 0 && 394 !getCommonReturnValue(Ret, CI)) 395 return false; 396 397 // OK! We can transform this tail call. If this is the first one found, 398 // create the new entry block, allowing us to branch back to the old entry. 399 if (OldEntry == 0) { 400 OldEntry = &F->getEntryBlock(); 401 BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry); 402 NewEntry->takeName(OldEntry); 403 OldEntry->setName("tailrecurse"); 404 BranchInst::Create(OldEntry, NewEntry); 405 406 // If this tail call is marked 'tail' and if there are any allocas in the 407 // entry block, move them up to the new entry block. 408 TailCallsAreMarkedTail = CI->isTailCall(); 409 if (TailCallsAreMarkedTail) 410 // Move all fixed sized allocas from OldEntry to NewEntry. 411 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(), 412 NEBI = NewEntry->begin(); OEBI != E; ) 413 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++)) 414 if (isa<ConstantInt>(AI->getArraySize())) 415 AI->moveBefore(NEBI); 416 417 // Now that we have created a new block, which jumps to the entry 418 // block, insert a PHI node for each argument of the function. 419 // For now, we initialize each PHI to only have the real arguments 420 // which are passed in. 421 Instruction *InsertPos = OldEntry->begin(); 422 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 423 I != E; ++I) { 424 PHINode *PN = PHINode::Create(I->getType(), 425 I->getName() + ".tr", InsertPos); 426 I->replaceAllUsesWith(PN); // Everyone use the PHI node now! 427 PN->addIncoming(I, NewEntry); 428 ArgumentPHIs.push_back(PN); 429 } 430 } 431 432 // If this function has self recursive calls in the tail position where some 433 // are marked tail and some are not, only transform one flavor or another. We 434 // have to choose whether we move allocas in the entry block to the new entry 435 // block or not, so we can't make a good choice for both. NOTE: We could do 436 // slightly better here in the case that the function has no entry block 437 // allocas. 438 if (TailCallsAreMarkedTail && !CI->isTailCall()) 439 return false; 440 441 // Ok, now that we know we have a pseudo-entry block WITH all of the 442 // required PHI nodes, add entries into the PHI node for the actual 443 // parameters passed into the tail-recursive call. 444 for (unsigned i = 0, e = CI->getNumOperands()-1; i != e; ++i) 445 ArgumentPHIs[i]->addIncoming(CI->getOperand(i+1), BB); 446 447 // If we are introducing an accumulator variable to eliminate the recursion, 448 // do so now. Note that we _know_ that no subsequent tail recursion 449 // eliminations will happen on this function because of the way the 450 // accumulator recursion predicate is set up. 451 // 452 if (AccumulatorRecursionEliminationInitVal) { 453 Instruction *AccRecInstr = AccumulatorRecursionInstr; 454 // Start by inserting a new PHI node for the accumulator. 455 PHINode *AccPN = PHINode::Create(AccRecInstr->getType(), "accumulator.tr", 456 OldEntry->begin()); 457 458 // Loop over all of the predecessors of the tail recursion block. For the 459 // real entry into the function we seed the PHI with the initial value, 460 // computed earlier. For any other existing branches to this block (due to 461 // other tail recursions eliminated) the accumulator is not modified. 462 // Because we haven't added the branch in the current block to OldEntry yet, 463 // it will not show up as a predecessor. 464 for (pred_iterator PI = pred_begin(OldEntry), PE = pred_end(OldEntry); 465 PI != PE; ++PI) { 466 if (*PI == &F->getEntryBlock()) 467 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, *PI); 468 else 469 AccPN->addIncoming(AccPN, *PI); 470 } 471 472 // Add an incoming argument for the current block, which is computed by our 473 // associative accumulator instruction. 474 AccPN->addIncoming(AccRecInstr, BB); 475 476 // Next, rewrite the accumulator recursion instruction so that it does not 477 // use the result of the call anymore, instead, use the PHI node we just 478 // inserted. 479 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN); 480 481 // Finally, rewrite any return instructions in the program to return the PHI 482 // node instead of the "initval" that they do currently. This loop will 483 // actually rewrite the return value we are destroying, but that's ok. 484 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) 485 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator())) 486 RI->setOperand(0, AccPN); 487 ++NumAccumAdded; 488 } 489 490 // Now that all of the PHI nodes are in place, remove the call and 491 // ret instructions, replacing them with an unconditional branch. 492 BranchInst::Create(OldEntry, Ret); 493 BB->getInstList().erase(Ret); // Remove return. 494 BB->getInstList().erase(CI); // Remove call. 495 ++NumEliminated; 496 return true; 497} 498