TailRecursionElimination.cpp revision d64152a70842b2f4186aa912938e69ca09c1434c
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// 29// There are several improvements that could be made: 30// 31// 1. If the function has any alloca instructions, these instructions will be 32// moved out of the entry block of the function, causing them to be 33// evaluated each time through the tail recursion. Safely keeping allocas 34// in the entry block requires analysis to proves that the tail-called 35// function does not read or write the stack object. 36// 2. Tail recursion is only performed if the call immediately preceeds the 37// return instruction. It's possible that there could be a jump between 38// the call and the return. 39// 3. There can be intervening operations between the call and the return that 40// prevent the TRE from occurring. For example, there could be GEP's and 41// stores to memory that will not be read or written by the call. This 42// requires some substantial analysis (such as with DSA) to prove safe to 43// move ahead of the call, but doing so could allow many more TREs to be 44// performed, for example in TreeAdd/TreeAlloc from the treeadd benchmark. 45// 46//===----------------------------------------------------------------------===// 47 48#include "llvm/Transforms/Scalar.h" 49#include "llvm/DerivedTypes.h" 50#include "llvm/Function.h" 51#include "llvm/Instructions.h" 52#include "llvm/Pass.h" 53#include "llvm/Support/CFG.h" 54#include "Support/Statistic.h" 55using namespace llvm; 56 57namespace { 58 Statistic<> NumEliminated("tailcallelim", "Number of tail calls removed"); 59 Statistic<> NumAccumAdded("tailcallelim","Number of accumulators introduced"); 60 61 struct TailCallElim : public FunctionPass { 62 virtual bool runOnFunction(Function &F); 63 64 private: 65 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry, 66 std::vector<PHINode*> &ArgumentPHIs); 67 bool CanMoveAboveCall(Instruction *I, CallInst *CI); 68 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI); 69 }; 70 RegisterOpt<TailCallElim> X("tailcallelim", "Tail Call Elimination"); 71} 72 73// Public interface to the TailCallElimination pass 74FunctionPass *llvm::createTailCallEliminationPass() { 75 return new TailCallElim(); 76} 77 78 79bool TailCallElim::runOnFunction(Function &F) { 80 // If this function is a varargs function, we won't be able to PHI the args 81 // right, so don't even try to convert it... 82 if (F.getFunctionType()->isVarArg()) return false; 83 84 BasicBlock *OldEntry = 0; 85 std::vector<PHINode*> ArgumentPHIs; 86 bool MadeChange = false; 87 88 // Loop over the function, looking for any returning blocks... 89 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 90 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) 91 MadeChange |= ProcessReturningBlock(Ret, OldEntry, ArgumentPHIs); 92 93 // If we eliminated any tail recursions, it's possible that we inserted some 94 // silly PHI nodes which just merge an initial value (the incoming operand) 95 // with themselves. Check to see if we did and clean up our mess if so. This 96 // occurs when a function passes an argument straight through to its tail 97 // call. 98 if (!ArgumentPHIs.empty()) { 99 unsigned NumIncoming = ArgumentPHIs[0]->getNumIncomingValues(); 100 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) { 101 PHINode *PN = ArgumentPHIs[i]; 102 Value *V = 0; 103 for (unsigned op = 0, e = NumIncoming; op != e; ++op) { 104 Value *Op = PN->getIncomingValue(op); 105 if (Op != PN) { 106 if (V == 0) { 107 V = Op; // First value seen? 108 } else if (V != Op) { 109 V = 0; 110 break; 111 } 112 } 113 } 114 115 // If the PHI Node is a dynamic constant, replace it with the value it is. 116 if (V) { 117 PN->replaceAllUsesWith(V); 118 PN->getParent()->getInstList().erase(PN); 119 } 120 } 121 } 122 123 return MadeChange; 124} 125 126 127/// CanMoveAboveCall - Return true if it is safe to move the specified 128/// instruction from after the call to before the call, assuming that all 129/// instructions between the call and this instruction are movable. 130/// 131bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) { 132 // FIXME: We can move load/store/call/free instructions above the call if the 133 // call does not mod/ref the memory location being processed. 134 if (I->mayWriteToMemory() || isa<LoadInst>(I)) 135 return false; 136 137 // Otherwise, if this is a side-effect free instruction, check to make sure 138 // that it does not use the return value of the call. If it doesn't use the 139 // return value of the call, it must only use things that are defined before 140 // the call, or movable instructions between the call and the instruction 141 // itself. 142 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 143 if (I->getOperand(i) == CI) 144 return false; 145 return true; 146} 147 148// isDynamicConstant - Return true if the specified value is the same when the 149// return would exit as it was when the initial iteration of the recursive 150// function was executed. 151// 152// We currently handle static constants and arguments that are not modified as 153// part of the recursion. 154// 155static bool isDynamicConstant(Value *V, CallInst *CI) { 156 if (isa<Constant>(V)) return true; // Static constants are always dyn consts 157 158 // Check to see if this is an immutable argument, if so, the value 159 // will be available to initialize the accumulator. 160 if (Argument *Arg = dyn_cast<Argument>(V)) { 161 // Figure out which argument number this is... 162 unsigned ArgNo = 0; 163 Function *F = CI->getParent()->getParent(); 164 for (Function::aiterator AI = F->abegin(); &*AI != Arg; ++AI) 165 ++ArgNo; 166 167 // If we are passing this argument into call as the corresponding 168 // argument operand, then the argument is dynamically constant. 169 // Otherwise, we cannot transform this function safely. 170 if (CI->getOperand(ArgNo+1) == Arg) 171 return true; 172 } 173 // Not a constant or immutable argument, we can't safely transform. 174 return false; 175} 176 177// getCommonReturnValue - Check to see if the function containing the specified 178// return instruction and tail call consistently returns the same 179// runtime-constant value at all exit points. If so, return the returned value. 180// 181static Value *getCommonReturnValue(ReturnInst *TheRI, CallInst *CI) { 182 Function *F = TheRI->getParent()->getParent(); 183 Value *ReturnedValue = 0; 184 185 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) 186 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator())) 187 if (RI != TheRI) { 188 Value *RetOp = RI->getOperand(0); 189 190 // We can only perform this transformation if the value returned is 191 // evaluatable at the start of the initial invocation of the function, 192 // instead of at the end of the evaluation. 193 // 194 if (!isDynamicConstant(RetOp, CI)) 195 return 0; 196 197 if (ReturnedValue && RetOp != ReturnedValue) 198 return 0; // Cannot transform if differing values are returned. 199 ReturnedValue = RetOp; 200 } 201 return ReturnedValue; 202} 203 204/// CanTransformAccumulatorRecursion - If the specified instruction can be 205/// transformed using accumulator recursion elimination, return the constant 206/// which is the start of the accumulator value. Otherwise return null. 207/// 208Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I, 209 CallInst *CI) { 210 if (!I->isAssociative()) return 0; 211 assert(I->getNumOperands() == 2 && 212 "Associative operations should have 2 args!"); 213 214 // Exactly one operand should be the result of the call instruction... 215 if (I->getOperand(0) == CI && I->getOperand(1) == CI || 216 I->getOperand(0) != CI && I->getOperand(1) != CI) 217 return 0; 218 219 // The only user of this instruction we allow is a single return instruction. 220 if (!I->hasOneUse() || !isa<ReturnInst>(I->use_back())) 221 return 0; 222 223 // Ok, now we have to check all of the other return instructions in this 224 // function. If they return non-constants or differing values, then we cannot 225 // transform the function safely. 226 return getCommonReturnValue(cast<ReturnInst>(I->use_back()), CI); 227} 228 229bool TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry, 230 std::vector<PHINode*> &ArgumentPHIs) { 231 BasicBlock *BB = Ret->getParent(); 232 Function *F = BB->getParent(); 233 234 if (&BB->front() == Ret) // Make sure there is something before the ret... 235 return false; 236 237 // Scan backwards from the return, checking to see if there is a tail call in 238 // this block. If so, set CI to it. 239 CallInst *CI; 240 BasicBlock::iterator BBI = Ret; 241 while (1) { 242 CI = dyn_cast<CallInst>(BBI); 243 if (CI && CI->getCalledFunction() == F) 244 break; 245 246 if (BBI == BB->begin()) 247 return false; // Didn't find a potential tail call. 248 --BBI; 249 } 250 251 // If we are introducing accumulator recursion to eliminate associative 252 // operations after the call instruction, this variable contains the initial 253 // value for the accumulator. If this value is set, we actually perform 254 // accumulator recursion elimination instead of simple tail recursion 255 // elimination. 256 Value *AccumulatorRecursionEliminationInitVal = 0; 257 Instruction *AccumulatorRecursionInstr = 0; 258 259 // Ok, we found a potential tail call. We can currently only transform the 260 // tail call if all of the instructions between the call and the return are 261 // movable to above the call itself, leaving the call next to the return. 262 // Check that this is the case now. 263 for (BBI = CI, ++BBI; &*BBI != Ret; ++BBI) 264 if (!CanMoveAboveCall(BBI, CI)) { 265 // If we can't move the instruction above the call, it might be because it 266 // is an associative operation that could be tranformed using accumulator 267 // recursion elimination. Check to see if this is the case, and if so, 268 // remember the initial accumulator value for later. 269 if ((AccumulatorRecursionEliminationInitVal = 270 CanTransformAccumulatorRecursion(BBI, CI))) { 271 // Yes, this is accumulator recursion. Remember which instruction 272 // accumulates. 273 AccumulatorRecursionInstr = BBI; 274 } else { 275 return false; // Otherwise, we cannot eliminate the tail recursion! 276 } 277 } 278 279 // We can only transform call/return pairs that either ignore the return value 280 // of the call and return void, ignore the value of the call and return a 281 // constant, return the value returned by the tail call, or that are being 282 // accumulator recursion variable eliminated. 283 if (Ret->getNumOperands() != 0 && Ret->getReturnValue() != CI && 284 AccumulatorRecursionEliminationInitVal == 0 && 285 !getCommonReturnValue(Ret, CI)) 286 return false; 287 288 // OK! We can transform this tail call. If this is the first one found, 289 // create the new entry block, allowing us to branch back to the old entry. 290 if (OldEntry == 0) { 291 OldEntry = &F->getEntryBlock(); 292 std::string OldName = OldEntry->getName(); OldEntry->setName("tailrecurse"); 293 BasicBlock *NewEntry = new BasicBlock(OldName, OldEntry); 294 new BranchInst(OldEntry, NewEntry); 295 296 // Now that we have created a new block, which jumps to the entry 297 // block, insert a PHI node for each argument of the function. 298 // For now, we initialize each PHI to only have the real arguments 299 // which are passed in. 300 Instruction *InsertPos = OldEntry->begin(); 301 for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I) { 302 PHINode *PN = new PHINode(I->getType(), I->getName()+".tr", InsertPos); 303 I->replaceAllUsesWith(PN); // Everyone use the PHI node now! 304 PN->addIncoming(I, NewEntry); 305 ArgumentPHIs.push_back(PN); 306 } 307 } 308 309 // Ok, now that we know we have a pseudo-entry block WITH all of the 310 // required PHI nodes, add entries into the PHI node for the actual 311 // parameters passed into the tail-recursive call. 312 for (unsigned i = 0, e = CI->getNumOperands()-1; i != e; ++i) 313 ArgumentPHIs[i]->addIncoming(CI->getOperand(i+1), BB); 314 315 // If we are introducing an accumulator variable to eliminate the recursion, 316 // do so now. Note that we _know_ that no subsequent tail recursion 317 // eliminations will happen on this function because of the way the 318 // accumulator recursion predicate is set up. 319 // 320 if (AccumulatorRecursionEliminationInitVal) { 321 Instruction *AccRecInstr = AccumulatorRecursionInstr; 322 // Start by inserting a new PHI node for the accumulator. 323 PHINode *AccPN = new PHINode(AccRecInstr->getType(), "accumulator.tr", 324 OldEntry->begin()); 325 326 // Loop over all of the predecessors of the tail recursion block. For the 327 // real entry into the function we seed the PHI with the initial value, 328 // computed earlier. For any other existing branches to this block (due to 329 // other tail recursions eliminated) the accumulator is not modified. 330 // Because we haven't added the branch in the current block to OldEntry yet, 331 // it will not show up as a predecessor. 332 for (pred_iterator PI = pred_begin(OldEntry), PE = pred_end(OldEntry); 333 PI != PE; ++PI) { 334 if (*PI == &F->getEntryBlock()) 335 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, *PI); 336 else 337 AccPN->addIncoming(AccPN, *PI); 338 } 339 340 // Add an incoming argument for the current block, which is computed by our 341 // associative accumulator instruction. 342 AccPN->addIncoming(AccRecInstr, BB); 343 344 // Next, rewrite the accumulator recursion instruction so that it does not 345 // use the result of the call anymore, instead, use the PHI node we just 346 // inserted. 347 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN); 348 349 // Finally, rewrite any return instructions in the program to return the PHI 350 // node instead of the "initval" that they do currently. This loop will 351 // actually rewrite the return value we are destroying, but that's ok. 352 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) 353 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator())) 354 RI->setOperand(0, AccPN); 355 ++NumAccumAdded; 356 } 357 358 // Now that all of the PHI nodes are in place, remove the call and 359 // ret instructions, replacing them with an unconditional branch. 360 new BranchInst(OldEntry, Ret); 361 BB->getInstList().erase(Ret); // Remove return. 362 BB->getInstList().erase(CI); // Remove call. 363 ++NumEliminated; 364 return true; 365} 366