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 precedes 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#include "llvm/Transforms/Scalar/TailRecursionElimination.h" 54#include "llvm/Transforms/Scalar.h" 55#include "llvm/ADT/STLExtras.h" 56#include "llvm/ADT/SmallPtrSet.h" 57#include "llvm/ADT/Statistic.h" 58#include "llvm/Analysis/GlobalsModRef.h" 59#include "llvm/Analysis/CFG.h" 60#include "llvm/Analysis/CaptureTracking.h" 61#include "llvm/Analysis/InlineCost.h" 62#include "llvm/Analysis/InstructionSimplify.h" 63#include "llvm/Analysis/Loads.h" 64#include "llvm/Analysis/TargetTransformInfo.h" 65#include "llvm/IR/CFG.h" 66#include "llvm/IR/CallSite.h" 67#include "llvm/IR/Constants.h" 68#include "llvm/IR/DataLayout.h" 69#include "llvm/IR/DerivedTypes.h" 70#include "llvm/IR/DiagnosticInfo.h" 71#include "llvm/IR/Function.h" 72#include "llvm/IR/Instructions.h" 73#include "llvm/IR/IntrinsicInst.h" 74#include "llvm/IR/Module.h" 75#include "llvm/IR/ValueHandle.h" 76#include "llvm/Pass.h" 77#include "llvm/Support/Debug.h" 78#include "llvm/Support/raw_ostream.h" 79#include "llvm/Transforms/Utils/BasicBlockUtils.h" 80#include "llvm/Transforms/Utils/Local.h" 81using namespace llvm; 82 83#define DEBUG_TYPE "tailcallelim" 84 85STATISTIC(NumEliminated, "Number of tail calls removed"); 86STATISTIC(NumRetDuped, "Number of return duplicated"); 87STATISTIC(NumAccumAdded, "Number of accumulators introduced"); 88 89/// \brief Scan the specified function for alloca instructions. 90/// If it contains any dynamic allocas, returns false. 91static bool canTRE(Function &F) { 92 // Because of PR962, we don't TRE dynamic allocas. 93 for (auto &BB : F) { 94 for (auto &I : BB) { 95 if (AllocaInst *AI = dyn_cast<AllocaInst>(&I)) { 96 if (!AI->isStaticAlloca()) 97 return false; 98 } 99 } 100 } 101 102 return true; 103} 104 105namespace { 106struct AllocaDerivedValueTracker { 107 // Start at a root value and walk its use-def chain to mark calls that use the 108 // value or a derived value in AllocaUsers, and places where it may escape in 109 // EscapePoints. 110 void walk(Value *Root) { 111 SmallVector<Use *, 32> Worklist; 112 SmallPtrSet<Use *, 32> Visited; 113 114 auto AddUsesToWorklist = [&](Value *V) { 115 for (auto &U : V->uses()) { 116 if (!Visited.insert(&U).second) 117 continue; 118 Worklist.push_back(&U); 119 } 120 }; 121 122 AddUsesToWorklist(Root); 123 124 while (!Worklist.empty()) { 125 Use *U = Worklist.pop_back_val(); 126 Instruction *I = cast<Instruction>(U->getUser()); 127 128 switch (I->getOpcode()) { 129 case Instruction::Call: 130 case Instruction::Invoke: { 131 CallSite CS(I); 132 bool IsNocapture = 133 CS.isDataOperand(U) && CS.doesNotCapture(CS.getDataOperandNo(U)); 134 callUsesLocalStack(CS, IsNocapture); 135 if (IsNocapture) { 136 // If the alloca-derived argument is passed in as nocapture, then it 137 // can't propagate to the call's return. That would be capturing. 138 continue; 139 } 140 break; 141 } 142 case Instruction::Load: { 143 // The result of a load is not alloca-derived (unless an alloca has 144 // otherwise escaped, but this is a local analysis). 145 continue; 146 } 147 case Instruction::Store: { 148 if (U->getOperandNo() == 0) 149 EscapePoints.insert(I); 150 continue; // Stores have no users to analyze. 151 } 152 case Instruction::BitCast: 153 case Instruction::GetElementPtr: 154 case Instruction::PHI: 155 case Instruction::Select: 156 case Instruction::AddrSpaceCast: 157 break; 158 default: 159 EscapePoints.insert(I); 160 break; 161 } 162 163 AddUsesToWorklist(I); 164 } 165 } 166 167 void callUsesLocalStack(CallSite CS, bool IsNocapture) { 168 // Add it to the list of alloca users. 169 AllocaUsers.insert(CS.getInstruction()); 170 171 // If it's nocapture then it can't capture this alloca. 172 if (IsNocapture) 173 return; 174 175 // If it can write to memory, it can leak the alloca value. 176 if (!CS.onlyReadsMemory()) 177 EscapePoints.insert(CS.getInstruction()); 178 } 179 180 SmallPtrSet<Instruction *, 32> AllocaUsers; 181 SmallPtrSet<Instruction *, 32> EscapePoints; 182}; 183} 184 185static bool markTails(Function &F, bool &AllCallsAreTailCalls) { 186 if (F.callsFunctionThatReturnsTwice()) 187 return false; 188 AllCallsAreTailCalls = true; 189 190 // The local stack holds all alloca instructions and all byval arguments. 191 AllocaDerivedValueTracker Tracker; 192 for (Argument &Arg : F.args()) { 193 if (Arg.hasByValAttr()) 194 Tracker.walk(&Arg); 195 } 196 for (auto &BB : F) { 197 for (auto &I : BB) 198 if (AllocaInst *AI = dyn_cast<AllocaInst>(&I)) 199 Tracker.walk(AI); 200 } 201 202 bool Modified = false; 203 204 // Track whether a block is reachable after an alloca has escaped. Blocks that 205 // contain the escaping instruction will be marked as being visited without an 206 // escaped alloca, since that is how the block began. 207 enum VisitType { 208 UNVISITED, 209 UNESCAPED, 210 ESCAPED 211 }; 212 DenseMap<BasicBlock *, VisitType> Visited; 213 214 // We propagate the fact that an alloca has escaped from block to successor. 215 // Visit the blocks that are propagating the escapedness first. To do this, we 216 // maintain two worklists. 217 SmallVector<BasicBlock *, 32> WorklistUnescaped, WorklistEscaped; 218 219 // We may enter a block and visit it thinking that no alloca has escaped yet, 220 // then see an escape point and go back around a loop edge and come back to 221 // the same block twice. Because of this, we defer setting tail on calls when 222 // we first encounter them in a block. Every entry in this list does not 223 // statically use an alloca via use-def chain analysis, but may find an alloca 224 // through other means if the block turns out to be reachable after an escape 225 // point. 226 SmallVector<CallInst *, 32> DeferredTails; 227 228 BasicBlock *BB = &F.getEntryBlock(); 229 VisitType Escaped = UNESCAPED; 230 do { 231 for (auto &I : *BB) { 232 if (Tracker.EscapePoints.count(&I)) 233 Escaped = ESCAPED; 234 235 CallInst *CI = dyn_cast<CallInst>(&I); 236 if (!CI || CI->isTailCall()) 237 continue; 238 239 bool IsNoTail = CI->isNoTailCall(); 240 241 if (!IsNoTail && CI->doesNotAccessMemory()) { 242 // A call to a readnone function whose arguments are all things computed 243 // outside this function can be marked tail. Even if you stored the 244 // alloca address into a global, a readnone function can't load the 245 // global anyhow. 246 // 247 // Note that this runs whether we know an alloca has escaped or not. If 248 // it has, then we can't trust Tracker.AllocaUsers to be accurate. 249 bool SafeToTail = true; 250 for (auto &Arg : CI->arg_operands()) { 251 if (isa<Constant>(Arg.getUser())) 252 continue; 253 if (Argument *A = dyn_cast<Argument>(Arg.getUser())) 254 if (!A->hasByValAttr()) 255 continue; 256 SafeToTail = false; 257 break; 258 } 259 if (SafeToTail) { 260 emitOptimizationRemark( 261 F.getContext(), "tailcallelim", F, CI->getDebugLoc(), 262 "marked this readnone call a tail call candidate"); 263 CI->setTailCall(); 264 Modified = true; 265 continue; 266 } 267 } 268 269 if (!IsNoTail && Escaped == UNESCAPED && !Tracker.AllocaUsers.count(CI)) { 270 DeferredTails.push_back(CI); 271 } else { 272 AllCallsAreTailCalls = false; 273 } 274 } 275 276 for (auto *SuccBB : make_range(succ_begin(BB), succ_end(BB))) { 277 auto &State = Visited[SuccBB]; 278 if (State < Escaped) { 279 State = Escaped; 280 if (State == ESCAPED) 281 WorklistEscaped.push_back(SuccBB); 282 else 283 WorklistUnescaped.push_back(SuccBB); 284 } 285 } 286 287 if (!WorklistEscaped.empty()) { 288 BB = WorklistEscaped.pop_back_val(); 289 Escaped = ESCAPED; 290 } else { 291 BB = nullptr; 292 while (!WorklistUnescaped.empty()) { 293 auto *NextBB = WorklistUnescaped.pop_back_val(); 294 if (Visited[NextBB] == UNESCAPED) { 295 BB = NextBB; 296 Escaped = UNESCAPED; 297 break; 298 } 299 } 300 } 301 } while (BB); 302 303 for (CallInst *CI : DeferredTails) { 304 if (Visited[CI->getParent()] != ESCAPED) { 305 // If the escape point was part way through the block, calls after the 306 // escape point wouldn't have been put into DeferredTails. 307 emitOptimizationRemark(F.getContext(), "tailcallelim", F, 308 CI->getDebugLoc(), 309 "marked this call a tail call candidate"); 310 CI->setTailCall(); 311 Modified = true; 312 } else { 313 AllCallsAreTailCalls = false; 314 } 315 } 316 317 return Modified; 318} 319 320/// Return true if it is safe to move the specified 321/// instruction from after the call to before the call, assuming that all 322/// instructions between the call and this instruction are movable. 323/// 324static bool canMoveAboveCall(Instruction *I, CallInst *CI) { 325 // FIXME: We can move load/store/call/free instructions above the call if the 326 // call does not mod/ref the memory location being processed. 327 if (I->mayHaveSideEffects()) // This also handles volatile loads. 328 return false; 329 330 if (LoadInst *L = dyn_cast<LoadInst>(I)) { 331 // Loads may always be moved above calls without side effects. 332 if (CI->mayHaveSideEffects()) { 333 // Non-volatile loads may be moved above a call with side effects if it 334 // does not write to memory and the load provably won't trap. 335 // FIXME: Writes to memory only matter if they may alias the pointer 336 // being loaded from. 337 const DataLayout &DL = L->getModule()->getDataLayout(); 338 if (CI->mayWriteToMemory() || 339 !isSafeToLoadUnconditionally(L->getPointerOperand(), 340 L->getAlignment(), DL, L)) 341 return false; 342 } 343 } 344 345 // Otherwise, if this is a side-effect free instruction, check to make sure 346 // that it does not use the return value of the call. If it doesn't use the 347 // return value of the call, it must only use things that are defined before 348 // the call, or movable instructions between the call and the instruction 349 // itself. 350 return std::find(I->op_begin(), I->op_end(), CI) == I->op_end(); 351} 352 353/// Return true if the specified value is the same when the return would exit 354/// as it was when the initial iteration of the recursive function was executed. 355/// 356/// We currently handle static constants and arguments that are not modified as 357/// part of the recursion. 358static bool isDynamicConstant(Value *V, CallInst *CI, ReturnInst *RI) { 359 if (isa<Constant>(V)) return true; // Static constants are always dyn consts 360 361 // Check to see if this is an immutable argument, if so, the value 362 // will be available to initialize the accumulator. 363 if (Argument *Arg = dyn_cast<Argument>(V)) { 364 // Figure out which argument number this is... 365 unsigned ArgNo = 0; 366 Function *F = CI->getParent()->getParent(); 367 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI) 368 ++ArgNo; 369 370 // If we are passing this argument into call as the corresponding 371 // argument operand, then the argument is dynamically constant. 372 // Otherwise, we cannot transform this function safely. 373 if (CI->getArgOperand(ArgNo) == Arg) 374 return true; 375 } 376 377 // Switch cases are always constant integers. If the value is being switched 378 // on and the return is only reachable from one of its cases, it's 379 // effectively constant. 380 if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor()) 381 if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator())) 382 if (SI->getCondition() == V) 383 return SI->getDefaultDest() != RI->getParent(); 384 385 // Not a constant or immutable argument, we can't safely transform. 386 return false; 387} 388 389/// Check to see if the function containing the specified tail call consistently 390/// returns the same runtime-constant value at all exit points except for 391/// IgnoreRI. If so, return the returned value. 392static Value *getCommonReturnValue(ReturnInst *IgnoreRI, CallInst *CI) { 393 Function *F = CI->getParent()->getParent(); 394 Value *ReturnedValue = nullptr; 395 396 for (BasicBlock &BBI : *F) { 397 ReturnInst *RI = dyn_cast<ReturnInst>(BBI.getTerminator()); 398 if (RI == nullptr || RI == IgnoreRI) continue; 399 400 // We can only perform this transformation if the value returned is 401 // evaluatable at the start of the initial invocation of the function, 402 // instead of at the end of the evaluation. 403 // 404 Value *RetOp = RI->getOperand(0); 405 if (!isDynamicConstant(RetOp, CI, RI)) 406 return nullptr; 407 408 if (ReturnedValue && RetOp != ReturnedValue) 409 return nullptr; // Cannot transform if differing values are returned. 410 ReturnedValue = RetOp; 411 } 412 return ReturnedValue; 413} 414 415/// If the specified instruction can be transformed using accumulator recursion 416/// elimination, return the constant which is the start of the accumulator 417/// value. Otherwise return null. 418static Value *canTransformAccumulatorRecursion(Instruction *I, CallInst *CI) { 419 if (!I->isAssociative() || !I->isCommutative()) return nullptr; 420 assert(I->getNumOperands() == 2 && 421 "Associative/commutative operations should have 2 args!"); 422 423 // Exactly one operand should be the result of the call instruction. 424 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) || 425 (I->getOperand(0) != CI && I->getOperand(1) != CI)) 426 return nullptr; 427 428 // The only user of this instruction we allow is a single return instruction. 429 if (!I->hasOneUse() || !isa<ReturnInst>(I->user_back())) 430 return nullptr; 431 432 // Ok, now we have to check all of the other return instructions in this 433 // function. If they return non-constants or differing values, then we cannot 434 // transform the function safely. 435 return getCommonReturnValue(cast<ReturnInst>(I->user_back()), CI); 436} 437 438static Instruction *firstNonDbg(BasicBlock::iterator I) { 439 while (isa<DbgInfoIntrinsic>(I)) 440 ++I; 441 return &*I; 442} 443 444static CallInst *findTRECandidate(Instruction *TI, 445 bool CannotTailCallElimCallsMarkedTail, 446 const TargetTransformInfo *TTI) { 447 BasicBlock *BB = TI->getParent(); 448 Function *F = BB->getParent(); 449 450 if (&BB->front() == TI) // Make sure there is something before the terminator. 451 return nullptr; 452 453 // Scan backwards from the return, checking to see if there is a tail call in 454 // this block. If so, set CI to it. 455 CallInst *CI = nullptr; 456 BasicBlock::iterator BBI(TI); 457 while (true) { 458 CI = dyn_cast<CallInst>(BBI); 459 if (CI && CI->getCalledFunction() == F) 460 break; 461 462 if (BBI == BB->begin()) 463 return nullptr; // Didn't find a potential tail call. 464 --BBI; 465 } 466 467 // If this call is marked as a tail call, and if there are dynamic allocas in 468 // the function, we cannot perform this optimization. 469 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail) 470 return nullptr; 471 472 // As a special case, detect code like this: 473 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call 474 // and disable this xform in this case, because the code generator will 475 // lower the call to fabs into inline code. 476 if (BB == &F->getEntryBlock() && 477 firstNonDbg(BB->front().getIterator()) == CI && 478 firstNonDbg(std::next(BB->begin())) == TI && CI->getCalledFunction() && 479 !TTI->isLoweredToCall(CI->getCalledFunction())) { 480 // A single-block function with just a call and a return. Check that 481 // the arguments match. 482 CallSite::arg_iterator I = CallSite(CI).arg_begin(), 483 E = CallSite(CI).arg_end(); 484 Function::arg_iterator FI = F->arg_begin(), 485 FE = F->arg_end(); 486 for (; I != E && FI != FE; ++I, ++FI) 487 if (*I != &*FI) break; 488 if (I == E && FI == FE) 489 return nullptr; 490 } 491 492 return CI; 493} 494 495static bool eliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret, 496 BasicBlock *&OldEntry, 497 bool &TailCallsAreMarkedTail, 498 SmallVectorImpl<PHINode *> &ArgumentPHIs, 499 bool CannotTailCallElimCallsMarkedTail) { 500 // If we are introducing accumulator recursion to eliminate operations after 501 // the call instruction that are both associative and commutative, the initial 502 // value for the accumulator is placed in this variable. If this value is set 503 // then we actually perform accumulator recursion elimination instead of 504 // simple tail recursion elimination. If the operation is an LLVM instruction 505 // (eg: "add") then it is recorded in AccumulatorRecursionInstr. If not, then 506 // we are handling the case when the return instruction returns a constant C 507 // which is different to the constant returned by other return instructions 508 // (which is recorded in AccumulatorRecursionEliminationInitVal). This is a 509 // special case of accumulator recursion, the operation being "return C". 510 Value *AccumulatorRecursionEliminationInitVal = nullptr; 511 Instruction *AccumulatorRecursionInstr = nullptr; 512 513 // Ok, we found a potential tail call. We can currently only transform the 514 // tail call if all of the instructions between the call and the return are 515 // movable to above the call itself, leaving the call next to the return. 516 // Check that this is the case now. 517 BasicBlock::iterator BBI(CI); 518 for (++BBI; &*BBI != Ret; ++BBI) { 519 if (canMoveAboveCall(&*BBI, CI)) continue; 520 521 // If we can't move the instruction above the call, it might be because it 522 // is an associative and commutative operation that could be transformed 523 // using accumulator recursion elimination. Check to see if this is the 524 // case, and if so, remember the initial accumulator value for later. 525 if ((AccumulatorRecursionEliminationInitVal = 526 canTransformAccumulatorRecursion(&*BBI, CI))) { 527 // Yes, this is accumulator recursion. Remember which instruction 528 // accumulates. 529 AccumulatorRecursionInstr = &*BBI; 530 } else { 531 return false; // Otherwise, we cannot eliminate the tail recursion! 532 } 533 } 534 535 // We can only transform call/return pairs that either ignore the return value 536 // of the call and return void, ignore the value of the call and return a 537 // constant, return the value returned by the tail call, or that are being 538 // accumulator recursion variable eliminated. 539 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI && 540 !isa<UndefValue>(Ret->getReturnValue()) && 541 AccumulatorRecursionEliminationInitVal == nullptr && 542 !getCommonReturnValue(nullptr, CI)) { 543 // One case remains that we are able to handle: the current return 544 // instruction returns a constant, and all other return instructions 545 // return a different constant. 546 if (!isDynamicConstant(Ret->getReturnValue(), CI, Ret)) 547 return false; // Current return instruction does not return a constant. 548 // Check that all other return instructions return a common constant. If 549 // so, record it in AccumulatorRecursionEliminationInitVal. 550 AccumulatorRecursionEliminationInitVal = getCommonReturnValue(Ret, CI); 551 if (!AccumulatorRecursionEliminationInitVal) 552 return false; 553 } 554 555 BasicBlock *BB = Ret->getParent(); 556 Function *F = BB->getParent(); 557 558 emitOptimizationRemark(F->getContext(), "tailcallelim", *F, CI->getDebugLoc(), 559 "transforming tail recursion to loop"); 560 561 // OK! We can transform this tail call. If this is the first one found, 562 // create the new entry block, allowing us to branch back to the old entry. 563 if (!OldEntry) { 564 OldEntry = &F->getEntryBlock(); 565 BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry); 566 NewEntry->takeName(OldEntry); 567 OldEntry->setName("tailrecurse"); 568 BranchInst::Create(OldEntry, NewEntry); 569 570 // If this tail call is marked 'tail' and if there are any allocas in the 571 // entry block, move them up to the new entry block. 572 TailCallsAreMarkedTail = CI->isTailCall(); 573 if (TailCallsAreMarkedTail) 574 // Move all fixed sized allocas from OldEntry to NewEntry. 575 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(), 576 NEBI = NewEntry->begin(); OEBI != E; ) 577 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++)) 578 if (isa<ConstantInt>(AI->getArraySize())) 579 AI->moveBefore(&*NEBI); 580 581 // Now that we have created a new block, which jumps to the entry 582 // block, insert a PHI node for each argument of the function. 583 // For now, we initialize each PHI to only have the real arguments 584 // which are passed in. 585 Instruction *InsertPos = &OldEntry->front(); 586 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 587 I != E; ++I) { 588 PHINode *PN = PHINode::Create(I->getType(), 2, 589 I->getName() + ".tr", InsertPos); 590 I->replaceAllUsesWith(PN); // Everyone use the PHI node now! 591 PN->addIncoming(&*I, NewEntry); 592 ArgumentPHIs.push_back(PN); 593 } 594 } 595 596 // If this function has self recursive calls in the tail position where some 597 // are marked tail and some are not, only transform one flavor or another. We 598 // have to choose whether we move allocas in the entry block to the new entry 599 // block or not, so we can't make a good choice for both. NOTE: We could do 600 // slightly better here in the case that the function has no entry block 601 // allocas. 602 if (TailCallsAreMarkedTail && !CI->isTailCall()) 603 return false; 604 605 // Ok, now that we know we have a pseudo-entry block WITH all of the 606 // required PHI nodes, add entries into the PHI node for the actual 607 // parameters passed into the tail-recursive call. 608 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) 609 ArgumentPHIs[i]->addIncoming(CI->getArgOperand(i), BB); 610 611 // If we are introducing an accumulator variable to eliminate the recursion, 612 // do so now. Note that we _know_ that no subsequent tail recursion 613 // eliminations will happen on this function because of the way the 614 // accumulator recursion predicate is set up. 615 // 616 if (AccumulatorRecursionEliminationInitVal) { 617 Instruction *AccRecInstr = AccumulatorRecursionInstr; 618 // Start by inserting a new PHI node for the accumulator. 619 pred_iterator PB = pred_begin(OldEntry), PE = pred_end(OldEntry); 620 PHINode *AccPN = PHINode::Create( 621 AccumulatorRecursionEliminationInitVal->getType(), 622 std::distance(PB, PE) + 1, "accumulator.tr", &OldEntry->front()); 623 624 // Loop over all of the predecessors of the tail recursion block. For the 625 // real entry into the function we seed the PHI with the initial value, 626 // computed earlier. For any other existing branches to this block (due to 627 // other tail recursions eliminated) the accumulator is not modified. 628 // Because we haven't added the branch in the current block to OldEntry yet, 629 // it will not show up as a predecessor. 630 for (pred_iterator PI = PB; PI != PE; ++PI) { 631 BasicBlock *P = *PI; 632 if (P == &F->getEntryBlock()) 633 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, P); 634 else 635 AccPN->addIncoming(AccPN, P); 636 } 637 638 if (AccRecInstr) { 639 // Add an incoming argument for the current block, which is computed by 640 // our associative and commutative accumulator instruction. 641 AccPN->addIncoming(AccRecInstr, BB); 642 643 // Next, rewrite the accumulator recursion instruction so that it does not 644 // use the result of the call anymore, instead, use the PHI node we just 645 // inserted. 646 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN); 647 } else { 648 // Add an incoming argument for the current block, which is just the 649 // constant returned by the current return instruction. 650 AccPN->addIncoming(Ret->getReturnValue(), BB); 651 } 652 653 // Finally, rewrite any return instructions in the program to return the PHI 654 // node instead of the "initval" that they do currently. This loop will 655 // actually rewrite the return value we are destroying, but that's ok. 656 for (BasicBlock &BBI : *F) 657 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI.getTerminator())) 658 RI->setOperand(0, AccPN); 659 ++NumAccumAdded; 660 } 661 662 // Now that all of the PHI nodes are in place, remove the call and 663 // ret instructions, replacing them with an unconditional branch. 664 BranchInst *NewBI = BranchInst::Create(OldEntry, Ret); 665 NewBI->setDebugLoc(CI->getDebugLoc()); 666 667 BB->getInstList().erase(Ret); // Remove return. 668 BB->getInstList().erase(CI); // Remove call. 669 ++NumEliminated; 670 return true; 671} 672 673static bool foldReturnAndProcessPred(BasicBlock *BB, ReturnInst *Ret, 674 BasicBlock *&OldEntry, 675 bool &TailCallsAreMarkedTail, 676 SmallVectorImpl<PHINode *> &ArgumentPHIs, 677 bool CannotTailCallElimCallsMarkedTail, 678 const TargetTransformInfo *TTI) { 679 bool Change = false; 680 681 // If the return block contains nothing but the return and PHI's, 682 // there might be an opportunity to duplicate the return in its 683 // predecessors and perform TRC there. Look for predecessors that end 684 // in unconditional branch and recursive call(s). 685 SmallVector<BranchInst*, 8> UncondBranchPreds; 686 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 687 BasicBlock *Pred = *PI; 688 TerminatorInst *PTI = Pred->getTerminator(); 689 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) 690 if (BI->isUnconditional()) 691 UncondBranchPreds.push_back(BI); 692 } 693 694 while (!UncondBranchPreds.empty()) { 695 BranchInst *BI = UncondBranchPreds.pop_back_val(); 696 BasicBlock *Pred = BI->getParent(); 697 if (CallInst *CI = findTRECandidate(BI, CannotTailCallElimCallsMarkedTail, TTI)){ 698 DEBUG(dbgs() << "FOLDING: " << *BB 699 << "INTO UNCOND BRANCH PRED: " << *Pred); 700 ReturnInst *RI = FoldReturnIntoUncondBranch(Ret, BB, Pred); 701 702 // Cleanup: if all predecessors of BB have been eliminated by 703 // FoldReturnIntoUncondBranch, delete it. It is important to empty it, 704 // because the ret instruction in there is still using a value which 705 // eliminateRecursiveTailCall will attempt to remove. 706 if (!BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB)) 707 BB->eraseFromParent(); 708 709 eliminateRecursiveTailCall(CI, RI, OldEntry, TailCallsAreMarkedTail, 710 ArgumentPHIs, 711 CannotTailCallElimCallsMarkedTail); 712 ++NumRetDuped; 713 Change = true; 714 } 715 } 716 717 return Change; 718} 719 720static bool processReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry, 721 bool &TailCallsAreMarkedTail, 722 SmallVectorImpl<PHINode *> &ArgumentPHIs, 723 bool CannotTailCallElimCallsMarkedTail, 724 const TargetTransformInfo *TTI) { 725 CallInst *CI = findTRECandidate(Ret, CannotTailCallElimCallsMarkedTail, TTI); 726 if (!CI) 727 return false; 728 729 return eliminateRecursiveTailCall(CI, Ret, OldEntry, TailCallsAreMarkedTail, 730 ArgumentPHIs, 731 CannotTailCallElimCallsMarkedTail); 732} 733 734static bool eliminateTailRecursion(Function &F, const TargetTransformInfo *TTI) { 735 if (F.getFnAttribute("disable-tail-calls").getValueAsString() == "true") 736 return false; 737 738 bool MadeChange = false; 739 bool AllCallsAreTailCalls = false; 740 MadeChange |= markTails(F, AllCallsAreTailCalls); 741 if (!AllCallsAreTailCalls) 742 return MadeChange; 743 744 // If this function is a varargs function, we won't be able to PHI the args 745 // right, so don't even try to convert it... 746 if (F.getFunctionType()->isVarArg()) 747 return false; 748 749 BasicBlock *OldEntry = nullptr; 750 bool TailCallsAreMarkedTail = false; 751 SmallVector<PHINode*, 8> ArgumentPHIs; 752 753 // If false, we cannot perform TRE on tail calls marked with the 'tail' 754 // attribute, because doing so would cause the stack size to increase (real 755 // TRE would deallocate variable sized allocas, TRE doesn't). 756 bool CanTRETailMarkedCall = canTRE(F); 757 758 // Change any tail recursive calls to loops. 759 // 760 // FIXME: The code generator produces really bad code when an 'escaping 761 // alloca' is changed from being a static alloca to being a dynamic alloca. 762 // Until this is resolved, disable this transformation if that would ever 763 // happen. This bug is PR962. 764 for (Function::iterator BBI = F.begin(), E = F.end(); BBI != E; /*in loop*/) { 765 BasicBlock *BB = &*BBI++; // foldReturnAndProcessPred may delete BB. 766 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) { 767 bool Change = 768 processReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail, 769 ArgumentPHIs, !CanTRETailMarkedCall, TTI); 770 if (!Change && BB->getFirstNonPHIOrDbg() == Ret) 771 Change = 772 foldReturnAndProcessPred(BB, Ret, OldEntry, TailCallsAreMarkedTail, 773 ArgumentPHIs, !CanTRETailMarkedCall, TTI); 774 MadeChange |= Change; 775 } 776 } 777 778 // If we eliminated any tail recursions, it's possible that we inserted some 779 // silly PHI nodes which just merge an initial value (the incoming operand) 780 // with themselves. Check to see if we did and clean up our mess if so. This 781 // occurs when a function passes an argument straight through to its tail 782 // call. 783 for (PHINode *PN : ArgumentPHIs) { 784 // If the PHI Node is a dynamic constant, replace it with the value it is. 785 if (Value *PNV = SimplifyInstruction(PN, F.getParent()->getDataLayout())) { 786 PN->replaceAllUsesWith(PNV); 787 PN->eraseFromParent(); 788 } 789 } 790 791 return MadeChange; 792} 793 794namespace { 795struct TailCallElim : public FunctionPass { 796 static char ID; // Pass identification, replacement for typeid 797 TailCallElim() : FunctionPass(ID) { 798 initializeTailCallElimPass(*PassRegistry::getPassRegistry()); 799 } 800 801 void getAnalysisUsage(AnalysisUsage &AU) const override { 802 AU.addRequired<TargetTransformInfoWrapperPass>(); 803 AU.addPreserved<GlobalsAAWrapperPass>(); 804 } 805 806 bool runOnFunction(Function &F) override { 807 if (skipFunction(F)) 808 return false; 809 810 return eliminateTailRecursion( 811 F, &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F)); 812 } 813}; 814} 815 816char TailCallElim::ID = 0; 817INITIALIZE_PASS_BEGIN(TailCallElim, "tailcallelim", "Tail Call Elimination", 818 false, false) 819INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 820INITIALIZE_PASS_END(TailCallElim, "tailcallelim", "Tail Call Elimination", 821 false, false) 822 823// Public interface to the TailCallElimination pass 824FunctionPass *llvm::createTailCallEliminationPass() { 825 return new TailCallElim(); 826} 827 828PreservedAnalyses TailCallElimPass::run(Function &F, 829 FunctionAnalysisManager &AM) { 830 831 TargetTransformInfo &TTI = AM.getResult<TargetIRAnalysis>(F); 832 833 bool Changed = eliminateTailRecursion(F, &TTI); 834 835 if (!Changed) 836 return PreservedAnalyses::all(); 837 PreservedAnalyses PA; 838 PA.preserve<GlobalsAA>(); 839 return PA; 840} 841