1 //===- TailRecursionElimination.cpp - Eliminate Tail Calls ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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
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
32 // There are several improvements that could be made:
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.
51 //===----------------------------------------------------------------------===//
53 #include "llvm/Transforms/Scalar.h"
54 #include "llvm/ADT/STLExtras.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/Statistic.h"
57 #include "llvm/Analysis/CFG.h"
58 #include "llvm/Analysis/CaptureTracking.h"
59 #include "llvm/Analysis/InlineCost.h"
60 #include "llvm/Analysis/InstructionSimplify.h"
61 #include "llvm/Analysis/Loads.h"
62 #include "llvm/Analysis/TargetTransformInfo.h"
63 #include "llvm/IR/CFG.h"
64 #include "llvm/IR/CallSite.h"
65 #include "llvm/IR/Constants.h"
66 #include "llvm/IR/DataLayout.h"
67 #include "llvm/IR/DerivedTypes.h"
68 #include "llvm/IR/DiagnosticInfo.h"
69 #include "llvm/IR/Function.h"
70 #include "llvm/IR/Instructions.h"
71 #include "llvm/IR/IntrinsicInst.h"
72 #include "llvm/IR/Module.h"
73 #include "llvm/IR/ValueHandle.h"
74 #include "llvm/Pass.h"
75 #include "llvm/Support/Debug.h"
76 #include "llvm/Support/raw_ostream.h"
77 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
78 #include "llvm/Transforms/Utils/Local.h"
81 #define DEBUG_TYPE "tailcallelim"
83 STATISTIC(NumEliminated, "Number of tail calls removed");
84 STATISTIC(NumRetDuped, "Number of return duplicated");
85 STATISTIC(NumAccumAdded, "Number of accumulators introduced");
88 struct TailCallElim : public FunctionPass {
89 const TargetTransformInfo *TTI;
91 static char ID; // Pass identification, replacement for typeid
92 TailCallElim() : FunctionPass(ID) {
93 initializeTailCallElimPass(*PassRegistry::getPassRegistry());
96 void getAnalysisUsage(AnalysisUsage &AU) const override;
98 bool runOnFunction(Function &F) override;
101 bool runTRE(Function &F);
102 bool markTails(Function &F, bool &AllCallsAreTailCalls);
104 CallInst *FindTRECandidate(Instruction *I,
105 bool CannotTailCallElimCallsMarkedTail);
106 bool EliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret,
107 BasicBlock *&OldEntry,
108 bool &TailCallsAreMarkedTail,
109 SmallVectorImpl<PHINode *> &ArgumentPHIs,
110 bool CannotTailCallElimCallsMarkedTail);
111 bool FoldReturnAndProcessPred(BasicBlock *BB,
112 ReturnInst *Ret, BasicBlock *&OldEntry,
113 bool &TailCallsAreMarkedTail,
114 SmallVectorImpl<PHINode *> &ArgumentPHIs,
115 bool CannotTailCallElimCallsMarkedTail);
116 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry,
117 bool &TailCallsAreMarkedTail,
118 SmallVectorImpl<PHINode *> &ArgumentPHIs,
119 bool CannotTailCallElimCallsMarkedTail);
120 bool CanMoveAboveCall(Instruction *I, CallInst *CI);
121 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI);
125 char TailCallElim::ID = 0;
126 INITIALIZE_PASS_BEGIN(TailCallElim, "tailcallelim",
127 "Tail Call Elimination", false, false)
128 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
129 INITIALIZE_PASS_END(TailCallElim, "tailcallelim",
130 "Tail Call Elimination", false, false)
132 // Public interface to the TailCallElimination pass
133 FunctionPass *llvm::createTailCallEliminationPass() {
134 return new TailCallElim();
137 void TailCallElim::getAnalysisUsage(AnalysisUsage &AU) const {
138 AU.addRequired<TargetTransformInfoWrapperPass>();
141 /// \brief Scan the specified function for alloca instructions.
142 /// If it contains any dynamic allocas, returns false.
143 static bool CanTRE(Function &F) {
144 // Because of PR962, we don't TRE dynamic allocas.
147 if (AllocaInst *AI = dyn_cast<AllocaInst>(&I)) {
148 if (!AI->isStaticAlloca())
157 bool TailCallElim::runOnFunction(Function &F) {
158 if (skipOptnoneFunction(F))
161 bool AllCallsAreTailCalls = false;
162 bool Modified = markTails(F, AllCallsAreTailCalls);
163 if (AllCallsAreTailCalls)
164 Modified |= runTRE(F);
169 struct AllocaDerivedValueTracker {
170 // Start at a root value and walk its use-def chain to mark calls that use the
171 // value or a derived value in AllocaUsers, and places where it may escape in
173 void walk(Value *Root) {
174 SmallVector<Use *, 32> Worklist;
175 SmallPtrSet<Use *, 32> Visited;
177 auto AddUsesToWorklist = [&](Value *V) {
178 for (auto &U : V->uses()) {
179 if (!Visited.insert(&U).second)
181 Worklist.push_back(&U);
185 AddUsesToWorklist(Root);
187 while (!Worklist.empty()) {
188 Use *U = Worklist.pop_back_val();
189 Instruction *I = cast<Instruction>(U->getUser());
191 switch (I->getOpcode()) {
192 case Instruction::Call:
193 case Instruction::Invoke: {
195 bool IsNocapture = !CS.isCallee(U) &&
196 CS.doesNotCapture(CS.getArgumentNo(U));
197 callUsesLocalStack(CS, IsNocapture);
199 // If the alloca-derived argument is passed in as nocapture, then it
200 // can't propagate to the call's return. That would be capturing.
205 case Instruction::Load: {
206 // The result of a load is not alloca-derived (unless an alloca has
207 // otherwise escaped, but this is a local analysis).
210 case Instruction::Store: {
211 if (U->getOperandNo() == 0)
212 EscapePoints.insert(I);
213 continue; // Stores have no users to analyze.
215 case Instruction::BitCast:
216 case Instruction::GetElementPtr:
217 case Instruction::PHI:
218 case Instruction::Select:
219 case Instruction::AddrSpaceCast:
222 EscapePoints.insert(I);
226 AddUsesToWorklist(I);
230 void callUsesLocalStack(CallSite CS, bool IsNocapture) {
231 // Add it to the list of alloca users.
232 AllocaUsers.insert(CS.getInstruction());
234 // If it's nocapture then it can't capture this alloca.
238 // If it can write to memory, it can leak the alloca value.
239 if (!CS.onlyReadsMemory())
240 EscapePoints.insert(CS.getInstruction());
243 SmallPtrSet<Instruction *, 32> AllocaUsers;
244 SmallPtrSet<Instruction *, 32> EscapePoints;
248 bool TailCallElim::markTails(Function &F, bool &AllCallsAreTailCalls) {
249 if (F.callsFunctionThatReturnsTwice())
251 AllCallsAreTailCalls = true;
253 // The local stack holds all alloca instructions and all byval arguments.
254 AllocaDerivedValueTracker Tracker;
255 for (Argument &Arg : F.args()) {
256 if (Arg.hasByValAttr())
261 if (AllocaInst *AI = dyn_cast<AllocaInst>(&I))
265 bool Modified = false;
267 // Track whether a block is reachable after an alloca has escaped. Blocks that
268 // contain the escaping instruction will be marked as being visited without an
269 // escaped alloca, since that is how the block began.
275 DenseMap<BasicBlock *, VisitType> Visited;
277 // We propagate the fact that an alloca has escaped from block to successor.
278 // Visit the blocks that are propagating the escapedness first. To do this, we
279 // maintain two worklists.
280 SmallVector<BasicBlock *, 32> WorklistUnescaped, WorklistEscaped;
282 // We may enter a block and visit it thinking that no alloca has escaped yet,
283 // then see an escape point and go back around a loop edge and come back to
284 // the same block twice. Because of this, we defer setting tail on calls when
285 // we first encounter them in a block. Every entry in this list does not
286 // statically use an alloca via use-def chain analysis, but may find an alloca
287 // through other means if the block turns out to be reachable after an escape
289 SmallVector<CallInst *, 32> DeferredTails;
291 BasicBlock *BB = &F.getEntryBlock();
292 VisitType Escaped = UNESCAPED;
294 for (auto &I : *BB) {
295 if (Tracker.EscapePoints.count(&I))
298 CallInst *CI = dyn_cast<CallInst>(&I);
299 if (!CI || CI->isTailCall())
302 if (CI->doesNotAccessMemory()) {
303 // A call to a readnone function whose arguments are all things computed
304 // outside this function can be marked tail. Even if you stored the
305 // alloca address into a global, a readnone function can't load the
308 // Note that this runs whether we know an alloca has escaped or not. If
309 // it has, then we can't trust Tracker.AllocaUsers to be accurate.
310 bool SafeToTail = true;
311 for (auto &Arg : CI->arg_operands()) {
312 if (isa<Constant>(Arg.getUser()))
314 if (Argument *A = dyn_cast<Argument>(Arg.getUser()))
315 if (!A->hasByValAttr())
321 emitOptimizationRemark(
322 F.getContext(), "tailcallelim", F, CI->getDebugLoc(),
323 "marked this readnone call a tail call candidate");
330 if (Escaped == UNESCAPED && !Tracker.AllocaUsers.count(CI)) {
331 DeferredTails.push_back(CI);
333 AllCallsAreTailCalls = false;
337 for (auto *SuccBB : make_range(succ_begin(BB), succ_end(BB))) {
338 auto &State = Visited[SuccBB];
339 if (State < Escaped) {
341 if (State == ESCAPED)
342 WorklistEscaped.push_back(SuccBB);
344 WorklistUnescaped.push_back(SuccBB);
348 if (!WorklistEscaped.empty()) {
349 BB = WorklistEscaped.pop_back_val();
353 while (!WorklistUnescaped.empty()) {
354 auto *NextBB = WorklistUnescaped.pop_back_val();
355 if (Visited[NextBB] == UNESCAPED) {
364 for (CallInst *CI : DeferredTails) {
365 if (Visited[CI->getParent()] != ESCAPED) {
366 // If the escape point was part way through the block, calls after the
367 // escape point wouldn't have been put into DeferredTails.
368 emitOptimizationRemark(F.getContext(), "tailcallelim", F,
370 "marked this call a tail call candidate");
374 AllCallsAreTailCalls = false;
381 bool TailCallElim::runTRE(Function &F) {
382 // If this function is a varargs function, we won't be able to PHI the args
383 // right, so don't even try to convert it...
384 if (F.getFunctionType()->isVarArg()) return false;
386 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
387 BasicBlock *OldEntry = nullptr;
388 bool TailCallsAreMarkedTail = false;
389 SmallVector<PHINode*, 8> ArgumentPHIs;
390 bool MadeChange = false;
392 // If false, we cannot perform TRE on tail calls marked with the 'tail'
393 // attribute, because doing so would cause the stack size to increase (real
394 // TRE would deallocate variable sized allocas, TRE doesn't).
395 bool CanTRETailMarkedCall = CanTRE(F);
397 // Change any tail recursive calls to loops.
399 // FIXME: The code generator produces really bad code when an 'escaping
400 // alloca' is changed from being a static alloca to being a dynamic alloca.
401 // Until this is resolved, disable this transformation if that would ever
402 // happen. This bug is PR962.
403 for (Function::iterator BBI = F.begin(), E = F.end(); BBI != E; /*in loop*/) {
404 BasicBlock *BB = BBI++; // FoldReturnAndProcessPred may delete BB.
405 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) {
406 bool Change = ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail,
407 ArgumentPHIs, !CanTRETailMarkedCall);
408 if (!Change && BB->getFirstNonPHIOrDbg() == Ret)
409 Change = FoldReturnAndProcessPred(BB, Ret, OldEntry,
410 TailCallsAreMarkedTail, ArgumentPHIs,
411 !CanTRETailMarkedCall);
412 MadeChange |= Change;
416 // If we eliminated any tail recursions, it's possible that we inserted some
417 // silly PHI nodes which just merge an initial value (the incoming operand)
418 // with themselves. Check to see if we did and clean up our mess if so. This
419 // occurs when a function passes an argument straight through to its tail
421 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) {
422 PHINode *PN = ArgumentPHIs[i];
424 // If the PHI Node is a dynamic constant, replace it with the value it is.
425 if (Value *PNV = SimplifyInstruction(PN, F.getParent()->getDataLayout())) {
426 PN->replaceAllUsesWith(PNV);
427 PN->eraseFromParent();
435 /// Return true if it is safe to move the specified
436 /// instruction from after the call to before the call, assuming that all
437 /// instructions between the call and this instruction are movable.
439 bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) {
440 // FIXME: We can move load/store/call/free instructions above the call if the
441 // call does not mod/ref the memory location being processed.
442 if (I->mayHaveSideEffects()) // This also handles volatile loads.
445 if (LoadInst *L = dyn_cast<LoadInst>(I)) {
446 // Loads may always be moved above calls without side effects.
447 if (CI->mayHaveSideEffects()) {
448 // Non-volatile loads may be moved above a call with side effects if it
449 // does not write to memory and the load provably won't trap.
450 // FIXME: Writes to memory only matter if they may alias the pointer
451 // being loaded from.
452 if (CI->mayWriteToMemory() ||
453 !isSafeToLoadUnconditionally(L->getPointerOperand(), L,
459 // Otherwise, if this is a side-effect free instruction, check to make sure
460 // that it does not use the return value of the call. If it doesn't use the
461 // return value of the call, it must only use things that are defined before
462 // the call, or movable instructions between the call and the instruction
464 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
465 if (I->getOperand(i) == CI)
470 /// Return true if the specified value is the same when the return would exit
471 /// as it was when the initial iteration of the recursive function was executed.
473 /// We currently handle static constants and arguments that are not modified as
474 /// part of the recursion.
475 static bool isDynamicConstant(Value *V, CallInst *CI, ReturnInst *RI) {
476 if (isa<Constant>(V)) return true; // Static constants are always dyn consts
478 // Check to see if this is an immutable argument, if so, the value
479 // will be available to initialize the accumulator.
480 if (Argument *Arg = dyn_cast<Argument>(V)) {
481 // Figure out which argument number this is...
483 Function *F = CI->getParent()->getParent();
484 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI)
487 // If we are passing this argument into call as the corresponding
488 // argument operand, then the argument is dynamically constant.
489 // Otherwise, we cannot transform this function safely.
490 if (CI->getArgOperand(ArgNo) == Arg)
494 // Switch cases are always constant integers. If the value is being switched
495 // on and the return is only reachable from one of its cases, it's
496 // effectively constant.
497 if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor())
498 if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator()))
499 if (SI->getCondition() == V)
500 return SI->getDefaultDest() != RI->getParent();
502 // Not a constant or immutable argument, we can't safely transform.
506 /// Check to see if the function containing the specified tail call consistently
507 /// returns the same runtime-constant value at all exit points except for
508 /// IgnoreRI. If so, return the returned value.
509 static Value *getCommonReturnValue(ReturnInst *IgnoreRI, CallInst *CI) {
510 Function *F = CI->getParent()->getParent();
511 Value *ReturnedValue = nullptr;
513 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) {
514 ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator());
515 if (RI == nullptr || RI == IgnoreRI) continue;
517 // We can only perform this transformation if the value returned is
518 // evaluatable at the start of the initial invocation of the function,
519 // instead of at the end of the evaluation.
521 Value *RetOp = RI->getOperand(0);
522 if (!isDynamicConstant(RetOp, CI, RI))
525 if (ReturnedValue && RetOp != ReturnedValue)
526 return nullptr; // Cannot transform if differing values are returned.
527 ReturnedValue = RetOp;
529 return ReturnedValue;
532 /// If the specified instruction can be transformed using accumulator recursion
533 /// elimination, return the constant which is the start of the accumulator
534 /// value. Otherwise return null.
535 Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I,
537 if (!I->isAssociative() || !I->isCommutative()) return nullptr;
538 assert(I->getNumOperands() == 2 &&
539 "Associative/commutative operations should have 2 args!");
541 // Exactly one operand should be the result of the call instruction.
542 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) ||
543 (I->getOperand(0) != CI && I->getOperand(1) != CI))
546 // The only user of this instruction we allow is a single return instruction.
547 if (!I->hasOneUse() || !isa<ReturnInst>(I->user_back()))
550 // Ok, now we have to check all of the other return instructions in this
551 // function. If they return non-constants or differing values, then we cannot
552 // transform the function safely.
553 return getCommonReturnValue(cast<ReturnInst>(I->user_back()), CI);
556 static Instruction *FirstNonDbg(BasicBlock::iterator I) {
557 while (isa<DbgInfoIntrinsic>(I))
563 TailCallElim::FindTRECandidate(Instruction *TI,
564 bool CannotTailCallElimCallsMarkedTail) {
565 BasicBlock *BB = TI->getParent();
566 Function *F = BB->getParent();
568 if (&BB->front() == TI) // Make sure there is something before the terminator.
571 // Scan backwards from the return, checking to see if there is a tail call in
572 // this block. If so, set CI to it.
573 CallInst *CI = nullptr;
574 BasicBlock::iterator BBI = TI;
576 CI = dyn_cast<CallInst>(BBI);
577 if (CI && CI->getCalledFunction() == F)
580 if (BBI == BB->begin())
581 return nullptr; // Didn't find a potential tail call.
585 // If this call is marked as a tail call, and if there are dynamic allocas in
586 // the function, we cannot perform this optimization.
587 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail)
590 // As a special case, detect code like this:
591 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call
592 // and disable this xform in this case, because the code generator will
593 // lower the call to fabs into inline code.
594 if (BB == &F->getEntryBlock() &&
595 FirstNonDbg(BB->front()) == CI &&
596 FirstNonDbg(std::next(BB->begin())) == TI &&
597 CI->getCalledFunction() &&
598 !TTI->isLoweredToCall(CI->getCalledFunction())) {
599 // A single-block function with just a call and a return. Check that
600 // the arguments match.
601 CallSite::arg_iterator I = CallSite(CI).arg_begin(),
602 E = CallSite(CI).arg_end();
603 Function::arg_iterator FI = F->arg_begin(),
605 for (; I != E && FI != FE; ++I, ++FI)
606 if (*I != &*FI) break;
607 if (I == E && FI == FE)
614 bool TailCallElim::EliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret,
615 BasicBlock *&OldEntry,
616 bool &TailCallsAreMarkedTail,
617 SmallVectorImpl<PHINode *> &ArgumentPHIs,
618 bool CannotTailCallElimCallsMarkedTail) {
619 // If we are introducing accumulator recursion to eliminate operations after
620 // the call instruction that are both associative and commutative, the initial
621 // value for the accumulator is placed in this variable. If this value is set
622 // then we actually perform accumulator recursion elimination instead of
623 // simple tail recursion elimination. If the operation is an LLVM instruction
624 // (eg: "add") then it is recorded in AccumulatorRecursionInstr. If not, then
625 // we are handling the case when the return instruction returns a constant C
626 // which is different to the constant returned by other return instructions
627 // (which is recorded in AccumulatorRecursionEliminationInitVal). This is a
628 // special case of accumulator recursion, the operation being "return C".
629 Value *AccumulatorRecursionEliminationInitVal = nullptr;
630 Instruction *AccumulatorRecursionInstr = nullptr;
632 // Ok, we found a potential tail call. We can currently only transform the
633 // tail call if all of the instructions between the call and the return are
634 // movable to above the call itself, leaving the call next to the return.
635 // Check that this is the case now.
636 BasicBlock::iterator BBI = CI;
637 for (++BBI; &*BBI != Ret; ++BBI) {
638 if (CanMoveAboveCall(BBI, CI)) continue;
640 // If we can't move the instruction above the call, it might be because it
641 // is an associative and commutative operation that could be transformed
642 // using accumulator recursion elimination. Check to see if this is the
643 // case, and if so, remember the initial accumulator value for later.
644 if ((AccumulatorRecursionEliminationInitVal =
645 CanTransformAccumulatorRecursion(BBI, CI))) {
646 // Yes, this is accumulator recursion. Remember which instruction
648 AccumulatorRecursionInstr = BBI;
650 return false; // Otherwise, we cannot eliminate the tail recursion!
654 // We can only transform call/return pairs that either ignore the return value
655 // of the call and return void, ignore the value of the call and return a
656 // constant, return the value returned by the tail call, or that are being
657 // accumulator recursion variable eliminated.
658 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI &&
659 !isa<UndefValue>(Ret->getReturnValue()) &&
660 AccumulatorRecursionEliminationInitVal == nullptr &&
661 !getCommonReturnValue(nullptr, CI)) {
662 // One case remains that we are able to handle: the current return
663 // instruction returns a constant, and all other return instructions
664 // return a different constant.
665 if (!isDynamicConstant(Ret->getReturnValue(), CI, Ret))
666 return false; // Current return instruction does not return a constant.
667 // Check that all other return instructions return a common constant. If
668 // so, record it in AccumulatorRecursionEliminationInitVal.
669 AccumulatorRecursionEliminationInitVal = getCommonReturnValue(Ret, CI);
670 if (!AccumulatorRecursionEliminationInitVal)
674 BasicBlock *BB = Ret->getParent();
675 Function *F = BB->getParent();
677 emitOptimizationRemark(F->getContext(), "tailcallelim", *F, CI->getDebugLoc(),
678 "transforming tail recursion to loop");
680 // OK! We can transform this tail call. If this is the first one found,
681 // create the new entry block, allowing us to branch back to the old entry.
683 OldEntry = &F->getEntryBlock();
684 BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry);
685 NewEntry->takeName(OldEntry);
686 OldEntry->setName("tailrecurse");
687 BranchInst::Create(OldEntry, NewEntry);
689 // If this tail call is marked 'tail' and if there are any allocas in the
690 // entry block, move them up to the new entry block.
691 TailCallsAreMarkedTail = CI->isTailCall();
692 if (TailCallsAreMarkedTail)
693 // Move all fixed sized allocas from OldEntry to NewEntry.
694 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(),
695 NEBI = NewEntry->begin(); OEBI != E; )
696 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++))
697 if (isa<ConstantInt>(AI->getArraySize()))
698 AI->moveBefore(NEBI);
700 // Now that we have created a new block, which jumps to the entry
701 // block, insert a PHI node for each argument of the function.
702 // For now, we initialize each PHI to only have the real arguments
703 // which are passed in.
704 Instruction *InsertPos = OldEntry->begin();
705 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
707 PHINode *PN = PHINode::Create(I->getType(), 2,
708 I->getName() + ".tr", InsertPos);
709 I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
710 PN->addIncoming(I, NewEntry);
711 ArgumentPHIs.push_back(PN);
715 // If this function has self recursive calls in the tail position where some
716 // are marked tail and some are not, only transform one flavor or another. We
717 // have to choose whether we move allocas in the entry block to the new entry
718 // block or not, so we can't make a good choice for both. NOTE: We could do
719 // slightly better here in the case that the function has no entry block
721 if (TailCallsAreMarkedTail && !CI->isTailCall())
724 // Ok, now that we know we have a pseudo-entry block WITH all of the
725 // required PHI nodes, add entries into the PHI node for the actual
726 // parameters passed into the tail-recursive call.
727 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i)
728 ArgumentPHIs[i]->addIncoming(CI->getArgOperand(i), BB);
730 // If we are introducing an accumulator variable to eliminate the recursion,
731 // do so now. Note that we _know_ that no subsequent tail recursion
732 // eliminations will happen on this function because of the way the
733 // accumulator recursion predicate is set up.
735 if (AccumulatorRecursionEliminationInitVal) {
736 Instruction *AccRecInstr = AccumulatorRecursionInstr;
737 // Start by inserting a new PHI node for the accumulator.
738 pred_iterator PB = pred_begin(OldEntry), PE = pred_end(OldEntry);
740 PHINode::Create(AccumulatorRecursionEliminationInitVal->getType(),
741 std::distance(PB, PE) + 1,
742 "accumulator.tr", OldEntry->begin());
744 // Loop over all of the predecessors of the tail recursion block. For the
745 // real entry into the function we seed the PHI with the initial value,
746 // computed earlier. For any other existing branches to this block (due to
747 // other tail recursions eliminated) the accumulator is not modified.
748 // Because we haven't added the branch in the current block to OldEntry yet,
749 // it will not show up as a predecessor.
750 for (pred_iterator PI = PB; PI != PE; ++PI) {
752 if (P == &F->getEntryBlock())
753 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, P);
755 AccPN->addIncoming(AccPN, P);
759 // Add an incoming argument for the current block, which is computed by
760 // our associative and commutative accumulator instruction.
761 AccPN->addIncoming(AccRecInstr, BB);
763 // Next, rewrite the accumulator recursion instruction so that it does not
764 // use the result of the call anymore, instead, use the PHI node we just
766 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN);
768 // Add an incoming argument for the current block, which is just the
769 // constant returned by the current return instruction.
770 AccPN->addIncoming(Ret->getReturnValue(), BB);
773 // Finally, rewrite any return instructions in the program to return the PHI
774 // node instead of the "initval" that they do currently. This loop will
775 // actually rewrite the return value we are destroying, but that's ok.
776 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
777 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
778 RI->setOperand(0, AccPN);
782 // Now that all of the PHI nodes are in place, remove the call and
783 // ret instructions, replacing them with an unconditional branch.
784 BranchInst *NewBI = BranchInst::Create(OldEntry, Ret);
785 NewBI->setDebugLoc(CI->getDebugLoc());
787 BB->getInstList().erase(Ret); // Remove return.
788 BB->getInstList().erase(CI); // Remove call.
793 bool TailCallElim::FoldReturnAndProcessPred(BasicBlock *BB,
794 ReturnInst *Ret, BasicBlock *&OldEntry,
795 bool &TailCallsAreMarkedTail,
796 SmallVectorImpl<PHINode *> &ArgumentPHIs,
797 bool CannotTailCallElimCallsMarkedTail) {
800 // If the return block contains nothing but the return and PHI's,
801 // there might be an opportunity to duplicate the return in its
802 // predecessors and perform TRC there. Look for predecessors that end
803 // in unconditional branch and recursive call(s).
804 SmallVector<BranchInst*, 8> UncondBranchPreds;
805 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
806 BasicBlock *Pred = *PI;
807 TerminatorInst *PTI = Pred->getTerminator();
808 if (BranchInst *BI = dyn_cast<BranchInst>(PTI))
809 if (BI->isUnconditional())
810 UncondBranchPreds.push_back(BI);
813 while (!UncondBranchPreds.empty()) {
814 BranchInst *BI = UncondBranchPreds.pop_back_val();
815 BasicBlock *Pred = BI->getParent();
816 if (CallInst *CI = FindTRECandidate(BI, CannotTailCallElimCallsMarkedTail)){
817 DEBUG(dbgs() << "FOLDING: " << *BB
818 << "INTO UNCOND BRANCH PRED: " << *Pred);
819 ReturnInst *RI = FoldReturnIntoUncondBranch(Ret, BB, Pred);
821 // Cleanup: if all predecessors of BB have been eliminated by
822 // FoldReturnIntoUncondBranch, delete it. It is important to empty it,
823 // because the ret instruction in there is still using a value which
824 // EliminateRecursiveTailCall will attempt to remove.
825 if (!BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
826 BB->eraseFromParent();
828 EliminateRecursiveTailCall(CI, RI, OldEntry, TailCallsAreMarkedTail,
830 CannotTailCallElimCallsMarkedTail);
840 TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry,
841 bool &TailCallsAreMarkedTail,
842 SmallVectorImpl<PHINode *> &ArgumentPHIs,
843 bool CannotTailCallElimCallsMarkedTail) {
844 CallInst *CI = FindTRECandidate(Ret, CannotTailCallElimCallsMarkedTail);
848 return EliminateRecursiveTailCall(CI, Ret, OldEntry, TailCallsAreMarkedTail,
850 CannotTailCallElimCallsMarkedTail);