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 expression to use an accumulator variable,
20 // thus compiling the typical naive factorial or 'fib' implementation into
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 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.
51 //===----------------------------------------------------------------------===//
53 #define DEBUG_TYPE "tailcallelim"
54 #include "llvm/Transforms/Scalar.h"
55 #include "llvm/Transforms/Utils/Local.h"
56 #include "llvm/Constants.h"
57 #include "llvm/DerivedTypes.h"
58 #include "llvm/Function.h"
59 #include "llvm/Instructions.h"
60 #include "llvm/Pass.h"
61 #include "llvm/Analysis/CaptureTracking.h"
62 #include "llvm/Analysis/InlineCost.h"
63 #include "llvm/Support/CallSite.h"
64 #include "llvm/Support/CFG.h"
65 #include "llvm/ADT/Statistic.h"
68 STATISTIC(NumEliminated, "Number of tail calls removed");
69 STATISTIC(NumAccumAdded, "Number of accumulators introduced");
72 struct TailCallElim : public FunctionPass {
73 static char ID; // Pass identification, replacement for typeid
74 TailCallElim() : FunctionPass(&ID) {}
76 virtual bool runOnFunction(Function &F);
79 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry,
80 bool &TailCallsAreMarkedTail,
81 SmallVector<PHINode*, 8> &ArgumentPHIs,
82 bool CannotTailCallElimCallsMarkedTail);
83 bool CanMoveAboveCall(Instruction *I, CallInst *CI);
84 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI);
88 char TailCallElim::ID = 0;
89 static RegisterPass<TailCallElim> X("tailcallelim", "Tail Call Elimination");
91 // Public interface to the TailCallElimination pass
92 FunctionPass *llvm::createTailCallEliminationPass() {
93 return new TailCallElim();
96 /// AllocaMightEscapeToCalls - Return true if this alloca may be accessed by
97 /// callees of this function. We only do very simple analysis right now, this
98 /// could be expanded in the future to use mod/ref information for particular
99 /// call sites if desired.
100 static bool AllocaMightEscapeToCalls(AllocaInst *AI) {
101 // FIXME: do simple 'address taken' analysis.
105 /// CheckForEscapingAllocas - Scan the specified basic block for alloca
106 /// instructions. If it contains any that might be accessed by calls, return
108 static bool CheckForEscapingAllocas(BasicBlock *BB,
109 bool &CannotTCETailMarkedCall) {
111 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
112 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
113 RetVal |= AllocaMightEscapeToCalls(AI);
115 // If this alloca is in the body of the function, or if it is a variable
116 // sized allocation, we cannot tail call eliminate calls marked 'tail'
117 // with this mechanism.
118 if (BB != &BB->getParent()->getEntryBlock() ||
119 !isa<ConstantInt>(AI->getArraySize()))
120 CannotTCETailMarkedCall = true;
125 bool TailCallElim::runOnFunction(Function &F) {
126 // If this function is a varargs function, we won't be able to PHI the args
127 // right, so don't even try to convert it...
128 if (F.getFunctionType()->isVarArg()) return false;
130 BasicBlock *OldEntry = 0;
131 bool TailCallsAreMarkedTail = false;
132 SmallVector<PHINode*, 8> ArgumentPHIs;
133 bool MadeChange = false;
135 bool FunctionContainsEscapingAllocas = false;
137 // CannotTCETailMarkedCall - If true, we cannot perform TCE on tail calls
138 // marked with the 'tail' attribute, because doing so would cause the stack
139 // size to increase (real TCE would deallocate variable sized allocas, TCE
141 bool CannotTCETailMarkedCall = false;
143 // Loop over the function, looking for any returning blocks, and keeping track
144 // of whether this function has any non-trivially used allocas.
145 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
146 if (FunctionContainsEscapingAllocas && CannotTCETailMarkedCall)
149 FunctionContainsEscapingAllocas |=
150 CheckForEscapingAllocas(BB, CannotTCETailMarkedCall);
153 /// FIXME: The code generator produces really bad code when an 'escaping
154 /// alloca' is changed from being a static alloca to being a dynamic alloca.
155 /// Until this is resolved, disable this transformation if that would ever
156 /// happen. This bug is PR962.
157 if (FunctionContainsEscapingAllocas)
160 // Second pass, change any tail calls to loops.
161 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
162 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator()))
163 MadeChange |= ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail,
164 ArgumentPHIs,CannotTCETailMarkedCall);
166 // If we eliminated any tail recursions, it's possible that we inserted some
167 // silly PHI nodes which just merge an initial value (the incoming operand)
168 // with themselves. Check to see if we did and clean up our mess if so. This
169 // occurs when a function passes an argument straight through to its tail
171 if (!ArgumentPHIs.empty()) {
172 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) {
173 PHINode *PN = ArgumentPHIs[i];
175 // If the PHI Node is a dynamic constant, replace it with the value it is.
176 if (Value *PNV = PN->hasConstantValue()) {
177 PN->replaceAllUsesWith(PNV);
178 PN->eraseFromParent();
183 // Finally, if this function contains no non-escaping allocas, mark all calls
184 // in the function as eligible for tail calls (there is no stack memory for
186 if (!FunctionContainsEscapingAllocas)
187 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
188 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
189 if (CallInst *CI = dyn_cast<CallInst>(I)) {
198 /// CanMoveAboveCall - Return true if it is safe to move the specified
199 /// instruction from after the call to before the call, assuming that all
200 /// instructions between the call and this instruction are movable.
202 bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) {
203 // FIXME: We can move load/store/call/free instructions above the call if the
204 // call does not mod/ref the memory location being processed.
205 if (I->mayHaveSideEffects()) // This also handles volatile loads.
208 if (LoadInst *L = dyn_cast<LoadInst>(I)) {
209 // Loads may always be moved above calls without side effects.
210 if (CI->mayHaveSideEffects()) {
211 // Non-volatile loads may be moved above a call with side effects if it
212 // does not write to memory and the load provably won't trap.
213 // FIXME: Writes to memory only matter if they may alias the pointer
214 // being loaded from.
215 if (CI->mayWriteToMemory() ||
216 !isSafeToLoadUnconditionally(L->getPointerOperand(), L,
222 // Otherwise, if this is a side-effect free instruction, check to make sure
223 // that it does not use the return value of the call. If it doesn't use the
224 // return value of the call, it must only use things that are defined before
225 // the call, or movable instructions between the call and the instruction
227 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
228 if (I->getOperand(i) == CI)
233 // isDynamicConstant - Return true if the specified value is the same when the
234 // return would exit as it was when the initial iteration of the recursive
235 // function was executed.
237 // We currently handle static constants and arguments that are not modified as
238 // part of the recursion.
240 static bool isDynamicConstant(Value *V, CallInst *CI, ReturnInst *RI) {
241 if (isa<Constant>(V)) return true; // Static constants are always dyn consts
243 // Check to see if this is an immutable argument, if so, the value
244 // will be available to initialize the accumulator.
245 if (Argument *Arg = dyn_cast<Argument>(V)) {
246 // Figure out which argument number this is...
248 Function *F = CI->getParent()->getParent();
249 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI)
252 // If we are passing this argument into call as the corresponding
253 // argument operand, then the argument is dynamically constant.
254 // Otherwise, we cannot transform this function safely.
255 if (CI->getOperand(ArgNo) == Arg)
259 // Switch cases are always constant integers. If the value is being switched
260 // on and the return is only reachable from one of its cases, it's
261 // effectively constant.
262 if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor())
263 if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator()))
264 if (SI->getCondition() == V)
265 return SI->getDefaultDest() != RI->getParent();
267 // Not a constant or immutable argument, we can't safely transform.
271 // getCommonReturnValue - Check to see if the function containing the specified
272 // return instruction and tail call consistently returns the same
273 // runtime-constant value at all exit points. If so, return the returned value.
275 static Value *getCommonReturnValue(ReturnInst *TheRI, CallInst *CI) {
276 Function *F = TheRI->getParent()->getParent();
277 Value *ReturnedValue = 0;
279 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
280 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
282 Value *RetOp = RI->getOperand(0);
284 // We can only perform this transformation if the value returned is
285 // evaluatable at the start of the initial invocation of the function,
286 // instead of at the end of the evaluation.
288 if (!isDynamicConstant(RetOp, CI, RI))
291 if (ReturnedValue && RetOp != ReturnedValue)
292 return 0; // Cannot transform if differing values are returned.
293 ReturnedValue = RetOp;
295 return ReturnedValue;
298 /// CanTransformAccumulatorRecursion - If the specified instruction can be
299 /// transformed using accumulator recursion elimination, return the constant
300 /// which is the start of the accumulator value. Otherwise return null.
302 Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I,
304 if (!I->isAssociative()) return 0;
305 assert(I->getNumOperands() == 2 &&
306 "Associative operations should have 2 args!");
308 // Exactly one operand should be the result of the call instruction...
309 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) ||
310 (I->getOperand(0) != CI && I->getOperand(1) != CI))
313 // The only user of this instruction we allow is a single return instruction.
314 if (!I->hasOneUse() || !isa<ReturnInst>(I->use_back()))
317 // Ok, now we have to check all of the other return instructions in this
318 // function. If they return non-constants or differing values, then we cannot
319 // transform the function safely.
320 return getCommonReturnValue(cast<ReturnInst>(I->use_back()), CI);
323 bool TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry,
324 bool &TailCallsAreMarkedTail,
325 SmallVector<PHINode*, 8> &ArgumentPHIs,
326 bool CannotTailCallElimCallsMarkedTail) {
327 BasicBlock *BB = Ret->getParent();
328 Function *F = BB->getParent();
330 if (&BB->front() == Ret) // Make sure there is something before the ret...
333 // Scan backwards from the return, checking to see if there is a tail call in
334 // this block. If so, set CI to it.
336 BasicBlock::iterator BBI = Ret;
338 CI = dyn_cast<CallInst>(BBI);
339 if (CI && CI->getCalledFunction() == F)
342 if (BBI == BB->begin())
343 return false; // Didn't find a potential tail call.
347 // If this call is marked as a tail call, and if there are dynamic allocas in
348 // the function, we cannot perform this optimization.
349 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail)
352 // As a special case, detect code like this:
353 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call
354 // and disable this xform in this case, because the code generator will
355 // lower the call to fabs into inline code.
356 if (BB == &F->getEntryBlock() &&
357 &BB->front() == CI && &*++BB->begin() == Ret &&
359 // A single-block function with just a call and a return. Check that
360 // the arguments match.
361 CallSite::arg_iterator I = CallSite(CI).arg_begin(),
362 E = CallSite(CI).arg_end();
363 Function::arg_iterator FI = F->arg_begin(),
365 for (; I != E && FI != FE; ++I, ++FI)
366 if (*I != &*FI) break;
367 if (I == E && FI == FE)
371 // If we are introducing accumulator recursion to eliminate associative
372 // operations after the call instruction, this variable contains the initial
373 // value for the accumulator. If this value is set, we actually perform
374 // accumulator recursion elimination instead of simple tail recursion
376 Value *AccumulatorRecursionEliminationInitVal = 0;
377 Instruction *AccumulatorRecursionInstr = 0;
379 // Ok, we found a potential tail call. We can currently only transform the
380 // tail call if all of the instructions between the call and the return are
381 // movable to above the call itself, leaving the call next to the return.
382 // Check that this is the case now.
383 for (BBI = CI, ++BBI; &*BBI != Ret; ++BBI)
384 if (!CanMoveAboveCall(BBI, CI)) {
385 // If we can't move the instruction above the call, it might be because it
386 // is an associative operation that could be tranformed using accumulator
387 // recursion elimination. Check to see if this is the case, and if so,
388 // remember the initial accumulator value for later.
389 if ((AccumulatorRecursionEliminationInitVal =
390 CanTransformAccumulatorRecursion(BBI, CI))) {
391 // Yes, this is accumulator recursion. Remember which instruction
393 AccumulatorRecursionInstr = BBI;
395 return false; // Otherwise, we cannot eliminate the tail recursion!
399 // We can only transform call/return pairs that either ignore the return value
400 // of the call and return void, ignore the value of the call and return a
401 // constant, return the value returned by the tail call, or that are being
402 // accumulator recursion variable eliminated.
403 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI &&
404 !isa<UndefValue>(Ret->getReturnValue()) &&
405 AccumulatorRecursionEliminationInitVal == 0 &&
406 !getCommonReturnValue(Ret, CI))
409 // OK! We can transform this tail call. If this is the first one found,
410 // create the new entry block, allowing us to branch back to the old entry.
412 OldEntry = &F->getEntryBlock();
413 BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry);
414 NewEntry->takeName(OldEntry);
415 OldEntry->setName("tailrecurse");
416 BranchInst::Create(OldEntry, NewEntry);
418 // If this tail call is marked 'tail' and if there are any allocas in the
419 // entry block, move them up to the new entry block.
420 TailCallsAreMarkedTail = CI->isTailCall();
421 if (TailCallsAreMarkedTail)
422 // Move all fixed sized allocas from OldEntry to NewEntry.
423 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(),
424 NEBI = NewEntry->begin(); OEBI != E; )
425 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++))
426 if (isa<ConstantInt>(AI->getArraySize()))
427 AI->moveBefore(NEBI);
429 // Now that we have created a new block, which jumps to the entry
430 // block, insert a PHI node for each argument of the function.
431 // For now, we initialize each PHI to only have the real arguments
432 // which are passed in.
433 Instruction *InsertPos = OldEntry->begin();
434 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
436 PHINode *PN = PHINode::Create(I->getType(),
437 I->getName() + ".tr", InsertPos);
438 I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
439 PN->addIncoming(I, NewEntry);
440 ArgumentPHIs.push_back(PN);
444 // If this function has self recursive calls in the tail position where some
445 // are marked tail and some are not, only transform one flavor or another. We
446 // have to choose whether we move allocas in the entry block to the new entry
447 // block or not, so we can't make a good choice for both. NOTE: We could do
448 // slightly better here in the case that the function has no entry block
450 if (TailCallsAreMarkedTail && !CI->isTailCall())
453 // Ok, now that we know we have a pseudo-entry block WITH all of the
454 // required PHI nodes, add entries into the PHI node for the actual
455 // parameters passed into the tail-recursive call.
456 for (unsigned i = 0, e = CI->getNumOperands()-1; i != e; ++i)
457 ArgumentPHIs[i]->addIncoming(CI->getOperand(i), BB);
459 // If we are introducing an accumulator variable to eliminate the recursion,
460 // do so now. Note that we _know_ that no subsequent tail recursion
461 // eliminations will happen on this function because of the way the
462 // accumulator recursion predicate is set up.
464 if (AccumulatorRecursionEliminationInitVal) {
465 Instruction *AccRecInstr = AccumulatorRecursionInstr;
466 // Start by inserting a new PHI node for the accumulator.
467 PHINode *AccPN = PHINode::Create(AccRecInstr->getType(), "accumulator.tr",
470 // Loop over all of the predecessors of the tail recursion block. For the
471 // real entry into the function we seed the PHI with the initial value,
472 // computed earlier. For any other existing branches to this block (due to
473 // other tail recursions eliminated) the accumulator is not modified.
474 // Because we haven't added the branch in the current block to OldEntry yet,
475 // it will not show up as a predecessor.
476 for (pred_iterator PI = pred_begin(OldEntry), PE = pred_end(OldEntry);
478 if (*PI == &F->getEntryBlock())
479 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, *PI);
481 AccPN->addIncoming(AccPN, *PI);
484 // Add an incoming argument for the current block, which is computed by our
485 // associative accumulator instruction.
486 AccPN->addIncoming(AccRecInstr, BB);
488 // Next, rewrite the accumulator recursion instruction so that it does not
489 // use the result of the call anymore, instead, use the PHI node we just
491 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN);
493 // Finally, rewrite any return instructions in the program to return the PHI
494 // node instead of the "initval" that they do currently. This loop will
495 // actually rewrite the return value we are destroying, but that's ok.
496 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
497 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
498 RI->setOperand(0, AccPN);
502 // Now that all of the PHI nodes are in place, remove the call and
503 // ret instructions, replacing them with an unconditional branch.
504 BranchInst::Create(OldEntry, Ret);
505 BB->getInstList().erase(Ret); // Remove return.
506 BB->getInstList().erase(CI); // Remove call.