1 //===- TailRecursionElimination.cpp - Eliminate Tail Calls ----------------===//
3 // The LLVM Compiler Infrastructure
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.
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 theier 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 #include "llvm/Transforms/Scalar.h"
54 #include "llvm/DerivedTypes.h"
55 #include "llvm/Function.h"
56 #include "llvm/Instructions.h"
57 #include "llvm/Pass.h"
58 #include "llvm/Support/CFG.h"
59 #include "llvm/ADT/Statistic.h"
63 Statistic<> NumEliminated("tailcallelim", "Number of tail calls removed");
64 Statistic<> NumAccumAdded("tailcallelim","Number of accumulators introduced");
66 struct TailCallElim : public FunctionPass {
67 virtual bool runOnFunction(Function &F);
70 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry,
71 std::vector<PHINode*> &ArgumentPHIs);
72 bool CanMoveAboveCall(Instruction *I, CallInst *CI);
73 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI);
75 RegisterOpt<TailCallElim> X("tailcallelim", "Tail Call Elimination");
78 // Public interface to the TailCallElimination pass
79 FunctionPass *llvm::createTailCallEliminationPass() {
80 return new TailCallElim();
84 /// AllocaMightEscapeToCalls - Return true if this alloca may be accessed by
85 /// callees of this function. We only do very simple analysis right now, this
86 /// could be expanded in the future to use mod/ref information for particular
87 /// call sites if desired.
88 static bool AllocaMightEscapeToCalls(AllocaInst *AI) {
89 // FIXME: do simple 'address taken' analysis.
93 /// FunctionContainsAllocas - Scan the specified basic block for alloca
94 /// instructions. If it contains any that might be accessed by calls, return
96 static bool CheckForEscapingAllocas(BasicBlock *BB) {
97 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
98 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
99 if (AllocaMightEscapeToCalls(AI))
104 bool TailCallElim::runOnFunction(Function &F) {
105 // If this function is a varargs function, we won't be able to PHI the args
106 // right, so don't even try to convert it...
107 if (F.getFunctionType()->isVarArg()) return false;
109 BasicBlock *OldEntry = 0;
110 std::vector<PHINode*> ArgumentPHIs;
111 bool MadeChange = false;
113 bool FunctionContainsEscapingAllocas = false;
115 // Loop over the function, looking for any returning blocks, and keeping track
116 // of whether this function has any non-trivially used allocas.
117 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
118 if (!FunctionContainsEscapingAllocas)
119 FunctionContainsEscapingAllocas = CheckForEscapingAllocas(BB);
121 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator()))
122 MadeChange |= ProcessReturningBlock(Ret, OldEntry, ArgumentPHIs);
125 // If we eliminated any tail recursions, it's possible that we inserted some
126 // silly PHI nodes which just merge an initial value (the incoming operand)
127 // with themselves. Check to see if we did and clean up our mess if so. This
128 // occurs when a function passes an argument straight through to its tail
130 if (!ArgumentPHIs.empty()) {
131 unsigned NumIncoming = ArgumentPHIs[0]->getNumIncomingValues();
132 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) {
133 PHINode *PN = ArgumentPHIs[i];
135 for (unsigned op = 0, e = NumIncoming; op != e; ++op) {
136 Value *Op = PN->getIncomingValue(op);
139 V = Op; // First value seen?
140 } else if (V != Op) {
147 // If the PHI Node is a dynamic constant, replace it with the value it is.
149 PN->replaceAllUsesWith(V);
150 PN->getParent()->getInstList().erase(PN);
155 // Finally, if this function contains no non-escaping allocas, mark all calls
156 // in the function as eligible for tail calls (there is no stack memory for
158 if (!FunctionContainsEscapingAllocas)
159 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
160 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
161 if (CallInst *CI = dyn_cast<CallInst>(I))
168 /// CanMoveAboveCall - Return true if it is safe to move the specified
169 /// instruction from after the call to before the call, assuming that all
170 /// instructions between the call and this instruction are movable.
172 bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) {
173 // FIXME: We can move load/store/call/free instructions above the call if the
174 // call does not mod/ref the memory location being processed.
175 if (I->mayWriteToMemory() || isa<LoadInst>(I))
178 // Otherwise, if this is a side-effect free instruction, check to make sure
179 // that it does not use the return value of the call. If it doesn't use the
180 // return value of the call, it must only use things that are defined before
181 // the call, or movable instructions between the call and the instruction
183 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
184 if (I->getOperand(i) == CI)
189 // isDynamicConstant - Return true if the specified value is the same when the
190 // return would exit as it was when the initial iteration of the recursive
191 // function was executed.
193 // We currently handle static constants and arguments that are not modified as
194 // part of the recursion.
196 static bool isDynamicConstant(Value *V, CallInst *CI) {
197 if (isa<Constant>(V)) return true; // Static constants are always dyn consts
199 // Check to see if this is an immutable argument, if so, the value
200 // will be available to initialize the accumulator.
201 if (Argument *Arg = dyn_cast<Argument>(V)) {
202 // Figure out which argument number this is...
204 Function *F = CI->getParent()->getParent();
205 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI)
208 // If we are passing this argument into call as the corresponding
209 // argument operand, then the argument is dynamically constant.
210 // Otherwise, we cannot transform this function safely.
211 if (CI->getOperand(ArgNo+1) == Arg)
214 // Not a constant or immutable argument, we can't safely transform.
218 // getCommonReturnValue - Check to see if the function containing the specified
219 // return instruction and tail call consistently returns the same
220 // runtime-constant value at all exit points. If so, return the returned value.
222 static Value *getCommonReturnValue(ReturnInst *TheRI, CallInst *CI) {
223 Function *F = TheRI->getParent()->getParent();
224 Value *ReturnedValue = 0;
226 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
227 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
229 Value *RetOp = RI->getOperand(0);
231 // We can only perform this transformation if the value returned is
232 // evaluatable at the start of the initial invocation of the function,
233 // instead of at the end of the evaluation.
235 if (!isDynamicConstant(RetOp, CI))
238 if (ReturnedValue && RetOp != ReturnedValue)
239 return 0; // Cannot transform if differing values are returned.
240 ReturnedValue = RetOp;
242 return ReturnedValue;
245 /// CanTransformAccumulatorRecursion - If the specified instruction can be
246 /// transformed using accumulator recursion elimination, return the constant
247 /// which is the start of the accumulator value. Otherwise return null.
249 Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I,
251 if (!I->isAssociative()) return 0;
252 assert(I->getNumOperands() == 2 &&
253 "Associative operations should have 2 args!");
255 // Exactly one operand should be the result of the call instruction...
256 if (I->getOperand(0) == CI && I->getOperand(1) == CI ||
257 I->getOperand(0) != CI && I->getOperand(1) != CI)
260 // The only user of this instruction we allow is a single return instruction.
261 if (!I->hasOneUse() || !isa<ReturnInst>(I->use_back()))
264 // Ok, now we have to check all of the other return instructions in this
265 // function. If they return non-constants or differing values, then we cannot
266 // transform the function safely.
267 return getCommonReturnValue(cast<ReturnInst>(I->use_back()), CI);
270 bool TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry,
271 std::vector<PHINode*> &ArgumentPHIs) {
272 BasicBlock *BB = Ret->getParent();
273 Function *F = BB->getParent();
275 if (&BB->front() == Ret) // Make sure there is something before the ret...
278 // Scan backwards from the return, checking to see if there is a tail call in
279 // this block. If so, set CI to it.
281 BasicBlock::iterator BBI = Ret;
283 CI = dyn_cast<CallInst>(BBI);
284 if (CI && CI->getCalledFunction() == F)
287 if (BBI == BB->begin())
288 return false; // Didn't find a potential tail call.
292 // If we are introducing accumulator recursion to eliminate associative
293 // operations after the call instruction, this variable contains the initial
294 // value for the accumulator. If this value is set, we actually perform
295 // accumulator recursion elimination instead of simple tail recursion
297 Value *AccumulatorRecursionEliminationInitVal = 0;
298 Instruction *AccumulatorRecursionInstr = 0;
300 // Ok, we found a potential tail call. We can currently only transform the
301 // tail call if all of the instructions between the call and the return are
302 // movable to above the call itself, leaving the call next to the return.
303 // Check that this is the case now.
304 for (BBI = CI, ++BBI; &*BBI != Ret; ++BBI)
305 if (!CanMoveAboveCall(BBI, CI)) {
306 // If we can't move the instruction above the call, it might be because it
307 // is an associative operation that could be tranformed using accumulator
308 // recursion elimination. Check to see if this is the case, and if so,
309 // remember the initial accumulator value for later.
310 if ((AccumulatorRecursionEliminationInitVal =
311 CanTransformAccumulatorRecursion(BBI, CI))) {
312 // Yes, this is accumulator recursion. Remember which instruction
314 AccumulatorRecursionInstr = BBI;
316 return false; // Otherwise, we cannot eliminate the tail recursion!
320 // We can only transform call/return pairs that either ignore the return value
321 // of the call and return void, ignore the value of the call and return a
322 // constant, return the value returned by the tail call, or that are being
323 // accumulator recursion variable eliminated.
324 if (Ret->getNumOperands() != 0 && Ret->getReturnValue() != CI &&
325 AccumulatorRecursionEliminationInitVal == 0 &&
326 !getCommonReturnValue(Ret, CI))
329 // OK! We can transform this tail call. If this is the first one found,
330 // create the new entry block, allowing us to branch back to the old entry.
332 OldEntry = &F->getEntryBlock();
333 std::string OldName = OldEntry->getName(); OldEntry->setName("tailrecurse");
334 BasicBlock *NewEntry = new BasicBlock(OldName, F, OldEntry);
335 new BranchInst(OldEntry, NewEntry);
337 // Now that we have created a new block, which jumps to the entry
338 // block, insert a PHI node for each argument of the function.
339 // For now, we initialize each PHI to only have the real arguments
340 // which are passed in.
341 Instruction *InsertPos = OldEntry->begin();
342 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
344 PHINode *PN = new PHINode(I->getType(), I->getName()+".tr", InsertPos);
345 I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
346 PN->addIncoming(I, NewEntry);
347 ArgumentPHIs.push_back(PN);
351 // Ok, now that we know we have a pseudo-entry block WITH all of the
352 // required PHI nodes, add entries into the PHI node for the actual
353 // parameters passed into the tail-recursive call.
354 for (unsigned i = 0, e = CI->getNumOperands()-1; i != e; ++i)
355 ArgumentPHIs[i]->addIncoming(CI->getOperand(i+1), BB);
357 // If we are introducing an accumulator variable to eliminate the recursion,
358 // do so now. Note that we _know_ that no subsequent tail recursion
359 // eliminations will happen on this function because of the way the
360 // accumulator recursion predicate is set up.
362 if (AccumulatorRecursionEliminationInitVal) {
363 Instruction *AccRecInstr = AccumulatorRecursionInstr;
364 // Start by inserting a new PHI node for the accumulator.
365 PHINode *AccPN = new PHINode(AccRecInstr->getType(), "accumulator.tr",
368 // Loop over all of the predecessors of the tail recursion block. For the
369 // real entry into the function we seed the PHI with the initial value,
370 // computed earlier. For any other existing branches to this block (due to
371 // other tail recursions eliminated) the accumulator is not modified.
372 // Because we haven't added the branch in the current block to OldEntry yet,
373 // it will not show up as a predecessor.
374 for (pred_iterator PI = pred_begin(OldEntry), PE = pred_end(OldEntry);
376 if (*PI == &F->getEntryBlock())
377 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, *PI);
379 AccPN->addIncoming(AccPN, *PI);
382 // Add an incoming argument for the current block, which is computed by our
383 // associative accumulator instruction.
384 AccPN->addIncoming(AccRecInstr, BB);
386 // Next, rewrite the accumulator recursion instruction so that it does not
387 // use the result of the call anymore, instead, use the PHI node we just
389 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN);
391 // Finally, rewrite any return instructions in the program to return the PHI
392 // node instead of the "initval" that they do currently. This loop will
393 // actually rewrite the return value we are destroying, but that's ok.
394 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
395 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
396 RI->setOperand(0, AccPN);
400 // Now that all of the PHI nodes are in place, remove the call and
401 // ret instructions, replacing them with an unconditional branch.
402 new BranchInst(OldEntry, Ret);
403 BB->getInstList().erase(Ret); // Remove return.
404 BB->getInstList().erase(CI); // Remove call.