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