1 //===- TailDuplication.cpp - Simplify CFG through tail duplication --------===//
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 pass performs a limited form of tail duplication, intended to simplify
11 // CFGs by removing some unconditional branches. This pass is necessary to
12 // straighten out loops created by the C front-end, but also is capable of
13 // making other code nicer. After this pass is run, the CFG simplify pass
14 // should be run to clean up the mess.
16 // This pass could be enhanced in the future to use profile information to be
19 //===----------------------------------------------------------------------===//
21 #include "llvm/Transforms/Scalar.h"
22 #include "llvm/Constant.h"
23 #include "llvm/Function.h"
24 #include "llvm/iPHINode.h"
25 #include "llvm/iTerminators.h"
26 #include "llvm/Pass.h"
27 #include "llvm/Type.h"
28 #include "llvm/Support/CFG.h"
29 #include "llvm/Support/ValueHolder.h"
30 #include "llvm/Transforms/Utils/Local.h"
31 #include "Support/Debug.h"
32 #include "Support/Statistic.h"
36 Statistic<> NumEliminated("tailduplicate",
37 "Number of unconditional branches eliminated");
38 Statistic<> NumPHINodes("tailduplicate", "Number of phi nodes inserted");
40 class TailDup : public FunctionPass {
41 bool runOnFunction(Function &F);
43 inline bool shouldEliminateUnconditionalBranch(TerminatorInst *TI);
44 inline bool canEliminateUnconditionalBranch(TerminatorInst *TI);
45 inline void eliminateUnconditionalBranch(BranchInst *BI);
46 inline void InsertPHINodesIfNecessary(Instruction *OrigInst, Value *NewInst,
47 BasicBlock *NewBlock);
48 inline Value *GetValueInBlock(BasicBlock *BB, Value *OrigVal,
49 std::map<BasicBlock*, ValueHolder> &ValueMap,
50 std::map<BasicBlock*, ValueHolder> &OutValueMap);
51 inline Value *GetValueOutBlock(BasicBlock *BB, Value *OrigVal,
52 std::map<BasicBlock*, ValueHolder> &ValueMap,
53 std::map<BasicBlock*, ValueHolder> &OutValueMap);
55 RegisterOpt<TailDup> X("tailduplicate", "Tail Duplication");
58 // Public interface to the Tail Duplication pass
59 Pass *llvm::createTailDuplicationPass() { return new TailDup(); }
61 /// runOnFunction - Top level algorithm - Loop over each unconditional branch in
62 /// the function, eliminating it if it looks attractive enough.
64 bool TailDup::runOnFunction(Function &F) {
66 for (Function::iterator I = F.begin(), E = F.end(); I != E; )
67 if (shouldEliminateUnconditionalBranch(I->getTerminator()) &&
68 canEliminateUnconditionalBranch(I->getTerminator())) {
69 eliminateUnconditionalBranch(cast<BranchInst>(I->getTerminator()));
77 /// shouldEliminateUnconditionalBranch - Return true if this branch looks
78 /// attractive to eliminate. We eliminate the branch if the destination basic
79 /// block has <= 5 instructions in it, not counting PHI nodes. In practice,
80 /// since one of these is a terminator instruction, this means that we will add
81 /// up to 4 instructions to the new block.
83 /// We don't count PHI nodes in the count since they will be removed when the
84 /// contents of the block are copied over.
86 bool TailDup::shouldEliminateUnconditionalBranch(TerminatorInst *TI) {
87 BranchInst *BI = dyn_cast<BranchInst>(TI);
88 if (!BI || !BI->isUnconditional()) return false; // Not an uncond branch!
90 BasicBlock *Dest = BI->getSuccessor(0);
91 if (Dest == BI->getParent()) return false; // Do not loop infinitely!
93 // Do not inline a block if we will just get another branch to the same block!
94 if (BranchInst *DBI = dyn_cast<BranchInst>(Dest->getTerminator()))
95 if (DBI->isUnconditional() && DBI->getSuccessor(0) == Dest)
96 return false; // Do not loop infinitely!
98 // Do not bother working on dead blocks...
99 pred_iterator PI = pred_begin(Dest), PE = pred_end(Dest);
100 if (PI == PE && Dest != Dest->getParent()->begin())
101 return false; // It's just a dead block, ignore it...
103 // Also, do not bother with blocks with only a single predecessor: simplify
104 // CFG will fold these two blocks together!
106 if (PI == PE) return false; // Exactly one predecessor!
108 BasicBlock::iterator I = Dest->begin();
109 while (isa<PHINode>(*I)) ++I;
111 for (unsigned Size = 0; I != Dest->end(); ++Size, ++I)
112 if (Size == 6) return false; // The block is too large...
116 /// canEliminateUnconditionalBranch - Unfortunately, the general form of tail
117 /// duplication can do very bad things to SSA form, by destroying arbitrary
118 /// relationships between dominators and dominator frontiers as it processes the
119 /// program. The right solution for this is to have an incrementally updating
120 /// dominator data structure, which can gracefully react to arbitrary
121 /// "addEdge/removeEdge" changes to the CFG. Implementing this is nontrivial,
122 /// however, so we just disable the transformation in cases where it is not
125 bool TailDup::canEliminateUnconditionalBranch(TerminatorInst *TI) {
126 // Basically, we refuse to make the transformation if any of the values
127 // computed in the 'tail' are used in any other basic blocks.
128 BasicBlock *Tail = TI->getSuccessor(0);
129 assert(isa<BranchInst>(TI) && cast<BranchInst>(TI)->isUnconditional());
131 for (BasicBlock::iterator I = Tail->begin(), E = Tail->end(); I != E; ++I)
132 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
134 Instruction *User = cast<Instruction>(*UI);
135 if (User->getParent() != Tail || isa<PHINode>(User))
142 /// eliminateUnconditionalBranch - Clone the instructions from the destination
143 /// block into the source block, eliminating the specified unconditional branch.
144 /// If the destination block defines values used by successors of the dest
145 /// block, we may need to insert PHI nodes.
147 void TailDup::eliminateUnconditionalBranch(BranchInst *Branch) {
148 BasicBlock *SourceBlock = Branch->getParent();
149 BasicBlock *DestBlock = Branch->getSuccessor(0);
150 assert(SourceBlock != DestBlock && "Our predicate is broken!");
152 DEBUG(std::cerr << "TailDuplication[" << SourceBlock->getParent()->getName()
153 << "]: Eliminating branch: " << *Branch);
155 // We are going to have to map operands from the original block B to the new
156 // copy of the block B'. If there are PHI nodes in the DestBlock, these PHI
157 // nodes also define part of this mapping. Loop over these PHI nodes, adding
158 // them to our mapping.
160 std::map<Value*, Value*> ValueMapping;
162 BasicBlock::iterator BI = DestBlock->begin();
163 bool HadPHINodes = isa<PHINode>(BI);
164 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
165 ValueMapping[PN] = PN->getIncomingValueForBlock(SourceBlock);
167 // Clone the non-phi instructions of the dest block into the source block,
168 // keeping track of the mapping...
170 for (; BI != DestBlock->end(); ++BI) {
171 Instruction *New = BI->clone();
172 New->setName(BI->getName());
173 SourceBlock->getInstList().push_back(New);
174 ValueMapping[BI] = New;
177 // Now that we have built the mapping information and cloned all of the
178 // instructions (giving us a new terminator, among other things), walk the new
179 // instructions, rewriting references of old instructions to use new
182 BI = Branch; ++BI; // Get an iterator to the first new instruction
183 for (; BI != SourceBlock->end(); ++BI)
184 for (unsigned i = 0, e = BI->getNumOperands(); i != e; ++i)
185 if (Value *Remapped = ValueMapping[BI->getOperand(i)])
186 BI->setOperand(i, Remapped);
188 // Next we check to see if any of the successors of DestBlock had PHI nodes.
189 // If so, we need to add entries to the PHI nodes for SourceBlock now.
190 for (succ_iterator SI = succ_begin(DestBlock), SE = succ_end(DestBlock);
192 BasicBlock *Succ = *SI;
193 for (BasicBlock::iterator PNI = Succ->begin();
194 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
195 // Ok, we have a PHI node. Figure out what the incoming value was for the
197 Value *IV = PN->getIncomingValueForBlock(DestBlock);
199 // Remap the value if necessary...
200 if (Value *MappedIV = ValueMapping[IV])
202 PN->addIncoming(IV, SourceBlock);
206 // Now that all of the instructions are correctly copied into the SourceBlock,
207 // we have one more minor problem: the successors of the original DestBB may
208 // use the values computed in DestBB either directly (if DestBB dominated the
209 // block), or through a PHI node. In either case, we need to insert PHI nodes
210 // into any successors of DestBB (which are now our successors) for each value
211 // that is computed in DestBB, but is used outside of it. All of these uses
212 // we have to rewrite with the new PHI node.
214 if (succ_begin(SourceBlock) != succ_end(SourceBlock)) // Avoid wasting time...
215 for (BI = DestBlock->begin(); BI != DestBlock->end(); ++BI)
216 if (BI->getType() != Type::VoidTy)
217 InsertPHINodesIfNecessary(BI, ValueMapping[BI], SourceBlock);
219 // Final step: now that we have finished everything up, walk the cloned
220 // instructions one last time, constant propagating and DCE'ing them, because
221 // they may not be needed anymore.
223 BI = Branch; ++BI; // Get an iterator to the first new instruction
225 while (BI != SourceBlock->end())
226 if (!dceInstruction(BI) && !doConstantPropagation(BI))
229 DestBlock->removePredecessor(SourceBlock); // Remove entries in PHI nodes...
230 SourceBlock->getInstList().erase(Branch); // Destroy the uncond branch...
232 ++NumEliminated; // We just killed a branch!
235 /// InsertPHINodesIfNecessary - So at this point, we cloned the OrigInst
236 /// instruction into the NewBlock with the value of NewInst. If OrigInst was
237 /// used outside of its defining basic block, we need to insert a PHI nodes into
240 void TailDup::InsertPHINodesIfNecessary(Instruction *OrigInst, Value *NewInst,
241 BasicBlock *NewBlock) {
242 // Loop over all of the uses of OrigInst, rewriting them to be newly inserted
243 // PHI nodes, unless they are in the same basic block as OrigInst.
244 BasicBlock *OrigBlock = OrigInst->getParent();
245 std::vector<Instruction*> Users;
246 Users.reserve(OrigInst->use_size());
247 for (Value::use_iterator I = OrigInst->use_begin(), E = OrigInst->use_end();
249 Instruction *In = cast<Instruction>(*I);
250 if (In->getParent() != OrigBlock || // Don't modify uses in the orig block!
255 // The common case is that the instruction is only used within the block that
256 // defines it. If we have this case, quick exit.
258 if (Users.empty()) return;
260 // Otherwise, we have a more complex case, handle it now. This requires the
261 // construction of a mapping between a basic block and the value to use when
262 // in the scope of that basic block. This map will map to the original and
263 // new values when in the original or new block, but will map to inserted PHI
264 // nodes when in other blocks.
266 std::map<BasicBlock*, ValueHolder> ValueMap;
267 std::map<BasicBlock*, ValueHolder> OutValueMap; // The outgoing value map
268 OutValueMap[OrigBlock] = OrigInst;
269 OutValueMap[NewBlock ] = NewInst; // Seed the initial values...
271 DEBUG(std::cerr << " ** Inserting PHI nodes for " << OrigInst);
272 while (!Users.empty()) {
273 Instruction *User = Users.back(); Users.pop_back();
275 if (PHINode *PN = dyn_cast<PHINode>(User)) {
276 // PHI nodes must be handled specially here, because their operands are
277 // actually defined in predecessor basic blocks, NOT in the block that the
278 // PHI node lives in. Note that we have already added entries to PHI nods
279 // which are in blocks that are immediate successors of OrigBlock, so
280 // don't modify them again.
281 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
282 if (PN->getIncomingValue(i) == OrigInst &&
283 PN->getIncomingBlock(i) != OrigBlock) {
284 Value *V = GetValueOutBlock(PN->getIncomingBlock(i), OrigInst,
285 ValueMap, OutValueMap);
286 PN->setIncomingValue(i, V);
290 // Any other user of the instruction can just replace any uses with the
291 // new value defined in the block it resides in.
292 Value *V = GetValueInBlock(User->getParent(), OrigInst, ValueMap,
294 User->replaceUsesOfWith(OrigInst, V);
299 /// GetValueInBlock - This is a recursive method which inserts PHI nodes into
300 /// the function until there is a value available in basic block BB.
302 Value *TailDup::GetValueInBlock(BasicBlock *BB, Value *OrigVal,
303 std::map<BasicBlock*, ValueHolder> &ValueMap,
304 std::map<BasicBlock*,ValueHolder> &OutValueMap){
305 ValueHolder &BBVal = ValueMap[BB];
306 if (BBVal) return BBVal; // Value already computed for this block?
308 // If this block has no predecessors, then it must be unreachable, thus, it
309 // doesn't matter which value we use.
310 if (pred_begin(BB) == pred_end(BB))
311 return BBVal = Constant::getNullValue(OrigVal->getType());
313 // If there is no value already available in this basic block, we need to
314 // either reuse a value from an incoming, dominating, basic block, or we need
315 // to create a new PHI node to merge in different incoming values. Because we
316 // don't know if we're part of a loop at this point or not, we create a PHI
317 // node, even if we will ultimately eliminate it.
318 PHINode *PN = new PHINode(OrigVal->getType(), OrigVal->getName()+".pn",
320 BBVal = PN; // Insert this into the BBVal slot in case of cycles...
322 ValueHolder &BBOutVal = OutValueMap[BB];
323 if (BBOutVal == 0) BBOutVal = PN;
325 // Now that we have created the PHI node, loop over all of the predecessors of
326 // this block, computing an incoming value for the predecessor.
327 std::vector<BasicBlock*> Preds(pred_begin(BB), pred_end(BB));
328 for (unsigned i = 0, e = Preds.size(); i != e; ++i)
329 PN->addIncoming(GetValueOutBlock(Preds[i], OrigVal, ValueMap, OutValueMap),
332 // The PHI node is complete. In many cases, however the PHI node was
333 // ultimately unnecessary: we could have just reused a dominating incoming
334 // value. If this is the case, nuke the PHI node and replace the map entry
335 // with the dominating value.
337 assert(PN->getNumIncomingValues() > 0 && "No predecessors?");
339 // Check to see if all of the elements in the PHI node are either the PHI node
340 // itself or ONE particular value.
342 Value *ReplVal = PN->getIncomingValue(i);
343 for (; ReplVal == PN && i != PN->getNumIncomingValues(); ++i)
344 ReplVal = PN->getIncomingValue(i); // Skip values equal to the PN
346 for (; i != PN->getNumIncomingValues(); ++i)
347 if (PN->getIncomingValue(i) != PN && PN->getIncomingValue(i) != ReplVal) {
352 // Found a value to replace the PHI node with?
353 if (ReplVal && ReplVal != PN) {
354 PN->replaceAllUsesWith(ReplVal);
355 BB->getInstList().erase(PN); // Erase the PHI node...
363 Value *TailDup::GetValueOutBlock(BasicBlock *BB, Value *OrigVal,
364 std::map<BasicBlock*, ValueHolder> &ValueMap,
365 std::map<BasicBlock*, ValueHolder> &OutValueMap) {
366 ValueHolder &BBVal = OutValueMap[BB];
367 if (BBVal) return BBVal; // Value already computed for this block?
369 return GetValueInBlock(BB, OrigVal, ValueMap, OutValueMap);