1 //===- TailDuplication.cpp - Simplify CFG through tail duplication --------===//
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 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 #define DEBUG_TYPE "tailduplicate"
22 #include "llvm/Transforms/Scalar.h"
23 #include "llvm/Constant.h"
24 #include "llvm/Function.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/Pass.h"
28 #include "llvm/Type.h"
29 #include "llvm/Support/CFG.h"
30 #include "llvm/Transforms/Utils/Local.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Compiler.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/ADT/SmallPtrSet.h"
39 STATISTIC(NumEliminated, "Number of unconditional branches eliminated");
41 static cl::opt<unsigned>
42 Threshold("taildup-threshold", cl::desc("Max block size to tail duplicate"),
43 cl::init(6), cl::Hidden);
46 class VISIBILITY_HIDDEN TailDup : public FunctionPass {
47 bool runOnFunction(Function &F);
49 static char ID; // Pass identification, replacement for typeid
50 TailDup() : FunctionPass((intptr_t)&ID) {}
53 inline bool shouldEliminateUnconditionalBranch(TerminatorInst *TI);
54 inline void eliminateUnconditionalBranch(BranchInst *BI);
55 SmallPtrSet<BasicBlock*, 4> CycleDetector;
60 static RegisterPass<TailDup> X("tailduplicate", "Tail Duplication");
62 // Public interface to the Tail Duplication pass
63 FunctionPass *llvm::createTailDuplicationPass() { return new TailDup(); }
65 /// runOnFunction - Top level algorithm - Loop over each unconditional branch in
66 /// the function, eliminating it if it looks attractive enough. CycleDetector
67 /// prevents infinite loops by checking that we aren't redirecting a branch to
68 /// a place it already pointed to earlier; see PR 2323.
69 bool TailDup::runOnFunction(Function &F) {
71 CycleDetector.clear();
72 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
73 if (shouldEliminateUnconditionalBranch(I->getTerminator())) {
74 eliminateUnconditionalBranch(cast<BranchInst>(I->getTerminator()));
78 CycleDetector.clear();
84 /// shouldEliminateUnconditionalBranch - Return true if this branch looks
85 /// attractive to eliminate. We eliminate the branch if the destination basic
86 /// block has <= 5 instructions in it, not counting PHI nodes. In practice,
87 /// since one of these is a terminator instruction, this means that we will add
88 /// up to 4 instructions to the new block.
90 /// We don't count PHI nodes in the count since they will be removed when the
91 /// contents of the block are copied over.
93 bool TailDup::shouldEliminateUnconditionalBranch(TerminatorInst *TI) {
94 BranchInst *BI = dyn_cast<BranchInst>(TI);
95 if (!BI || !BI->isUnconditional()) return false; // Not an uncond branch!
97 BasicBlock *Dest = BI->getSuccessor(0);
98 if (Dest == BI->getParent()) return false; // Do not loop infinitely!
100 // Do not inline a block if we will just get another branch to the same block!
101 TerminatorInst *DTI = Dest->getTerminator();
102 if (BranchInst *DBI = dyn_cast<BranchInst>(DTI))
103 if (DBI->isUnconditional() && DBI->getSuccessor(0) == Dest)
104 return false; // Do not loop infinitely!
106 // FIXME: DemoteRegToStack cannot yet demote invoke instructions to the stack,
107 // because doing so would require breaking critical edges. This should be
109 if (!DTI->use_empty())
112 // Do not bother with blocks with only a single predecessor: simplify
113 // CFG will fold these two blocks together!
114 pred_iterator PI = pred_begin(Dest), PE = pred_end(Dest);
116 if (PI == PE) return false; // Exactly one predecessor!
118 BasicBlock::iterator I = Dest->begin();
119 while (isa<PHINode>(*I)) ++I;
121 for (unsigned Size = 0; I != Dest->end(); ++I) {
122 if (Size == Threshold) return false; // The block is too large.
124 // Don't tail duplicate call instructions. They are very large compared to
125 // other instructions.
126 if (isa<CallInst>(I) || isa<InvokeInst>(I)) return false;
128 // Only count instructions that are not debugger intrinsics.
129 if (!isa<DbgInfoIntrinsic>(I)) ++Size;
132 // Do not tail duplicate a block that has thousands of successors into a block
133 // with a single successor if the block has many other predecessors. This can
134 // cause an N^2 explosion in CFG edges (and PHI node entries), as seen in
135 // cases that have a large number of indirect gotos.
136 unsigned NumSuccs = DTI->getNumSuccessors();
138 unsigned TooMany = 128;
139 if (NumSuccs >= TooMany) return false;
140 TooMany = TooMany/NumSuccs;
141 for (; PI != PE; ++PI)
142 if (TooMany-- == 0) return false;
145 // If this unconditional branch is a fall-through, be careful about
146 // tail duplicating it. In particular, we don't want to taildup it if the
147 // original block will still be there after taildup is completed: doing so
148 // would eliminate the fall-through, requiring unconditional branches.
149 Function::iterator DestI = Dest;
150 if (&*--DestI == BI->getParent()) {
151 // The uncond branch is a fall-through. Tail duplication of the block is
152 // will eliminate the fall-through-ness and end up cloning the terminator
153 // at the end of the Dest block. Since the original Dest block will
154 // continue to exist, this means that one or the other will not be able to
155 // fall through. One typical example that this helps with is code like:
160 // Cloning the 'if b' block into the end of the first foo block is messy.
162 // The messy case is when the fall-through block falls through to other
163 // blocks. This is what we would be preventing if we cloned the block.
165 if (++DestI != Dest->getParent()->end()) {
166 BasicBlock *DestSucc = DestI;
167 // If any of Dest's successors are fall-throughs, don't do this xform.
168 for (succ_iterator SI = succ_begin(Dest), SE = succ_end(Dest);
175 // Finally, check that we haven't redirected to this target block earlier;
176 // there are cases where we loop forever if we don't check this (PR 2323).
177 if (!CycleDetector.insert(Dest))
183 /// FindObviousSharedDomOf - We know there is a branch from SrcBlock to
184 /// DestBlock, and that SrcBlock is not the only predecessor of DstBlock. If we
185 /// can find a predecessor of SrcBlock that is a dominator of both SrcBlock and
186 /// DstBlock, return it.
187 static BasicBlock *FindObviousSharedDomOf(BasicBlock *SrcBlock,
188 BasicBlock *DstBlock) {
189 // SrcBlock must have a single predecessor.
190 pred_iterator PI = pred_begin(SrcBlock), PE = pred_end(SrcBlock);
191 if (PI == PE || ++PI != PE) return 0;
193 BasicBlock *SrcPred = *pred_begin(SrcBlock);
195 // Look at the predecessors of DstBlock. One of them will be SrcBlock. If
196 // there is only one other pred, get it, otherwise we can't handle it.
197 PI = pred_begin(DstBlock); PE = pred_end(DstBlock);
198 BasicBlock *DstOtherPred = 0;
199 if (*PI == SrcBlock) {
200 if (++PI == PE) return 0;
202 if (++PI != PE) return 0;
205 if (++PI == PE || *PI != SrcBlock || ++PI != PE) return 0;
208 // We can handle two situations here: "if then" and "if then else" blocks. An
209 // 'if then' situation is just where DstOtherPred == SrcPred.
210 if (DstOtherPred == SrcPred)
213 // Check to see if we have an "if then else" situation, which means that
214 // DstOtherPred will have a single predecessor and it will be SrcPred.
215 PI = pred_begin(DstOtherPred); PE = pred_end(DstOtherPred);
216 if (PI != PE && *PI == SrcPred) {
217 if (++PI != PE) return 0; // Not a single pred.
218 return SrcPred; // Otherwise, it's an "if then" situation. Return the if.
221 // Otherwise, this is something we can't handle.
226 /// eliminateUnconditionalBranch - Clone the instructions from the destination
227 /// block into the source block, eliminating the specified unconditional branch.
228 /// If the destination block defines values used by successors of the dest
229 /// block, we may need to insert PHI nodes.
231 void TailDup::eliminateUnconditionalBranch(BranchInst *Branch) {
232 BasicBlock *SourceBlock = Branch->getParent();
233 BasicBlock *DestBlock = Branch->getSuccessor(0);
234 assert(SourceBlock != DestBlock && "Our predicate is broken!");
236 DOUT << "TailDuplication[" << SourceBlock->getParent()->getName()
237 << "]: Eliminating branch: " << *Branch;
239 // See if we can avoid duplicating code by moving it up to a dominator of both
241 if (BasicBlock *DomBlock = FindObviousSharedDomOf(SourceBlock, DestBlock)) {
242 DOUT << "Found shared dominator: " << DomBlock->getName() << "\n";
244 // If there are non-phi instructions in DestBlock that have no operands
245 // defined in DestBlock, and if the instruction has no side effects, we can
246 // move the instruction to DomBlock instead of duplicating it.
247 BasicBlock::iterator BBI = DestBlock->begin();
248 while (isa<PHINode>(BBI)) ++BBI;
249 while (!isa<TerminatorInst>(BBI)) {
250 Instruction *I = BBI++;
252 bool CanHoist = !I->isTrapping() && !I->mayWriteToMemory();
254 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op)
255 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(op)))
256 if (OpI->getParent() == DestBlock ||
257 (isa<InvokeInst>(OpI) && OpI->getParent() == DomBlock)) {
262 // Remove from DestBlock, move right before the term in DomBlock.
263 DestBlock->getInstList().remove(I);
264 DomBlock->getInstList().insert(DomBlock->getTerminator(), I);
265 DOUT << "Hoisted: " << *I;
271 // Tail duplication can not update SSA properties correctly if the values
272 // defined in the duplicated tail are used outside of the tail itself. For
273 // this reason, we spill all values that are used outside of the tail to the
275 for (BasicBlock::iterator I = DestBlock->begin(); I != DestBlock->end(); ++I)
276 if (I->isUsedOutsideOfBlock(DestBlock)) {
277 // We found a use outside of the tail. Create a new stack slot to
278 // break this inter-block usage pattern.
279 DemoteRegToStack(*I);
282 // We are going to have to map operands from the original block B to the new
283 // copy of the block B'. If there are PHI nodes in the DestBlock, these PHI
284 // nodes also define part of this mapping. Loop over these PHI nodes, adding
285 // them to our mapping.
287 std::map<Value*, Value*> ValueMapping;
289 BasicBlock::iterator BI = DestBlock->begin();
290 bool HadPHINodes = isa<PHINode>(BI);
291 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
292 ValueMapping[PN] = PN->getIncomingValueForBlock(SourceBlock);
294 // Clone the non-phi instructions of the dest block into the source block,
295 // keeping track of the mapping...
297 for (; BI != DestBlock->end(); ++BI) {
298 Instruction *New = BI->clone();
299 New->setName(BI->getName());
300 SourceBlock->getInstList().push_back(New);
301 ValueMapping[BI] = New;
304 // Now that we have built the mapping information and cloned all of the
305 // instructions (giving us a new terminator, among other things), walk the new
306 // instructions, rewriting references of old instructions to use new
309 BI = Branch; ++BI; // Get an iterator to the first new instruction
310 for (; BI != SourceBlock->end(); ++BI)
311 for (unsigned i = 0, e = BI->getNumOperands(); i != e; ++i)
312 if (Value *Remapped = ValueMapping[BI->getOperand(i)])
313 BI->setOperand(i, Remapped);
315 // Next we check to see if any of the successors of DestBlock had PHI nodes.
316 // If so, we need to add entries to the PHI nodes for SourceBlock now.
317 for (succ_iterator SI = succ_begin(DestBlock), SE = succ_end(DestBlock);
319 BasicBlock *Succ = *SI;
320 for (BasicBlock::iterator PNI = Succ->begin(); isa<PHINode>(PNI); ++PNI) {
321 PHINode *PN = cast<PHINode>(PNI);
322 // Ok, we have a PHI node. Figure out what the incoming value was for the
324 Value *IV = PN->getIncomingValueForBlock(DestBlock);
326 // Remap the value if necessary...
327 if (Value *MappedIV = ValueMapping[IV])
329 PN->addIncoming(IV, SourceBlock);
333 // Next, remove the old branch instruction, and any PHI node entries that we
335 BI = Branch; ++BI; // Get an iterator to the first new instruction
336 DestBlock->removePredecessor(SourceBlock); // Remove entries in PHI nodes...
337 SourceBlock->getInstList().erase(Branch); // Destroy the uncond branch...
339 // Final step: now that we have finished everything up, walk the cloned
340 // instructions one last time, constant propagating and DCE'ing them, because
341 // they may not be needed anymore.
344 while (BI != SourceBlock->end())
345 if (!dceInstruction(BI) && !doConstantPropagation(BI))
348 ++NumEliminated; // We just killed a branch!