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 TailDupThreshold("taildup-threshold",
43 cl::desc("Max block size to tail duplicate"),
44 cl::init(1), cl::Hidden);
47 class VISIBILITY_HIDDEN TailDup : public FunctionPass {
48 bool runOnFunction(Function &F);
50 static char ID; // Pass identification, replacement for typeid
51 TailDup() : FunctionPass((intptr_t)&ID) {}
54 inline bool shouldEliminateUnconditionalBranch(TerminatorInst *, unsigned);
55 inline void eliminateUnconditionalBranch(BranchInst *BI);
56 SmallPtrSet<BasicBlock*, 4> CycleDetector;
61 static RegisterPass<TailDup> X("tailduplicate", "Tail Duplication");
63 // Public interface to the Tail Duplication pass
64 FunctionPass *llvm::createTailDuplicationPass() { return new TailDup(); }
66 /// runOnFunction - Top level algorithm - Loop over each unconditional branch in
67 /// the function, eliminating it if it looks attractive enough. CycleDetector
68 /// prevents infinite loops by checking that we aren't redirecting a branch to
69 /// a place it already pointed to earlier; see PR 2323.
70 bool TailDup::runOnFunction(Function &F) {
72 CycleDetector.clear();
73 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
74 if (shouldEliminateUnconditionalBranch(I->getTerminator(),
76 eliminateUnconditionalBranch(cast<BranchInst>(I->getTerminator()));
80 CycleDetector.clear();
86 /// shouldEliminateUnconditionalBranch - Return true if this branch looks
87 /// attractive to eliminate. We eliminate the branch if the destination basic
88 /// block has <= 5 instructions in it, not counting PHI nodes. In practice,
89 /// since one of these is a terminator instruction, this means that we will add
90 /// up to 4 instructions to the new block.
92 /// We don't count PHI nodes in the count since they will be removed when the
93 /// contents of the block are copied over.
95 bool TailDup::shouldEliminateUnconditionalBranch(TerminatorInst *TI,
97 BranchInst *BI = dyn_cast<BranchInst>(TI);
98 if (!BI || !BI->isUnconditional()) return false; // Not an uncond branch!
100 BasicBlock *Dest = BI->getSuccessor(0);
101 if (Dest == BI->getParent()) return false; // Do not loop infinitely!
103 // Do not inline a block if we will just get another branch to the same block!
104 TerminatorInst *DTI = Dest->getTerminator();
105 if (BranchInst *DBI = dyn_cast<BranchInst>(DTI))
106 if (DBI->isUnconditional() && DBI->getSuccessor(0) == Dest)
107 return false; // Do not loop infinitely!
109 // FIXME: DemoteRegToStack cannot yet demote invoke instructions to the stack,
110 // because doing so would require breaking critical edges. This should be
112 if (!DTI->use_empty())
115 // Do not bother with blocks with only a single predecessor: simplify
116 // CFG will fold these two blocks together!
117 pred_iterator PI = pred_begin(Dest), PE = pred_end(Dest);
119 if (PI == PE) return false; // Exactly one predecessor!
121 BasicBlock::iterator I = Dest->getFirstNonPHI();
123 for (unsigned Size = 0; I != Dest->end(); ++I) {
124 if (Size == Threshold) return false; // The block is too large.
126 // Don't tail duplicate call instructions. They are very large compared to
127 // other instructions.
128 if (isa<CallInst>(I) || isa<InvokeInst>(I)) return false;
130 // Allso alloca and malloc.
131 if (isa<AllocationInst>(I)) return false;
133 // Some vector instructions can expand into a number of instructions.
134 if (isa<ShuffleVectorInst>(I) || isa<ExtractElementInst>(I) ||
135 isa<InsertElementInst>(I)) return false;
137 // Only count instructions that are not debugger intrinsics.
138 if (!isa<DbgInfoIntrinsic>(I)) ++Size;
141 // Do not tail duplicate a block that has thousands of successors into a block
142 // with a single successor if the block has many other predecessors. This can
143 // cause an N^2 explosion in CFG edges (and PHI node entries), as seen in
144 // cases that have a large number of indirect gotos.
145 unsigned NumSuccs = DTI->getNumSuccessors();
147 unsigned TooMany = 128;
148 if (NumSuccs >= TooMany) return false;
149 TooMany = TooMany/NumSuccs;
150 for (; PI != PE; ++PI)
151 if (TooMany-- == 0) return false;
154 // If this unconditional branch is a fall-through, be careful about
155 // tail duplicating it. In particular, we don't want to taildup it if the
156 // original block will still be there after taildup is completed: doing so
157 // would eliminate the fall-through, requiring unconditional branches.
158 Function::iterator DestI = Dest;
159 if (&*--DestI == BI->getParent()) {
160 // The uncond branch is a fall-through. Tail duplication of the block is
161 // will eliminate the fall-through-ness and end up cloning the terminator
162 // at the end of the Dest block. Since the original Dest block will
163 // continue to exist, this means that one or the other will not be able to
164 // fall through. One typical example that this helps with is code like:
169 // Cloning the 'if b' block into the end of the first foo block is messy.
171 // The messy case is when the fall-through block falls through to other
172 // blocks. This is what we would be preventing if we cloned the block.
174 if (++DestI != Dest->getParent()->end()) {
175 BasicBlock *DestSucc = DestI;
176 // If any of Dest's successors are fall-throughs, don't do this xform.
177 for (succ_iterator SI = succ_begin(Dest), SE = succ_end(Dest);
184 // Finally, check that we haven't redirected to this target block earlier;
185 // there are cases where we loop forever if we don't check this (PR 2323).
186 if (!CycleDetector.insert(Dest))
192 /// FindObviousSharedDomOf - We know there is a branch from SrcBlock to
193 /// DestBlock, and that SrcBlock is not the only predecessor of DstBlock. If we
194 /// can find a predecessor of SrcBlock that is a dominator of both SrcBlock and
195 /// DstBlock, return it.
196 static BasicBlock *FindObviousSharedDomOf(BasicBlock *SrcBlock,
197 BasicBlock *DstBlock) {
198 // SrcBlock must have a single predecessor.
199 pred_iterator PI = pred_begin(SrcBlock), PE = pred_end(SrcBlock);
200 if (PI == PE || ++PI != PE) return 0;
202 BasicBlock *SrcPred = *pred_begin(SrcBlock);
204 // Look at the predecessors of DstBlock. One of them will be SrcBlock. If
205 // there is only one other pred, get it, otherwise we can't handle it.
206 PI = pred_begin(DstBlock); PE = pred_end(DstBlock);
207 BasicBlock *DstOtherPred = 0;
208 if (*PI == SrcBlock) {
209 if (++PI == PE) return 0;
211 if (++PI != PE) return 0;
214 if (++PI == PE || *PI != SrcBlock || ++PI != PE) return 0;
217 // We can handle two situations here: "if then" and "if then else" blocks. An
218 // 'if then' situation is just where DstOtherPred == SrcPred.
219 if (DstOtherPred == SrcPred)
222 // Check to see if we have an "if then else" situation, which means that
223 // DstOtherPred will have a single predecessor and it will be SrcPred.
224 PI = pred_begin(DstOtherPred); PE = pred_end(DstOtherPred);
225 if (PI != PE && *PI == SrcPred) {
226 if (++PI != PE) return 0; // Not a single pred.
227 return SrcPred; // Otherwise, it's an "if then" situation. Return the if.
230 // Otherwise, this is something we can't handle.
235 /// eliminateUnconditionalBranch - Clone the instructions from the destination
236 /// block into the source block, eliminating the specified unconditional branch.
237 /// If the destination block defines values used by successors of the dest
238 /// block, we may need to insert PHI nodes.
240 void TailDup::eliminateUnconditionalBranch(BranchInst *Branch) {
241 BasicBlock *SourceBlock = Branch->getParent();
242 BasicBlock *DestBlock = Branch->getSuccessor(0);
243 assert(SourceBlock != DestBlock && "Our predicate is broken!");
245 DOUT << "TailDuplication[" << SourceBlock->getParent()->getName()
246 << "]: Eliminating branch: " << *Branch;
248 // See if we can avoid duplicating code by moving it up to a dominator of both
250 if (BasicBlock *DomBlock = FindObviousSharedDomOf(SourceBlock, DestBlock)) {
251 DOUT << "Found shared dominator: " << DomBlock->getName() << "\n";
253 // If there are non-phi instructions in DestBlock that have no operands
254 // defined in DestBlock, and if the instruction has no side effects, we can
255 // move the instruction to DomBlock instead of duplicating it.
256 BasicBlock::iterator BBI = DestBlock->getFirstNonPHI();
257 while (!isa<TerminatorInst>(BBI)) {
258 Instruction *I = BBI++;
260 bool CanHoist = !I->isTrapping() && !I->mayWriteToMemory();
262 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op)
263 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(op)))
264 if (OpI->getParent() == DestBlock ||
265 (isa<InvokeInst>(OpI) && OpI->getParent() == DomBlock)) {
270 // Remove from DestBlock, move right before the term in DomBlock.
271 DestBlock->getInstList().remove(I);
272 DomBlock->getInstList().insert(DomBlock->getTerminator(), I);
273 DOUT << "Hoisted: " << *I;
279 // Tail duplication can not update SSA properties correctly if the values
280 // defined in the duplicated tail are used outside of the tail itself. For
281 // this reason, we spill all values that are used outside of the tail to the
283 for (BasicBlock::iterator I = DestBlock->begin(); I != DestBlock->end(); ++I)
284 if (I->isUsedOutsideOfBlock(DestBlock)) {
285 // We found a use outside of the tail. Create a new stack slot to
286 // break this inter-block usage pattern.
287 DemoteRegToStack(*I);
290 // We are going to have to map operands from the original block B to the new
291 // copy of the block B'. If there are PHI nodes in the DestBlock, these PHI
292 // nodes also define part of this mapping. Loop over these PHI nodes, adding
293 // them to our mapping.
295 std::map<Value*, Value*> ValueMapping;
297 BasicBlock::iterator BI = DestBlock->begin();
298 bool HadPHINodes = isa<PHINode>(BI);
299 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
300 ValueMapping[PN] = PN->getIncomingValueForBlock(SourceBlock);
302 // Clone the non-phi instructions of the dest block into the source block,
303 // keeping track of the mapping...
305 for (; BI != DestBlock->end(); ++BI) {
306 Instruction *New = BI->clone();
307 New->setName(BI->getName());
308 SourceBlock->getInstList().push_back(New);
309 ValueMapping[BI] = New;
312 // Now that we have built the mapping information and cloned all of the
313 // instructions (giving us a new terminator, among other things), walk the new
314 // instructions, rewriting references of old instructions to use new
317 BI = Branch; ++BI; // Get an iterator to the first new instruction
318 for (; BI != SourceBlock->end(); ++BI)
319 for (unsigned i = 0, e = BI->getNumOperands(); i != e; ++i)
320 if (Value *Remapped = ValueMapping[BI->getOperand(i)])
321 BI->setOperand(i, Remapped);
323 // Next we check to see if any of the successors of DestBlock had PHI nodes.
324 // If so, we need to add entries to the PHI nodes for SourceBlock now.
325 for (succ_iterator SI = succ_begin(DestBlock), SE = succ_end(DestBlock);
327 BasicBlock *Succ = *SI;
328 for (BasicBlock::iterator PNI = Succ->begin(); isa<PHINode>(PNI); ++PNI) {
329 PHINode *PN = cast<PHINode>(PNI);
330 // Ok, we have a PHI node. Figure out what the incoming value was for the
332 Value *IV = PN->getIncomingValueForBlock(DestBlock);
334 // Remap the value if necessary...
335 if (Value *MappedIV = ValueMapping[IV])
337 PN->addIncoming(IV, SourceBlock);
341 // Next, remove the old branch instruction, and any PHI node entries that we
343 BI = Branch; ++BI; // Get an iterator to the first new instruction
344 DestBlock->removePredecessor(SourceBlock); // Remove entries in PHI nodes...
345 SourceBlock->getInstList().erase(Branch); // Destroy the uncond branch...
347 // Final step: now that we have finished everything up, walk the cloned
348 // instructions one last time, constant propagating and DCE'ing them, because
349 // they may not be needed anymore.
352 while (BI != SourceBlock->end())
353 if (!dceInstruction(BI) && !doConstantPropagation(BI))
356 ++NumEliminated; // We just killed a branch!