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"
38 STATISTIC(NumEliminated, "Number of unconditional branches eliminated");
40 static cl::opt<unsigned>
41 Threshold("taildup-threshold", cl::desc("Max block size to tail duplicate"),
42 cl::init(6), cl::Hidden);
45 class VISIBILITY_HIDDEN TailDup : public FunctionPass {
46 bool runOnFunction(Function &F);
48 static char ID; // Pass identification, replacement for typeid
49 TailDup() : FunctionPass((intptr_t)&ID) {}
52 inline bool shouldEliminateUnconditionalBranch(TerminatorInst *TI);
53 inline void eliminateUnconditionalBranch(BranchInst *BI);
58 static RegisterPass<TailDup> X("tailduplicate", "Tail Duplication");
60 // Public interface to the Tail Duplication pass
61 FunctionPass *llvm::createTailDuplicationPass() { return new TailDup(); }
63 /// runOnFunction - Top level algorithm - Loop over each unconditional branch in
64 /// the function, eliminating it if it looks attractive enough.
66 bool TailDup::runOnFunction(Function &F) {
68 for (Function::iterator I = F.begin(), E = F.end(); I != E; )
69 if (shouldEliminateUnconditionalBranch(I->getTerminator())) {
70 eliminateUnconditionalBranch(cast<BranchInst>(I->getTerminator()));
78 /// shouldEliminateUnconditionalBranch - Return true if this branch looks
79 /// attractive to eliminate. We eliminate the branch if the destination basic
80 /// block has <= 5 instructions in it, not counting PHI nodes. In practice,
81 /// since one of these is a terminator instruction, this means that we will add
82 /// up to 4 instructions to the new block.
84 /// We don't count PHI nodes in the count since they will be removed when the
85 /// contents of the block are copied over.
87 bool TailDup::shouldEliminateUnconditionalBranch(TerminatorInst *TI) {
88 BranchInst *BI = dyn_cast<BranchInst>(TI);
89 if (!BI || !BI->isUnconditional()) return false; // Not an uncond branch!
91 BasicBlock *Dest = BI->getSuccessor(0);
92 if (Dest == BI->getParent()) return false; // Do not loop infinitely!
94 // Do not inline a block if we will just get another branch to the same block!
95 TerminatorInst *DTI = Dest->getTerminator();
96 if (BranchInst *DBI = dyn_cast<BranchInst>(DTI))
97 if (DBI->isUnconditional() && DBI->getSuccessor(0) == Dest)
98 return false; // Do not loop infinitely!
100 // FIXME: DemoteRegToStack cannot yet demote invoke instructions to the stack,
101 // because doing so would require breaking critical edges. This should be
103 if (!DTI->use_empty())
106 // Do not bother working on dead blocks...
107 pred_iterator PI = pred_begin(Dest), PE = pred_end(Dest);
108 if (PI == PE && Dest != Dest->getParent()->begin())
109 return false; // It's just a dead block, ignore it...
111 // Also, do not bother with blocks with only a single predecessor: simplify
112 // CFG will fold these two blocks together!
114 if (PI == PE) return false; // Exactly one predecessor!
116 BasicBlock::iterator I = Dest->begin();
117 while (isa<PHINode>(*I)) ++I;
119 for (unsigned Size = 0; I != Dest->end(); ++I) {
120 if (Size == Threshold) return false; // The block is too large.
122 // Don't tail duplicate call instructions. They are very large compared to
123 // other instructions.
124 if (isa<CallInst>(I) || isa<InvokeInst>(I)) return false;
126 // Only count instructions that are not debugger intrinsics.
127 if (!isa<DbgInfoIntrinsic>(I)) ++Size;
130 // Do not tail duplicate a block that has thousands of successors into a block
131 // with a single successor if the block has many other predecessors. This can
132 // cause an N^2 explosion in CFG edges (and PHI node entries), as seen in
133 // cases that have a large number of indirect gotos.
134 unsigned NumSuccs = DTI->getNumSuccessors();
136 unsigned TooMany = 128;
137 if (NumSuccs >= TooMany) return false;
138 TooMany = TooMany/NumSuccs;
139 for (; PI != PE; ++PI)
140 if (TooMany-- == 0) return false;
143 // Finally, if this unconditional branch is a fall-through, be careful about
144 // tail duplicating it. In particular, we don't want to taildup it if the
145 // original block will still be there after taildup is completed: doing so
146 // would eliminate the fall-through, requiring unconditional branches.
147 Function::iterator DestI = Dest;
148 if (&*--DestI == BI->getParent()) {
149 // The uncond branch is a fall-through. Tail duplication of the block is
150 // will eliminate the fall-through-ness and end up cloning the terminator
151 // at the end of the Dest block. Since the original Dest block will
152 // continue to exist, this means that one or the other will not be able to
153 // fall through. One typical example that this helps with is code like:
158 // Cloning the 'if b' block into the end of the first foo block is messy.
160 // The messy case is when the fall-through block falls through to other
161 // blocks. This is what we would be preventing if we cloned the block.
163 if (++DestI != Dest->getParent()->end()) {
164 BasicBlock *DestSucc = DestI;
165 // If any of Dest's successors are fall-throughs, don't do this xform.
166 for (succ_iterator SI = succ_begin(Dest), SE = succ_end(Dest);
176 /// FindObviousSharedDomOf - We know there is a branch from SrcBlock to
177 /// DestBlock, and that SrcBlock is not the only predecessor of DstBlock. If we
178 /// can find a predecessor of SrcBlock that is a dominator of both SrcBlock and
179 /// DstBlock, return it.
180 static BasicBlock *FindObviousSharedDomOf(BasicBlock *SrcBlock,
181 BasicBlock *DstBlock) {
182 // SrcBlock must have a single predecessor.
183 pred_iterator PI = pred_begin(SrcBlock), PE = pred_end(SrcBlock);
184 if (PI == PE || ++PI != PE) return 0;
186 BasicBlock *SrcPred = *pred_begin(SrcBlock);
188 // Look at the predecessors of DstBlock. One of them will be SrcBlock. If
189 // there is only one other pred, get it, otherwise we can't handle it.
190 PI = pred_begin(DstBlock); PE = pred_end(DstBlock);
191 BasicBlock *DstOtherPred = 0;
192 if (*PI == SrcBlock) {
193 if (++PI == PE) return 0;
195 if (++PI != PE) return 0;
198 if (++PI == PE || *PI != SrcBlock || ++PI != PE) return 0;
201 // We can handle two situations here: "if then" and "if then else" blocks. An
202 // 'if then' situation is just where DstOtherPred == SrcPred.
203 if (DstOtherPred == SrcPred)
206 // Check to see if we have an "if then else" situation, which means that
207 // DstOtherPred will have a single predecessor and it will be SrcPred.
208 PI = pred_begin(DstOtherPred); PE = pred_end(DstOtherPred);
209 if (PI != PE && *PI == SrcPred) {
210 if (++PI != PE) return 0; // Not a single pred.
211 return SrcPred; // Otherwise, it's an "if then" situation. Return the if.
214 // Otherwise, this is something we can't handle.
219 /// eliminateUnconditionalBranch - Clone the instructions from the destination
220 /// block into the source block, eliminating the specified unconditional branch.
221 /// If the destination block defines values used by successors of the dest
222 /// block, we may need to insert PHI nodes.
224 void TailDup::eliminateUnconditionalBranch(BranchInst *Branch) {
225 BasicBlock *SourceBlock = Branch->getParent();
226 BasicBlock *DestBlock = Branch->getSuccessor(0);
227 assert(SourceBlock != DestBlock && "Our predicate is broken!");
229 DOUT << "TailDuplication[" << SourceBlock->getParent()->getName()
230 << "]: Eliminating branch: " << *Branch;
232 // See if we can avoid duplicating code by moving it up to a dominator of both
234 if (BasicBlock *DomBlock = FindObviousSharedDomOf(SourceBlock, DestBlock)) {
235 DOUT << "Found shared dominator: " << DomBlock->getName() << "\n";
237 // If there are non-phi instructions in DestBlock that have no operands
238 // defined in DestBlock, and if the instruction has no side effects, we can
239 // move the instruction to DomBlock instead of duplicating it.
240 BasicBlock::iterator BBI = DestBlock->begin();
241 while (isa<PHINode>(BBI)) ++BBI;
242 while (!isa<TerminatorInst>(BBI)) {
243 Instruction *I = BBI++;
245 bool CanHoist = !I->isTrapping() && !I->mayWriteToMemory();
247 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op)
248 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(op)))
249 if (OpI->getParent() == DestBlock ||
250 (isa<InvokeInst>(OpI) && OpI->getParent() == DomBlock)) {
255 // Remove from DestBlock, move right before the term in DomBlock.
256 DestBlock->getInstList().remove(I);
257 DomBlock->getInstList().insert(DomBlock->getTerminator(), I);
258 DOUT << "Hoisted: " << *I;
264 // Tail duplication can not update SSA properties correctly if the values
265 // defined in the duplicated tail are used outside of the tail itself. For
266 // this reason, we spill all values that are used outside of the tail to the
268 for (BasicBlock::iterator I = DestBlock->begin(); I != DestBlock->end(); ++I)
269 if (I->isUsedOutsideOfBlock(DestBlock)) {
270 // We found a use outside of the tail. Create a new stack slot to
271 // break this inter-block usage pattern.
272 DemoteRegToStack(*I);
275 // We are going to have to map operands from the original block B to the new
276 // copy of the block B'. If there are PHI nodes in the DestBlock, these PHI
277 // nodes also define part of this mapping. Loop over these PHI nodes, adding
278 // them to our mapping.
280 std::map<Value*, Value*> ValueMapping;
282 BasicBlock::iterator BI = DestBlock->begin();
283 bool HadPHINodes = isa<PHINode>(BI);
284 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
285 ValueMapping[PN] = PN->getIncomingValueForBlock(SourceBlock);
287 // Clone the non-phi instructions of the dest block into the source block,
288 // keeping track of the mapping...
290 for (; BI != DestBlock->end(); ++BI) {
291 Instruction *New = BI->clone();
292 New->setName(BI->getName());
293 SourceBlock->getInstList().push_back(New);
294 ValueMapping[BI] = New;
297 // Now that we have built the mapping information and cloned all of the
298 // instructions (giving us a new terminator, among other things), walk the new
299 // instructions, rewriting references of old instructions to use new
302 BI = Branch; ++BI; // Get an iterator to the first new instruction
303 for (; BI != SourceBlock->end(); ++BI)
304 for (unsigned i = 0, e = BI->getNumOperands(); i != e; ++i)
305 if (Value *Remapped = ValueMapping[BI->getOperand(i)])
306 BI->setOperand(i, Remapped);
308 // Next we check to see if any of the successors of DestBlock had PHI nodes.
309 // If so, we need to add entries to the PHI nodes for SourceBlock now.
310 for (succ_iterator SI = succ_begin(DestBlock), SE = succ_end(DestBlock);
312 BasicBlock *Succ = *SI;
313 for (BasicBlock::iterator PNI = Succ->begin(); isa<PHINode>(PNI); ++PNI) {
314 PHINode *PN = cast<PHINode>(PNI);
315 // Ok, we have a PHI node. Figure out what the incoming value was for the
317 Value *IV = PN->getIncomingValueForBlock(DestBlock);
319 // Remap the value if necessary...
320 if (Value *MappedIV = ValueMapping[IV])
322 PN->addIncoming(IV, SourceBlock);
326 // Next, remove the old branch instruction, and any PHI node entries that we
328 BI = Branch; ++BI; // Get an iterator to the first new instruction
329 DestBlock->removePredecessor(SourceBlock); // Remove entries in PHI nodes...
330 SourceBlock->getInstList().erase(Branch); // Destroy the uncond branch...
332 // Final step: now that we have finished everything up, walk the cloned
333 // instructions one last time, constant propagating and DCE'ing them, because
334 // they may not be needed anymore.
337 while (BI != SourceBlock->end())
338 if (!dceInstruction(BI) && !doConstantPropagation(BI))
341 ++NumEliminated; // We just killed a branch!