1 //===- JumpThreading.cpp - Thread control through conditional blocks ------===//
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 file implements the Jump Threading pass.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "jump-threading"
15 #include "llvm/Transforms/Scalar.h"
16 #include "llvm/IntrinsicInst.h"
17 #include "llvm/Pass.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
21 #include "llvm/Transforms/Utils/Local.h"
22 #include "llvm/Support/CommandLine.h"
23 #include "llvm/Support/Compiler.h"
24 #include "llvm/Support/Debug.h"
27 STATISTIC(NumThreads, "Number of jumps threaded");
28 STATISTIC(NumFolds, "Number of terminators folded");
30 static cl::opt<unsigned>
31 Threshold("jump-threading-threshold",
32 cl::desc("Max block size to duplicate for jump threading"),
33 cl::init(6), cl::Hidden);
36 /// This pass performs 'jump threading', which looks at blocks that have
37 /// multiple predecessors and multiple successors. If one or more of the
38 /// predecessors of the block can be proven to always jump to one of the
39 /// successors, we forward the edge from the predecessor to the successor by
40 /// duplicating the contents of this block.
42 /// An example of when this can occur is code like this:
49 /// In this case, the unconditional branch at the end of the first if can be
50 /// revectored to the false side of the second if.
52 class VISIBILITY_HIDDEN JumpThreading : public FunctionPass {
54 static char ID; // Pass identification
55 JumpThreading() : FunctionPass((intptr_t)&ID) {}
57 bool runOnFunction(Function &F);
58 bool ThreadBlock(BasicBlock *BB);
59 void ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, BasicBlock *SuccBB);
60 BasicBlock *FactorCommonPHIPreds(PHINode *PN, Constant *CstVal);
62 bool ProcessJumpOnPHI(PHINode *PN);
63 bool ProcessBranchOnLogical(Value *V, BasicBlock *BB, bool isAnd);
64 bool ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB);
66 char JumpThreading::ID = 0;
67 RegisterPass<JumpThreading> X("jump-threading", "Jump Threading");
70 // Public interface to the Jump Threading pass
71 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
73 /// runOnFunction - Top level algorithm.
75 bool JumpThreading::runOnFunction(Function &F) {
76 DOUT << "Jump threading on function '" << F.getNameStart() << "'\n";
78 bool AnotherIteration = true, EverChanged = false;
79 while (AnotherIteration) {
80 AnotherIteration = false;
82 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
83 while (ThreadBlock(I))
85 AnotherIteration = Changed;
86 EverChanged |= Changed;
91 /// FactorCommonPHIPreds - If there are multiple preds with the same incoming
92 /// value for the PHI, factor them together so we get one block to thread for
94 /// This is important for things like "phi i1 [true, true, false, true, x]"
95 /// where we only need to clone the block for the true blocks once.
97 BasicBlock *JumpThreading::FactorCommonPHIPreds(PHINode *PN, Constant *CstVal) {
98 SmallVector<BasicBlock*, 16> CommonPreds;
99 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
100 if (PN->getIncomingValue(i) == CstVal)
101 CommonPreds.push_back(PN->getIncomingBlock(i));
103 if (CommonPreds.size() == 1)
104 return CommonPreds[0];
106 DOUT << " Factoring out " << CommonPreds.size()
107 << " common predecessors.\n";
108 return SplitBlockPredecessors(PN->getParent(),
109 &CommonPreds[0], CommonPreds.size(),
114 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
115 /// thread across it.
116 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
117 BasicBlock::const_iterator I = BB->begin();
118 /// Ignore PHI nodes, these will be flattened when duplication happens.
119 while (isa<PHINode>(*I)) ++I;
121 // Sum up the cost of each instruction until we get to the terminator. Don't
122 // include the terminator because the copy won't include it.
124 for (; !isa<TerminatorInst>(I); ++I) {
125 // Debugger intrinsics don't incur code size.
126 if (isa<DbgInfoIntrinsic>(I)) continue;
128 // If this is a pointer->pointer bitcast, it is free.
129 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType()))
132 // All other instructions count for at least one unit.
135 // Calls are more expensive. If they are non-intrinsic calls, we model them
136 // as having cost of 4. If they are a non-vector intrinsic, we model them
137 // as having cost of 2 total, and if they are a vector intrinsic, we model
138 // them as having cost 1.
139 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
140 if (!isa<IntrinsicInst>(CI))
142 else if (isa<VectorType>(CI->getType()))
147 // Threading through a switch statement is particularly profitable. If this
148 // block ends in a switch, decrease its cost to make it more likely to happen.
149 if (isa<SwitchInst>(I))
150 Size = Size > 6 ? Size-6 : 0;
156 /// ThreadBlock - If there are any predecessors whose control can be threaded
157 /// through to a successor, transform them now.
158 bool JumpThreading::ThreadBlock(BasicBlock *BB) {
159 // See if this block ends with a branch of switch. If so, see if the
160 // condition is a phi node. If so, and if an entry of the phi node is a
161 // constant, we can thread the block.
163 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
164 // Can't thread an unconditional jump.
165 if (BI->isUnconditional()) return false;
166 Condition = BI->getCondition();
167 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
168 Condition = SI->getCondition();
170 return false; // Must be an invoke.
172 // If the terminator of this block is branching on a constant, simplify the
173 // terminator to an unconditional branch. This can occur due to threading in
175 if (isa<ConstantInt>(Condition)) {
176 DOUT << " In block '" << BB->getNameStart()
177 << "' folding terminator: " << *BB->getTerminator();
179 ConstantFoldTerminator(BB);
183 // If there is only a single predecessor of this block, nothing to fold.
184 if (BB->getSinglePredecessor())
187 // See if this is a phi node in the current block.
188 PHINode *PN = dyn_cast<PHINode>(Condition);
189 if (PN && PN->getParent() == BB)
190 return ProcessJumpOnPHI(PN);
192 // If this is a conditional branch whose condition is and/or of a phi, try to
194 if (BinaryOperator *CondI = dyn_cast<BinaryOperator>(Condition)) {
195 if ((CondI->getOpcode() == Instruction::And ||
196 CondI->getOpcode() == Instruction::Or) &&
197 isa<BranchInst>(BB->getTerminator()) &&
198 ProcessBranchOnLogical(CondI, BB,
199 CondI->getOpcode() == Instruction::And))
203 // If we have "br (phi != 42)" and the phi node has any constant values as
204 // operands, we can thread through this block.
205 if (CmpInst *CondCmp = dyn_cast<CmpInst>(Condition))
206 if (isa<PHINode>(CondCmp->getOperand(0)) &&
207 isa<Constant>(CondCmp->getOperand(1)) &&
208 ProcessBranchOnCompare(CondCmp, BB))
214 /// ProcessJumpOnPHI - We have a conditional branch of switch on a PHI node in
215 /// the current block. See if there are any simplifications we can do based on
216 /// inputs to the phi node.
218 bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) {
219 // See if the phi node has any constant values. If so, we can determine where
220 // the corresponding predecessor will branch.
221 ConstantInt *PredCst = 0;
222 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
223 if ((PredCst = dyn_cast<ConstantInt>(PN->getIncomingValue(i))))
226 // If no incoming value has a constant, we don't know the destination of any
231 // See if the cost of duplicating this block is low enough.
232 BasicBlock *BB = PN->getParent();
233 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
234 if (JumpThreadCost > Threshold) {
235 DOUT << " Not threading BB '" << BB->getNameStart()
236 << "' - Cost is too high: " << JumpThreadCost << "\n";
240 // If so, we can actually do this threading. Merge any common predecessors
241 // that will act the same.
242 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst);
244 // Next, figure out which successor we are threading to.
246 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
247 SuccBB = BI->getSuccessor(PredCst == ConstantInt::getFalse());
249 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
250 SuccBB = SI->getSuccessor(SI->findCaseValue(PredCst));
253 // If threading to the same block as we come from, we would infinite loop.
255 DOUT << " Not threading BB '" << BB->getNameStart()
256 << "' - would thread to self!\n";
260 // And finally, do it!
261 DOUT << " Threading edge from '" << PredBB->getNameStart() << "' to '"
262 << SuccBB->getNameStart() << "' with cost: " << JumpThreadCost
263 << ", across block:\n "
266 ThreadEdge(BB, PredBB, SuccBB);
271 /// ProcessJumpOnLogicalPHI - PN's basic block contains a conditional branch
272 /// whose condition is an AND/OR where one side is PN. If PN has constant
273 /// operands that permit us to evaluate the condition for some operand, thread
274 /// through the block. For example with:
275 /// br (and X, phi(Y, Z, false))
276 /// the predecessor corresponding to the 'false' will always jump to the false
277 /// destination of the branch.
279 bool JumpThreading::ProcessBranchOnLogical(Value *V, BasicBlock *BB,
281 // If this is a binary operator tree of the same AND/OR opcode, check the
283 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V))
284 if (isAnd && BO->getOpcode() == Instruction::And ||
285 !isAnd && BO->getOpcode() == Instruction::Or) {
286 if (ProcessBranchOnLogical(BO->getOperand(0), BB, isAnd))
288 if (ProcessBranchOnLogical(BO->getOperand(1), BB, isAnd))
292 // If this isn't a PHI node, we can't handle it.
293 PHINode *PN = dyn_cast<PHINode>(V);
294 if (!PN || PN->getParent() != BB) return false;
296 // We can only do the simplification for phi nodes of 'false' with AND or
297 // 'true' with OR. See if we have any entries in the phi for this.
298 unsigned PredNo = ~0U;
299 ConstantInt *PredCst = ConstantInt::get(Type::Int1Ty, !isAnd);
300 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
301 if (PN->getIncomingValue(i) == PredCst) {
307 // If no match, bail out.
311 // See if the cost of duplicating this block is low enough.
312 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
313 if (JumpThreadCost > Threshold) {
314 DOUT << " Not threading BB '" << BB->getNameStart()
315 << "' - Cost is too high: " << JumpThreadCost << "\n";
319 // If so, we can actually do this threading. Merge any common predecessors
320 // that will act the same.
321 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst);
323 // Next, figure out which successor we are threading to. If this was an AND,
324 // the constant must be FALSE, and we must be targeting the 'false' block.
325 // If this is an OR, the constant must be TRUE, and we must be targeting the
327 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(isAnd);
329 // If threading to the same block as we come from, we would infinite loop.
331 DOUT << " Not threading BB '" << BB->getNameStart()
332 << "' - would thread to self!\n";
336 // And finally, do it!
337 DOUT << " Threading edge through bool from '" << PredBB->getNameStart()
338 << "' to '" << SuccBB->getNameStart() << "' with cost: "
339 << JumpThreadCost << ", across block:\n "
342 ThreadEdge(BB, PredBB, SuccBB);
347 /// ProcessBranchOnCompare - We found a branch on a comparison between a phi
348 /// node and a constant. If the PHI node contains any constants as inputs, we
349 /// can fold the compare for that edge and thread through it.
350 bool JumpThreading::ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB) {
351 PHINode *PN = cast<PHINode>(Cmp->getOperand(0));
352 Constant *RHS = cast<Constant>(Cmp->getOperand(1));
354 // If the phi isn't in the current block, an incoming edge to this block
355 // doesn't control the destination.
356 if (PN->getParent() != BB)
359 // We can do this simplification if any comparisons fold to true or false.
361 Constant *PredCst = 0;
362 bool TrueDirection = false;
363 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
364 PredCst = dyn_cast<Constant>(PN->getIncomingValue(i));
365 if (PredCst == 0) continue;
368 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Cmp))
369 Res = ConstantExpr::getICmp(ICI->getPredicate(), PredCst, RHS);
371 Res = ConstantExpr::getFCmp(cast<FCmpInst>(Cmp)->getPredicate(),
373 // If this folded to a constant expr, we can't do anything.
374 if (ConstantInt *ResC = dyn_cast<ConstantInt>(Res)) {
375 TrueDirection = ResC->getZExtValue();
378 // If this folded to undef, just go the false way.
379 if (isa<UndefValue>(Res)) {
380 TrueDirection = false;
384 // Otherwise, we can't fold this input.
388 // If no match, bail out.
392 // See if the cost of duplicating this block is low enough.
393 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
394 if (JumpThreadCost > Threshold) {
395 DOUT << " Not threading BB '" << BB->getNameStart()
396 << "' - Cost is too high: " << JumpThreadCost << "\n";
400 // If so, we can actually do this threading. Merge any common predecessors
401 // that will act the same.
402 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst);
404 // Next, get our successor.
405 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(!TrueDirection);
407 // If threading to the same block as we come from, we would infinite loop.
409 DOUT << " Not threading BB '" << BB->getNameStart()
410 << "' - would thread to self!\n";
415 // And finally, do it!
416 DOUT << " Threading edge through bool from '" << PredBB->getNameStart()
417 << "' to '" << SuccBB->getNameStart() << "' with cost: "
418 << JumpThreadCost << ", across block:\n "
421 ThreadEdge(BB, PredBB, SuccBB);
427 /// ThreadEdge - We have decided that it is safe and profitable to thread an
428 /// edge from PredBB to SuccBB across BB. Transform the IR to reflect this
430 void JumpThreading::ThreadEdge(BasicBlock *BB, BasicBlock *PredBB,
431 BasicBlock *SuccBB) {
433 // Jump Threading can not update SSA properties correctly if the values
434 // defined in the duplicated block are used outside of the block itself. For
435 // this reason, we spill all values that are used outside of BB to the stack.
436 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
437 if (!I->isUsedOutsideOfBlock(BB))
440 // We found a use of I outside of BB. Create a new stack slot to
441 // break this inter-block usage pattern.
442 if (!isa<StructType>(I->getType())) {
443 DemoteRegToStack(*I);
447 // Alternatively, I must be a call or invoke that returns multiple retvals.
448 // We can't use 'DemoteRegToStack' because that will create loads and
449 // stores of aggregates which is not valid yet. If I is a call, we can just
450 // pull all the getresult instructions up to this block. If I is an invoke,
451 // we are out of luck.
452 BasicBlock::iterator IP = I; ++IP;
453 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
455 cast<GetResultInst>(UI)->moveBefore(IP);
458 // We are going to have to map operands from the original BB block to the new
459 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
460 // account for entry from PredBB.
461 DenseMap<Instruction*, Value*> ValueMapping;
464 BasicBlock::Create(BB->getName()+".thread", BB->getParent(), BB);
465 NewBB->moveAfter(PredBB);
467 BasicBlock::iterator BI = BB->begin();
468 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
469 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
471 // Clone the non-phi instructions of BB into NewBB, keeping track of the
472 // mapping and using it to remap operands in the cloned instructions.
473 for (; !isa<TerminatorInst>(BI); ++BI) {
474 Instruction *New = BI->clone();
475 New->setName(BI->getNameStart());
476 NewBB->getInstList().push_back(New);
477 ValueMapping[BI] = New;
479 // Remap operands to patch up intra-block references.
480 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
481 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i)))
482 if (Value *Remapped = ValueMapping[Inst])
483 New->setOperand(i, Remapped);
486 // We didn't copy the terminator from BB over to NewBB, because there is now
487 // an unconditional jump to SuccBB. Insert the unconditional jump.
488 BranchInst::Create(SuccBB, NewBB);
490 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
491 // PHI nodes for NewBB now.
492 for (BasicBlock::iterator PNI = SuccBB->begin(); isa<PHINode>(PNI); ++PNI) {
493 PHINode *PN = cast<PHINode>(PNI);
494 // Ok, we have a PHI node. Figure out what the incoming value was for the
496 Value *IV = PN->getIncomingValueForBlock(BB);
498 // Remap the value if necessary.
499 if (Instruction *Inst = dyn_cast<Instruction>(IV))
500 if (Value *MappedIV = ValueMapping[Inst])
502 PN->addIncoming(IV, NewBB);
505 // Finally, NewBB is good to go. Update the terminator of PredBB to jump to
506 // NewBB instead of BB. This eliminates predecessors from BB, which requires
507 // us to simplify any PHI nodes in BB.
508 TerminatorInst *PredTerm = PredBB->getTerminator();
509 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
510 if (PredTerm->getSuccessor(i) == BB) {
511 BB->removePredecessor(PredBB);
512 PredTerm->setSuccessor(i, NewBB);