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/LLVMContext.h"
18 #include "llvm/Pass.h"
19 #include "llvm/Analysis/ConstantFolding.h"
20 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
21 #include "llvm/Transforms/Utils/Local.h"
22 #include "llvm/Target/TargetData.h"
23 #include "llvm/ADT/DenseMap.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/STLExtras.h"
26 #include "llvm/ADT/SmallPtrSet.h"
27 #include "llvm/ADT/SmallSet.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/ValueHandle.h"
31 #include "llvm/Support/raw_ostream.h"
34 STATISTIC(NumThreads
, "Number of jumps threaded");
35 STATISTIC(NumFolds
, "Number of terminators folded");
37 static cl::opt
<unsigned>
38 Threshold("jump-threading-threshold",
39 cl::desc("Max block size to duplicate for jump threading"),
40 cl::init(6), cl::Hidden
);
43 /// This pass performs 'jump threading', which looks at blocks that have
44 /// multiple predecessors and multiple successors. If one or more of the
45 /// predecessors of the block can be proven to always jump to one of the
46 /// successors, we forward the edge from the predecessor to the successor by
47 /// duplicating the contents of this block.
49 /// An example of when this can occur is code like this:
56 /// In this case, the unconditional branch at the end of the first if can be
57 /// revectored to the false side of the second if.
59 class JumpThreading
: public FunctionPass
{
62 SmallPtrSet
<BasicBlock
*, 16> LoopHeaders
;
64 SmallSet
<AssertingVH
<BasicBlock
>, 16> LoopHeaders
;
67 static char ID
; // Pass identification
68 JumpThreading() : FunctionPass(&ID
) {}
70 virtual void getAnalysisUsage(AnalysisUsage
&AU
) const {
73 bool runOnFunction(Function
&F
);
74 void FindLoopHeaders(Function
&F
);
76 bool ProcessBlock(BasicBlock
*BB
);
77 bool ThreadEdge(BasicBlock
*BB
, BasicBlock
*PredBB
, BasicBlock
*SuccBB
,
78 unsigned JumpThreadCost
);
79 BasicBlock
*FactorCommonPHIPreds(PHINode
*PN
, Value
*Val
);
80 bool ProcessBranchOnDuplicateCond(BasicBlock
*PredBB
, BasicBlock
*DestBB
);
81 bool ProcessSwitchOnDuplicateCond(BasicBlock
*PredBB
, BasicBlock
*DestBB
);
83 bool ProcessJumpOnPHI(PHINode
*PN
);
84 bool ProcessBranchOnLogical(Value
*V
, BasicBlock
*BB
, bool isAnd
);
85 bool ProcessBranchOnCompare(CmpInst
*Cmp
, BasicBlock
*BB
);
87 bool SimplifyPartiallyRedundantLoad(LoadInst
*LI
);
91 char JumpThreading::ID
= 0;
92 static RegisterPass
<JumpThreading
>
93 X("jump-threading", "Jump Threading");
95 // Public interface to the Jump Threading pass
96 FunctionPass
*llvm::createJumpThreadingPass() { return new JumpThreading(); }
98 /// runOnFunction - Top level algorithm.
100 bool JumpThreading::runOnFunction(Function
&F
) {
101 DEBUG(errs() << "Jump threading on function '" << F
.getName() << "'\n");
102 TD
= getAnalysisIfAvailable
<TargetData
>();
106 bool AnotherIteration
= true, EverChanged
= false;
107 while (AnotherIteration
) {
108 AnotherIteration
= false;
109 bool Changed
= false;
110 for (Function::iterator I
= F
.begin(), E
= F
.end(); I
!= E
;) {
112 while (ProcessBlock(BB
))
117 // If the block is trivially dead, zap it. This eliminates the successor
118 // edges which simplifies the CFG.
119 if (pred_begin(BB
) == pred_end(BB
) &&
120 BB
!= &BB
->getParent()->getEntryBlock()) {
121 DEBUG(errs() << " JT: Deleting dead block '" << BB
->getName()
122 << "' with terminator: " << *BB
->getTerminator());
123 LoopHeaders
.erase(BB
);
128 AnotherIteration
= Changed
;
129 EverChanged
|= Changed
;
136 /// FindLoopHeaders - We do not want jump threading to turn proper loop
137 /// structures into irreducible loops. Doing this breaks up the loop nesting
138 /// hierarchy and pessimizes later transformations. To prevent this from
139 /// happening, we first have to find the loop headers. Here we approximate this
140 /// by finding targets of backedges in the CFG.
142 /// Note that there definitely are cases when we want to allow threading of
143 /// edges across a loop header. For example, threading a jump from outside the
144 /// loop (the preheader) to an exit block of the loop is definitely profitable.
145 /// It is also almost always profitable to thread backedges from within the loop
146 /// to exit blocks, and is often profitable to thread backedges to other blocks
147 /// within the loop (forming a nested loop). This simple analysis is not rich
148 /// enough to track all of these properties and keep it up-to-date as the CFG
149 /// mutates, so we don't allow any of these transformations.
151 void JumpThreading::FindLoopHeaders(Function
&F
) {
152 SmallVector
<std::pair
<const BasicBlock
*,const BasicBlock
*>, 32> Edges
;
153 FindFunctionBackedges(F
, Edges
);
155 for (unsigned i
= 0, e
= Edges
.size(); i
!= e
; ++i
)
156 LoopHeaders
.insert(const_cast<BasicBlock
*>(Edges
[i
].second
));
160 /// FactorCommonPHIPreds - If there are multiple preds with the same incoming
161 /// value for the PHI, factor them together so we get one block to thread for
163 /// This is important for things like "phi i1 [true, true, false, true, x]"
164 /// where we only need to clone the block for the true blocks once.
166 BasicBlock
*JumpThreading::FactorCommonPHIPreds(PHINode
*PN
, Value
*Val
) {
167 SmallVector
<BasicBlock
*, 16> CommonPreds
;
168 for (unsigned i
= 0, e
= PN
->getNumIncomingValues(); i
!= e
; ++i
)
169 if (PN
->getIncomingValue(i
) == Val
)
170 CommonPreds
.push_back(PN
->getIncomingBlock(i
));
172 if (CommonPreds
.size() == 1)
173 return CommonPreds
[0];
175 DEBUG(errs() << " Factoring out " << CommonPreds
.size()
176 << " common predecessors.\n");
177 return SplitBlockPredecessors(PN
->getParent(),
178 &CommonPreds
[0], CommonPreds
.size(),
183 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
184 /// thread across it.
185 static unsigned getJumpThreadDuplicationCost(const BasicBlock
*BB
) {
186 /// Ignore PHI nodes, these will be flattened when duplication happens.
187 BasicBlock::const_iterator I
= BB
->getFirstNonPHI();
189 // Sum up the cost of each instruction until we get to the terminator. Don't
190 // include the terminator because the copy won't include it.
192 for (; !isa
<TerminatorInst
>(I
); ++I
) {
193 // Debugger intrinsics don't incur code size.
194 if (isa
<DbgInfoIntrinsic
>(I
)) continue;
196 // If this is a pointer->pointer bitcast, it is free.
197 if (isa
<BitCastInst
>(I
) && isa
<PointerType
>(I
->getType()))
200 // All other instructions count for at least one unit.
203 // Calls are more expensive. If they are non-intrinsic calls, we model them
204 // as having cost of 4. If they are a non-vector intrinsic, we model them
205 // as having cost of 2 total, and if they are a vector intrinsic, we model
206 // them as having cost 1.
207 if (const CallInst
*CI
= dyn_cast
<CallInst
>(I
)) {
208 if (!isa
<IntrinsicInst
>(CI
))
210 else if (!isa
<VectorType
>(CI
->getType()))
215 // Threading through a switch statement is particularly profitable. If this
216 // block ends in a switch, decrease its cost to make it more likely to happen.
217 if (isa
<SwitchInst
>(I
))
218 Size
= Size
> 6 ? Size
-6 : 0;
223 /// ProcessBlock - If there are any predecessors whose control can be threaded
224 /// through to a successor, transform them now.
225 bool JumpThreading::ProcessBlock(BasicBlock
*BB
) {
226 // If this block has a single predecessor, and if that pred has a single
227 // successor, merge the blocks. This encourages recursive jump threading
228 // because now the condition in this block can be threaded through
229 // predecessors of our predecessor block.
230 if (BasicBlock
*SinglePred
= BB
->getSinglePredecessor())
231 if (SinglePred
->getTerminator()->getNumSuccessors() == 1 &&
233 // If SinglePred was a loop header, BB becomes one.
234 if (LoopHeaders
.erase(SinglePred
))
235 LoopHeaders
.insert(BB
);
237 // Remember if SinglePred was the entry block of the function. If so, we
238 // will need to move BB back to the entry position.
239 bool isEntry
= SinglePred
== &SinglePred
->getParent()->getEntryBlock();
240 MergeBasicBlockIntoOnlyPred(BB
);
242 if (isEntry
&& BB
!= &BB
->getParent()->getEntryBlock())
243 BB
->moveBefore(&BB
->getParent()->getEntryBlock());
247 // See if this block ends with a branch or switch. If so, see if the
248 // condition is a phi node. If so, and if an entry of the phi node is a
249 // constant, we can thread the block.
251 if (BranchInst
*BI
= dyn_cast
<BranchInst
>(BB
->getTerminator())) {
252 // Can't thread an unconditional jump.
253 if (BI
->isUnconditional()) return false;
254 Condition
= BI
->getCondition();
255 } else if (SwitchInst
*SI
= dyn_cast
<SwitchInst
>(BB
->getTerminator()))
256 Condition
= SI
->getCondition();
258 return false; // Must be an invoke.
260 // If the terminator of this block is branching on a constant, simplify the
261 // terminator to an unconditional branch. This can occur due to threading in
263 if (isa
<ConstantInt
>(Condition
)) {
264 DEBUG(errs() << " In block '" << BB
->getName()
265 << "' folding terminator: " << *BB
->getTerminator());
267 ConstantFoldTerminator(BB
);
271 // If the terminator is branching on an undef, we can pick any of the
272 // successors to branch to. Since this is arbitrary, we pick the successor
273 // with the fewest predecessors. This should reduce the in-degree of the
275 if (isa
<UndefValue
>(Condition
)) {
276 TerminatorInst
*BBTerm
= BB
->getTerminator();
277 unsigned MinSucc
= 0;
278 BasicBlock
*TestBB
= BBTerm
->getSuccessor(MinSucc
);
279 // Compute the successor with the minimum number of predecessors.
280 unsigned MinNumPreds
= std::distance(pred_begin(TestBB
), pred_end(TestBB
));
281 for (unsigned i
= 1, e
= BBTerm
->getNumSuccessors(); i
!= e
; ++i
) {
282 TestBB
= BBTerm
->getSuccessor(i
);
283 unsigned NumPreds
= std::distance(pred_begin(TestBB
), pred_end(TestBB
));
284 if (NumPreds
< MinNumPreds
)
288 // Fold the branch/switch.
289 for (unsigned i
= 0, e
= BBTerm
->getNumSuccessors(); i
!= e
; ++i
) {
290 if (i
== MinSucc
) continue;
291 BBTerm
->getSuccessor(i
)->removePredecessor(BB
);
294 DEBUG(errs() << " In block '" << BB
->getName()
295 << "' folding undef terminator: " << *BBTerm
);
296 BranchInst::Create(BBTerm
->getSuccessor(MinSucc
), BBTerm
);
297 BBTerm
->eraseFromParent();
301 Instruction
*CondInst
= dyn_cast
<Instruction
>(Condition
);
303 // If the condition is an instruction defined in another block, see if a
304 // predecessor has the same condition:
308 if (!Condition
->hasOneUse() && // Multiple uses.
309 (CondInst
== 0 || CondInst
->getParent() != BB
)) { // Non-local definition.
310 pred_iterator PI
= pred_begin(BB
), E
= pred_end(BB
);
311 if (isa
<BranchInst
>(BB
->getTerminator())) {
312 for (; PI
!= E
; ++PI
)
313 if (BranchInst
*PBI
= dyn_cast
<BranchInst
>((*PI
)->getTerminator()))
314 if (PBI
->isConditional() && PBI
->getCondition() == Condition
&&
315 ProcessBranchOnDuplicateCond(*PI
, BB
))
318 assert(isa
<SwitchInst
>(BB
->getTerminator()) && "Unknown jump terminator");
319 for (; PI
!= E
; ++PI
)
320 if (SwitchInst
*PSI
= dyn_cast
<SwitchInst
>((*PI
)->getTerminator()))
321 if (PSI
->getCondition() == Condition
&&
322 ProcessSwitchOnDuplicateCond(*PI
, BB
))
327 // All the rest of our checks depend on the condition being an instruction.
331 // See if this is a phi node in the current block.
332 if (PHINode
*PN
= dyn_cast
<PHINode
>(CondInst
))
333 if (PN
->getParent() == BB
)
334 return ProcessJumpOnPHI(PN
);
336 // If this is a conditional branch whose condition is and/or of a phi, try to
338 if ((CondInst
->getOpcode() == Instruction::And
||
339 CondInst
->getOpcode() == Instruction::Or
) &&
340 isa
<BranchInst
>(BB
->getTerminator()) &&
341 ProcessBranchOnLogical(CondInst
, BB
,
342 CondInst
->getOpcode() == Instruction::And
))
345 if (CmpInst
*CondCmp
= dyn_cast
<CmpInst
>(CondInst
)) {
346 if (isa
<PHINode
>(CondCmp
->getOperand(0))) {
347 // If we have "br (phi != 42)" and the phi node has any constant values
348 // as operands, we can thread through this block.
350 // If we have "br (cmp phi, x)" and the phi node contains x such that the
351 // comparison uniquely identifies the branch target, we can thread
352 // through this block.
354 if (ProcessBranchOnCompare(CondCmp
, BB
))
358 // If we have a comparison, loop over the predecessors to see if there is
359 // a condition with the same value.
360 pred_iterator PI
= pred_begin(BB
), E
= pred_end(BB
);
361 for (; PI
!= E
; ++PI
)
362 if (BranchInst
*PBI
= dyn_cast
<BranchInst
>((*PI
)->getTerminator()))
363 if (PBI
->isConditional() && *PI
!= BB
) {
364 if (CmpInst
*CI
= dyn_cast
<CmpInst
>(PBI
->getCondition())) {
365 if (CI
->getOperand(0) == CondCmp
->getOperand(0) &&
366 CI
->getOperand(1) == CondCmp
->getOperand(1) &&
367 CI
->getPredicate() == CondCmp
->getPredicate()) {
368 // TODO: Could handle things like (x != 4) --> (x == 17)
369 if (ProcessBranchOnDuplicateCond(*PI
, BB
))
376 // Check for some cases that are worth simplifying. Right now we want to look
377 // for loads that are used by a switch or by the condition for the branch. If
378 // we see one, check to see if it's partially redundant. If so, insert a PHI
379 // which can then be used to thread the values.
381 // This is particularly important because reg2mem inserts loads and stores all
382 // over the place, and this blocks jump threading if we don't zap them.
383 Value
*SimplifyValue
= CondInst
;
384 if (CmpInst
*CondCmp
= dyn_cast
<CmpInst
>(SimplifyValue
))
385 if (isa
<Constant
>(CondCmp
->getOperand(1)))
386 SimplifyValue
= CondCmp
->getOperand(0);
388 if (LoadInst
*LI
= dyn_cast
<LoadInst
>(SimplifyValue
))
389 if (SimplifyPartiallyRedundantLoad(LI
))
392 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
393 // "(X == 4)" thread through this block.
398 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
399 /// block that jump on exactly the same condition. This means that we almost
400 /// always know the direction of the edge in the DESTBB:
402 /// br COND, DESTBB, BBY
404 /// br COND, BBZ, BBW
406 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
407 /// in DESTBB, we have to thread over it.
408 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock
*PredBB
,
410 BranchInst
*PredBI
= cast
<BranchInst
>(PredBB
->getTerminator());
412 // If both successors of PredBB go to DESTBB, we don't know anything. We can
413 // fold the branch to an unconditional one, which allows other recursive
416 if (PredBI
->getSuccessor(1) != BB
)
418 else if (PredBI
->getSuccessor(0) != BB
)
421 DEBUG(errs() << " In block '" << PredBB
->getName()
422 << "' folding terminator: " << *PredBB
->getTerminator());
424 ConstantFoldTerminator(PredBB
);
428 BranchInst
*DestBI
= cast
<BranchInst
>(BB
->getTerminator());
430 // If the dest block has one predecessor, just fix the branch condition to a
431 // constant and fold it.
432 if (BB
->getSinglePredecessor()) {
433 DEBUG(errs() << " In block '" << BB
->getName()
434 << "' folding condition to '" << BranchDir
<< "': "
435 << *BB
->getTerminator());
437 DestBI
->setCondition(ConstantInt::get(Type::getInt1Ty(BB
->getContext()),
439 ConstantFoldTerminator(BB
);
443 // Otherwise we need to thread from PredBB to DestBB's successor which
444 // involves code duplication. Check to see if it is worth it.
445 unsigned JumpThreadCost
= getJumpThreadDuplicationCost(BB
);
446 if (JumpThreadCost
> Threshold
) {
447 DEBUG(errs() << " Not threading BB '" << BB
->getName()
448 << "' - Cost is too high: " << JumpThreadCost
<< "\n");
452 // Next, figure out which successor we are threading to.
453 BasicBlock
*SuccBB
= DestBI
->getSuccessor(!BranchDir
);
455 // Ok, try to thread it!
456 return ThreadEdge(BB
, PredBB
, SuccBB
, JumpThreadCost
);
459 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
460 /// block that switch on exactly the same condition. This means that we almost
461 /// always know the direction of the edge in the DESTBB:
463 /// switch COND [... DESTBB, BBY ... ]
465 /// switch COND [... BBZ, BBW ]
467 /// Optimizing switches like this is very important, because simplifycfg builds
468 /// switches out of repeated 'if' conditions.
469 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock
*PredBB
,
470 BasicBlock
*DestBB
) {
471 // Can't thread edge to self.
472 if (PredBB
== DestBB
)
476 SwitchInst
*PredSI
= cast
<SwitchInst
>(PredBB
->getTerminator());
477 SwitchInst
*DestSI
= cast
<SwitchInst
>(DestBB
->getTerminator());
479 // There are a variety of optimizations that we can potentially do on these
480 // blocks: we order them from most to least preferable.
482 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
483 // directly to their destination. This does not introduce *any* code size
484 // growth. Skip debug info first.
485 BasicBlock::iterator BBI
= DestBB
->begin();
486 while (isa
<DbgInfoIntrinsic
>(BBI
))
489 // FIXME: Thread if it just contains a PHI.
490 if (isa
<SwitchInst
>(BBI
)) {
491 bool MadeChange
= false;
492 // Ignore the default edge for now.
493 for (unsigned i
= 1, e
= DestSI
->getNumSuccessors(); i
!= e
; ++i
) {
494 ConstantInt
*DestVal
= DestSI
->getCaseValue(i
);
495 BasicBlock
*DestSucc
= DestSI
->getSuccessor(i
);
497 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
498 // PredSI has an explicit case for it. If so, forward. If it is covered
499 // by the default case, we can't update PredSI.
500 unsigned PredCase
= PredSI
->findCaseValue(DestVal
);
501 if (PredCase
== 0) continue;
503 // If PredSI doesn't go to DestBB on this value, then it won't reach the
504 // case on this condition.
505 if (PredSI
->getSuccessor(PredCase
) != DestBB
&&
506 DestSI
->getSuccessor(i
) != DestBB
)
509 // Otherwise, we're safe to make the change. Make sure that the edge from
510 // DestSI to DestSucc is not critical and has no PHI nodes.
511 DEBUG(errs() << "FORWARDING EDGE " << *DestVal
<< " FROM: " << *PredSI
);
512 DEBUG(errs() << "THROUGH: " << *DestSI
);
514 // If the destination has PHI nodes, just split the edge for updating
516 if (isa
<PHINode
>(DestSucc
->begin()) && !DestSucc
->getSinglePredecessor()){
517 SplitCriticalEdge(DestSI
, i
, this);
518 DestSucc
= DestSI
->getSuccessor(i
);
520 FoldSingleEntryPHINodes(DestSucc
);
521 PredSI
->setSuccessor(PredCase
, DestSucc
);
533 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
534 /// load instruction, eliminate it by replacing it with a PHI node. This is an
535 /// important optimization that encourages jump threading, and needs to be run
536 /// interlaced with other jump threading tasks.
537 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst
*LI
) {
538 // Don't hack volatile loads.
539 if (LI
->isVolatile()) return false;
541 // If the load is defined in a block with exactly one predecessor, it can't be
542 // partially redundant.
543 BasicBlock
*LoadBB
= LI
->getParent();
544 if (LoadBB
->getSinglePredecessor())
547 Value
*LoadedPtr
= LI
->getOperand(0);
549 // If the loaded operand is defined in the LoadBB, it can't be available.
550 // FIXME: Could do PHI translation, that would be fun :)
551 if (Instruction
*PtrOp
= dyn_cast
<Instruction
>(LoadedPtr
))
552 if (PtrOp
->getParent() == LoadBB
)
555 // Scan a few instructions up from the load, to see if it is obviously live at
556 // the entry to its block.
557 BasicBlock::iterator BBIt
= LI
;
559 if (Value
*AvailableVal
= FindAvailableLoadedValue(LoadedPtr
, LoadBB
,
561 // If the value if the load is locally available within the block, just use
562 // it. This frequently occurs for reg2mem'd allocas.
563 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
565 // If the returned value is the load itself, replace with an undef. This can
566 // only happen in dead loops.
567 if (AvailableVal
== LI
) AvailableVal
= UndefValue::get(LI
->getType());
568 LI
->replaceAllUsesWith(AvailableVal
);
569 LI
->eraseFromParent();
573 // Otherwise, if we scanned the whole block and got to the top of the block,
574 // we know the block is locally transparent to the load. If not, something
575 // might clobber its value.
576 if (BBIt
!= LoadBB
->begin())
580 SmallPtrSet
<BasicBlock
*, 8> PredsScanned
;
581 typedef SmallVector
<std::pair
<BasicBlock
*, Value
*>, 8> AvailablePredsTy
;
582 AvailablePredsTy AvailablePreds
;
583 BasicBlock
*OneUnavailablePred
= 0;
585 // If we got here, the loaded value is transparent through to the start of the
586 // block. Check to see if it is available in any of the predecessor blocks.
587 for (pred_iterator PI
= pred_begin(LoadBB
), PE
= pred_end(LoadBB
);
589 BasicBlock
*PredBB
= *PI
;
591 // If we already scanned this predecessor, skip it.
592 if (!PredsScanned
.insert(PredBB
))
595 // Scan the predecessor to see if the value is available in the pred.
596 BBIt
= PredBB
->end();
597 Value
*PredAvailable
= FindAvailableLoadedValue(LoadedPtr
, PredBB
, BBIt
, 6);
598 if (!PredAvailable
) {
599 OneUnavailablePred
= PredBB
;
603 // If so, this load is partially redundant. Remember this info so that we
604 // can create a PHI node.
605 AvailablePreds
.push_back(std::make_pair(PredBB
, PredAvailable
));
608 // If the loaded value isn't available in any predecessor, it isn't partially
610 if (AvailablePreds
.empty()) return false;
612 // Okay, the loaded value is available in at least one (and maybe all!)
613 // predecessors. If the value is unavailable in more than one unique
614 // predecessor, we want to insert a merge block for those common predecessors.
615 // This ensures that we only have to insert one reload, thus not increasing
617 BasicBlock
*UnavailablePred
= 0;
619 // If there is exactly one predecessor where the value is unavailable, the
620 // already computed 'OneUnavailablePred' block is it. If it ends in an
621 // unconditional branch, we know that it isn't a critical edge.
622 if (PredsScanned
.size() == AvailablePreds
.size()+1 &&
623 OneUnavailablePred
->getTerminator()->getNumSuccessors() == 1) {
624 UnavailablePred
= OneUnavailablePred
;
625 } else if (PredsScanned
.size() != AvailablePreds
.size()) {
626 // Otherwise, we had multiple unavailable predecessors or we had a critical
627 // edge from the one.
628 SmallVector
<BasicBlock
*, 8> PredsToSplit
;
629 SmallPtrSet
<BasicBlock
*, 8> AvailablePredSet
;
631 for (unsigned i
= 0, e
= AvailablePreds
.size(); i
!= e
; ++i
)
632 AvailablePredSet
.insert(AvailablePreds
[i
].first
);
634 // Add all the unavailable predecessors to the PredsToSplit list.
635 for (pred_iterator PI
= pred_begin(LoadBB
), PE
= pred_end(LoadBB
);
637 if (!AvailablePredSet
.count(*PI
))
638 PredsToSplit
.push_back(*PI
);
640 // Split them out to their own block.
642 SplitBlockPredecessors(LoadBB
, &PredsToSplit
[0], PredsToSplit
.size(),
643 "thread-split", this);
646 // If the value isn't available in all predecessors, then there will be
647 // exactly one where it isn't available. Insert a load on that edge and add
648 // it to the AvailablePreds list.
649 if (UnavailablePred
) {
650 assert(UnavailablePred
->getTerminator()->getNumSuccessors() == 1 &&
651 "Can't handle critical edge here!");
652 Value
*NewVal
= new LoadInst(LoadedPtr
, LI
->getName()+".pr",
653 UnavailablePred
->getTerminator());
654 AvailablePreds
.push_back(std::make_pair(UnavailablePred
, NewVal
));
657 // Now we know that each predecessor of this block has a value in
658 // AvailablePreds, sort them for efficient access as we're walking the preds.
659 array_pod_sort(AvailablePreds
.begin(), AvailablePreds
.end());
661 // Create a PHI node at the start of the block for the PRE'd load value.
662 PHINode
*PN
= PHINode::Create(LI
->getType(), "", LoadBB
->begin());
665 // Insert new entries into the PHI for each predecessor. A single block may
666 // have multiple entries here.
667 for (pred_iterator PI
= pred_begin(LoadBB
), E
= pred_end(LoadBB
); PI
!= E
;
669 AvailablePredsTy::iterator I
=
670 std::lower_bound(AvailablePreds
.begin(), AvailablePreds
.end(),
671 std::make_pair(*PI
, (Value
*)0));
673 assert(I
!= AvailablePreds
.end() && I
->first
== *PI
&&
674 "Didn't find entry for predecessor!");
676 PN
->addIncoming(I
->second
, I
->first
);
679 //cerr << "PRE: " << *LI << *PN << "\n";
681 LI
->replaceAllUsesWith(PN
);
682 LI
->eraseFromParent();
688 /// ProcessJumpOnPHI - We have a conditional branch of switch on a PHI node in
689 /// the current block. See if there are any simplifications we can do based on
690 /// inputs to the phi node.
692 bool JumpThreading::ProcessJumpOnPHI(PHINode
*PN
) {
693 // See if the phi node has any constant values. If so, we can determine where
694 // the corresponding predecessor will branch.
695 ConstantInt
*PredCst
= 0;
696 for (unsigned i
= 0, e
= PN
->getNumIncomingValues(); i
!= e
; ++i
)
697 if ((PredCst
= dyn_cast
<ConstantInt
>(PN
->getIncomingValue(i
))))
700 // If no incoming value has a constant, we don't know the destination of any
705 // See if the cost of duplicating this block is low enough.
706 BasicBlock
*BB
= PN
->getParent();
707 unsigned JumpThreadCost
= getJumpThreadDuplicationCost(BB
);
708 if (JumpThreadCost
> Threshold
) {
709 DEBUG(errs() << " Not threading BB '" << BB
->getName()
710 << "' - Cost is too high: " << JumpThreadCost
<< "\n");
714 // If so, we can actually do this threading. Merge any common predecessors
715 // that will act the same.
716 BasicBlock
*PredBB
= FactorCommonPHIPreds(PN
, PredCst
);
718 // Next, figure out which successor we are threading to.
720 if (BranchInst
*BI
= dyn_cast
<BranchInst
>(BB
->getTerminator()))
721 SuccBB
= BI
->getSuccessor(PredCst
==
722 ConstantInt::getFalse(PredBB
->getContext()));
724 SwitchInst
*SI
= cast
<SwitchInst
>(BB
->getTerminator());
725 SuccBB
= SI
->getSuccessor(SI
->findCaseValue(PredCst
));
728 // Ok, try to thread it!
729 return ThreadEdge(BB
, PredBB
, SuccBB
, JumpThreadCost
);
732 /// ProcessJumpOnLogicalPHI - PN's basic block contains a conditional branch
733 /// whose condition is an AND/OR where one side is PN. If PN has constant
734 /// operands that permit us to evaluate the condition for some operand, thread
735 /// through the block. For example with:
736 /// br (and X, phi(Y, Z, false))
737 /// the predecessor corresponding to the 'false' will always jump to the false
738 /// destination of the branch.
740 bool JumpThreading::ProcessBranchOnLogical(Value
*V
, BasicBlock
*BB
,
742 // If this is a binary operator tree of the same AND/OR opcode, check the
744 if (BinaryOperator
*BO
= dyn_cast
<BinaryOperator
>(V
))
745 if ((isAnd
&& BO
->getOpcode() == Instruction::And
) ||
746 (!isAnd
&& BO
->getOpcode() == Instruction::Or
)) {
747 if (ProcessBranchOnLogical(BO
->getOperand(0), BB
, isAnd
))
749 if (ProcessBranchOnLogical(BO
->getOperand(1), BB
, isAnd
))
753 // If this isn't a PHI node, we can't handle it.
754 PHINode
*PN
= dyn_cast
<PHINode
>(V
);
755 if (!PN
|| PN
->getParent() != BB
) return false;
757 // We can only do the simplification for phi nodes of 'false' with AND or
758 // 'true' with OR. See if we have any entries in the phi for this.
759 unsigned PredNo
= ~0U;
760 ConstantInt
*PredCst
= ConstantInt::get(Type::getInt1Ty(BB
->getContext()),
762 for (unsigned i
= 0, e
= PN
->getNumIncomingValues(); i
!= e
; ++i
) {
763 if (PN
->getIncomingValue(i
) == PredCst
) {
769 // If no match, bail out.
773 // See if the cost of duplicating this block is low enough.
774 unsigned JumpThreadCost
= getJumpThreadDuplicationCost(BB
);
775 if (JumpThreadCost
> Threshold
) {
776 DEBUG(errs() << " Not threading BB '" << BB
->getName()
777 << "' - Cost is too high: " << JumpThreadCost
<< "\n");
781 // If so, we can actually do this threading. Merge any common predecessors
782 // that will act the same.
783 BasicBlock
*PredBB
= FactorCommonPHIPreds(PN
, PredCst
);
785 // Next, figure out which successor we are threading to. If this was an AND,
786 // the constant must be FALSE, and we must be targeting the 'false' block.
787 // If this is an OR, the constant must be TRUE, and we must be targeting the
789 BasicBlock
*SuccBB
= BB
->getTerminator()->getSuccessor(isAnd
);
791 // Ok, try to thread it!
792 return ThreadEdge(BB
, PredBB
, SuccBB
, JumpThreadCost
);
795 /// GetResultOfComparison - Given an icmp/fcmp predicate and the left and right
796 /// hand sides of the compare instruction, try to determine the result. If the
797 /// result can not be determined, a null pointer is returned.
798 static Constant
*GetResultOfComparison(CmpInst::Predicate pred
,
799 Value
*LHS
, Value
*RHS
,
800 LLVMContext
&Context
) {
801 if (Constant
*CLHS
= dyn_cast
<Constant
>(LHS
))
802 if (Constant
*CRHS
= dyn_cast
<Constant
>(RHS
))
803 return ConstantExpr::getCompare(pred
, CLHS
, CRHS
);
806 if (isa
<IntegerType
>(LHS
->getType()) || isa
<PointerType
>(LHS
->getType()))
807 return ICmpInst::isTrueWhenEqual(pred
) ?
808 ConstantInt::getTrue(Context
) : ConstantInt::getFalse(Context
);
813 /// ProcessBranchOnCompare - We found a branch on a comparison between a phi
814 /// node and a value. If we can identify when the comparison is true between
815 /// the phi inputs and the value, we can fold the compare for that edge and
816 /// thread through it.
817 bool JumpThreading::ProcessBranchOnCompare(CmpInst
*Cmp
, BasicBlock
*BB
) {
818 PHINode
*PN
= cast
<PHINode
>(Cmp
->getOperand(0));
819 Value
*RHS
= Cmp
->getOperand(1);
821 // If the phi isn't in the current block, an incoming edge to this block
822 // doesn't control the destination.
823 if (PN
->getParent() != BB
)
826 // We can do this simplification if any comparisons fold to true or false.
829 bool TrueDirection
= false;
830 for (unsigned i
= 0, e
= PN
->getNumIncomingValues(); i
!= e
; ++i
) {
831 PredVal
= PN
->getIncomingValue(i
);
833 Constant
*Res
= GetResultOfComparison(Cmp
->getPredicate(), PredVal
,
834 RHS
, Cmp
->getContext());
840 // If this folded to a constant expr, we can't do anything.
841 if (ConstantInt
*ResC
= dyn_cast
<ConstantInt
>(Res
)) {
842 TrueDirection
= ResC
->getZExtValue();
845 // If this folded to undef, just go the false way.
846 if (isa
<UndefValue
>(Res
)) {
847 TrueDirection
= false;
851 // Otherwise, we can't fold this input.
855 // If no match, bail out.
859 // See if the cost of duplicating this block is low enough.
860 unsigned JumpThreadCost
= getJumpThreadDuplicationCost(BB
);
861 if (JumpThreadCost
> Threshold
) {
862 DEBUG(errs() << " Not threading BB '" << BB
->getName()
863 << "' - Cost is too high: " << JumpThreadCost
<< "\n");
867 // If so, we can actually do this threading. Merge any common predecessors
868 // that will act the same.
869 BasicBlock
*PredBB
= FactorCommonPHIPreds(PN
, PredVal
);
871 // Next, get our successor.
872 BasicBlock
*SuccBB
= BB
->getTerminator()->getSuccessor(!TrueDirection
);
874 // Ok, try to thread it!
875 return ThreadEdge(BB
, PredBB
, SuccBB
, JumpThreadCost
);
879 /// ThreadEdge - We have decided that it is safe and profitable to thread an
880 /// edge from PredBB to SuccBB across BB. Transform the IR to reflect this
882 bool JumpThreading::ThreadEdge(BasicBlock
*BB
, BasicBlock
*PredBB
,
883 BasicBlock
*SuccBB
, unsigned JumpThreadCost
) {
885 // If threading to the same block as we come from, we would infinite loop.
887 DEBUG(errs() << " Not threading across BB '" << BB
->getName()
888 << "' - would thread to self!\n");
892 // If threading this would thread across a loop header, don't thread the edge.
893 // See the comments above FindLoopHeaders for justifications and caveats.
894 if (LoopHeaders
.count(BB
)) {
895 DEBUG(errs() << " Not threading from '" << PredBB
->getName()
896 << "' across loop header BB '" << BB
->getName()
897 << "' to dest BB '" << SuccBB
->getName()
898 << "' - it might create an irreducible loop!\n");
902 // And finally, do it!
903 DEBUG(errs() << " Threading edge from '" << PredBB
->getName() << "' to '"
904 << SuccBB
->getName() << "' with cost: " << JumpThreadCost
905 << ", across block:\n "
908 // Jump Threading can not update SSA properties correctly if the values
909 // defined in the duplicated block are used outside of the block itself. For
910 // this reason, we spill all values that are used outside of BB to the stack.
911 for (BasicBlock::iterator I
= BB
->begin(); I
!= BB
->end(); ++I
) {
912 if (!I
->isUsedOutsideOfBlock(BB
))
915 // We found a use of I outside of BB. Create a new stack slot to
916 // break this inter-block usage pattern.
917 DemoteRegToStack(*I
);
920 // We are going to have to map operands from the original BB block to the new
921 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
922 // account for entry from PredBB.
923 DenseMap
<Instruction
*, Value
*> ValueMapping
;
925 BasicBlock
*NewBB
= BasicBlock::Create(BB
->getContext(),
926 BB
->getName()+".thread",
927 BB
->getParent(), BB
);
928 NewBB
->moveAfter(PredBB
);
930 BasicBlock::iterator BI
= BB
->begin();
931 for (; PHINode
*PN
= dyn_cast
<PHINode
>(BI
); ++BI
)
932 ValueMapping
[PN
] = PN
->getIncomingValueForBlock(PredBB
);
934 // Clone the non-phi instructions of BB into NewBB, keeping track of the
935 // mapping and using it to remap operands in the cloned instructions.
936 for (; !isa
<TerminatorInst
>(BI
); ++BI
) {
937 Instruction
*New
= BI
->clone(BI
->getContext());
938 New
->setName(BI
->getName());
939 NewBB
->getInstList().push_back(New
);
940 ValueMapping
[BI
] = New
;
942 // Remap operands to patch up intra-block references.
943 for (unsigned i
= 0, e
= New
->getNumOperands(); i
!= e
; ++i
)
944 if (Instruction
*Inst
= dyn_cast
<Instruction
>(New
->getOperand(i
))) {
945 DenseMap
<Instruction
*, Value
*>::iterator I
= ValueMapping
.find(Inst
);
946 if (I
!= ValueMapping
.end())
947 New
->setOperand(i
, I
->second
);
951 // We didn't copy the terminator from BB over to NewBB, because there is now
952 // an unconditional jump to SuccBB. Insert the unconditional jump.
953 BranchInst::Create(SuccBB
, NewBB
);
955 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
956 // PHI nodes for NewBB now.
957 for (BasicBlock::iterator PNI
= SuccBB
->begin(); isa
<PHINode
>(PNI
); ++PNI
) {
958 PHINode
*PN
= cast
<PHINode
>(PNI
);
959 // Ok, we have a PHI node. Figure out what the incoming value was for the
961 Value
*IV
= PN
->getIncomingValueForBlock(BB
);
963 // Remap the value if necessary.
964 if (Instruction
*Inst
= dyn_cast
<Instruction
>(IV
)) {
965 DenseMap
<Instruction
*, Value
*>::iterator I
= ValueMapping
.find(Inst
);
966 if (I
!= ValueMapping
.end())
969 PN
->addIncoming(IV
, NewBB
);
972 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
973 // NewBB instead of BB. This eliminates predecessors from BB, which requires
974 // us to simplify any PHI nodes in BB.
975 TerminatorInst
*PredTerm
= PredBB
->getTerminator();
976 for (unsigned i
= 0, e
= PredTerm
->getNumSuccessors(); i
!= e
; ++i
)
977 if (PredTerm
->getSuccessor(i
) == BB
) {
978 BB
->removePredecessor(PredBB
);
979 PredTerm
->setSuccessor(i
, NewBB
);
982 // At this point, the IR is fully up to date and consistent. Do a quick scan
983 // over the new instructions and zap any that are constants or dead. This
984 // frequently happens because of phi translation.
986 for (BasicBlock::iterator E
= NewBB
->end(); BI
!= E
; ) {
987 Instruction
*Inst
= BI
++;
988 if (Constant
*C
= ConstantFoldInstruction(Inst
, BB
->getContext(), TD
)) {
989 Inst
->replaceAllUsesWith(C
);
990 Inst
->eraseFromParent();
994 RecursivelyDeleteTriviallyDeadInstructions(Inst
);