1 //===-- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop ------===//
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 transforms loops that contain branches on loop-invariant conditions
11 // to have multiple loops. For example, it turns the left into the right code:
20 // This can increase the size of the code exponentially (doubling it every time
21 // a loop is unswitched) so we only unswitch if the resultant code will be
22 // smaller than a threshold.
24 // This pass expects LICM to be run before it to hoist invariant conditions out
25 // of the loop, to make the unswitching opportunity obvious.
27 //===----------------------------------------------------------------------===//
29 #define DEBUG_TYPE "loop-unswitch"
30 #include "llvm/Transforms/Scalar.h"
31 #include "llvm/Constants.h"
32 #include "llvm/DerivedTypes.h"
33 #include "llvm/Function.h"
34 #include "llvm/Instructions.h"
35 #include "llvm/Analysis/InlineCost.h"
36 #include "llvm/Analysis/InstructionSimplify.h"
37 #include "llvm/Analysis/LoopInfo.h"
38 #include "llvm/Analysis/LoopPass.h"
39 #include "llvm/Analysis/Dominators.h"
40 #include "llvm/Analysis/ScalarEvolution.h"
41 #include "llvm/Transforms/Utils/Cloning.h"
42 #include "llvm/Transforms/Utils/Local.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/ADT/SmallPtrSet.h"
46 #include "llvm/ADT/STLExtras.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/raw_ostream.h"
54 STATISTIC(NumBranches
, "Number of branches unswitched");
55 STATISTIC(NumSwitches
, "Number of switches unswitched");
56 STATISTIC(NumSelects
, "Number of selects unswitched");
57 STATISTIC(NumTrivial
, "Number of unswitches that are trivial");
58 STATISTIC(NumSimplify
, "Number of simplifications of unswitched code");
60 // The specific value of 50 here was chosen based only on intuition and a
61 // few specific examples.
62 static cl::opt
<unsigned>
63 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
64 cl::init(50), cl::Hidden
);
67 class LoopUnswitch
: public LoopPass
{
68 LoopInfo
*LI
; // Loop information
71 // LoopProcessWorklist - Used to check if second loop needs processing
72 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
73 std::vector
<Loop
*> LoopProcessWorklist
;
74 SmallPtrSet
<Value
*,8> UnswitchedVals
;
81 BasicBlock
*loopHeader
;
82 BasicBlock
*loopPreheader
;
84 // LoopBlocks contains all of the basic blocks of the loop, including the
85 // preheader of the loop, the body of the loop, and the exit blocks of the
86 // loop, in that order.
87 std::vector
<BasicBlock
*> LoopBlocks
;
88 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
89 std::vector
<BasicBlock
*> NewBlocks
;
92 static char ID
; // Pass ID, replacement for typeid
93 explicit LoopUnswitch(bool Os
= false) :
94 LoopPass(ID
), OptimizeForSize(Os
), redoLoop(false),
95 currentLoop(NULL
), DT(NULL
), loopHeader(NULL
),
97 initializeLoopUnswitchPass(*PassRegistry::getPassRegistry());
100 bool runOnLoop(Loop
*L
, LPPassManager
&LPM
);
101 bool processCurrentLoop();
103 /// This transformation requires natural loop information & requires that
104 /// loop preheaders be inserted into the CFG.
106 virtual void getAnalysisUsage(AnalysisUsage
&AU
) const {
107 AU
.addRequiredID(LoopSimplifyID
);
108 AU
.addPreservedID(LoopSimplifyID
);
109 AU
.addRequired
<LoopInfo
>();
110 AU
.addPreserved
<LoopInfo
>();
111 AU
.addRequiredID(LCSSAID
);
112 AU
.addPreservedID(LCSSAID
);
113 AU
.addPreserved
<DominatorTree
>();
114 AU
.addPreserved
<ScalarEvolution
>();
119 virtual void releaseMemory() {
120 UnswitchedVals
.clear();
123 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
125 void RemoveLoopFromWorklist(Loop
*L
) {
126 std::vector
<Loop
*>::iterator I
= std::find(LoopProcessWorklist
.begin(),
127 LoopProcessWorklist
.end(), L
);
128 if (I
!= LoopProcessWorklist
.end())
129 LoopProcessWorklist
.erase(I
);
132 void initLoopData() {
133 loopHeader
= currentLoop
->getHeader();
134 loopPreheader
= currentLoop
->getLoopPreheader();
137 /// Split all of the edges from inside the loop to their exit blocks.
138 /// Update the appropriate Phi nodes as we do so.
139 void SplitExitEdges(Loop
*L
, const SmallVector
<BasicBlock
*, 8> &ExitBlocks
);
141 bool UnswitchIfProfitable(Value
*LoopCond
, Constant
*Val
);
142 void UnswitchTrivialCondition(Loop
*L
, Value
*Cond
, Constant
*Val
,
143 BasicBlock
*ExitBlock
);
144 void UnswitchNontrivialCondition(Value
*LIC
, Constant
*OnVal
, Loop
*L
);
146 void RewriteLoopBodyWithConditionConstant(Loop
*L
, Value
*LIC
,
147 Constant
*Val
, bool isEqual
);
149 void EmitPreheaderBranchOnCondition(Value
*LIC
, Constant
*Val
,
150 BasicBlock
*TrueDest
,
151 BasicBlock
*FalseDest
,
152 Instruction
*InsertPt
);
154 void SimplifyCode(std::vector
<Instruction
*> &Worklist
, Loop
*L
);
155 void RemoveBlockIfDead(BasicBlock
*BB
,
156 std::vector
<Instruction
*> &Worklist
, Loop
*l
);
157 void RemoveLoopFromHierarchy(Loop
*L
);
158 bool IsTrivialUnswitchCondition(Value
*Cond
, Constant
**Val
= 0,
159 BasicBlock
**LoopExit
= 0);
163 char LoopUnswitch::ID
= 0;
164 INITIALIZE_PASS_BEGIN(LoopUnswitch
, "loop-unswitch", "Unswitch loops",
166 INITIALIZE_PASS_DEPENDENCY(LoopSimplify
)
167 INITIALIZE_PASS_DEPENDENCY(LoopInfo
)
168 INITIALIZE_PASS_DEPENDENCY(LCSSA
)
169 INITIALIZE_PASS_END(LoopUnswitch
, "loop-unswitch", "Unswitch loops",
172 Pass
*llvm::createLoopUnswitchPass(bool Os
) {
173 return new LoopUnswitch(Os
);
176 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
177 /// invariant in the loop, or has an invariant piece, return the invariant.
178 /// Otherwise, return null.
179 static Value
*FindLIVLoopCondition(Value
*Cond
, Loop
*L
, bool &Changed
) {
180 // We can never unswitch on vector conditions.
181 if (Cond
->getType()->isVectorTy())
184 // Constants should be folded, not unswitched on!
185 if (isa
<Constant
>(Cond
)) return 0;
187 // TODO: Handle: br (VARIANT|INVARIANT).
189 // Hoist simple values out.
190 if (L
->makeLoopInvariant(Cond
, Changed
))
193 if (BinaryOperator
*BO
= dyn_cast
<BinaryOperator
>(Cond
))
194 if (BO
->getOpcode() == Instruction::And
||
195 BO
->getOpcode() == Instruction::Or
) {
196 // If either the left or right side is invariant, we can unswitch on this,
197 // which will cause the branch to go away in one loop and the condition to
198 // simplify in the other one.
199 if (Value
*LHS
= FindLIVLoopCondition(BO
->getOperand(0), L
, Changed
))
201 if (Value
*RHS
= FindLIVLoopCondition(BO
->getOperand(1), L
, Changed
))
208 bool LoopUnswitch::runOnLoop(Loop
*L
, LPPassManager
&LPM_Ref
) {
209 LI
= &getAnalysis
<LoopInfo
>();
211 DT
= getAnalysisIfAvailable
<DominatorTree
>();
213 Function
*F
= currentLoop
->getHeader()->getParent();
214 bool Changed
= false;
216 assert(currentLoop
->isLCSSAForm(*DT
));
218 Changed
|= processCurrentLoop();
222 // FIXME: Reconstruct dom info, because it is not preserved properly.
224 DT
->runOnFunction(*F
);
229 /// processCurrentLoop - Do actual work and unswitch loop if possible
231 bool LoopUnswitch::processCurrentLoop() {
232 bool Changed
= false;
233 LLVMContext
&Context
= currentLoop
->getHeader()->getContext();
235 // Loop over all of the basic blocks in the loop. If we find an interior
236 // block that is branching on a loop-invariant condition, we can unswitch this
238 for (Loop::block_iterator I
= currentLoop
->block_begin(),
239 E
= currentLoop
->block_end(); I
!= E
; ++I
) {
240 TerminatorInst
*TI
= (*I
)->getTerminator();
241 if (BranchInst
*BI
= dyn_cast
<BranchInst
>(TI
)) {
242 // If this isn't branching on an invariant condition, we can't unswitch
244 if (BI
->isConditional()) {
245 // See if this, or some part of it, is loop invariant. If so, we can
246 // unswitch on it if we desire.
247 Value
*LoopCond
= FindLIVLoopCondition(BI
->getCondition(),
248 currentLoop
, Changed
);
249 if (LoopCond
&& UnswitchIfProfitable(LoopCond
,
250 ConstantInt::getTrue(Context
))) {
255 } else if (SwitchInst
*SI
= dyn_cast
<SwitchInst
>(TI
)) {
256 Value
*LoopCond
= FindLIVLoopCondition(SI
->getCondition(),
257 currentLoop
, Changed
);
258 if (LoopCond
&& SI
->getNumCases() > 1) {
259 // Find a value to unswitch on:
260 // FIXME: this should chose the most expensive case!
261 Constant
*UnswitchVal
= SI
->getCaseValue(1);
262 // Do not process same value again and again.
263 if (!UnswitchedVals
.insert(UnswitchVal
))
266 if (UnswitchIfProfitable(LoopCond
, UnswitchVal
)) {
273 // Scan the instructions to check for unswitchable values.
274 for (BasicBlock::iterator BBI
= (*I
)->begin(), E
= (*I
)->end();
276 if (SelectInst
*SI
= dyn_cast
<SelectInst
>(BBI
)) {
277 Value
*LoopCond
= FindLIVLoopCondition(SI
->getCondition(),
278 currentLoop
, Changed
);
279 if (LoopCond
&& UnswitchIfProfitable(LoopCond
,
280 ConstantInt::getTrue(Context
))) {
289 /// isTrivialLoopExitBlock - Check to see if all paths from BB exit the
290 /// loop with no side effects (including infinite loops).
292 /// If true, we return true and set ExitBB to the block we
295 static bool isTrivialLoopExitBlockHelper(Loop
*L
, BasicBlock
*BB
,
297 std::set
<BasicBlock
*> &Visited
) {
298 if (!Visited
.insert(BB
).second
) {
299 // Already visited. Without more analysis, this could indicate an infinte loop.
301 } else if (!L
->contains(BB
)) {
302 // Otherwise, this is a loop exit, this is fine so long as this is the
304 if (ExitBB
!= 0) return false;
309 // Otherwise, this is an unvisited intra-loop node. Check all successors.
310 for (succ_iterator SI
= succ_begin(BB
), E
= succ_end(BB
); SI
!= E
; ++SI
) {
311 // Check to see if the successor is a trivial loop exit.
312 if (!isTrivialLoopExitBlockHelper(L
, *SI
, ExitBB
, Visited
))
316 // Okay, everything after this looks good, check to make sure that this block
317 // doesn't include any side effects.
318 for (BasicBlock::iterator I
= BB
->begin(), E
= BB
->end(); I
!= E
; ++I
)
319 if (I
->mayHaveSideEffects())
325 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
326 /// leads to an exit from the specified loop, and has no side-effects in the
327 /// process. If so, return the block that is exited to, otherwise return null.
328 static BasicBlock
*isTrivialLoopExitBlock(Loop
*L
, BasicBlock
*BB
) {
329 std::set
<BasicBlock
*> Visited
;
330 Visited
.insert(L
->getHeader()); // Branches to header make infinite loops.
331 BasicBlock
*ExitBB
= 0;
332 if (isTrivialLoopExitBlockHelper(L
, BB
, ExitBB
, Visited
))
337 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
338 /// trivial: that is, that the condition controls whether or not the loop does
339 /// anything at all. If this is a trivial condition, unswitching produces no
340 /// code duplications (equivalently, it produces a simpler loop and a new empty
341 /// loop, which gets deleted).
343 /// If this is a trivial condition, return true, otherwise return false. When
344 /// returning true, this sets Cond and Val to the condition that controls the
345 /// trivial condition: when Cond dynamically equals Val, the loop is known to
346 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
349 bool LoopUnswitch::IsTrivialUnswitchCondition(Value
*Cond
, Constant
**Val
,
350 BasicBlock
**LoopExit
) {
351 BasicBlock
*Header
= currentLoop
->getHeader();
352 TerminatorInst
*HeaderTerm
= Header
->getTerminator();
353 LLVMContext
&Context
= Header
->getContext();
355 BasicBlock
*LoopExitBB
= 0;
356 if (BranchInst
*BI
= dyn_cast
<BranchInst
>(HeaderTerm
)) {
357 // If the header block doesn't end with a conditional branch on Cond, we
359 if (!BI
->isConditional() || BI
->getCondition() != Cond
)
362 // Check to see if a successor of the branch is guaranteed to
363 // exit through a unique exit block without having any
364 // side-effects. If so, determine the value of Cond that causes it to do
366 if ((LoopExitBB
= isTrivialLoopExitBlock(currentLoop
,
367 BI
->getSuccessor(0)))) {
368 if (Val
) *Val
= ConstantInt::getTrue(Context
);
369 } else if ((LoopExitBB
= isTrivialLoopExitBlock(currentLoop
,
370 BI
->getSuccessor(1)))) {
371 if (Val
) *Val
= ConstantInt::getFalse(Context
);
373 } else if (SwitchInst
*SI
= dyn_cast
<SwitchInst
>(HeaderTerm
)) {
374 // If this isn't a switch on Cond, we can't handle it.
375 if (SI
->getCondition() != Cond
) return false;
377 // Check to see if a successor of the switch is guaranteed to go to the
378 // latch block or exit through a one exit block without having any
379 // side-effects. If so, determine the value of Cond that causes it to do
380 // this. Note that we can't trivially unswitch on the default case.
381 for (unsigned i
= 1, e
= SI
->getNumSuccessors(); i
!= e
; ++i
)
382 if ((LoopExitBB
= isTrivialLoopExitBlock(currentLoop
,
383 SI
->getSuccessor(i
)))) {
384 // Okay, we found a trivial case, remember the value that is trivial.
385 if (Val
) *Val
= SI
->getCaseValue(i
);
390 // If we didn't find a single unique LoopExit block, or if the loop exit block
391 // contains phi nodes, this isn't trivial.
392 if (!LoopExitBB
|| isa
<PHINode
>(LoopExitBB
->begin()))
393 return false; // Can't handle this.
395 if (LoopExit
) *LoopExit
= LoopExitBB
;
397 // We already know that nothing uses any scalar values defined inside of this
398 // loop. As such, we just have to check to see if this loop will execute any
399 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
400 // part of the loop that the code *would* execute. We already checked the
401 // tail, check the header now.
402 for (BasicBlock::iterator I
= Header
->begin(), E
= Header
->end(); I
!= E
; ++I
)
403 if (I
->mayHaveSideEffects())
408 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
409 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
410 /// unswitch the loop, reprocess the pieces, then return true.
411 bool LoopUnswitch::UnswitchIfProfitable(Value
*LoopCond
, Constant
*Val
) {
415 // If LoopSimplify was unable to form a preheader, don't do any unswitching.
419 Function
*F
= loopHeader
->getParent();
421 Constant
*CondVal
= 0;
422 BasicBlock
*ExitBlock
= 0;
423 if (IsTrivialUnswitchCondition(LoopCond
, &CondVal
, &ExitBlock
)) {
424 // If the condition is trivial, always unswitch. There is no code growth
426 UnswitchTrivialCondition(currentLoop
, LoopCond
, CondVal
, ExitBlock
);
430 // Check to see if it would be profitable to unswitch current loop.
432 // Do not do non-trivial unswitch while optimizing for size.
433 if (OptimizeForSize
|| F
->hasFnAttr(Attribute::OptimizeForSize
))
436 // FIXME: This is overly conservative because it does not take into
437 // consideration code simplification opportunities and code that can
438 // be shared by the resultant unswitched loops.
440 for (Loop::block_iterator I
= currentLoop
->block_begin(),
441 E
= currentLoop
->block_end();
443 Metrics
.analyzeBasicBlock(*I
);
445 // Limit the number of instructions to avoid causing significant code
446 // expansion, and the number of basic blocks, to avoid loops with
447 // large numbers of branches which cause loop unswitching to go crazy.
448 // This is a very ad-hoc heuristic.
449 if (Metrics
.NumInsts
> Threshold
||
450 Metrics
.NumBlocks
* 5 > Threshold
||
451 Metrics
.containsIndirectBr
|| Metrics
.isRecursive
) {
452 DEBUG(dbgs() << "NOT unswitching loop %"
453 << currentLoop
->getHeader()->getName() << ", cost too high: "
454 << currentLoop
->getBlocks().size() << "\n");
458 UnswitchNontrivialCondition(LoopCond
, Val
, currentLoop
);
462 /// CloneLoop - Recursively clone the specified loop and all of its children,
463 /// mapping the blocks with the specified map.
464 static Loop
*CloneLoop(Loop
*L
, Loop
*PL
, ValueToValueMapTy
&VM
,
465 LoopInfo
*LI
, LPPassManager
*LPM
) {
466 Loop
*New
= new Loop();
467 LPM
->insertLoop(New
, PL
);
469 // Add all of the blocks in L to the new loop.
470 for (Loop::block_iterator I
= L
->block_begin(), E
= L
->block_end();
472 if (LI
->getLoopFor(*I
) == L
)
473 New
->addBasicBlockToLoop(cast
<BasicBlock
>(VM
[*I
]), LI
->getBase());
475 // Add all of the subloops to the new loop.
476 for (Loop::iterator I
= L
->begin(), E
= L
->end(); I
!= E
; ++I
)
477 CloneLoop(*I
, New
, VM
, LI
, LPM
);
482 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
483 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
484 /// code immediately before InsertPt.
485 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value
*LIC
, Constant
*Val
,
486 BasicBlock
*TrueDest
,
487 BasicBlock
*FalseDest
,
488 Instruction
*InsertPt
) {
489 // Insert a conditional branch on LIC to the two preheaders. The original
490 // code is the true version and the new code is the false version.
491 Value
*BranchVal
= LIC
;
492 if (!isa
<ConstantInt
>(Val
) ||
493 Val
->getType() != Type::getInt1Ty(LIC
->getContext()))
494 BranchVal
= new ICmpInst(InsertPt
, ICmpInst::ICMP_EQ
, LIC
, Val
, "tmp");
495 else if (Val
!= ConstantInt::getTrue(Val
->getContext()))
496 // We want to enter the new loop when the condition is true.
497 std::swap(TrueDest
, FalseDest
);
499 // Insert the new branch.
500 BranchInst
*BI
= BranchInst::Create(TrueDest
, FalseDest
, BranchVal
, InsertPt
);
502 // If either edge is critical, split it. This helps preserve LoopSimplify
503 // form for enclosing loops.
504 SplitCriticalEdge(BI
, 0, this);
505 SplitCriticalEdge(BI
, 1, this);
508 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
509 /// condition in it (a cond branch from its header block to its latch block,
510 /// where the path through the loop that doesn't execute its body has no
511 /// side-effects), unswitch it. This doesn't involve any code duplication, just
512 /// moving the conditional branch outside of the loop and updating loop info.
513 void LoopUnswitch::UnswitchTrivialCondition(Loop
*L
, Value
*Cond
,
515 BasicBlock
*ExitBlock
) {
516 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %"
517 << loopHeader
->getName() << " [" << L
->getBlocks().size()
518 << " blocks] in Function " << L
->getHeader()->getParent()->getName()
519 << " on cond: " << *Val
<< " == " << *Cond
<< "\n");
521 // First step, split the preheader, so that we know that there is a safe place
522 // to insert the conditional branch. We will change loopPreheader to have a
523 // conditional branch on Cond.
524 BasicBlock
*NewPH
= SplitEdge(loopPreheader
, loopHeader
, this);
526 // Now that we have a place to insert the conditional branch, create a place
527 // to branch to: this is the exit block out of the loop that we should
530 // Split this block now, so that the loop maintains its exit block, and so
531 // that the jump from the preheader can execute the contents of the exit block
532 // without actually branching to it (the exit block should be dominated by the
533 // loop header, not the preheader).
534 assert(!L
->contains(ExitBlock
) && "Exit block is in the loop?");
535 BasicBlock
*NewExit
= SplitBlock(ExitBlock
, ExitBlock
->begin(), this);
537 // Okay, now we have a position to branch from and a position to branch to,
538 // insert the new conditional branch.
539 EmitPreheaderBranchOnCondition(Cond
, Val
, NewExit
, NewPH
,
540 loopPreheader
->getTerminator());
541 LPM
->deleteSimpleAnalysisValue(loopPreheader
->getTerminator(), L
);
542 loopPreheader
->getTerminator()->eraseFromParent();
544 // We need to reprocess this loop, it could be unswitched again.
547 // Now that we know that the loop is never entered when this condition is a
548 // particular value, rewrite the loop with this info. We know that this will
549 // at least eliminate the old branch.
550 RewriteLoopBodyWithConditionConstant(L
, Cond
, Val
, false);
554 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
555 /// blocks. Update the appropriate Phi nodes as we do so.
556 void LoopUnswitch::SplitExitEdges(Loop
*L
,
557 const SmallVector
<BasicBlock
*, 8> &ExitBlocks
){
559 for (unsigned i
= 0, e
= ExitBlocks
.size(); i
!= e
; ++i
) {
560 BasicBlock
*ExitBlock
= ExitBlocks
[i
];
561 SmallVector
<BasicBlock
*, 4> Preds(pred_begin(ExitBlock
),
562 pred_end(ExitBlock
));
563 SplitBlockPredecessors(ExitBlock
, Preds
.data(), Preds
.size(),
568 /// UnswitchNontrivialCondition - We determined that the loop is profitable
569 /// to unswitch when LIC equal Val. Split it into loop versions and test the
570 /// condition outside of either loop. Return the loops created as Out1/Out2.
571 void LoopUnswitch::UnswitchNontrivialCondition(Value
*LIC
, Constant
*Val
,
573 Function
*F
= loopHeader
->getParent();
574 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
575 << loopHeader
->getName() << " [" << L
->getBlocks().size()
576 << " blocks] in Function " << F
->getName()
577 << " when '" << *Val
<< "' == " << *LIC
<< "\n");
579 if (ScalarEvolution
*SE
= getAnalysisIfAvailable
<ScalarEvolution
>())
585 // First step, split the preheader and exit blocks, and add these blocks to
586 // the LoopBlocks list.
587 BasicBlock
*NewPreheader
= SplitEdge(loopPreheader
, loopHeader
, this);
588 LoopBlocks
.push_back(NewPreheader
);
590 // We want the loop to come after the preheader, but before the exit blocks.
591 LoopBlocks
.insert(LoopBlocks
.end(), L
->block_begin(), L
->block_end());
593 SmallVector
<BasicBlock
*, 8> ExitBlocks
;
594 L
->getUniqueExitBlocks(ExitBlocks
);
596 // Split all of the edges from inside the loop to their exit blocks. Update
597 // the appropriate Phi nodes as we do so.
598 SplitExitEdges(L
, ExitBlocks
);
600 // The exit blocks may have been changed due to edge splitting, recompute.
602 L
->getUniqueExitBlocks(ExitBlocks
);
604 // Add exit blocks to the loop blocks.
605 LoopBlocks
.insert(LoopBlocks
.end(), ExitBlocks
.begin(), ExitBlocks
.end());
607 // Next step, clone all of the basic blocks that make up the loop (including
608 // the loop preheader and exit blocks), keeping track of the mapping between
609 // the instructions and blocks.
610 NewBlocks
.reserve(LoopBlocks
.size());
611 ValueToValueMapTy VMap
;
612 for (unsigned i
= 0, e
= LoopBlocks
.size(); i
!= e
; ++i
) {
613 BasicBlock
*NewBB
= CloneBasicBlock(LoopBlocks
[i
], VMap
, ".us", F
);
614 NewBlocks
.push_back(NewBB
);
615 VMap
[LoopBlocks
[i
]] = NewBB
; // Keep the BB mapping.
616 LPM
->cloneBasicBlockSimpleAnalysis(LoopBlocks
[i
], NewBB
, L
);
619 // Splice the newly inserted blocks into the function right before the
620 // original preheader.
621 F
->getBasicBlockList().splice(NewPreheader
, F
->getBasicBlockList(),
622 NewBlocks
[0], F
->end());
624 // Now we create the new Loop object for the versioned loop.
625 Loop
*NewLoop
= CloneLoop(L
, L
->getParentLoop(), VMap
, LI
, LPM
);
626 Loop
*ParentLoop
= L
->getParentLoop();
628 // Make sure to add the cloned preheader and exit blocks to the parent loop
630 ParentLoop
->addBasicBlockToLoop(NewBlocks
[0], LI
->getBase());
633 for (unsigned i
= 0, e
= ExitBlocks
.size(); i
!= e
; ++i
) {
634 BasicBlock
*NewExit
= cast
<BasicBlock
>(VMap
[ExitBlocks
[i
]]);
635 // The new exit block should be in the same loop as the old one.
636 if (Loop
*ExitBBLoop
= LI
->getLoopFor(ExitBlocks
[i
]))
637 ExitBBLoop
->addBasicBlockToLoop(NewExit
, LI
->getBase());
639 assert(NewExit
->getTerminator()->getNumSuccessors() == 1 &&
640 "Exit block should have been split to have one successor!");
641 BasicBlock
*ExitSucc
= NewExit
->getTerminator()->getSuccessor(0);
643 // If the successor of the exit block had PHI nodes, add an entry for
646 for (BasicBlock::iterator I
= ExitSucc
->begin(); isa
<PHINode
>(I
); ++I
) {
647 PN
= cast
<PHINode
>(I
);
648 Value
*V
= PN
->getIncomingValueForBlock(ExitBlocks
[i
]);
649 ValueToValueMapTy::iterator It
= VMap
.find(V
);
650 if (It
!= VMap
.end()) V
= It
->second
;
651 PN
->addIncoming(V
, NewExit
);
655 // Rewrite the code to refer to itself.
656 for (unsigned i
= 0, e
= NewBlocks
.size(); i
!= e
; ++i
)
657 for (BasicBlock::iterator I
= NewBlocks
[i
]->begin(),
658 E
= NewBlocks
[i
]->end(); I
!= E
; ++I
)
659 RemapInstruction(I
, VMap
,RF_NoModuleLevelChanges
|RF_IgnoreMissingEntries
);
661 // Rewrite the original preheader to select between versions of the loop.
662 BranchInst
*OldBR
= cast
<BranchInst
>(loopPreheader
->getTerminator());
663 assert(OldBR
->isUnconditional() && OldBR
->getSuccessor(0) == LoopBlocks
[0] &&
664 "Preheader splitting did not work correctly!");
666 // Emit the new branch that selects between the two versions of this loop.
667 EmitPreheaderBranchOnCondition(LIC
, Val
, NewBlocks
[0], LoopBlocks
[0], OldBR
);
668 LPM
->deleteSimpleAnalysisValue(OldBR
, L
);
669 OldBR
->eraseFromParent();
671 LoopProcessWorklist
.push_back(NewLoop
);
674 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody
675 // deletes the instruction (for example by simplifying a PHI that feeds into
676 // the condition that we're unswitching on), we don't rewrite the second
678 WeakVH
LICHandle(LIC
);
680 // Now we rewrite the original code to know that the condition is true and the
681 // new code to know that the condition is false.
682 RewriteLoopBodyWithConditionConstant(L
, LIC
, Val
, false);
684 // It's possible that simplifying one loop could cause the other to be
685 // changed to another value or a constant. If its a constant, don't simplify
687 if (!LoopProcessWorklist
.empty() && LoopProcessWorklist
.back() == NewLoop
&&
688 LICHandle
&& !isa
<Constant
>(LICHandle
))
689 RewriteLoopBodyWithConditionConstant(NewLoop
, LICHandle
, Val
, true);
692 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
694 static void RemoveFromWorklist(Instruction
*I
,
695 std::vector
<Instruction
*> &Worklist
) {
696 std::vector
<Instruction
*>::iterator WI
= std::find(Worklist
.begin(),
698 while (WI
!= Worklist
.end()) {
699 unsigned Offset
= WI
-Worklist
.begin();
701 WI
= std::find(Worklist
.begin()+Offset
, Worklist
.end(), I
);
705 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
706 /// program, replacing all uses with V and update the worklist.
707 static void ReplaceUsesOfWith(Instruction
*I
, Value
*V
,
708 std::vector
<Instruction
*> &Worklist
,
709 Loop
*L
, LPPassManager
*LPM
) {
710 DEBUG(dbgs() << "Replace with '" << *V
<< "': " << *I
);
712 // Add uses to the worklist, which may be dead now.
713 for (unsigned i
= 0, e
= I
->getNumOperands(); i
!= e
; ++i
)
714 if (Instruction
*Use
= dyn_cast
<Instruction
>(I
->getOperand(i
)))
715 Worklist
.push_back(Use
);
717 // Add users to the worklist which may be simplified now.
718 for (Value::use_iterator UI
= I
->use_begin(), E
= I
->use_end();
720 Worklist
.push_back(cast
<Instruction
>(*UI
));
721 LPM
->deleteSimpleAnalysisValue(I
, L
);
722 RemoveFromWorklist(I
, Worklist
);
723 I
->replaceAllUsesWith(V
);
724 I
->eraseFromParent();
728 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
729 /// information, and remove any dead successors it has.
731 void LoopUnswitch::RemoveBlockIfDead(BasicBlock
*BB
,
732 std::vector
<Instruction
*> &Worklist
,
734 if (pred_begin(BB
) != pred_end(BB
)) {
735 // This block isn't dead, since an edge to BB was just removed, see if there
736 // are any easy simplifications we can do now.
737 if (BasicBlock
*Pred
= BB
->getSinglePredecessor()) {
738 // If it has one pred, fold phi nodes in BB.
739 while (isa
<PHINode
>(BB
->begin()))
740 ReplaceUsesOfWith(BB
->begin(),
741 cast
<PHINode
>(BB
->begin())->getIncomingValue(0),
744 // If this is the header of a loop and the only pred is the latch, we now
745 // have an unreachable loop.
746 if (Loop
*L
= LI
->getLoopFor(BB
))
747 if (loopHeader
== BB
&& L
->contains(Pred
)) {
748 // Remove the branch from the latch to the header block, this makes
749 // the header dead, which will make the latch dead (because the header
750 // dominates the latch).
751 LPM
->deleteSimpleAnalysisValue(Pred
->getTerminator(), L
);
752 Pred
->getTerminator()->eraseFromParent();
753 new UnreachableInst(BB
->getContext(), Pred
);
755 // The loop is now broken, remove it from LI.
756 RemoveLoopFromHierarchy(L
);
758 // Reprocess the header, which now IS dead.
759 RemoveBlockIfDead(BB
, Worklist
, L
);
763 // If pred ends in a uncond branch, add uncond branch to worklist so that
764 // the two blocks will get merged.
765 if (BranchInst
*BI
= dyn_cast
<BranchInst
>(Pred
->getTerminator()))
766 if (BI
->isUnconditional())
767 Worklist
.push_back(BI
);
772 DEBUG(dbgs() << "Nuking dead block: " << *BB
);
774 // Remove the instructions in the basic block from the worklist.
775 for (BasicBlock::iterator I
= BB
->begin(), E
= BB
->end(); I
!= E
; ++I
) {
776 RemoveFromWorklist(I
, Worklist
);
778 // Anything that uses the instructions in this basic block should have their
779 // uses replaced with undefs.
780 // If I is not void type then replaceAllUsesWith undef.
781 // This allows ValueHandlers and custom metadata to adjust itself.
782 if (!I
->getType()->isVoidTy())
783 I
->replaceAllUsesWith(UndefValue::get(I
->getType()));
786 // If this is the edge to the header block for a loop, remove the loop and
787 // promote all subloops.
788 if (Loop
*BBLoop
= LI
->getLoopFor(BB
)) {
789 if (BBLoop
->getLoopLatch() == BB
)
790 RemoveLoopFromHierarchy(BBLoop
);
793 // Remove the block from the loop info, which removes it from any loops it
798 // Remove phi node entries in successors for this block.
799 TerminatorInst
*TI
= BB
->getTerminator();
800 SmallVector
<BasicBlock
*, 4> Succs
;
801 for (unsigned i
= 0, e
= TI
->getNumSuccessors(); i
!= e
; ++i
) {
802 Succs
.push_back(TI
->getSuccessor(i
));
803 TI
->getSuccessor(i
)->removePredecessor(BB
);
806 // Unique the successors, remove anything with multiple uses.
807 array_pod_sort(Succs
.begin(), Succs
.end());
808 Succs
.erase(std::unique(Succs
.begin(), Succs
.end()), Succs
.end());
810 // Remove the basic block, including all of the instructions contained in it.
811 LPM
->deleteSimpleAnalysisValue(BB
, L
);
812 BB
->eraseFromParent();
813 // Remove successor blocks here that are not dead, so that we know we only
814 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
815 // then getting removed before we revisit them, which is badness.
817 for (unsigned i
= 0; i
!= Succs
.size(); ++i
)
818 if (pred_begin(Succs
[i
]) != pred_end(Succs
[i
])) {
819 // One exception is loop headers. If this block was the preheader for a
820 // loop, then we DO want to visit the loop so the loop gets deleted.
821 // We know that if the successor is a loop header, that this loop had to
822 // be the preheader: the case where this was the latch block was handled
823 // above and headers can only have two predecessors.
824 if (!LI
->isLoopHeader(Succs
[i
])) {
825 Succs
.erase(Succs
.begin()+i
);
830 for (unsigned i
= 0, e
= Succs
.size(); i
!= e
; ++i
)
831 RemoveBlockIfDead(Succs
[i
], Worklist
, L
);
834 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
835 /// become unwrapped, either because the backedge was deleted, or because the
836 /// edge into the header was removed. If the edge into the header from the
837 /// latch block was removed, the loop is unwrapped but subloops are still alive,
838 /// so they just reparent loops. If the loops are actually dead, they will be
840 void LoopUnswitch::RemoveLoopFromHierarchy(Loop
*L
) {
841 LPM
->deleteLoopFromQueue(L
);
842 RemoveLoopFromWorklist(L
);
845 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
846 // the value specified by Val in the specified loop, or we know it does NOT have
847 // that value. Rewrite any uses of LIC or of properties correlated to it.
848 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop
*L
, Value
*LIC
,
851 assert(!isa
<Constant
>(LIC
) && "Why are we unswitching on a constant?");
853 // FIXME: Support correlated properties, like:
860 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
861 // selects, switches.
862 std::vector
<User
*> Users(LIC
->use_begin(), LIC
->use_end());
863 std::vector
<Instruction
*> Worklist
;
864 LLVMContext
&Context
= Val
->getContext();
867 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
868 // in the loop with the appropriate one directly.
869 if (IsEqual
|| (isa
<ConstantInt
>(Val
) &&
870 Val
->getType()->isIntegerTy(1))) {
875 Replacement
= ConstantInt::get(Type::getInt1Ty(Val
->getContext()),
876 !cast
<ConstantInt
>(Val
)->getZExtValue());
878 for (unsigned i
= 0, e
= Users
.size(); i
!= e
; ++i
)
879 if (Instruction
*U
= cast
<Instruction
>(Users
[i
])) {
882 U
->replaceUsesOfWith(LIC
, Replacement
);
883 Worklist
.push_back(U
);
885 SimplifyCode(Worklist
, L
);
889 // Otherwise, we don't know the precise value of LIC, but we do know that it
890 // is certainly NOT "Val". As such, simplify any uses in the loop that we
891 // can. This case occurs when we unswitch switch statements.
892 for (unsigned i
= 0, e
= Users
.size(); i
!= e
; ++i
) {
893 Instruction
*U
= cast
<Instruction
>(Users
[i
]);
897 Worklist
.push_back(U
);
899 // TODO: We could do other simplifications, for example, turning
900 // 'icmp eq LIC, Val' -> false.
902 // If we know that LIC is not Val, use this info to simplify code.
903 SwitchInst
*SI
= dyn_cast
<SwitchInst
>(U
);
904 if (SI
== 0 || !isa
<ConstantInt
>(Val
)) continue;
906 unsigned DeadCase
= SI
->findCaseValue(cast
<ConstantInt
>(Val
));
907 if (DeadCase
== 0) continue; // Default case is live for multiple values.
909 // Found a dead case value. Don't remove PHI nodes in the
910 // successor if they become single-entry, those PHI nodes may
911 // be in the Users list.
913 // FIXME: This is a hack. We need to keep the successor around
914 // and hooked up so as to preserve the loop structure, because
915 // trying to update it is complicated. So instead we preserve the
916 // loop structure and put the block on a dead code path.
917 BasicBlock
*Switch
= SI
->getParent();
918 SplitEdge(Switch
, SI
->getSuccessor(DeadCase
), this);
919 // Compute the successors instead of relying on the return value
920 // of SplitEdge, since it may have split the switch successor
922 BasicBlock
*NewSISucc
= SI
->getSuccessor(DeadCase
);
923 BasicBlock
*OldSISucc
= *succ_begin(NewSISucc
);
924 // Create an "unreachable" destination.
925 BasicBlock
*Abort
= BasicBlock::Create(Context
, "us-unreachable",
928 new UnreachableInst(Context
, Abort
);
929 // Force the new case destination to branch to the "unreachable"
930 // block while maintaining a (dead) CFG edge to the old block.
931 NewSISucc
->getTerminator()->eraseFromParent();
932 BranchInst::Create(Abort
, OldSISucc
,
933 ConstantInt::getTrue(Context
), NewSISucc
);
934 // Release the PHI operands for this edge.
935 for (BasicBlock::iterator II
= NewSISucc
->begin();
936 PHINode
*PN
= dyn_cast
<PHINode
>(II
); ++II
)
937 PN
->setIncomingValue(PN
->getBasicBlockIndex(Switch
),
938 UndefValue::get(PN
->getType()));
939 // Tell the domtree about the new block. We don't fully update the
940 // domtree here -- instead we force it to do a full recomputation
941 // after the pass is complete -- but we do need to inform it of
944 DT
->addNewBlock(Abort
, NewSISucc
);
947 SimplifyCode(Worklist
, L
);
950 /// SimplifyCode - Okay, now that we have simplified some instructions in the
951 /// loop, walk over it and constant prop, dce, and fold control flow where
952 /// possible. Note that this is effectively a very simple loop-structure-aware
953 /// optimizer. During processing of this loop, L could very well be deleted, so
954 /// it must not be used.
956 /// FIXME: When the loop optimizer is more mature, separate this out to a new
959 void LoopUnswitch::SimplifyCode(std::vector
<Instruction
*> &Worklist
, Loop
*L
) {
960 while (!Worklist
.empty()) {
961 Instruction
*I
= Worklist
.back();
965 if (isInstructionTriviallyDead(I
)) {
966 DEBUG(dbgs() << "Remove dead instruction '" << *I
);
968 // Add uses to the worklist, which may be dead now.
969 for (unsigned i
= 0, e
= I
->getNumOperands(); i
!= e
; ++i
)
970 if (Instruction
*Use
= dyn_cast
<Instruction
>(I
->getOperand(i
)))
971 Worklist
.push_back(Use
);
972 LPM
->deleteSimpleAnalysisValue(I
, L
);
973 RemoveFromWorklist(I
, Worklist
);
974 I
->eraseFromParent();
979 // See if instruction simplification can hack this up. This is common for
980 // things like "select false, X, Y" after unswitching made the condition be
982 if (Value
*V
= SimplifyInstruction(I
, 0, DT
))
983 if (LI
->replacementPreservesLCSSAForm(I
, V
)) {
984 ReplaceUsesOfWith(I
, V
, Worklist
, L
, LPM
);
988 // Special case hacks that appear commonly in unswitched code.
989 if (BranchInst
*BI
= dyn_cast
<BranchInst
>(I
)) {
990 if (BI
->isUnconditional()) {
991 // If BI's parent is the only pred of the successor, fold the two blocks
993 BasicBlock
*Pred
= BI
->getParent();
994 BasicBlock
*Succ
= BI
->getSuccessor(0);
995 BasicBlock
*SinglePred
= Succ
->getSinglePredecessor();
996 if (!SinglePred
) continue; // Nothing to do.
997 assert(SinglePred
== Pred
&& "CFG broken");
999 DEBUG(dbgs() << "Merging blocks: " << Pred
->getName() << " <- "
1000 << Succ
->getName() << "\n");
1002 // Resolve any single entry PHI nodes in Succ.
1003 while (PHINode
*PN
= dyn_cast
<PHINode
>(Succ
->begin()))
1004 ReplaceUsesOfWith(PN
, PN
->getIncomingValue(0), Worklist
, L
, LPM
);
1006 // Move all of the successor contents from Succ to Pred.
1007 Pred
->getInstList().splice(BI
, Succ
->getInstList(), Succ
->begin(),
1009 LPM
->deleteSimpleAnalysisValue(BI
, L
);
1010 BI
->eraseFromParent();
1011 RemoveFromWorklist(BI
, Worklist
);
1013 // If Succ has any successors with PHI nodes, update them to have
1014 // entries coming from Pred instead of Succ.
1015 Succ
->replaceAllUsesWith(Pred
);
1017 // Remove Succ from the loop tree.
1018 LI
->removeBlock(Succ
);
1019 LPM
->deleteSimpleAnalysisValue(Succ
, L
);
1020 Succ
->eraseFromParent();
1025 if (ConstantInt
*CB
= dyn_cast
<ConstantInt
>(BI
->getCondition())){
1026 // Conditional branch. Turn it into an unconditional branch, then
1027 // remove dead blocks.
1028 continue; // FIXME: Enable.
1030 DEBUG(dbgs() << "Folded branch: " << *BI
);
1031 BasicBlock
*DeadSucc
= BI
->getSuccessor(CB
->getZExtValue());
1032 BasicBlock
*LiveSucc
= BI
->getSuccessor(!CB
->getZExtValue());
1033 DeadSucc
->removePredecessor(BI
->getParent(), true);
1034 Worklist
.push_back(BranchInst::Create(LiveSucc
, BI
));
1035 LPM
->deleteSimpleAnalysisValue(BI
, L
);
1036 BI
->eraseFromParent();
1037 RemoveFromWorklist(BI
, Worklist
);
1040 RemoveBlockIfDead(DeadSucc
, Worklist
, L
);