1 //===-- UnrollLoop.cpp - Loop unrolling utilities -------------------------===//
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 some loop unrolling utilities. It does not define any
11 // actual pass or policy, but provides a single function to perform loop
14 // It works best when loops have been canonicalized by the -indvars pass,
15 // allowing it to determine the trip counts of loops easily.
17 // The process of unrolling can produce extraneous basic blocks linked with
18 // unconditional branches. This will be corrected in the future.
19 //===----------------------------------------------------------------------===//
21 #define DEBUG_TYPE "loop-unroll"
22 #include "llvm/Transforms/Utils/UnrollLoop.h"
23 #include "llvm/BasicBlock.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Analysis/InstructionSimplify.h"
26 #include "llvm/Analysis/LoopPass.h"
27 #include "llvm/Analysis/ScalarEvolution.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/raw_ostream.h"
30 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
31 #include "llvm/Transforms/Utils/Cloning.h"
32 #include "llvm/Transforms/Utils/Local.h"
36 // TODO: Should these be here or in LoopUnroll?
37 STATISTIC(NumCompletelyUnrolled
, "Number of loops completely unrolled");
38 STATISTIC(NumUnrolled
, "Number of loops unrolled (completely or otherwise)");
40 /// RemapInstruction - Convert the instruction operands from referencing the
41 /// current values into those specified by VMap.
42 static inline void RemapInstruction(Instruction
*I
,
43 ValueToValueMapTy
&VMap
) {
44 for (unsigned op
= 0, E
= I
->getNumOperands(); op
!= E
; ++op
) {
45 Value
*Op
= I
->getOperand(op
);
46 ValueToValueMapTy::iterator It
= VMap
.find(Op
);
48 I
->setOperand(op
, It
->second
);
52 /// FoldBlockIntoPredecessor - Folds a basic block into its predecessor if it
53 /// only has one predecessor, and that predecessor only has one successor.
54 /// The LoopInfo Analysis that is passed will be kept consistent.
55 /// Returns the new combined block.
56 static BasicBlock
*FoldBlockIntoPredecessor(BasicBlock
*BB
, LoopInfo
* LI
) {
57 // Merge basic blocks into their predecessor if there is only one distinct
58 // pred, and if there is only one distinct successor of the predecessor, and
59 // if there are no PHI nodes.
60 BasicBlock
*OnlyPred
= BB
->getSinglePredecessor();
61 if (!OnlyPred
) return 0;
63 if (OnlyPred
->getTerminator()->getNumSuccessors() != 1)
66 DEBUG(dbgs() << "Merging: " << *BB
<< "into: " << *OnlyPred
);
68 // Resolve any PHI nodes at the start of the block. They are all
69 // guaranteed to have exactly one entry if they exist, unless there are
70 // multiple duplicate (but guaranteed to be equal) entries for the
71 // incoming edges. This occurs when there are multiple edges from
72 // OnlyPred to OnlySucc.
73 FoldSingleEntryPHINodes(BB
);
75 // Delete the unconditional branch from the predecessor...
76 OnlyPred
->getInstList().pop_back();
78 // Move all definitions in the successor to the predecessor...
79 OnlyPred
->getInstList().splice(OnlyPred
->end(), BB
->getInstList());
81 // Make all PHI nodes that referred to BB now refer to Pred as their
83 BB
->replaceAllUsesWith(OnlyPred
);
85 std::string OldName
= BB
->getName();
87 // Erase basic block from the function...
89 BB
->eraseFromParent();
91 // Inherit predecessor's name if it exists...
92 if (!OldName
.empty() && !OnlyPred
->hasName())
93 OnlyPred
->setName(OldName
);
98 /// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true
99 /// if unrolling was succesful, or false if the loop was unmodified. Unrolling
100 /// can only fail when the loop's latch block is not terminated by a conditional
101 /// branch instruction. However, if the trip count (and multiple) are not known,
102 /// loop unrolling will mostly produce more code that is no faster.
104 /// The LoopInfo Analysis that is passed will be kept consistent.
106 /// If a LoopPassManager is passed in, and the loop is fully removed, it will be
107 /// removed from the LoopPassManager as well. LPM can also be NULL.
108 bool llvm::UnrollLoop(Loop
*L
, unsigned Count
, LoopInfo
* LI
, LPPassManager
* LPM
) {
109 BasicBlock
*Preheader
= L
->getLoopPreheader();
111 DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
115 BasicBlock
*LatchBlock
= L
->getLoopLatch();
117 DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
121 BasicBlock
*Header
= L
->getHeader();
122 BranchInst
*BI
= dyn_cast
<BranchInst
>(LatchBlock
->getTerminator());
124 if (!BI
|| BI
->isUnconditional()) {
125 // The loop-rotate pass can be helpful to avoid this in many cases.
127 " Can't unroll; loop not terminated by a conditional branch.\n");
131 // Notify ScalarEvolution that the loop will be substantially changed,
132 // if not outright eliminated.
133 if (ScalarEvolution
*SE
= LPM
->getAnalysisIfAvailable
<ScalarEvolution
>())
137 unsigned TripCount
= L
->getSmallConstantTripCount();
138 // Find trip multiple if count is not available
139 unsigned TripMultiple
= 1;
141 TripMultiple
= L
->getSmallConstantTripMultiple();
144 DEBUG(dbgs() << " Trip Count = " << TripCount
<< "\n");
145 if (TripMultiple
!= 1)
146 DEBUG(dbgs() << " Trip Multiple = " << TripMultiple
<< "\n");
148 // Effectively "DCE" unrolled iterations that are beyond the tripcount
149 // and will never be executed.
150 if (TripCount
!= 0 && Count
> TripCount
)
154 assert(TripMultiple
> 0);
155 assert(TripCount
== 0 || TripCount
% TripMultiple
== 0);
157 // Are we eliminating the loop control altogether?
158 bool CompletelyUnroll
= Count
== TripCount
;
160 // If we know the trip count, we know the multiple...
161 unsigned BreakoutTrip
= 0;
162 if (TripCount
!= 0) {
163 BreakoutTrip
= TripCount
% Count
;
166 // Figure out what multiple to use.
167 BreakoutTrip
= TripMultiple
=
168 (unsigned)GreatestCommonDivisor64(Count
, TripMultiple
);
171 if (CompletelyUnroll
) {
172 DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header
->getName()
173 << " with trip count " << TripCount
<< "!\n");
175 DEBUG(dbgs() << "UNROLLING loop %" << Header
->getName()
177 if (TripMultiple
== 0 || BreakoutTrip
!= TripMultiple
) {
178 DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip
);
179 } else if (TripMultiple
!= 1) {
180 DEBUG(dbgs() << " with " << TripMultiple
<< " trips per branch");
182 DEBUG(dbgs() << "!\n");
185 std::vector
<BasicBlock
*> LoopBlocks
= L
->getBlocks();
187 bool ContinueOnTrue
= L
->contains(BI
->getSuccessor(0));
188 BasicBlock
*LoopExit
= BI
->getSuccessor(ContinueOnTrue
);
190 // For the first iteration of the loop, we should use the precloned values for
191 // PHI nodes. Insert associations now.
192 ValueToValueMapTy LastValueMap
;
193 std::vector
<PHINode
*> OrigPHINode
;
194 for (BasicBlock::iterator I
= Header
->begin(); isa
<PHINode
>(I
); ++I
) {
195 PHINode
*PN
= cast
<PHINode
>(I
);
196 OrigPHINode
.push_back(PN
);
198 dyn_cast
<Instruction
>(PN
->getIncomingValueForBlock(LatchBlock
)))
203 std::vector
<BasicBlock
*> Headers
;
204 std::vector
<BasicBlock
*> Latches
;
205 Headers
.push_back(Header
);
206 Latches
.push_back(LatchBlock
);
208 for (unsigned It
= 1; It
!= Count
; ++It
) {
209 std::vector
<BasicBlock
*> NewBlocks
;
211 for (std::vector
<BasicBlock
*>::iterator BB
= LoopBlocks
.begin(),
212 E
= LoopBlocks
.end(); BB
!= E
; ++BB
) {
213 ValueToValueMapTy VMap
;
214 BasicBlock
*New
= CloneBasicBlock(*BB
, VMap
, "." + Twine(It
));
215 Header
->getParent()->getBasicBlockList().push_back(New
);
217 // Loop over all of the PHI nodes in the block, changing them to use the
218 // incoming values from the previous block.
220 for (unsigned i
= 0, e
= OrigPHINode
.size(); i
!= e
; ++i
) {
221 PHINode
*NewPHI
= cast
<PHINode
>(VMap
[OrigPHINode
[i
]]);
222 Value
*InVal
= NewPHI
->getIncomingValueForBlock(LatchBlock
);
223 if (Instruction
*InValI
= dyn_cast
<Instruction
>(InVal
))
224 if (It
> 1 && L
->contains(InValI
))
225 InVal
= LastValueMap
[InValI
];
226 VMap
[OrigPHINode
[i
]] = InVal
;
227 New
->getInstList().erase(NewPHI
);
230 // Update our running map of newest clones
231 LastValueMap
[*BB
] = New
;
232 for (ValueToValueMapTy::iterator VI
= VMap
.begin(), VE
= VMap
.end();
234 LastValueMap
[VI
->first
] = VI
->second
;
236 L
->addBasicBlockToLoop(New
, LI
->getBase());
238 // Add phi entries for newly created values to all exit blocks except
239 // the successor of the latch block. The successor of the exit block will
240 // be updated specially after unrolling all the way.
241 if (*BB
!= LatchBlock
)
242 for (Value::use_iterator UI
= (*BB
)->use_begin(), UE
= (*BB
)->use_end();
244 Instruction
*UseInst
= cast
<Instruction
>(*UI
);
246 if (isa
<PHINode
>(UseInst
) && !L
->contains(UseInst
)) {
247 PHINode
*phi
= cast
<PHINode
>(UseInst
);
248 Value
*Incoming
= phi
->getIncomingValueForBlock(*BB
);
249 phi
->addIncoming(Incoming
, New
);
253 // Keep track of new headers and latches as we create them, so that
254 // we can insert the proper branches later.
256 Headers
.push_back(New
);
257 if (*BB
== LatchBlock
) {
258 Latches
.push_back(New
);
260 // Also, clear out the new latch's back edge so that it doesn't look
261 // like a new loop, so that it's amenable to being merged with adjacent
263 TerminatorInst
*Term
= New
->getTerminator();
264 assert(L
->contains(Term
->getSuccessor(!ContinueOnTrue
)));
265 assert(Term
->getSuccessor(ContinueOnTrue
) == LoopExit
);
266 Term
->setSuccessor(!ContinueOnTrue
, NULL
);
269 NewBlocks
.push_back(New
);
272 // Remap all instructions in the most recent iteration
273 for (unsigned i
= 0; i
< NewBlocks
.size(); ++i
)
274 for (BasicBlock::iterator I
= NewBlocks
[i
]->begin(),
275 E
= NewBlocks
[i
]->end(); I
!= E
; ++I
)
276 ::RemapInstruction(I
, LastValueMap
);
279 // The latch block exits the loop. If there are any PHI nodes in the
280 // successor blocks, update them to use the appropriate values computed as the
281 // last iteration of the loop.
283 SmallPtrSet
<PHINode
*, 8> Users
;
284 for (Value::use_iterator UI
= LatchBlock
->use_begin(),
285 UE
= LatchBlock
->use_end(); UI
!= UE
; ++UI
)
286 if (PHINode
*phi
= dyn_cast
<PHINode
>(*UI
))
289 BasicBlock
*LastIterationBB
= cast
<BasicBlock
>(LastValueMap
[LatchBlock
]);
290 for (SmallPtrSet
<PHINode
*,8>::iterator SI
= Users
.begin(), SE
= Users
.end();
293 Value
*InVal
= PN
->removeIncomingValue(LatchBlock
, false);
294 // If this value was defined in the loop, take the value defined by the
295 // last iteration of the loop.
296 if (Instruction
*InValI
= dyn_cast
<Instruction
>(InVal
)) {
297 if (L
->contains(InValI
))
298 InVal
= LastValueMap
[InVal
];
300 PN
->addIncoming(InVal
, LastIterationBB
);
304 // Now, if we're doing complete unrolling, loop over the PHI nodes in the
305 // original block, setting them to their incoming values.
306 if (CompletelyUnroll
) {
307 BasicBlock
*Preheader
= L
->getLoopPreheader();
308 for (unsigned i
= 0, e
= OrigPHINode
.size(); i
!= e
; ++i
) {
309 PHINode
*PN
= OrigPHINode
[i
];
310 PN
->replaceAllUsesWith(PN
->getIncomingValueForBlock(Preheader
));
311 Header
->getInstList().erase(PN
);
315 // Now that all the basic blocks for the unrolled iterations are in place,
316 // set up the branches to connect them.
317 for (unsigned i
= 0, e
= Latches
.size(); i
!= e
; ++i
) {
318 // The original branch was replicated in each unrolled iteration.
319 BranchInst
*Term
= cast
<BranchInst
>(Latches
[i
]->getTerminator());
321 // The branch destination.
322 unsigned j
= (i
+ 1) % e
;
323 BasicBlock
*Dest
= Headers
[j
];
324 bool NeedConditional
= true;
326 // For a complete unroll, make the last iteration end with a branch
327 // to the exit block.
328 if (CompletelyUnroll
&& j
== 0) {
330 NeedConditional
= false;
333 // If we know the trip count or a multiple of it, we can safely use an
334 // unconditional branch for some iterations.
335 if (j
!= BreakoutTrip
&& (TripMultiple
== 0 || j
% TripMultiple
!= 0)) {
336 NeedConditional
= false;
339 if (NeedConditional
) {
340 // Update the conditional branch's successor for the following
342 Term
->setSuccessor(!ContinueOnTrue
, Dest
);
344 Term
->setUnconditionalDest(Dest
);
345 // Merge adjacent basic blocks, if possible.
346 if (BasicBlock
*Fold
= FoldBlockIntoPredecessor(Dest
, LI
)) {
347 std::replace(Latches
.begin(), Latches
.end(), Dest
, Fold
);
348 std::replace(Headers
.begin(), Headers
.end(), Dest
, Fold
);
353 // At this point, the code is well formed. We now do a quick sweep over the
354 // inserted code, doing constant propagation and dead code elimination as we
356 const std::vector
<BasicBlock
*> &NewLoopBlocks
= L
->getBlocks();
357 for (std::vector
<BasicBlock
*>::const_iterator BB
= NewLoopBlocks
.begin(),
358 BBE
= NewLoopBlocks
.end(); BB
!= BBE
; ++BB
)
359 for (BasicBlock::iterator I
= (*BB
)->begin(), E
= (*BB
)->end(); I
!= E
; ) {
360 Instruction
*Inst
= I
++;
362 if (isInstructionTriviallyDead(Inst
))
363 (*BB
)->getInstList().erase(Inst
);
364 else if (Value
*V
= SimplifyInstruction(Inst
))
365 if (LI
->replacementPreservesLCSSAForm(Inst
, V
)) {
366 Inst
->replaceAllUsesWith(V
);
367 (*BB
)->getInstList().erase(Inst
);
371 NumCompletelyUnrolled
+= CompletelyUnroll
;
373 // Remove the loop from the LoopPassManager if it's completely removed.
374 if (CompletelyUnroll
&& LPM
!= NULL
)
375 LPM
->deleteLoopFromQueue(L
);