1 //===-- UnrollLoop.cpp - Loop unrolling utilities -------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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
12 // unrolling.
13 //
14 // The process of unrolling can produce extraneous basic blocks linked with
15 // unconditional branches. This will be corrected in the future.
16 //
17 //===----------------------------------------------------------------------===//
18
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/AssumptionCache.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/LoopIterator.h"
24 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
25 #include "llvm/Analysis/ScalarEvolution.h"
26 #include "llvm/Transforms/Utils/Local.h"
27 #include "llvm/IR/BasicBlock.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DebugInfoMetadata.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
36 #include "llvm/Transforms/Utils/Cloning.h"
37 #include "llvm/Transforms/Utils/LoopSimplify.h"
38 #include "llvm/Transforms/Utils/LoopUtils.h"
39 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
40 #include "llvm/Transforms/Utils/UnrollLoop.h"
41 using namespace llvm;
42
43 #define DEBUG_TYPE "loop-unroll"
44
45 // TODO: Should these be here or in LoopUnroll?
46 STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
47 STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
48
49 static cl::opt<bool>
50 UnrollRuntimeEpilog("unroll-runtime-epilog", cl::init(false), cl::Hidden,
51 cl::desc("Allow runtime unrolled loops to be unrolled "
52 "with epilog instead of prolog."));
53
54 static cl::opt<bool>
55 UnrollVerifyDomtree("unroll-verify-domtree", cl::Hidden,
56 cl::desc("Verify domtree after unrolling"),
57 #ifdef EXPENSIVE_CHECKS
58 cl::init(true)
59 #else
60 cl::init(false)
61 #endif
62 );
63
64 /// Convert the instruction operands from referencing the current values into
65 /// those specified by VMap.
remapInstruction(Instruction * I,ValueToValueMapTy & VMap)66 void llvm::remapInstruction(Instruction *I, ValueToValueMapTy &VMap) {
67 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
68 Value *Op = I->getOperand(op);
69
70 // Unwrap arguments of dbg.value intrinsics.
71 bool Wrapped = false;
72 if (auto *V = dyn_cast<MetadataAsValue>(Op))
73 if (auto *Unwrapped = dyn_cast<ValueAsMetadata>(V->getMetadata())) {
74 Op = Unwrapped->getValue();
75 Wrapped = true;
76 }
77
78 auto wrap = [&](Value *V) {
79 auto &C = I->getContext();
80 return Wrapped ? MetadataAsValue::get(C, ValueAsMetadata::get(V)) : V;
81 };
82
83 ValueToValueMapTy::iterator It = VMap.find(Op);
84 if (It != VMap.end())
85 I->setOperand(op, wrap(It->second));
86 }
87
88 if (PHINode *PN = dyn_cast<PHINode>(I)) {
89 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
90 ValueToValueMapTy::iterator It = VMap.find(PN->getIncomingBlock(i));
91 if (It != VMap.end())
92 PN->setIncomingBlock(i, cast<BasicBlock>(It->second));
93 }
94 }
95 }
96
97 /// Folds a basic block into its predecessor if it only has one predecessor, and
98 /// that predecessor only has one successor.
99 /// The LoopInfo Analysis that is passed will be kept consistent.
foldBlockIntoPredecessor(BasicBlock * BB,LoopInfo * LI,ScalarEvolution * SE,DominatorTree * DT)100 BasicBlock *llvm::foldBlockIntoPredecessor(BasicBlock *BB, LoopInfo *LI,
101 ScalarEvolution *SE,
102 DominatorTree *DT) {
103 // Merge basic blocks into their predecessor if there is only one distinct
104 // pred, and if there is only one distinct successor of the predecessor, and
105 // if there are no PHI nodes.
106 BasicBlock *OnlyPred = BB->getSinglePredecessor();
107 if (!OnlyPred) return nullptr;
108
109 if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
110 return nullptr;
111
112 LLVM_DEBUG(dbgs() << "Merging: " << BB->getName() << " into "
113 << OnlyPred->getName() << "\n");
114
115 // Resolve any PHI nodes at the start of the block. They are all
116 // guaranteed to have exactly one entry if they exist, unless there are
117 // multiple duplicate (but guaranteed to be equal) entries for the
118 // incoming edges. This occurs when there are multiple edges from
119 // OnlyPred to OnlySucc.
120 FoldSingleEntryPHINodes(BB);
121
122 // Delete the unconditional branch from the predecessor...
123 OnlyPred->getInstList().pop_back();
124
125 // Make all PHI nodes that referred to BB now refer to Pred as their
126 // source...
127 BB->replaceAllUsesWith(OnlyPred);
128
129 // Move all definitions in the successor to the predecessor...
130 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
131
132 // OldName will be valid until erased.
133 StringRef OldName = BB->getName();
134
135 // Erase the old block and update dominator info.
136 if (DT)
137 if (DomTreeNode *DTN = DT->getNode(BB)) {
138 DomTreeNode *PredDTN = DT->getNode(OnlyPred);
139 SmallVector<DomTreeNode *, 8> Children(DTN->begin(), DTN->end());
140 for (auto *DI : Children)
141 DT->changeImmediateDominator(DI, PredDTN);
142
143 DT->eraseNode(BB);
144 }
145
146 LI->removeBlock(BB);
147
148 // Inherit predecessor's name if it exists...
149 if (!OldName.empty() && !OnlyPred->hasName())
150 OnlyPred->setName(OldName);
151
152 BB->eraseFromParent();
153
154 return OnlyPred;
155 }
156
157 /// Check if unrolling created a situation where we need to insert phi nodes to
158 /// preserve LCSSA form.
159 /// \param Blocks is a vector of basic blocks representing unrolled loop.
160 /// \param L is the outer loop.
161 /// It's possible that some of the blocks are in L, and some are not. In this
162 /// case, if there is a use is outside L, and definition is inside L, we need to
163 /// insert a phi-node, otherwise LCSSA will be broken.
164 /// The function is just a helper function for llvm::UnrollLoop that returns
165 /// true if this situation occurs, indicating that LCSSA needs to be fixed.
needToInsertPhisForLCSSA(Loop * L,std::vector<BasicBlock * > Blocks,LoopInfo * LI)166 static bool needToInsertPhisForLCSSA(Loop *L, std::vector<BasicBlock *> Blocks,
167 LoopInfo *LI) {
168 for (BasicBlock *BB : Blocks) {
169 if (LI->getLoopFor(BB) == L)
170 continue;
171 for (Instruction &I : *BB) {
172 for (Use &U : I.operands()) {
173 if (auto Def = dyn_cast<Instruction>(U)) {
174 Loop *DefLoop = LI->getLoopFor(Def->getParent());
175 if (!DefLoop)
176 continue;
177 if (DefLoop->contains(L))
178 return true;
179 }
180 }
181 }
182 }
183 return false;
184 }
185
186 /// Adds ClonedBB to LoopInfo, creates a new loop for ClonedBB if necessary
187 /// and adds a mapping from the original loop to the new loop to NewLoops.
188 /// Returns nullptr if no new loop was created and a pointer to the
189 /// original loop OriginalBB was part of otherwise.
addClonedBlockToLoopInfo(BasicBlock * OriginalBB,BasicBlock * ClonedBB,LoopInfo * LI,NewLoopsMap & NewLoops)190 const Loop* llvm::addClonedBlockToLoopInfo(BasicBlock *OriginalBB,
191 BasicBlock *ClonedBB, LoopInfo *LI,
192 NewLoopsMap &NewLoops) {
193 // Figure out which loop New is in.
194 const Loop *OldLoop = LI->getLoopFor(OriginalBB);
195 assert(OldLoop && "Should (at least) be in the loop being unrolled!");
196
197 Loop *&NewLoop = NewLoops[OldLoop];
198 if (!NewLoop) {
199 // Found a new sub-loop.
200 assert(OriginalBB == OldLoop->getHeader() &&
201 "Header should be first in RPO");
202
203 NewLoop = LI->AllocateLoop();
204 Loop *NewLoopParent = NewLoops.lookup(OldLoop->getParentLoop());
205
206 if (NewLoopParent)
207 NewLoopParent->addChildLoop(NewLoop);
208 else
209 LI->addTopLevelLoop(NewLoop);
210
211 NewLoop->addBasicBlockToLoop(ClonedBB, *LI);
212 return OldLoop;
213 } else {
214 NewLoop->addBasicBlockToLoop(ClonedBB, *LI);
215 return nullptr;
216 }
217 }
218
219 /// The function chooses which type of unroll (epilog or prolog) is more
220 /// profitabale.
221 /// Epilog unroll is more profitable when there is PHI that starts from
222 /// constant. In this case epilog will leave PHI start from constant,
223 /// but prolog will convert it to non-constant.
224 ///
225 /// loop:
226 /// PN = PHI [I, Latch], [CI, PreHeader]
227 /// I = foo(PN)
228 /// ...
229 ///
230 /// Epilog unroll case.
231 /// loop:
232 /// PN = PHI [I2, Latch], [CI, PreHeader]
233 /// I1 = foo(PN)
234 /// I2 = foo(I1)
235 /// ...
236 /// Prolog unroll case.
237 /// NewPN = PHI [PrologI, Prolog], [CI, PreHeader]
238 /// loop:
239 /// PN = PHI [I2, Latch], [NewPN, PreHeader]
240 /// I1 = foo(PN)
241 /// I2 = foo(I1)
242 /// ...
243 ///
isEpilogProfitable(Loop * L)244 static bool isEpilogProfitable(Loop *L) {
245 BasicBlock *PreHeader = L->getLoopPreheader();
246 BasicBlock *Header = L->getHeader();
247 assert(PreHeader && Header);
248 for (const PHINode &PN : Header->phis()) {
249 if (isa<ConstantInt>(PN.getIncomingValueForBlock(PreHeader)))
250 return true;
251 }
252 return false;
253 }
254
255 /// Perform some cleanup and simplifications on loops after unrolling. It is
256 /// useful to simplify the IV's in the new loop, as well as do a quick
257 /// simplify/dce pass of the instructions.
simplifyLoopAfterUnroll(Loop * L,bool SimplifyIVs,LoopInfo * LI,ScalarEvolution * SE,DominatorTree * DT,AssumptionCache * AC)258 void llvm::simplifyLoopAfterUnroll(Loop *L, bool SimplifyIVs, LoopInfo *LI,
259 ScalarEvolution *SE, DominatorTree *DT,
260 AssumptionCache *AC) {
261 // Simplify any new induction variables in the partially unrolled loop.
262 if (SE && SimplifyIVs) {
263 SmallVector<WeakTrackingVH, 16> DeadInsts;
264 simplifyLoopIVs(L, SE, DT, LI, DeadInsts);
265
266 // Aggressively clean up dead instructions that simplifyLoopIVs already
267 // identified. Any remaining should be cleaned up below.
268 while (!DeadInsts.empty())
269 if (Instruction *Inst =
270 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
271 RecursivelyDeleteTriviallyDeadInstructions(Inst);
272 }
273
274 // At this point, the code is well formed. We now do a quick sweep over the
275 // inserted code, doing constant propagation and dead code elimination as we
276 // go.
277 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
278 for (BasicBlock *BB : L->getBlocks()) {
279 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
280 Instruction *Inst = &*I++;
281
282 if (Value *V = SimplifyInstruction(Inst, {DL, nullptr, DT, AC}))
283 if (LI->replacementPreservesLCSSAForm(Inst, V))
284 Inst->replaceAllUsesWith(V);
285 if (isInstructionTriviallyDead(Inst))
286 BB->getInstList().erase(Inst);
287 }
288 }
289
290 // TODO: after peeling or unrolling, previously loop variant conditions are
291 // likely to fold to constants, eagerly propagating those here will require
292 // fewer cleanup passes to be run. Alternatively, a LoopEarlyCSE might be
293 // appropriate.
294 }
295
296 /// Unroll the given loop by Count. The loop must be in LCSSA form. Unrolling
297 /// can only fail when the loop's latch block is not terminated by a conditional
298 /// branch instruction. However, if the trip count (and multiple) are not known,
299 /// loop unrolling will mostly produce more code that is no faster.
300 ///
301 /// TripCount is the upper bound of the iteration on which control exits
302 /// LatchBlock. Control may exit the loop prior to TripCount iterations either
303 /// via an early branch in other loop block or via LatchBlock terminator. This
304 /// is relaxed from the general definition of trip count which is the number of
305 /// times the loop header executes. Note that UnrollLoop assumes that the loop
306 /// counter test is in LatchBlock in order to remove unnecesssary instances of
307 /// the test. If control can exit the loop from the LatchBlock's terminator
308 /// prior to TripCount iterations, flag PreserveCondBr needs to be set.
309 ///
310 /// PreserveCondBr indicates whether the conditional branch of the LatchBlock
311 /// needs to be preserved. It is needed when we use trip count upper bound to
312 /// fully unroll the loop. If PreserveOnlyFirst is also set then only the first
313 /// conditional branch needs to be preserved.
314 ///
315 /// Similarly, TripMultiple divides the number of times that the LatchBlock may
316 /// execute without exiting the loop.
317 ///
318 /// If AllowRuntime is true then UnrollLoop will consider unrolling loops that
319 /// have a runtime (i.e. not compile time constant) trip count. Unrolling these
320 /// loops require a unroll "prologue" that runs "RuntimeTripCount % Count"
321 /// iterations before branching into the unrolled loop. UnrollLoop will not
322 /// runtime-unroll the loop if computing RuntimeTripCount will be expensive and
323 /// AllowExpensiveTripCount is false.
324 ///
325 /// If we want to perform PGO-based loop peeling, PeelCount is set to the
326 /// number of iterations we want to peel off.
327 ///
328 /// The LoopInfo Analysis that is passed will be kept consistent.
329 ///
330 /// This utility preserves LoopInfo. It will also preserve ScalarEvolution and
331 /// DominatorTree if they are non-null.
332 ///
333 /// If RemainderLoop is non-null, it will receive the remainder loop (if
334 /// required and not fully unrolled).
UnrollLoop(Loop * L,unsigned Count,unsigned TripCount,bool Force,bool AllowRuntime,bool AllowExpensiveTripCount,bool PreserveCondBr,bool PreserveOnlyFirst,unsigned TripMultiple,unsigned PeelCount,bool UnrollRemainder,LoopInfo * LI,ScalarEvolution * SE,DominatorTree * DT,AssumptionCache * AC,OptimizationRemarkEmitter * ORE,bool PreserveLCSSA,Loop ** RemainderLoop)335 LoopUnrollResult llvm::UnrollLoop(
336 Loop *L, unsigned Count, unsigned TripCount, bool Force, bool AllowRuntime,
337 bool AllowExpensiveTripCount, bool PreserveCondBr, bool PreserveOnlyFirst,
338 unsigned TripMultiple, unsigned PeelCount, bool UnrollRemainder,
339 LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, AssumptionCache *AC,
340 OptimizationRemarkEmitter *ORE, bool PreserveLCSSA, Loop **RemainderLoop) {
341
342 BasicBlock *Preheader = L->getLoopPreheader();
343 if (!Preheader) {
344 LLVM_DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
345 return LoopUnrollResult::Unmodified;
346 }
347
348 BasicBlock *LatchBlock = L->getLoopLatch();
349 if (!LatchBlock) {
350 LLVM_DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
351 return LoopUnrollResult::Unmodified;
352 }
353
354 // Loops with indirectbr cannot be cloned.
355 if (!L->isSafeToClone()) {
356 LLVM_DEBUG(dbgs() << " Can't unroll; Loop body cannot be cloned.\n");
357 return LoopUnrollResult::Unmodified;
358 }
359
360 // The current loop unroll pass can only unroll loops with a single latch
361 // that's a conditional branch exiting the loop.
362 // FIXME: The implementation can be extended to work with more complicated
363 // cases, e.g. loops with multiple latches.
364 BasicBlock *Header = L->getHeader();
365 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
366
367 if (!BI || BI->isUnconditional()) {
368 // The loop-rotate pass can be helpful to avoid this in many cases.
369 LLVM_DEBUG(
370 dbgs()
371 << " Can't unroll; loop not terminated by a conditional branch.\n");
372 return LoopUnrollResult::Unmodified;
373 }
374
375 auto CheckSuccessors = [&](unsigned S1, unsigned S2) {
376 return BI->getSuccessor(S1) == Header && !L->contains(BI->getSuccessor(S2));
377 };
378
379 if (!CheckSuccessors(0, 1) && !CheckSuccessors(1, 0)) {
380 LLVM_DEBUG(dbgs() << "Can't unroll; only loops with one conditional latch"
381 " exiting the loop can be unrolled\n");
382 return LoopUnrollResult::Unmodified;
383 }
384
385 if (Header->hasAddressTaken()) {
386 // The loop-rotate pass can be helpful to avoid this in many cases.
387 LLVM_DEBUG(
388 dbgs() << " Won't unroll loop: address of header block is taken.\n");
389 return LoopUnrollResult::Unmodified;
390 }
391
392 if (TripCount != 0)
393 LLVM_DEBUG(dbgs() << " Trip Count = " << TripCount << "\n");
394 if (TripMultiple != 1)
395 LLVM_DEBUG(dbgs() << " Trip Multiple = " << TripMultiple << "\n");
396
397 // Effectively "DCE" unrolled iterations that are beyond the tripcount
398 // and will never be executed.
399 if (TripCount != 0 && Count > TripCount)
400 Count = TripCount;
401
402 // Don't enter the unroll code if there is nothing to do.
403 if (TripCount == 0 && Count < 2 && PeelCount == 0) {
404 LLVM_DEBUG(dbgs() << "Won't unroll; almost nothing to do\n");
405 return LoopUnrollResult::Unmodified;
406 }
407
408 assert(Count > 0);
409 assert(TripMultiple > 0);
410 assert(TripCount == 0 || TripCount % TripMultiple == 0);
411
412 // Are we eliminating the loop control altogether?
413 bool CompletelyUnroll = Count == TripCount;
414 SmallVector<BasicBlock *, 4> ExitBlocks;
415 L->getExitBlocks(ExitBlocks);
416 std::vector<BasicBlock*> OriginalLoopBlocks = L->getBlocks();
417
418 // Go through all exits of L and see if there are any phi-nodes there. We just
419 // conservatively assume that they're inserted to preserve LCSSA form, which
420 // means that complete unrolling might break this form. We need to either fix
421 // it in-place after the transformation, or entirely rebuild LCSSA. TODO: For
422 // now we just recompute LCSSA for the outer loop, but it should be possible
423 // to fix it in-place.
424 bool NeedToFixLCSSA = PreserveLCSSA && CompletelyUnroll &&
425 any_of(ExitBlocks, [](const BasicBlock *BB) {
426 return isa<PHINode>(BB->begin());
427 });
428
429 // We assume a run-time trip count if the compiler cannot
430 // figure out the loop trip count and the unroll-runtime
431 // flag is specified.
432 bool RuntimeTripCount = (TripCount == 0 && Count > 0 && AllowRuntime);
433
434 assert((!RuntimeTripCount || !PeelCount) &&
435 "Did not expect runtime trip-count unrolling "
436 "and peeling for the same loop");
437
438 bool Peeled = false;
439 if (PeelCount) {
440 Peeled = peelLoop(L, PeelCount, LI, SE, DT, AC, PreserveLCSSA);
441
442 // Successful peeling may result in a change in the loop preheader/trip
443 // counts. If we later unroll the loop, we want these to be updated.
444 if (Peeled) {
445 BasicBlock *ExitingBlock = L->getExitingBlock();
446 assert(ExitingBlock && "Loop without exiting block?");
447 Preheader = L->getLoopPreheader();
448 TripCount = SE->getSmallConstantTripCount(L, ExitingBlock);
449 TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock);
450 }
451 }
452
453 // Loops containing convergent instructions must have a count that divides
454 // their TripMultiple.
455 LLVM_DEBUG(
456 {
457 bool HasConvergent = false;
458 for (auto &BB : L->blocks())
459 for (auto &I : *BB)
460 if (auto CS = CallSite(&I))
461 HasConvergent |= CS.isConvergent();
462 assert((!HasConvergent || TripMultiple % Count == 0) &&
463 "Unroll count must divide trip multiple if loop contains a "
464 "convergent operation.");
465 });
466
467 bool EpilogProfitability =
468 UnrollRuntimeEpilog.getNumOccurrences() ? UnrollRuntimeEpilog
469 : isEpilogProfitable(L);
470
471 if (RuntimeTripCount && TripMultiple % Count != 0 &&
472 !UnrollRuntimeLoopRemainder(L, Count, AllowExpensiveTripCount,
473 EpilogProfitability, UnrollRemainder, LI, SE,
474 DT, AC, PreserveLCSSA, RemainderLoop)) {
475 if (Force)
476 RuntimeTripCount = false;
477 else {
478 LLVM_DEBUG(dbgs() << "Won't unroll; remainder loop could not be "
479 "generated when assuming runtime trip count\n");
480 return LoopUnrollResult::Unmodified;
481 }
482 }
483
484 // If we know the trip count, we know the multiple...
485 unsigned BreakoutTrip = 0;
486 if (TripCount != 0) {
487 BreakoutTrip = TripCount % Count;
488 TripMultiple = 0;
489 } else {
490 // Figure out what multiple to use.
491 BreakoutTrip = TripMultiple =
492 (unsigned)GreatestCommonDivisor64(Count, TripMultiple);
493 }
494
495 using namespace ore;
496 // Report the unrolling decision.
497 if (CompletelyUnroll) {
498 LLVM_DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
499 << " with trip count " << TripCount << "!\n");
500 if (ORE)
501 ORE->emit([&]() {
502 return OptimizationRemark(DEBUG_TYPE, "FullyUnrolled", L->getStartLoc(),
503 L->getHeader())
504 << "completely unrolled loop with "
505 << NV("UnrollCount", TripCount) << " iterations";
506 });
507 } else if (PeelCount) {
508 LLVM_DEBUG(dbgs() << "PEELING loop %" << Header->getName()
509 << " with iteration count " << PeelCount << "!\n");
510 if (ORE)
511 ORE->emit([&]() {
512 return OptimizationRemark(DEBUG_TYPE, "Peeled", L->getStartLoc(),
513 L->getHeader())
514 << " peeled loop by " << NV("PeelCount", PeelCount)
515 << " iterations";
516 });
517 } else {
518 auto DiagBuilder = [&]() {
519 OptimizationRemark Diag(DEBUG_TYPE, "PartialUnrolled", L->getStartLoc(),
520 L->getHeader());
521 return Diag << "unrolled loop by a factor of "
522 << NV("UnrollCount", Count);
523 };
524
525 LLVM_DEBUG(dbgs() << "UNROLLING loop %" << Header->getName() << " by "
526 << Count);
527 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
528 LLVM_DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip);
529 if (ORE)
530 ORE->emit([&]() {
531 return DiagBuilder() << " with a breakout at trip "
532 << NV("BreakoutTrip", BreakoutTrip);
533 });
534 } else if (TripMultiple != 1) {
535 LLVM_DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
536 if (ORE)
537 ORE->emit([&]() {
538 return DiagBuilder() << " with " << NV("TripMultiple", TripMultiple)
539 << " trips per branch";
540 });
541 } else if (RuntimeTripCount) {
542 LLVM_DEBUG(dbgs() << " with run-time trip count");
543 if (ORE)
544 ORE->emit(
545 [&]() { return DiagBuilder() << " with run-time trip count"; });
546 }
547 LLVM_DEBUG(dbgs() << "!\n");
548 }
549
550 // We are going to make changes to this loop. SCEV may be keeping cached info
551 // about it, in particular about backedge taken count. The changes we make
552 // are guaranteed to invalidate this information for our loop. It is tempting
553 // to only invalidate the loop being unrolled, but it is incorrect as long as
554 // all exiting branches from all inner loops have impact on the outer loops,
555 // and if something changes inside them then any of outer loops may also
556 // change. When we forget outermost loop, we also forget all contained loops
557 // and this is what we need here.
558 if (SE)
559 SE->forgetTopmostLoop(L);
560
561 bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
562 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
563
564 // For the first iteration of the loop, we should use the precloned values for
565 // PHI nodes. Insert associations now.
566 ValueToValueMapTy LastValueMap;
567 std::vector<PHINode*> OrigPHINode;
568 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
569 OrigPHINode.push_back(cast<PHINode>(I));
570 }
571
572 std::vector<BasicBlock*> Headers;
573 std::vector<BasicBlock*> Latches;
574 Headers.push_back(Header);
575 Latches.push_back(LatchBlock);
576
577 // The current on-the-fly SSA update requires blocks to be processed in
578 // reverse postorder so that LastValueMap contains the correct value at each
579 // exit.
580 LoopBlocksDFS DFS(L);
581 DFS.perform(LI);
582
583 // Stash the DFS iterators before adding blocks to the loop.
584 LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO();
585 LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO();
586
587 std::vector<BasicBlock*> UnrolledLoopBlocks = L->getBlocks();
588
589 // Loop Unrolling might create new loops. While we do preserve LoopInfo, we
590 // might break loop-simplified form for these loops (as they, e.g., would
591 // share the same exit blocks). We'll keep track of loops for which we can
592 // break this so that later we can re-simplify them.
593 SmallSetVector<Loop *, 4> LoopsToSimplify;
594 for (Loop *SubLoop : *L)
595 LoopsToSimplify.insert(SubLoop);
596
597 if (Header->getParent()->isDebugInfoForProfiling())
598 for (BasicBlock *BB : L->getBlocks())
599 for (Instruction &I : *BB)
600 if (!isa<DbgInfoIntrinsic>(&I))
601 if (const DILocation *DIL = I.getDebugLoc()) {
602 auto NewDIL = DIL->cloneWithDuplicationFactor(Count);
603 if (NewDIL)
604 I.setDebugLoc(NewDIL.getValue());
605 else
606 LLVM_DEBUG(dbgs()
607 << "Failed to create new discriminator: "
608 << DIL->getFilename() << " Line: " << DIL->getLine());
609 }
610
611 for (unsigned It = 1; It != Count; ++It) {
612 std::vector<BasicBlock*> NewBlocks;
613 SmallDenseMap<const Loop *, Loop *, 4> NewLoops;
614 NewLoops[L] = L;
615
616 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
617 ValueToValueMapTy VMap;
618 BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
619 Header->getParent()->getBasicBlockList().push_back(New);
620
621 assert((*BB != Header || LI->getLoopFor(*BB) == L) &&
622 "Header should not be in a sub-loop");
623 // Tell LI about New.
624 const Loop *OldLoop = addClonedBlockToLoopInfo(*BB, New, LI, NewLoops);
625 if (OldLoop)
626 LoopsToSimplify.insert(NewLoops[OldLoop]);
627
628 if (*BB == Header)
629 // Loop over all of the PHI nodes in the block, changing them to use
630 // the incoming values from the previous block.
631 for (PHINode *OrigPHI : OrigPHINode) {
632 PHINode *NewPHI = cast<PHINode>(VMap[OrigPHI]);
633 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
634 if (Instruction *InValI = dyn_cast<Instruction>(InVal))
635 if (It > 1 && L->contains(InValI))
636 InVal = LastValueMap[InValI];
637 VMap[OrigPHI] = InVal;
638 New->getInstList().erase(NewPHI);
639 }
640
641 // Update our running map of newest clones
642 LastValueMap[*BB] = New;
643 for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
644 VI != VE; ++VI)
645 LastValueMap[VI->first] = VI->second;
646
647 // Add phi entries for newly created values to all exit blocks.
648 for (BasicBlock *Succ : successors(*BB)) {
649 if (L->contains(Succ))
650 continue;
651 for (PHINode &PHI : Succ->phis()) {
652 Value *Incoming = PHI.getIncomingValueForBlock(*BB);
653 ValueToValueMapTy::iterator It = LastValueMap.find(Incoming);
654 if (It != LastValueMap.end())
655 Incoming = It->second;
656 PHI.addIncoming(Incoming, New);
657 }
658 }
659 // Keep track of new headers and latches as we create them, so that
660 // we can insert the proper branches later.
661 if (*BB == Header)
662 Headers.push_back(New);
663 if (*BB == LatchBlock)
664 Latches.push_back(New);
665
666 NewBlocks.push_back(New);
667 UnrolledLoopBlocks.push_back(New);
668
669 // Update DomTree: since we just copy the loop body, and each copy has a
670 // dedicated entry block (copy of the header block), this header's copy
671 // dominates all copied blocks. That means, dominance relations in the
672 // copied body are the same as in the original body.
673 if (DT) {
674 if (*BB == Header)
675 DT->addNewBlock(New, Latches[It - 1]);
676 else {
677 auto BBDomNode = DT->getNode(*BB);
678 auto BBIDom = BBDomNode->getIDom();
679 BasicBlock *OriginalBBIDom = BBIDom->getBlock();
680 DT->addNewBlock(
681 New, cast<BasicBlock>(LastValueMap[cast<Value>(OriginalBBIDom)]));
682 }
683 }
684 }
685
686 // Remap all instructions in the most recent iteration
687 for (BasicBlock *NewBlock : NewBlocks) {
688 for (Instruction &I : *NewBlock) {
689 ::remapInstruction(&I, LastValueMap);
690 if (auto *II = dyn_cast<IntrinsicInst>(&I))
691 if (II->getIntrinsicID() == Intrinsic::assume)
692 AC->registerAssumption(II);
693 }
694 }
695 }
696
697 // Loop over the PHI nodes in the original block, setting incoming values.
698 for (PHINode *PN : OrigPHINode) {
699 if (CompletelyUnroll) {
700 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
701 Header->getInstList().erase(PN);
702 }
703 else if (Count > 1) {
704 Value *InVal = PN->removeIncomingValue(LatchBlock, false);
705 // If this value was defined in the loop, take the value defined by the
706 // last iteration of the loop.
707 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
708 if (L->contains(InValI))
709 InVal = LastValueMap[InVal];
710 }
711 assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
712 PN->addIncoming(InVal, Latches.back());
713 }
714 }
715
716 // Now that all the basic blocks for the unrolled iterations are in place,
717 // set up the branches to connect them.
718 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
719 // The original branch was replicated in each unrolled iteration.
720 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
721
722 // The branch destination.
723 unsigned j = (i + 1) % e;
724 BasicBlock *Dest = Headers[j];
725 bool NeedConditional = true;
726
727 if (RuntimeTripCount && j != 0) {
728 NeedConditional = false;
729 }
730
731 // For a complete unroll, make the last iteration end with a branch
732 // to the exit block.
733 if (CompletelyUnroll) {
734 if (j == 0)
735 Dest = LoopExit;
736 // If using trip count upper bound to completely unroll, we need to keep
737 // the conditional branch except the last one because the loop may exit
738 // after any iteration.
739 assert(NeedConditional &&
740 "NeedCondition cannot be modified by both complete "
741 "unrolling and runtime unrolling");
742 NeedConditional = (PreserveCondBr && j && !(PreserveOnlyFirst && i != 0));
743 } else if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
744 // If we know the trip count or a multiple of it, we can safely use an
745 // unconditional branch for some iterations.
746 NeedConditional = false;
747 }
748
749 if (NeedConditional) {
750 // Update the conditional branch's successor for the following
751 // iteration.
752 Term->setSuccessor(!ContinueOnTrue, Dest);
753 } else {
754 // Remove phi operands at this loop exit
755 if (Dest != LoopExit) {
756 BasicBlock *BB = Latches[i];
757 for (BasicBlock *Succ: successors(BB)) {
758 if (Succ == Headers[i])
759 continue;
760 for (PHINode &Phi : Succ->phis())
761 Phi.removeIncomingValue(BB, false);
762 }
763 }
764 // Replace the conditional branch with an unconditional one.
765 BranchInst::Create(Dest, Term);
766 Term->eraseFromParent();
767 }
768 }
769
770 // Update dominators of blocks we might reach through exits.
771 // Immediate dominator of such block might change, because we add more
772 // routes which can lead to the exit: we can now reach it from the copied
773 // iterations too.
774 if (DT && Count > 1) {
775 for (auto *BB : OriginalLoopBlocks) {
776 auto *BBDomNode = DT->getNode(BB);
777 SmallVector<BasicBlock *, 16> ChildrenToUpdate;
778 for (auto *ChildDomNode : BBDomNode->getChildren()) {
779 auto *ChildBB = ChildDomNode->getBlock();
780 if (!L->contains(ChildBB))
781 ChildrenToUpdate.push_back(ChildBB);
782 }
783 BasicBlock *NewIDom;
784 if (BB == LatchBlock) {
785 // The latch is special because we emit unconditional branches in
786 // some cases where the original loop contained a conditional branch.
787 // Since the latch is always at the bottom of the loop, if the latch
788 // dominated an exit before unrolling, the new dominator of that exit
789 // must also be a latch. Specifically, the dominator is the first
790 // latch which ends in a conditional branch, or the last latch if
791 // there is no such latch.
792 NewIDom = Latches.back();
793 for (BasicBlock *IterLatch : Latches) {
794 Instruction *Term = IterLatch->getTerminator();
795 if (isa<BranchInst>(Term) && cast<BranchInst>(Term)->isConditional()) {
796 NewIDom = IterLatch;
797 break;
798 }
799 }
800 } else {
801 // The new idom of the block will be the nearest common dominator
802 // of all copies of the previous idom. This is equivalent to the
803 // nearest common dominator of the previous idom and the first latch,
804 // which dominates all copies of the previous idom.
805 NewIDom = DT->findNearestCommonDominator(BB, LatchBlock);
806 }
807 for (auto *ChildBB : ChildrenToUpdate)
808 DT->changeImmediateDominator(ChildBB, NewIDom);
809 }
810 }
811
812 assert(!DT || !UnrollVerifyDomtree ||
813 DT->verify(DominatorTree::VerificationLevel::Fast));
814
815 // Merge adjacent basic blocks, if possible.
816 for (BasicBlock *Latch : Latches) {
817 BranchInst *Term = cast<BranchInst>(Latch->getTerminator());
818 if (Term->isUnconditional()) {
819 BasicBlock *Dest = Term->getSuccessor(0);
820 if (BasicBlock *Fold = foldBlockIntoPredecessor(Dest, LI, SE, DT)) {
821 // Dest has been folded into Fold. Update our worklists accordingly.
822 std::replace(Latches.begin(), Latches.end(), Dest, Fold);
823 UnrolledLoopBlocks.erase(std::remove(UnrolledLoopBlocks.begin(),
824 UnrolledLoopBlocks.end(), Dest),
825 UnrolledLoopBlocks.end());
826 }
827 }
828 }
829
830 // At this point, the code is well formed. We now simplify the unrolled loop,
831 // doing constant propagation and dead code elimination as we go.
832 simplifyLoopAfterUnroll(L, !CompletelyUnroll && (Count > 1 || Peeled), LI, SE,
833 DT, AC);
834
835 NumCompletelyUnrolled += CompletelyUnroll;
836 ++NumUnrolled;
837
838 Loop *OuterL = L->getParentLoop();
839 // Update LoopInfo if the loop is completely removed.
840 if (CompletelyUnroll)
841 LI->erase(L);
842
843 // After complete unrolling most of the blocks should be contained in OuterL.
844 // However, some of them might happen to be out of OuterL (e.g. if they
845 // precede a loop exit). In this case we might need to insert PHI nodes in
846 // order to preserve LCSSA form.
847 // We don't need to check this if we already know that we need to fix LCSSA
848 // form.
849 // TODO: For now we just recompute LCSSA for the outer loop in this case, but
850 // it should be possible to fix it in-place.
851 if (PreserveLCSSA && OuterL && CompletelyUnroll && !NeedToFixLCSSA)
852 NeedToFixLCSSA |= ::needToInsertPhisForLCSSA(OuterL, UnrolledLoopBlocks, LI);
853
854 // If we have a pass and a DominatorTree we should re-simplify impacted loops
855 // to ensure subsequent analyses can rely on this form. We want to simplify
856 // at least one layer outside of the loop that was unrolled so that any
857 // changes to the parent loop exposed by the unrolling are considered.
858 if (DT) {
859 if (OuterL) {
860 // OuterL includes all loops for which we can break loop-simplify, so
861 // it's sufficient to simplify only it (it'll recursively simplify inner
862 // loops too).
863 if (NeedToFixLCSSA) {
864 // LCSSA must be performed on the outermost affected loop. The unrolled
865 // loop's last loop latch is guaranteed to be in the outermost loop
866 // after LoopInfo's been updated by LoopInfo::erase.
867 Loop *LatchLoop = LI->getLoopFor(Latches.back());
868 Loop *FixLCSSALoop = OuterL;
869 if (!FixLCSSALoop->contains(LatchLoop))
870 while (FixLCSSALoop->getParentLoop() != LatchLoop)
871 FixLCSSALoop = FixLCSSALoop->getParentLoop();
872
873 formLCSSARecursively(*FixLCSSALoop, *DT, LI, SE);
874 } else if (PreserveLCSSA) {
875 assert(OuterL->isLCSSAForm(*DT) &&
876 "Loops should be in LCSSA form after loop-unroll.");
877 }
878
879 // TODO: That potentially might be compile-time expensive. We should try
880 // to fix the loop-simplified form incrementally.
881 simplifyLoop(OuterL, DT, LI, SE, AC, PreserveLCSSA);
882 } else {
883 // Simplify loops for which we might've broken loop-simplify form.
884 for (Loop *SubLoop : LoopsToSimplify)
885 simplifyLoop(SubLoop, DT, LI, SE, AC, PreserveLCSSA);
886 }
887 }
888
889 return CompletelyUnroll ? LoopUnrollResult::FullyUnrolled
890 : LoopUnrollResult::PartiallyUnrolled;
891 }
892
893 /// Given an llvm.loop loop id metadata node, returns the loop hint metadata
894 /// node with the given name (for example, "llvm.loop.unroll.count"). If no
895 /// such metadata node exists, then nullptr is returned.
GetUnrollMetadata(MDNode * LoopID,StringRef Name)896 MDNode *llvm::GetUnrollMetadata(MDNode *LoopID, StringRef Name) {
897 // First operand should refer to the loop id itself.
898 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
899 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
900
901 for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
902 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
903 if (!MD)
904 continue;
905
906 MDString *S = dyn_cast<MDString>(MD->getOperand(0));
907 if (!S)
908 continue;
909
910 if (Name.equals(S->getString()))
911 return MD;
912 }
913 return nullptr;
914 }
915